CN116000945B - Intelligent control method of cable deicing robot - Google Patents
Intelligent control method of cable deicing robot Download PDFInfo
- Publication number
- CN116000945B CN116000945B CN202211552526.1A CN202211552526A CN116000945B CN 116000945 B CN116000945 B CN 116000945B CN 202211552526 A CN202211552526 A CN 202211552526A CN 116000945 B CN116000945 B CN 116000945B
- Authority
- CN
- China
- Prior art keywords
- data
- cable
- icing
- deicing
- ice
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 44
- 238000012549 training Methods 0.000 claims abstract description 51
- 238000013145 classification model Methods 0.000 claims abstract description 34
- 230000009467 reduction Effects 0.000 claims abstract description 15
- 239000011248 coating agent Substances 0.000 claims abstract description 11
- 238000000576 coating method Methods 0.000 claims abstract description 11
- 238000012545 processing Methods 0.000 claims abstract description 9
- 238000004422 calculation algorithm Methods 0.000 claims description 13
- 230000033001 locomotion Effects 0.000 claims description 11
- 230000006870 function Effects 0.000 claims description 9
- 238000002372 labelling Methods 0.000 claims description 9
- 230000009466 transformation Effects 0.000 claims description 9
- 230000004913 activation Effects 0.000 claims description 7
- 238000010606 normalization Methods 0.000 claims description 7
- 238000012360 testing method Methods 0.000 claims description 7
- 230000007246 mechanism Effects 0.000 claims description 5
- 239000013598 vector Substances 0.000 claims description 5
- 230000004927 fusion Effects 0.000 claims description 4
- 238000004458 analytical method Methods 0.000 claims description 3
- 238000013178 mathematical model Methods 0.000 claims description 3
- 238000005457 optimization Methods 0.000 claims 2
- 238000013528 artificial neural network Methods 0.000 abstract description 6
- 230000006872 improvement Effects 0.000 abstract description 4
- 239000011159 matrix material Substances 0.000 description 11
- 238000010586 diagram Methods 0.000 description 10
- 238000004364 calculation method Methods 0.000 description 9
- 238000000513 principal component analysis Methods 0.000 description 7
- 230000008569 process Effects 0.000 description 5
- 230000011218 segmentation Effects 0.000 description 5
- 230000000694 effects Effects 0.000 description 4
- 238000010801 machine learning Methods 0.000 description 4
- 230000008034 disappearance Effects 0.000 description 3
- 238000012546 transfer Methods 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 238000004880 explosion Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 235000015243 ice cream Nutrition 0.000 description 2
- 238000012804 iterative process Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000010224 classification analysis Methods 0.000 description 1
- 238000004140 cleaning Methods 0.000 description 1
- 238000013501 data transformation Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02E—REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS, RELATED TO ENERGY GENERATION, TRANSMISSION OR DISTRIBUTION
- Y02E10/00—Energy generation through renewable energy sources
- Y02E10/70—Wind energy
- Y02E10/72—Wind turbines with rotation axis in wind direction
Landscapes
- Image Processing (AREA)
Abstract
The invention provides an intelligent control method of a cable deicing robot, which comprises an icing morphology prediction model, a Bayesian classification model and a GBDT classification model based on an improved UNet neural network. The improvement mode of the UNet network model is as follows: adding batch standardized BN, optimizing convolution layer, adding residual error module and attention module, training the improved UNet network model, and obtaining ice coating form prediction graph according to original graph. And (3) performing PCA dimension reduction on the icing morphology prediction graph, and inputting the icing morphology prediction graph into a naive Bayesian classification model for training to realize the prediction of icing categories. And finally, establishing a GBDT classification model, acquiring characteristic data information of icing types and meteorological data, performing data modeling, performing dimension reduction processing, inputting the processed data information into the GBDT model for training, and predicting deicing parameters and kinematic parameters of the robot. The three models are combined, so that the purposes of controlling the walking and deicing of the robot according to the images and weather can be achieved.
Description
Technical Field
The invention relates to the technical field of intelligent construction, in particular to an intelligent control method of a cable deicing robot.
Background
With the increase of the number of bridges in China, the safety problem of road and bridge traffic in winter is also increasingly outstanding. It is counted that more than 70% of bridges in China have hidden danger of icing in winter, especially in Yangtze river basin. Due to the continuous new construction of the high-rise structure, the problem of damage to the bridge caused by icing in winter is transferred from the bridge deck to the upper structure of the bridge. In recent years, bridge superstructure such as cable icing, the frequent occurrence of slush falling events has seriously affected bridge deck vehicle traffic and pedestrian safety, and direct or indirect economic loss is caused to society. The conventional cable deicing method is generally manual deicing, but the deicing mode is generally challenging and dangerous, because the aloft work needs to overcome extremely strong psychology and has extremely high personal safety protection, and the manual deicing mode is not efficient, completely depends on human experience, and is not high in intelligence.
Disclosure of Invention
The invention aims to provide an intelligent control method of a cable deicing robot, which is used for solving the technical problems of low efficiency and low intelligence in the method in the prior art.
