CN116109879B - Control system and control method of leveling machine - Google Patents
Control system and control method of leveling machine Download PDFInfo
- Publication number
- CN116109879B CN116109879B CN202310389735.7A CN202310389735A CN116109879B CN 116109879 B CN116109879 B CN 116109879B CN 202310389735 A CN202310389735 A CN 202310389735A CN 116109879 B CN116109879 B CN 116109879B
- Authority
- CN
- China
- Prior art keywords
- training
- feature vector
- classification
- image block
- vector
- 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 27
- 239000000463 material Substances 0.000 claims abstract description 71
- 239000013598 vector Substances 0.000 claims description 413
- 238000012549 training Methods 0.000 claims description 153
- 230000001419 dependent effect Effects 0.000 claims description 129
- 230000006870 function Effects 0.000 claims description 50
- 239000011159 matrix material Substances 0.000 claims description 47
- 238000003062 neural network model Methods 0.000 claims description 34
- 230000015654 memory Effects 0.000 claims description 31
- 238000012937 correction Methods 0.000 claims description 18
- 230000029087 digestion Effects 0.000 claims description 15
- 239000000546 pharmaceutical excipient Substances 0.000 claims description 13
- 238000012545 processing Methods 0.000 claims description 12
- 230000001629 suppression Effects 0.000 claims description 11
- 230000004927 fusion Effects 0.000 claims description 10
- 238000004364 calculation method Methods 0.000 claims description 9
- 230000005764 inhibitory process Effects 0.000 claims description 9
- 230000007787 long-term memory Effects 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 4
- 230000006403 short-term memory Effects 0.000 claims description 4
- 238000004513 sizing Methods 0.000 claims description 3
- 230000000694 effects Effects 0.000 abstract description 24
- 238000013135 deep learning Methods 0.000 abstract description 13
- 238000013473 artificial intelligence Methods 0.000 abstract description 7
- 238000009826 distribution Methods 0.000 description 16
- 238000013528 artificial neural network Methods 0.000 description 9
- 238000010586 diagram Methods 0.000 description 6
- 239000002002 slurry Substances 0.000 description 6
- 238000010276 construction Methods 0.000 description 3
- 238000000605 extraction Methods 0.000 description 3
- 229910052751 metal Inorganic materials 0.000 description 3
- 239000002184 metal Substances 0.000 description 3
- 230000002159 abnormal effect Effects 0.000 description 2
- 230000000903 blocking effect Effects 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 238000003058 natural language processing Methods 0.000 description 2
- BASFCYQUMIYNBI-UHFFFAOYSA-N platinum Chemical compound [Pt] BASFCYQUMIYNBI-UHFFFAOYSA-N 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- 238000013519 translation Methods 0.000 description 2
- 230000008034 disappearance Effects 0.000 description 1
- 229910052697 platinum Inorganic materials 0.000 description 1
- 238000003825 pressing Methods 0.000 description 1
- 238000010008 shearing Methods 0.000 description 1
- 238000012876 topography Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B21—MECHANICAL METAL-WORKING WITHOUT ESSENTIALLY REMOVING MATERIAL; PUNCHING METAL
- B21D—WORKING OR PROCESSING OF SHEET METAL OR METAL TUBES, RODS OR PROFILES WITHOUT ESSENTIALLY REMOVING MATERIAL; PUNCHING METAL
- B21D1/00—Straightening, restoring form or removing local distortions of sheet metal or specific articles made therefrom; Stretching sheet metal combined with rolling
- B21D1/02—Straightening, restoring form or removing local distortions of sheet metal or specific articles made therefrom; Stretching sheet metal combined with rolling by rollers
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/7715—Feature extraction, e.g. by transforming the feature space, e.g. multi-dimensional scaling [MDS]; Mappings, e.g. subspace methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
- G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- 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
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/02—Total factory control, e.g. smart factories, flexible manufacturing systems [FMS] or integrated manufacturing systems [IMS]
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Computing Systems (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Multimedia (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Mechanical Engineering (AREA)
- Feedback Control In General (AREA)
- Image Analysis (AREA)
Abstract
A control system of a leveling machine and a control method thereof are disclosed, which adaptively select a roll shaft of an appropriate diameter to improve leveling efficiency and effect based on the surface condition of a material to be leveled. Specifically, in the technical scheme of the application, an artificial intelligence algorithm based on deep learning is adopted to extract long-distance context semantic association information and medium-distance context semantic association information from the surface image of the material to be leveled, and the type label classification of the triaxial roller is carried out by fusing the characteristic information. In this way, the roller shaft having an appropriate diameter can be adaptively selected based on the actual condition of the surface of the operated object to improve the leveling efficiency and effect.