In order to achieve the above purpose, the present invention adopts the following technical scheme:
the intelligent control method of the cable deicing robot comprises the following steps:
s1: acquiring multi-source data, wherein the multi-source data comprise cable icing morphology image data, meteorological data under the cable icing morphology, motion parameters of a robot under the cable icing morphology and the meteorological data and deicing parameters used, and marking the motion parameters of the robot and the deicing parameters used according to the cable icing morphology type and the meteorological data in the set cable icing morphology image data to obtain a robot motion parameter label and a deicing parameter label under the cable icing morphology type and the meteorological data;
s2: performing image enhancement processing on the cable ice coating form image data, and performing labeling to obtain an image data sample set, wherein the labeling result comprises ice coating form categories in the image data;
s3: dividing a training data set from the image data sample set, inputting the training data set into a pre-constructed UNet network model for training to obtain a trained UNet network model, wherein the trained UNet network model is used for obtaining an icing form prediction graph according to input image data;
s4: extracting ice coverage area and ice length from an ice coverage form prediction graph output by a trained UNet network model, constructing a ice feature data set, performing main component analysis on the ice feature data set, performing dimension reduction, and then inputting a pre-constructed naive Bayesian classification model for training to obtain a trained naive Bayesian classification model, wherein the trained naive Bayesian classification model is used for predicting ice category information comprising big ice, small ice and micro ice according to the input ice feature data set;
s5: fusing the ice type information obtained by prediction of the trained naive Bayesian classification model, the meteorological data in the cable icing state, the robot kinematics parameter labels and the deicing parameter labels in the cable icing state type and the meteorological data to obtain fused characteristic data, and dividing the fused characteristic data into training data and test data after dimension reduction treatment;
s6: training a GBDT model by using a GBDT regression algorithm based on training data, taking the trained GBDT model as a prediction model, and predicting according to input icing type information and meteorological data in a cable icing form by using the prediction model to obtain a robot kinematics parameter label and a deicing parameter label;
s7: and intelligently controlling the cable deicing robot according to the predicted kinematic parameter labels and the deicing parameter labels.
In one embodiment, the meteorological data in the form of cable ice coating is monitored by meteorological environment sensors, including temperature, humidity and wind speed.
In one embodiment, the step S2 of performing image enhancement processing on the cable ice-covering image data includes geometric transformation and color space transformation, and image labeling software labelme is adopted when labeling the image-enhanced data.
In one embodiment, the pre-built UNet network model in step S3 adds a batch of standardized BN, an optimized convolution layer, a residual module and an attention module on the basis of the original UNet network model, where the batch of standardized BN is used to normalize data input to the UNet network model, the optimized convolution layer is used to combine convolution kernels of different sizes, learn feature information in a larger receptive field range, the residual module is used to transfer feature data, and the attention module is used to perform information fusion and pay attention to spatial information.
In one embodiment, the batch normalization BN is processed as:
Y (k) ＝γ (k) X (k)+ β (k)
wherein k is the number of convolution kernels,is the result of standard normalization, x (k) A value representing the hidden layer before activation, +.>Represents x (k) Var represents the variance of the random variable, gamma (k) And beta (k) To adjust the parameters, Y (k) Results after BN treatment.
In one embodiment, the optimal convolution layer uses 3 x 3,5 x 5,7 x 7 convolution kernels to combine in parallel instead of a single 3 x 3 convolution kernel in the original UNet network model, with a step size of 1 for the combined convolution.
In one embodiment, the residual module adds a BN layer and a ReLU activation function before each convolution layer, and adds an additional 1×1 convolution layer and BN layer to the input data.
In one embodiment, the attention module fuses with the corresponding part in the downsampling during the upsampling of the feature map, adds an attention mechanism during the fusing, increases the spatial information of the learned feature map, and then upsamples.
In one embodiment, the naive bayes classification model adopts a classification method based on independent assumption of bayes theorem and characteristic conditions, wherein a mathematical model of the naive bayes classifier is expressed as follows:
P(c k |X)＝max{P(c 1 |X),P(c 2 |X),...P(c m |X)}
wherein, a feature vector X (X 1 ,x 2 ,...x n ) For a sample to be classified, the output space is a class mark set Y= { c 1 ,c 2 ,...,c m A Bayesian classification model for classifying samples X, P (c) k X) is c under the conditions where X occurs k Probability of occurrence, c k Sample class to be classified predicted for naive bayes classification model.
Compared with the prior art, the invention has the following advantages and beneficial technical effects:
(1) The method provided by the invention is an intelligent control method of the cable deicing robot fusing the image and the meteorological environment data. Three machine learning models are established, namely an improved UNet network model, namely a machine learning model for identifying cable icing images is established and used for outputting icing morphology prediction graphs, a naive Bayesian classification model is established and used for classifying the icing morphology, and a classification prediction model based on GBDT algorithm is established and used for carrying out classification prediction on robot kinematic parameters and required deicing parameters.
(2) The invention provides an intelligent control method of a cable deicing robot fusing images and meteorological environment data, which is used for improving an original UNet network model by adding batch standardized BN, optimizing a convolution layer, adding a residual error module and an attention module, wherein the adding of the BN layer can ensure that the data distribution of each layer is stable when the network is trained, and the change of data of a certain layer cannot be accumulated to the next layer. The residual error module is added to improve the generalization force of the network, solve the problems of gradient disappearance and gradient explosion, promote the transfer of characteristic information and further improve the network learning rate. The channel attention mechanism is added, so that the high-value information of the characteristics of different channels can be rapidly positioned, and overfitting is prevented, namely, hardware resource expenditure during network training is reduced while a large amount of repeated calculation of low-value information is avoided, and the training efficiency and generalization force of the network are improved.