Description
Technical Field
The present application relates to the field of automation equipment, and more particularly to a control system for a leveling machine and a control method thereof.
Background
The leveling machine extrudes the uneven metal plate through the upper and lower rollers to achieve the leveling effect. The general leveling machine is provided with a plurality of roll shafts, for example, a three-shaft roll leveling machine, the construction efficiency is higher by adopting a larger shaft diameter, the flatness is better, but the surface slurry is easier to separate, and the slurry is thinner. The smaller shaft diameter is adopted, so that the pulp lifting effect is better, but the shaft is easy to deform, and correction should be noted.
Accordingly, a control system of a leveling machine is desired that can select a roller shaft having an appropriate diameter based on the surface condition of an object to be operated to improve the leveling effect.
Disclosure of Invention
The present application has been made in order to solve the above technical problems. Embodiments of the present application provide a control system of a leveling machine and a control method thereof, which adaptively select a roll shaft with an appropriate diameter to improve the leveling effect based on the surface condition of a material to be leveled. Specifically, in the technical solution of the present application, considering that this is essentially a classification problem, that is, for example, for a three-axis roll leveler, an artificial intelligence algorithm based on deep learning is adopted to extract long-distance context semantic association information and mid-distance context semantic association information from a surface image of a material to be leveled, and the feature information is fused to perform type tag classification of the three-axis roll. In this way, the roller shaft having an appropriate diameter can be adaptively selected based on the actual condition of the surface of the operated object to improve the leveling effect.
According to one aspect of the present application, there is provided a control system for a trowel, comprising: the surface image acquisition unit is used for acquiring a surface image of the material to be leveled; an image dividing unit for dividing the surface image of the material to be leveled into a two-dimensional image block sequence; the image block vectorization unit is used for enabling each two-dimensional image block in the two-dimensional image block sequence to pass through a linear embedding layer respectively so as to obtain a sequence of image block feature vectors; a first scale context coding unit, configured to pass the sequence of image block feature vectors through a context encoder based on a converter to obtain a plurality of image block semantic feature vectors; the cascade unit is used for cascading the plurality of image block semantic feature vectors to obtain long-distance dependent image feature vectors; the second scale context coding unit is used for enabling the sequence of the image block feature vectors to pass through a two-way long-short-term memory neural network model to obtain a middle-distance dependent image feature vector; the multi-scale fusion unit is used for fusing the long-distance dependent image feature vector and the middle-distance dependent image feature vector to obtain a classification feature vector; the vector correction unit is used for carrying out geometric constraint re-parameterization of a positive sizing space on the classification feature vector so as to obtain a corrected classification feature vector; and a leveling control result generating unit for passing the corrected classification feature vector through a classifier to obtain a classification result, wherein the classification result is used for representing the type label of the roll shaft.
In the control system of the leveling machine, the image dividing unit is further configured to uniformly divide the surface image of the material to be leveled into the two-dimensional image block sequence.
In the control system of the leveling machine, the image block vectorization unit is further configured to perform linear projection on each two-dimensional image block in the two-dimensional image block sequence with a learning embedding matrix by using the linear embedding layer to obtain the sequence of image block feature vectors.
In the control system of the above leveling machine, the first scale context encoding unit is further configured to: arranging the sequence of the image block feature vectors into an input vector; converting the input vector into a query vector and a key vector through a learning embedding matrix respectively; calculating the product between the query vector and the transpose vector of the key vector to obtain a self-attention correlation matrix; carrying out standardization processing on the self-attention association matrix to obtain a standardized self-attention association matrix; inputting the standardized self-attention association matrix into a Softmax activation function to activate so as to obtain a self-attention feature matrix; and multiplying the self-attention feature matrix with each image block feature vector in the sequence of image block feature vectors to obtain the plurality of image block semantic feature vectors.