(3) The invention provides an intelligent control method of a cable deicing robot fusing images and meteorological environment data, wherein in the conventional data which are easy to extract about weather and the like, such as temperature and humidity collection, data are preprocessed by utilizing technologies such as data integration, data transformation, data protocol, data cleaning and the like, the processed data are converted into digital features which can be used for machine learning, namely feature valorization, and finally a data sample set is established.
(4) The invention provides an intelligent control method of a cable deicing robot fusing images and meteorological environment data. The GBDT algorithm is adopted for the robot operation mode identification method, the classification analysis is carried out on the residual error between the regression tree and the target value, and the residual error is continuously reduced by the gradual lifting algorithm, so that the calculated value gradually approaches the target value. Because different residuals can be processed differently under the regression tree, even if error points exist in the sample, the training result is not affected, and the GBDT algorithm has a reasonable training process and good robustness.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are required in the embodiments or the description of the prior art will be briefly described, and it is obvious that the drawings in the following description are some embodiments of the present invention, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a structural framework diagram of an intelligent control method of a cable deicing robot provided by the invention;
fig. 2 is a diagram of a UNet network according to the present invention;
fig. 3 is a diagram of an improved overall UNet network structure provided by the present invention;
FIG. 4 is a block diagram of a residual block provided by the present invention;
FIG. 5 is a block diagram of a channel attention module provided by the present invention;
fig. 6 is a morphology of ice coating obtained using UNet neural network according to the present invention;
FIG. 7 is a GBDT algorithm calculation flow chart for solving the problem of robot operation mode identification provided by the invention;
fig. 8 is a top view of the cable deicing robot provided by the present invention.
Detailed Description
The invention discloses an intelligent control method of a cable deicing robot, which constructs an icing morphology prediction model, a Bayesian classification model and a GBDT classification model based on an improved UNet neural network. The improvement mode of the UNet network model is as follows: adding batch standardized BN, optimizing convolution layer, adding residual error module and attention module, training the improved UNet network model, and obtaining ice coating form prediction graph according to original graph. And (3) performing PCA dimension reduction on the icing morphology prediction graph, and inputting the obtained icing morphology prediction graph into a naive Bayesian classification model for training, so that the icing type can be predicted. And finally, establishing a GBDT classification model, acquiring characteristic data information of icing types and meteorological data, performing data modeling, performing dimension reduction processing, inputting the processed data information into the GBDT model for training, and predicting deicing parameters and kinematic parameters of the robot. The three models are combined, so that the purposes of controlling the walking and deicing of the robot according to the images and weather can be achieved.
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments of the present invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Example 1
The embodiment of the invention provides an intelligent control method of a cable deicing robot, which comprises the following steps:
s1: acquiring multi-source data, wherein the multi-source data comprise cable icing morphology image data, meteorological data under the cable icing morphology, motion parameters of a robot under the cable icing morphology and the meteorological data and deicing parameters used, and marking the motion parameters of the robot and the deicing parameters used according to the cable icing morphology type and the meteorological data in the set cable icing morphology image data to obtain a robot motion parameter label and a deicing parameter label under the cable icing morphology type and the meteorological data;
s2: performing image enhancement processing on the cable ice coating form image data, and performing labeling to obtain an image data sample set, wherein the labeling result comprises ice coating form categories in the image data;
s3: dividing a training data set from the image data sample set, inputting the training data set into a pre-constructed UNet network model for training to obtain a trained UNet network model, wherein the trained UNet network model is used for obtaining an icing form prediction graph according to input image data;
s4: extracting ice coverage area and ice length from an ice coverage form prediction graph output by a trained UNet network model, constructing a ice feature data set, performing main component analysis on the ice feature data set, performing dimension reduction, and then inputting a pre-constructed naive Bayesian classification model for training to obtain a trained naive Bayesian classification model, wherein the trained naive Bayesian classification model is used for predicting ice category information comprising big ice, small ice and micro ice according to the input ice feature data set;
s5: fusing the ice type information obtained by prediction of the trained naive Bayesian classification model, the meteorological data in the cable icing state, the robot kinematics parameter labels and the deicing parameter labels in the cable icing state type and the meteorological data to obtain fused characteristic data, and dividing the fused characteristic data into training data and test data after dimension reduction treatment;
s6: training a GBDT model by using a GBDT regression algorithm based on training data, taking the trained GBDT model as a prediction model, and predicting according to input icing type information and meteorological data in a cable icing form by using the prediction model to obtain a robot kinematics parameter label and a deicing parameter label;
s7: and intelligently controlling the cable deicing robot according to the predicted kinematic parameter labels and the deicing parameter labels.
In general, step S1 is to collect and classify multi-source data, sort image data and weather environment data, and then obtain cable surrounding environment feature data under multi-region cable icing conditions, sort according to icing form and air image data set in experimental stage, and obtain corresponding robot kinematics parameter labels and deicing parameter labels. Step S2 is enhancement and annotation of the image data. Step S3 is to divide a training data set from the data obtained in step S2 and train a UNet network model, step S4 is to train a naive Bayesian classification model, step S5 is to fuse icing type information obtained through prediction of the naive Bayesian classification model, meteorological data in a cable icing form, a robot kinematics parameter label and a deicing parameter label in the cable icing form type and the meteorological data, then form feature data, and divide data for training a subsequent model. Step S6 is training GBDT model based on the training data obtained in step S5, and step S7 is controlling deicing robot according to the prediction result of the model in step S6.
Referring to fig. 1, a structural frame diagram of an intelligent control method of a cable deicing robot provided by the invention is provided.