In the control system of the above leveling machine, the multi-scale fusion unit is further configured to: fusing the long-distance dependent image feature vector and the medium-distance dependent image feature vector to obtain the classification feature vector by the following formula; wherein, the formula is:wherein (1)>Representing the long-range dependent image feature vector,representing the mid-range dependent image feature vector, < >>Representing a cascading function.
In the control system of the leveling machine, the vector correction unit is further configured to: performing geometric constraint re-parameterization of positive definite excipient space on the classification feature vector by the following formula to obtain a corrected classification feature vector; wherein the correction is commonThe formula is:wherein->Is the classification feature vector,/->And->Is feature set +.>Mean and variance of>Is a transpose of the classification feature vector,representing the square of the two norms of the vector, +.>Frobenius norms of the matrix are represented, < >>Is the classification feature vector->Is>Characteristic value of individual position->Is the +.f of the corrected classification feature vector>Characteristic values of the individual positions.
In the control system of the above leveling machine, the leveling control result generating unit is further configured to: using the classifier to classify the data according to the following formula Processing the corrected classification feature vector to generate the classification result; wherein, the formula is:wherein->For outputting result vector, ++>And->Respectively +.>Weights and bias vectors corresponding to the respective classifications, +.>An exponential operation representing a vector that represents a natural exponential function value that is a power of a eigenvalue of each position in the vector.
The control system of the leveling machine further comprises a training module for training the linear embedded layer, the context encoder based on the converter, the two-way long-short-term memory neural network model and the classifier.
In the control system of the leveling machine, the training module includes: the training device comprises a training surface image acquisition unit for training a material to be leveled, and a training data acquisition unit for acquiring training data, wherein the training data comprise training surface images of the material to be leveled and the true value of the type label of the roll shaft; the training image dividing unit is used for dividing the training surface image of the material to be leveled into a training two-dimensional image block sequence; the training image block vectorization unit is used for enabling each training two-dimensional image block in the training two-dimensional image block sequence to pass through the linear embedding layer respectively so as to obtain a sequence of training image block feature vectors; a training first scale context coding unit, configured to pass the sequence of training image block feature vectors through the converter-based context encoder to obtain a plurality of training image block semantic feature vectors; the training cascade unit is used for cascading the plurality of training image block semantic feature vectors to obtain training long-distance dependent image feature vectors; the training second scale context coding unit is used for enabling the sequence of the training image block feature vectors to pass through the two-way long-short-term memory neural network model to obtain the training middle-distance dependent image feature vectors; the training multi-scale fusion unit is used for fusing the training long-distance dependent image feature vector and the training medium-distance dependent image feature vector to obtain a training classification feature vector; a classification loss function value calculation unit, configured to pass the training classification feature vector through the classifier to obtain a classification loss function value; the classification mode digestion inhibition loss function value calculation unit is used for calculating the classification mode digestion inhibition loss function value of the training long-distance dependent image feature vector and the training medium-distance dependent image feature vector; and a training unit configured to perform training on the linear embedding layer, the converter-based context encoder, the two-way long-short-term memory neural network model, and the classifier as a loss function value by using a weighted sum of the classification loss function value and the classification loss function value.
In the control system of the above leveling machine, the classification mode digestion suppression loss function value calculation unit is further configured to: calculating the classification mode digestion inhibition loss function values of the training long-distance dependent image feature vector and the training medium-distance dependent image feature vector according to the following formula; wherein, the formula is:wherein->And->The training long-distance dependent image feature vector and the training medium-distance dependent image feature vector are respectively +.>And->The classifier is a weight matrix of the training long-distance dependent image feature vector and the training medium-distance dependent image feature vector respectively, < >>Representing the square of the two norms of the vector, +.>Frobenius norms of the matrix are represented, < >>Representing subtraction by position +.>Representing an exponential operation.
According to another aspect of the present application, there is also provided a control method of a leveling machine, including: acquiring a surface image of a material to be leveled; dividing the surface image of the material to be leveled into a two-dimensional image block sequence; respectively passing each two-dimensional image block in the two-dimensional image block sequence through a linear embedding layer to obtain a sequence of image block feature vectors; passing the sequence of image block feature vectors through a context encoder based on a converter to obtain a plurality of image block semantic feature vectors; cascading the plurality of image block semantic feature vectors to obtain a long-distance dependent image feature vector; the sequence of the image block feature vectors is processed through a two-way long-short-term memory neural network model to obtain a medium-distance dependent image feature vector; fusing the long-distance dependent image feature vector and the medium-distance dependent image feature vector to obtain a classification feature vector; performing geometric constraint re-parameterization of the positive assignment norm space on the classification feature vector to obtain a corrected classification feature vector; and passing the corrected classification feature vector through a classifier to obtain a classification result, wherein the classification result is used for representing the type label of the roll shaft.
Compared with the prior art, the control system and the control method of the leveling machine, provided by the application, are used for adaptively selecting the roll shaft with the proper diameter to improve the leveling effect based on the surface condition of the material to be leveled. Specifically, in the technical solution of the present application, considering that this is essentially a classification problem, that is, for example, for a three-axis roll leveler, an artificial intelligence algorithm based on deep learning is adopted to extract long-distance context semantic association information and mid-distance context semantic association information from a surface image of a material to be leveled, and the feature information is fused to perform type tag classification of the three-axis roll. In this way, the roller shaft having an appropriate diameter can be adaptively selected based on the actual condition of the surface of the operated object to improve the leveling effect.
Drawings
The foregoing and other objects, features and advantages of the present application will become more apparent from the following more particular description of embodiments of the present application, as illustrated in the accompanying drawings. The accompanying drawings are included to provide a further understanding of embodiments of the application and are incorporated in and constitute a part of this specification, illustrate the application and not constitute a limitation to the application. In the drawings, like reference numerals generally refer to like parts or steps.
FIG. 1 illustrates a block diagram of a control system of a screed according to an embodiment of the present application.
FIG. 2 illustrates a system architecture diagram of a control system of a screed according to an embodiment of the present application.
FIG. 3 illustrates a block diagram of a training module in a control system of a screed according to an embodiment of the present application.
Fig. 4 illustrates a flowchart of a control method of the screed according to an embodiment of the present application.
Fig. 5 illustrates a flowchart of a training phase in a control method of a screed according to an embodiment of the present application.
Detailed Description
Hereinafter, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. It should be apparent that the described embodiments are only some of the embodiments of the present application and not all of the embodiments of the present application, and it should be understood that the present application is not limited by the example embodiments described herein.
Summary of the application: as described above, the leveling machine is a machine for leveling an uneven metal plate by pressing a strip or plate having a certain thickness by upper and lower rolls. The general leveling machine is provided with a plurality of roll shafts, for example, a three-shaft roll leveling machine, the construction efficiency is higher by adopting a larger shaft diameter, the flatness is better, but the surface slurry is easier to separate, and the slurry is thinner. The smaller shaft diameter is adopted, so that the pulp lifting effect is better, but the shaft is easy to deform, and correction should be noted. Accordingly, a control system of a leveling machine is desired that can select a roller shaft having an appropriate diameter based on the surface condition of an object to be operated to improve the leveling effect.
At present, deep learning and neural networks have been widely used in the fields of computer vision, natural language processing, speech signal processing, and the like. In addition, deep learning and neural networks have also shown levels approaching and even exceeding humans in the fields of image classification, object detection, semantic segmentation, text translation, and the like.
In recent years, deep learning and development of a neural network provide new solutions and schemes for intelligent roll shaft selection of a leveling machine.
Accordingly, in the technical solution of the present application, it is considered that the selection of the roll shafts for different types needs to be performed according to the surface condition of the operated object, that is, the roll shaft with a suitable diameter is adaptively selected based on the surface condition of the material to be leveled to improve the leveling effect. Specifically, in the technical solution of the present application, considering that this is essentially a classification problem, that is, for example, for a three-axis roll leveler, an artificial intelligence algorithm based on deep learning is adopted to extract long-distance context semantic association information and mid-distance context semantic association information from a surface image of a material to be leveled, and the feature information is fused to perform type tag classification of the three-axis roll. In this way, the roller shaft having an appropriate diameter can be adaptively selected based on the actual condition of the surface of the operated object to improve the leveling effect.