Specifically, the image data is image data related to icing characteristics acquired in a laboratory using a depth camera.
According to the method disclosed by the invention, the established improved UNet network model for ice image recognition, the naive Bayesian classification model for ice category classification and the GBDT model for robot operation mode recognition are applied to an intelligent control system of the cable deicing robot. The method predicts various icing states of the cable surface timely and accurately, and can enable the deicing robot to select different deicing modes according to different icing states, so that the deicing effect of time and labor saving, safety and accuracy is achieved well.
In one embodiment, the meteorological data in the form of cable ice coating is monitored by meteorological environment sensors, including temperature, humidity and wind speed.
In one embodiment, the step S2 of performing image enhancement processing on the cable ice-covering image data includes geometric transformation and color space transformation, and image labeling software labelme is adopted when labeling the image-enhanced data.
In the specific implementation process, the adopted image enhancement parameters comprise: brightness, contrast, gamma values to improve sharpness of the cable ice image, and increase the difference between the slush and background features, e.g., brightness decrease, contrast enhancement, gamma values of 1.25. And marking the ice by using image marking software labelme.
In one embodiment, the pre-built UNet network model in step S3 adds a batch of standardized BN, an optimized convolution layer, a residual module and an attention module on the basis of the original UNet network model, where the batch of standardized BN is used to normalize data input to the UNet network model, the optimized convolution layer is used to combine convolution kernels of different sizes, learn feature information in a larger receptive field range, the residual module is used to transfer feature data, and the attention module is used to perform information fusion and pay attention to spatial information.
Specifically, the original U-net network structure is shown in fig. 2, the left side of the structure is a feature extraction and downsampling part, the right side of the structure is a decoding part, and the segmentation map is finally obtained through a series of upsampling. In the left feature extraction part, a new-scale feature map is generated after each pooling layer, and 5 scales are added to the original map. In the right up-sampling part, each time up-sampling is performed, a characteristic diagram corresponding to the scale of the left part is generated, and the characteristic diagram is spliced with the left characteristic diagram. And finally, outputting two layers, namely a foreground and a background.
The improved UNet network structure in this embodiment mainly comprises two parts of downsampling and upsampling, the UNet original convolution is replaced by a combined convolution, the downsampling part of the residual module added into the network solves the gradient disappearance phenomenon, the jump connection part adds the attention mechanism, so that the feature map is respectively fused with the downsampling corresponding part in the same-size feature map in the upsampling process, the method can enable aggregate features to be extracted by combining the information of the bottom layer and the high layer, the corresponding encoder feature map contains more spatial information, and the improved integral UNet network structure is shown in fig. 3.
In one embodiment, the batch normalization BN is processed as:
Y (k) ＝γ (k) X (k) +β (k)
wherein k is the number of convolution kernels,is the result of standard normalization, x (k) A value representing the hidden layer before activation, +.>Represents x (k) Var represents the variance of the random variable, gamma (k) And beta (k) To adjust the parameters, Y (k) Results after BN treatment.
Specifically, the purpose of adding the BN layer is mainly to normalize the data of the network, and the transformed X is processed by the method (k) The inverse transformation is activated, so that the expressive power of the network can be enhanced.
In one embodiment, the optimal convolution layer uses 3 x 3,5 x 5,7 x 7 convolution kernels to combine in parallel instead of a single 3 x 3 convolution kernel in the original UNet network model, with a step size of 1 for the combined convolution.
Specifically, the optimized convolution layer fuses convolution kernels with different sizes, and can learn characteristic information within a larger receptive field range.
In one embodiment, the residual module adds a BN layer and a ReLU activation function before each convolution layer, and adds an additional 1×1 convolution layer and BN layer to the input data.
In the implementation process, as shown in fig. 4, the residual error module adds a BN layer and a ReLU activation function before each convolution layer, and adds a 1×1 convolution layer and a BN layer to the input X. The nonlinear transformation added on the characteristics can enable the network to be continuously degenerated along with the increase of depth, is beneficial to solving the problems of gradient disappearance and gradient explosion, and can promote the transmission of characteristic information so as to further improve the network learning rate.
In one embodiment, the attention module fuses with the corresponding part in the downsampling during the upsampling of the feature map, adds an attention mechanism during the fusing, increases the spatial information of the learned feature map, and then upsamples.
In the implementation, the channel attention module (i.e., attention module) structure is shown in fig. 5. The input of the channel attention module is a feature map with the size of H multiplied by W multiplied by C, which is obtained by downsampling, the feature map with the size of H multiplied by W multiplied by C is obtained by performing Reshape operation twice to obtain feature maps with the size of H multiplied by W, then the two result feature maps are subjected to matrix multiplication, the feature maps with the sizes of (H multiplied by W) multiplied by C multiplied by H multiplied by W multiplied by C, and the feature maps with the sizes of C multiplied by H multiplied by W multiplied by C are obtained by C multiplied by C, and finally the result is added with the original feature map to obtain the output of the attention module.
In one embodiment, the image data sample set obtained in step S2 is randomly divided into a training set, a validation set and a test set 3 according to a ratio of 7:2:1, and the number of images in each portion is 7000, 2000 and 1000 respectively. The input data is an original picture acquired by the depth camera, the training of the sample can be completed through 5 layers of downsampling and 5 layers of upsampling and finally through a layer of 1×1 convolution layer output, a trained network model is automatically stored, a new cable icing image is input into a network for segmentation, and a recognized slush shape diagram is obtained as shown in fig. 6.
In one embodiment, principal Component Analysis (PCA) is a statistical idea-based dimension reduction method. PCA transforms the original data into a group of representations with linear independence of each dimension through linear transformation, and the representations are used for extracting main characteristic components of the data so as to achieve the purpose of dimension reduction of the high-dimension data. The basic process is as follows:
let the original dataset be represented as a matrix X m×n X is taken as m×n Zero-equalizing each row of the row, wherein the calculation expression is shown as a formula (1), and x is ij Representation matrix X m×n The elements of row i and column j,representation matrix X m×n Mean value of ith row, S i Representation matrix X m×n Standard deviation of row i.
Obtaining a covariance matrix C according to formula (2), wherein m represents the number of samples, k represents the number of columns, and x j Representation matrix X m×n The elements of the j-th column,representation matrix X m×n The mean value of column j. At the same time, the eigenvalue lambda of the covariance matrix C is obtained 1 ≥λ 2 ≥…≥λ k Corresponding feature vector d 1 ，d 2 ，…，d k 。
And (3) obtaining a characteristic contribution rate according to the formula (3), and arranging the characteristic vectors into a matrix P according to the corresponding characteristic values from top to bottom, wherein Y=PX is the k-dimensional matrix after dimension reduction.
In one embodiment, the naive bayes classification model adopts a classification method based on independent assumption of bayes theorem and characteristic conditions, wherein a mathematical model of the naive bayes classifier is expressed as follows:
P(c k |X)＝max{P(c 1 |X),P(c 2 |X),...P(c m |X)} (4)
wherein, a feature vector X (X 1 ,x 2 ,...x n ) For a sample to be classified, the output space is a class mark set Y= { c 1 ,c 2 ,...,c m A Bayesian classification model for classifying samples X, P (c) k X) is c under the conditions where X occurs k Probability of occurrence, c k Sample class to be classified predicted for naive bayes classification model.
Specifically, the conditional probability steps are as follows:
a) Constructing a training sample set of known class labels;
b) The conditional probability of individual features in a statistical training set in each class, e.g. P (x 1 |c 1 ),P(x 2 |c 1 ),...,P(x n |c 1 ),
c) Assuming that the feature attributes are independent of each other, the conditional probability expression obtained according to the bayesian theorem is:
d) In the formula (5), the denominator P (X) is the same for all categories, so that only the numerator needs to be maximized, and the simplified conditional probability expression is:
in one embodiment, the GBDT main idea is: training a new learner in the gradient direction to reduce the learning error rate of the former learner, and iteratively generating the new learner based on the former learner, wherein the calculation formula is as follows:
F m (x)＝F m-1 (x)+ρh(x) 1≤m≤M (7)
the GBDT strong learner of the present invention mainly comprises the following 4 steps:
1. firstly, initializing a learner, wherein the calculation formula of the initial learner is as follows:
wherein F is 0 () Represents the initial learner, the value of rho is the average value of all training sample label values, L (y) i ρ) represents a loss function, y i To learn samples, N represents the number of samples.
2. The negative gradient value of the current loss function is calculated, the iteration number is M, namely, the fitting target of the regression tree in the iteration is calculated, and the calculation formula is as follows:
wherein r is m,i Representing the negative gradient value after m iterations of sample i, delta represents the fit
3. Through the mth iteration, an optimal base classifier is obtained, and the calculation formula is as follows:
wherein alpha is m Represents the optimal basis classifier after m iterations, βh represents the learning rate, βh=0.1
Optimal learning rate rho based on linear optimal searching mode m Updating the next learner, wherein the calculation formula is as follows: f (F) m ()＝F m-1 ()+ρ m h(x i ； m )
If m=m iteration ends, otherwise repeating step 2-3.
4. A final strong learner G is generated.
Finally, the improved UNet network model, the naive Bayesian classifier and the GBDT classification model are fused, so that the cable deicing robot can achieve the effect of selecting proper deicing parameters and kinematic parameters according to the original pictures and the meteorological data.
As shown in fig. 8, the depth cameras 2 are provided on both sides of the deicing robot housing 1. The depth camera adopts a high-performance CMOS image sensor, the resolution can reach 2560 x 1440@25fps, the color is true to restore, and the image quality is clear and fine. And the plug-in card storage is supported, so that more images can be stored conveniently. The protection device is arranged under the appearance of the camera of the gun type, so that the damage of the camera caused by weather problems in use is prevented. The camera is located deicing machine side, and two machine positions can make the slush image of gathering have more angles, can make the recognition of slush image more accurate.
The cable temperature and humidity sensor 3 is located at the front side of the crawler belt, and the wind speed sensor is located at the upper side of the deicing robot, so that the temperature and humidity of the cable in front of the deicing robot and the wind speed around the robot can be monitored in real time. And then can be according to the humiture and wind speed data that monitor combine icing shape that the robot predicted to confirm suitable deicing parameter and kinematics parameter.
The invention collects about 1000 cable icing images altogether, the original image size is 1536×1024, the original image and the labeling image are cut into 4000 768×512 small images during training, a certain amount of operations such as translation, rotation, scaling and the like are performed to amplify the data before the cable icing images are sent into a network to 10000, and the amplified data set is labeled by labelme to obtain a slush label image after the slush shape is marked.