Specifically, in the technical scheme of the application, first, a surface image of a material to be leveled is acquired. The surface image of the material to be leveled is then divided into a two-dimensional sequence of image blocks. That is, the image of the surface of the material to be leveled is diced to obtain non-overlapping and fixed-size image blocks, wherein each image block represents the surface state of each local area of the material to be leveled.
And then, respectively passing each two-dimensional image block in the two-dimensional image block sequence through a linear embedding layer to convert each two-dimensional image block into a one-dimensional embedding vector so as to obtain a sequence of image block characteristic vectors, wherein the linear embedding layer linearly projects each two-dimensional image block in the two-dimensional image block sequence into a one-dimensional embedding vector through a learnable embedding matrix, and the one-dimensional embedding vector represents an embedding vector representation of the surface state of each local area of the material to be leveled. In other words, in the technical scheme of the application, the linear embedded layer is located at the front end of the network, the blocking operation is performed on the defect image, so that the non-overlapped image blocks with fixed sizes can be obtained, mapped into the embedded vectors, and then the category vectors and the position codes are added.
Further, the sequence of image block feature vectors may be passed through a context encoder based on a converter to obtain a plurality of image block semantic feature vectors, and the plurality of image block semantic feature vectors may be concatenated to obtain a long-range dependent image feature vector. That is, based on the transformer concept, the converter is used to capture the long-range context dependent characteristic, and the global context-based semantic coding is performed on each image block feature vector in the sequence of image block feature vectors to obtain a context semantic association feature representation with the overall semantic association of the sequence of image block feature vectors as a context, i.e. the plurality of image block semantic feature vectors. It should be appreciated that in the context of the present application, a long-range contextual semantic association feature representation of the surface state features of the respective localized areas of the material to be leveled relative to the overall surface state features of the material to be leveled may be captured by the transducer-based context encoder.
And then, the sequence of the image block feature vectors passes through a two-way long-short-term memory neural network model to obtain the medium-distance dependent image feature vectors, namely, the two-way long-term memory neural network model is utilized to capture the medium-distance context dependent characteristics. It should be understood that in the technical solution of the present application, the mid-distance context semantic association feature representation of the surface state features of each local area of the material to be leveled relative to the surface state features of other local areas of the material to be leveled may be captured by the two-way long-short term memory neural network model.
And then fusing the long-distance dependent image feature vector and the medium-distance dependent image feature vector to obtain the overall feature distribution information of the surface state of the material to be leveled, thereby obtaining the classification feature vector. Further, the corrected classification feature vector is passed through a classifier to obtain a classification result, and the classification result is used for representing the type label of the roll shaft. That is, the classification processing is performed with the classification feature vector, and a classification result for representing the type tag of the roll shaft can be obtained.
In particular, in the technical solution of the present application, since the classification feature vector is obtained by fusing the long-distance dependent image feature vector and the intermediate-distance dependent image feature vector, when the gradient of the loss function is calculated and counter-propagated from the classifier to the model in the model training, the gradient may pass through the context encoder based on the converter and the two-way long-short-term memory neural network model, respectively, and at this time, the resolution of the feature pattern extracted by the context encoder based on the converter and the two-way long-short-term memory neural network model may be caused due to abnormal gradient branching.
Thus, in addition to the classification loss function, a classification mode digestion inhibition loss function is further introduced to address digestion of the extracted feature mode, specifically expressed as:,/>and->The long-distance dependent image feature vector and the medium-distance dependent image feature vector, respectively, and +.>And->Respectively for the classifierAnd->Weight matrix of>Representing the square of the two norms of the vector.
That is, by introducing a classification mode digestion suppression loss function, the pseudo-difference of classifier weights is pushed to the true feature distribution difference of the long-distance dependent image feature vector and the medium-distance dependent image feature vector, so that the directional derivative in the process of gradient back propagation is guaranteed to be regularized near a gradient branching point, that is, the gradient is subjected to over-weighting between the context encoder based on the converter and the two-way long-short-term memory neural network model, so that the classification mode digestion of the features is suppressed, and the extraction capacity of classification features of the context encoder based on the converter and the two-way long-short-term memory neural network model is improved, so that the accuracy of classification results of the classification feature vectors is correspondingly improved.