The UNet network of the invention is built according to the pytorch, firstly, a UNet root directory is built, collected image data is placed under the UNet folder in a DRIVE name, and the image data is divided into a test set folder (test) and a training set folder (training). The images folder, the manual folder and the mask folder are respectively established under the test folder and the tracking folder. The images folder is used for storing original pictures for segmentation, and the mask folder is used for storing binary pictures for confirming segmented areas; the manual file is used for storing pictures segmented by using labelme manually. Training is directly carried out by using a train script which is set in advance after the data are well arranged, a result file is generated under the current folder after training is finished, and the folder records a verification result after training is finished. Meanwhile, after training is completed, the prediction script can be called to check the training effect.
After PCA dimension reduction is carried out on the data output by the improved UNet network model, a Bayesian conditional probability formula is utilized, as shown in a formula (8), the probability that P (x_j|c) is the characteristic attribute x_j of the known slush class c is utilized to calculate the posterior probability that the known image attribute belongs to different slush classes, as shown in a formula (9); finally, the image is classified into a slush category with the maximum posterior probability according to the maximum posterior probability, as shown in formula (10).
For classification of ice formation, the size of the ice displayed on the image cannot be equal to the actual size due to the angle problem of the depth camera, and the cable at the position of the ice can be selected as a reference object due to the fact that the width of the cable is unchanged. Let the length of the slush shown in the image be a, the cable width corresponding to the slush position be b, when a: b >1.5, setting the regional slush category as big slush; when 0.5< a: b is less than or equal to 1.5, setting the regional slush category as medium slush; when a: b is less than or equal to 0.5, setting the type of the regional slush as small slush; in addition, the ice covering area of the cable in the image is set as c, the area of the cable area without ice covering is set as d, and when the ratio of c to d is less than or equal to 0.7< 1, the ice type of the area is set as big ice covering; when the ratio of c to d is 0.3< 0.7, setting the ice type of the area as ice-on-center; when 0< c:d is less than or equal to 0.3, the regional slush category is set as small icing.
Through indoor experiments, different deicing parameters and kinematic parameters can be set according to different icing conditions. For the ice cream, the ambient temperature and the ambient humidity are very low, a primary deicing mode is adopted, and the deicing mode adopts a mode of low-speed forward movement, increased forward stress and high-speed rotation of the skates to remove the ice cream. And when the ambient temperature and humidity are lower, a secondary deicing mode is adopted for normal ice, and the mode adopts a mode of low-speed forward movement, increasing forward stress and normal speed rotation of the skates to carry out deicing. For small ice, when the ambient temperature and humidity are high, a three-stage deicing mode is adopted. The mode adopts the normal speed rotation of the skates, increases the forward stress and carries out deicing in a normal speed advancing mode. For the position without ice edges, and when the temperature and the humidity are high, deicing is not needed, so the ice-removing device directly and rapidly advances until the ice is identified. In addition, no matter what kind of ice, when the wind speed sensor senses that the wind speed is great, the forward stress is increased to prevent the robot from sliding off. And finally, establishing a sample by using the deicing parameters and the kinematic parameters.
A base learner, which becomes a classification regression tree (CART tree), is used in each iteration of the GBDT algorithm. CART tree applies to high bias, low variance and sufficient depth. In the regression problem, an iterative process is executed, each base learner is trained based on the learning error rate of the previous base learner, and a gradient descent technique is adopted to carry out negative gradient fitting on the loss function to the regression tree so as to ensure that the decision model is continuously improved.
According to the principle of the GBDT algorithm, the calculation flow of the GBDT algorithm for solving the problem of the robot operation mode identification is shown in FIG. 7. Because the robot operation mode is identified as a multi-classification problem, the loss function is preferably a cross entropy loss function; the optimizer selects an Adam optimizer for adaptively adjusting the learning rate in the iterative process to perform network learning so as to improve the network learning speed. To prevent the network from overfitting, dropout regularization is used to promote the generalization ability of the model. And finally constructing a robot operation mode identification model based on the GBDT algorithm.
One specific example is as follows:
semantic segmentation of frozen images using a modified UNet neural network is first determined. And building a corresponding module in the pytorch to establish a UNet network model. Taking 70% of the collected 10000 cable ice images as a training set and 30% as a verification set. Machine learning is performed to form a neural network. The input layer is a cable icing picture shot by the depth camera, and the output layer is a specific slush shape chart after segmentation is completed. A naive bayes classifier is built in Python. And (3) performing PCA (principal component analysis) dimensionality reduction on the data output by the improved UNet neural network, and then inputting the data into a naive Bayes classifier. And a classification of the slush shape can be obtained. And finally, carrying out data modeling on the ice category information after the naive Bayes classification is finished and the acquired meteorological data information, wherein the method comprises the steps of acquiring characteristic data information, acquiring kinematic parameters and acquiring deicing parameter label data information, training a GBDT model by using a GBDT regression algorithm, taking the trained GBDT model as a prediction model, and finally achieving the purpose of enabling the deicing robot to determine proper deicing parameters and kinematic parameters according to different ice shapes and different meteorological environment data.
The invention provides a method for intelligently controlling a cable deicing robot by fusing images and meteorological environment data, and the method and the way for realizing the technical scheme are numerous, the above is only a preferred embodiment of the invention, and it should be pointed out that a plurality of improvements and modifications can be made to those skilled in the art without departing from the principle of the invention, and the improvements and modifications are also regarded as the protection scope of the invention. The components not explicitly described in this embodiment can be implemented by using the prior art.