In particular, in the inference stage, when the classification feature vector is obtained by fusing the long-distance-dependent image feature vector and the intermediate-distance-dependent image feature vector, in order to make full use of the long-distance-dependent information and the intermediate-distance-dependent information of the image feature vector, the classification feature vector is preferably obtained by directly concatenating the long-distance-dependent image feature vector and the intermediate-distance-dependent image feature vector, but in the case of direct concatenation, the explicit difference of the feature distributions of each of the long-distance-dependent image feature vector and the intermediate-distance-dependent image feature vector still causes a problem of discretization of the overall feature distribution of the classification feature vector, so that the classification feature vector has poor convergence with respect to the probability of a predetermined target class when classification regression is performed by the classifier, which may affect the training speed of the classifier and the accuracy of the classification result.
Thus, in the solution of the present application, the classification feature vector is, for example, denoted asPerforming geometric constraint re-parameterization of a positive-localization space, wherein the geometric constraint re-parameterization specifically comprises the following steps: />,/>And->Is feature set +.>Mean and variance of>Representing the square of the two norms of the vector, +. >Is a transpose of the classification feature vector, < >>Frobenius norms of the matrix are represented, < >>And->Said classification feature vector +.>Is>Characteristic value of individual position, and->Is in the form of a row vector.
Here, the classification feature vectorThe geometric constrained repartitioning of the forward-defined excipient space of (2) may be based on a projection modulo length relation of the Bessel inequality by projecting the square of the vector norm expressed in the form of an inner product within the associated set space of vectors themselves such that the set of distributions of vectors has modulo length constraints within the geometric metric subspace of the forward-defined excipient space to repartitionize the distribution space to a bounded forward-defined excipient space having a closed subspace based on the geometric constraints of the feature distribution. Thus, the classification feature vector +.>The overall characteristic distribution has convergence relative to the target class probability value in the classification domain, so that the training speed of the classifier and the accuracy of the classification result are improved. In this way, the roller shaft having an appropriate diameter can be adaptively selected based on the actual condition of the surface of the operated object to improve the leveling effect.
Based on this, the present application proposes a control system of a screed comprising: the surface image acquisition unit is used for acquiring a surface image of the material to be leveled; an image dividing unit for dividing the surface image of the material to be leveled into a two-dimensional image block sequence; the image block vectorization unit is used for enabling each two-dimensional image block in the two-dimensional image block sequence to pass through a linear embedding layer respectively so as to obtain a sequence of image block feature vectors; a first scale context coding unit, configured to pass the sequence of image block feature vectors through a context encoder based on a converter to obtain a plurality of image block semantic feature vectors; the cascade unit is used for cascading the plurality of image block semantic feature vectors to obtain long-distance dependent image feature vectors; the second scale context coding unit is used for enabling the sequence of the image block feature vectors to pass through a two-way long-short-term memory neural network model to obtain a middle-distance dependent image feature vector; the multi-scale fusion unit is used for fusing the long-distance dependent image feature vector and the middle-distance dependent image feature vector to obtain a classification feature vector; the vector correction unit is used for carrying out geometric constraint re-parameterization of a positive sizing space on the classification feature vector so as to obtain a corrected classification feature vector; and a leveling control result generating unit for passing the corrected classification feature vector through a classifier to obtain a classification result, wherein the classification result is used for representing the type label of the roll shaft.
Having described the basic principles of the present application, various non-limiting embodiments of the present application will now be described in detail with reference to the accompanying drawings.