Claims (9)
1. An intelligent control method of a cable deicing robot is characterized by comprising the following steps:
s1: acquiring multi-source data, wherein the multi-source data comprise cable icing morphology image data, meteorological data under the cable icing morphology, motion parameters of a robot under the cable icing morphology and the meteorological data and deicing parameters used, and marking the motion parameters of the robot and the deicing parameters used according to the cable icing morphology type and the meteorological data in the set cable icing morphology image data to obtain a robot motion parameter label and a deicing parameter label under the cable icing morphology type and the meteorological data;
s2: performing image enhancement processing on the cable ice coating form image data, and performing labeling to obtain an image data sample set, wherein the labeling result comprises ice coating form categories in the image data;
s3: dividing a training data set from the image data sample set, inputting the training data set into a pre-constructed UNet network model for training to obtain a trained UNet network model, wherein the trained UNet network model is used for obtaining an icing form prediction graph according to input image data;
s4: extracting ice coverage area and ice length from an ice coverage form prediction graph output by a trained UNet network model, constructing a ice feature data set, performing main component analysis on the ice feature data set, performing dimension reduction, and then inputting a pre-constructed naive Bayesian classification model for training to obtain a trained naive Bayesian classification model, wherein the trained naive Bayesian classification model is used for predicting ice category information comprising big ice, small ice and micro ice according to the input ice feature data set;
s5: fusing the ice type information obtained by prediction of the trained naive Bayesian classification model, the meteorological data in the cable icing state, the robot kinematics parameter labels and the deicing parameter labels in the cable icing state type and the meteorological data to obtain fused characteristic data, and dividing the fused characteristic data into training data and test data after dimension reduction treatment;
s6: training a GBDT model by using a GBDT regression algorithm based on training data, taking the trained GBDT model as a prediction model, and predicting according to input icing type information and meteorological data in a cable icing form by using the prediction model to obtain a robot kinematics parameter label and a deicing parameter label;
s7: and intelligently controlling the cable deicing robot according to the predicted kinematic parameter labels and the deicing parameter labels.
2. The intelligent control method of a cable deicing robot of claim 1, wherein the meteorological data in the form of cable icing is monitored by a meteorological environment sensor, including temperature, humidity and wind speed.
3. The intelligent control method of the cable deicing robot according to claim 1, wherein the step S2 of performing image enhancement processing on the cable icing morphology image data comprises geometric transformation and color space transformation, and image annotation software labelme is adopted when the image enhanced data is annotated.
4. The intelligent control method of the cable deicing robot according to claim 1, wherein the pre-built UNet network model in step S3 is added with batch normalization BN, optimization convolution layer, residual error module and attention module based on the original UNet network model, wherein the batch normalization BN is used for normalizing data input to the UNet network model, the optimization convolution layer is used for merging convolution kernels of different sizes, learning characteristic information in a larger receptive field range, the residual error module is used for transmitting the characteristic data, and the attention module is used for information fusion and focusing on spatial information.
5. The intelligent control method of a cable deicing robot of claim 4, wherein the batch standardized BN has a processing formula:
Y (k) ＝Y (k) X (k) +β (k)
wherein k is the number of convolution kernels,is the result of standard normalization, x (k) A value representing the hidden layer before activation, ex (k) ]Represents x (k) Var represents the variance of the random variable, gamma (k) And beta (k) To adjust the parameters, Y (k) Results after BN treatment.
6. The intelligent control method of a cable deicing robot of claim 4, wherein the optimized convolution layers are combined in parallel by combining 3 x 3,5 x 5,7 x 7 convolution kernels instead of a single 3 x 3 convolution kernel in the original UNet network model, the step size of the combined convolution being 1.
7. The intelligent control method of a cable deicing robot of claim 4, wherein the residual module adds BN and ReLU activation functions before each convolution layer and adds an additional 1 x 1 convolution layer and BN layer to the input data.
8. The intelligent control method of a cable deicing robot according to claim 4, wherein the attention module fuses with a corresponding part in downsampling during upsampling of the feature map, adds an attention mechanism during the fusion to increase spatial information of the learned feature map, and then upsamples.
9. The intelligent control method of a cable deicing robot according to claim 1, wherein the naive bayes classification model adopts a classification method based on independent assumption of bayes theorem and characteristic conditions, wherein a mathematical model of the naive bayes classifier is expressed as follows:
P(c k |X)＝max{P(c 1 |X),P(c 2 |X),...P(c m |X)}
wherein, a feature vector X (X 1 ,x 2 ,...x n ) For a sample to be classified, the output space is a class mark set Y= { c 1 ,c 2 ,...,c m A Bayesian classification model for classifying samples X, P (c) k X) is c under the conditions where X occurs k Probability of occurrence, c k Sample class to be classified predicted for naive bayes classification model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211552526.1A CN116000945B (en) | 2022-12-05 | 2022-12-05 | Intelligent control method of cable deicing robot |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211552526.1A CN116000945B (en) | 2022-12-05 | 2022-12-05 | Intelligent control method of cable deicing robot |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116000945A CN116000945A (en) | 2023-04-25 |
| CN116000945B true CN116000945B (en) | 2024-04-12 |
Family
ID=86025660
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211552526.1A Active CN116000945B (en) | 2022-12-05 | 2022-12-05 | Intelligent control method of cable deicing robot |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116000945B (en) |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106595551A (en) * | 2016-12-13 | 2017-04-26 | 武汉大学 | Icing thickness detection method for power transmission line icing image based on deep learning |
| RU2658344C1 (en) * | 2017-09-08 | 2018-06-20 | Сергей Сергеевич Дементьев | Overhead transmission lines wires ice loads control system |
| CN108345955A (en) * | 2017-11-24 | 2018-07-31 | 北京国网富达科技发展有限责任公司 | A kind of powerline ice-covering short term prediction method, device and storage medium |
| CN108360442A (en) * | 2018-02-26 | 2018-08-03 | 深圳市易成自动驾驶技术有限公司 | Intelligent snow-removing method, intelligent snow sweeper and computer readable storage medium |
| CN110070530A (en) * | 2019-04-19 | 2019-07-30 | 山东大学 | A kind of powerline ice-covering detection method based on deep neural network |
| KR20210033262A (en) * | 2019-09-18 | 2021-03-26 | 한국전력공사 | Diagnostic apparatus for environmental infringement of power line |
| CN113435484A (en) * | 2021-06-15 | 2021-09-24 | 兰州交通大学 | U-Net neural network ice lake extraction method combined with self-attention mechanism |
| CN114022671A (en) * | 2021-11-11 | 2022-02-08 | 中国矿业大学 | Intelligent ropeway cable position identification method based on residual error neural network |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11580646B2 (en) * | 2021-03-26 | 2023-02-14 | Nanjing University Of Posts And Telecommunications | Medical image segmentation method based on U-Net |
-
2022
- 2022-12-05 CN CN202211552526.1A patent/CN116000945B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106595551A (en) * | 2016-12-13 | 2017-04-26 | 武汉大学 | Icing thickness detection method for power transmission line icing image based on deep learning |
| RU2658344C1 (en) * | 2017-09-08 | 2018-06-20 | Сергей Сергеевич Дементьев | Overhead transmission lines wires ice loads control system |
| CN108345955A (en) * | 2017-11-24 | 2018-07-31 | 北京国网富达科技发展有限责任公司 | A kind of powerline ice-covering short term prediction method, device and storage medium |
| CN108360442A (en) * | 2018-02-26 | 2018-08-03 | 深圳市易成自动驾驶技术有限公司 | Intelligent snow-removing method, intelligent snow sweeper and computer readable storage medium |
| CN110070530A (en) * | 2019-04-19 | 2019-07-30 | 山东大学 | A kind of powerline ice-covering detection method based on deep neural network |
| KR20210033262A (en) * | 2019-09-18 | 2021-03-26 | 한국전력공사 | Diagnostic apparatus for environmental infringement of power line |
| CN113435484A (en) * | 2021-06-15 | 2021-09-24 | 兰州交通大学 | U-Net neural network ice lake extraction method combined with self-attention mechanism |
| CN114022671A (en) * | 2021-11-11 | 2022-02-08 | 中国矿业大学 | Intelligent ropeway cable position identification method based on residual error neural network |
Non-Patent Citations (1)
| Title |
|---|
| 基于朴素贝叶斯分类模型的车型识别方法;孙青;刘智勇;;五邑大学学报(自然科学版);20080815(第03期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116000945A (en) | 2023-04-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108647742B (en) | Rapid target detection method based on lightweight neural network | |
| Bouti et al. | A robust system for road sign detection and classification using LeNet architecture based on convolutional neural network | |
| CN113421269B (en) | Real-time semantic segmentation method based on double-branch deep convolutional neural network | |
| CN111612008B (en) | Image segmentation method based on convolution network | |
| CN110796168A (en) | Improved YOLOv 3-based vehicle detection method | |
| CN109886066A (en) | Fast target detection method based on the fusion of multiple dimensioned and multilayer feature | |
| CN112396587B (en) | Method for detecting congestion degree in bus compartment based on collaborative training and density map | |
| CN111582059A (en) | Facial expression recognition method based on variational self-encoder | |
| CN111680655A (en) | Video target detection method for aerial images of unmanned aerial vehicle | |
| CN104134068B (en) | Monitoring vehicle feature representation and classification method based on sparse coding | |
| CN113902915A (en) | Semantic segmentation method and system based on low-illumination complex road scene | |
| CN112232371B (en) | American license plate recognition method based on YOLOv3 and text recognition | |
| CN110321862B (en) | Pedestrian re-identification method based on compact ternary loss | |
| CN113269054B (en) | Aerial video analysis method based on space-time 2D convolutional neural network | |
| CN108427919B (en) | Unsupervised oil tank target detection method based on shape-guided saliency model | |
| CN114048822A (en) | Attention mechanism feature fusion segmentation method for image | |
| CN112329533A (en) | Local pavement adhesion coefficient estimation method based on image segmentation | |
| CN115240024A (en) | Method and system for segmenting extraterrestrial pictures by combining self-supervised learning and semi-supervised learning | |
| CN114581486A (en) | Template updating target tracking algorithm based on full convolution twin network multilayer characteristics | |
| CN116883650A (en) | Image-level weak supervision semantic segmentation method based on attention and local stitching | |
| CN114743126A (en) | Lane line sign segmentation method based on graph attention machine mechanism network | |
| CN114972952A (en) | Industrial part defect identification method based on model lightweight | |
| CN110751005B (en) | Pedestrian detection method integrating depth perception features and kernel extreme learning machine | |
| CN113792631A (en) | Aircraft detection and tracking method based on multi-scale self-adaption and side-domain attention | |
| CN116000945B (en) | Intelligent control method of cable deicing robot |
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 |