Exemplary System: FIG. 1 illustrates a block diagram of a control system of a screed according to an embodiment of the present application. As shown in fig. 1, a control system 100 of a trowel according to an embodiment of the present application includes: a material surface image acquisition unit 110 for acquiring a surface image of the material to be leveled; an image dividing unit 120 for dividing the surface image of the material to be leveled into a two-dimensional image block sequence; an image block vectorization unit 130, configured to pass each two-dimensional image block in the two-dimensional image block sequence through a linear embedding layer to obtain a sequence of image block feature vectors; a first scale context encoding unit 140 for passing the sequence of image block feature vectors through a converter-based context encoder to obtain a plurality of image block semantic feature vectors; a concatenation unit 150, configured to concatenate the plurality of image block semantic feature vectors to obtain a long-distance dependent image feature vector; a second scale context encoding unit 160, configured to pass the sequence of image block feature vectors through a two-way long-short term memory neural network model to obtain a medium-distance dependent image feature vector; a multi-scale fusion unit 170 for fusing the long-distance dependent image feature vector and the middle-distance dependent image feature vector to obtain a classification feature vector; a vector correction unit 180, configured to perform geometric constraint re-parameterization of the normal vector space on the classification feature vector to obtain a corrected classification feature vector; and a leveling control result generating unit 190 for passing the corrected classification feature vector through a classifier to obtain a classification result for representing a type tag of the roll shaft.
FIG. 2 illustrates a system architecture diagram of a control system of a screed according to an embodiment of the present application. In this system architecture, as shown in fig. 2, a surface image of a material to be leveled is first acquired and divided into a two-dimensional image block sequence. And then, respectively passing each two-dimensional image block in the two-dimensional image block sequence through a linear embedding layer to obtain a sequence of image block feature vectors. Then, the sequence of image block feature vectors is passed through a context encoder based on a converter to obtain a plurality of image block semantic feature vectors, and the plurality of image block semantic feature vectors are concatenated to obtain a long-range dependent image feature vector. And then, the sequence of the image block feature vectors is passed through a two-way long-short-term memory neural network model to obtain the medium-distance dependent image feature vectors. Then, fusing the long-distance dependent image feature vector and the medium-distance dependent image feature vector to obtain a classification feature vector, and performing geometric constraint re-parameterization of a positive excipient space on the classification feature vector to obtain a corrected classification feature vector; and then, the corrected classification feature vector is passed through a classifier to obtain a classification result, wherein the classification result is used for representing the type label of the roll shaft.
In the control system 100 of the leveling machine, the surface image acquisition unit 110 is configured to acquire a surface image of the material to be leveled. The leveling machine is used for extruding the uneven metal plate through the upper roller and the lower roller to achieve the leveling effect. For example, the special three-in-one servo feeding leveling machine for the thin plate in the LMC2 can process materials to be leveled with the thickness of 0.3-3.2 mm, can be matched with hydraulic shearing to cut the materials to be leveled, and can be used in industries such as household appliances, automobiles, air conditioner platinum and the like.
The general leveling machine is provided with a plurality of roller shafts, for example, a three-roller leveling machine, the construction efficiency is higher by adopting a larger shaft diameter, the flatness is better, but the surface slurry is easier to separate, and the slurry is thinner. The smaller shaft diameter is adopted, so that the pulp lifting effect is better, but the shaft is easy to deform, and correction should be noted. Accordingly, a control system of a leveling machine is desired that can select a roller shaft having an appropriate diameter based on the surface condition of an object to be operated to improve the leveling effect. At present, deep learning and neural networks have been widely used in the fields of computer vision, natural language processing, speech signal processing, and the like. In addition, deep learning and neural networks have also shown levels approaching and even exceeding humans in the fields of image classification, object detection, semantic segmentation, text translation, and the like. In recent years, deep learning and development of a neural network provide new solutions and schemes for intelligent roll shaft selection of a leveling machine.
Accordingly, in the technical solution of the present application, it is considered that the selection of the roll shafts for different types needs to be performed according to the surface condition of the operated object, that is, the roll shaft with a suitable diameter is adaptively selected based on the surface condition of the material to be leveled to improve the leveling effect. Specifically, in the technical solution of the present application, considering that this is essentially a classification problem, that is, for example, for a control system of a three-axis roller leveling machine, an artificial intelligence algorithm based on deep learning is adopted to extract long-distance context semantic association information and mid-distance context semantic association information from a surface image of a material to be leveled, and the feature information is fused to perform type tag classification of the three-axis roller. In this way, the roller shaft having an appropriate diameter can be adaptively selected based on the actual condition of the surface of the operated object to improve the leveling effect. Specifically, in the technical scheme of the application, first, a surface image of a material to be leveled is acquired. Wherein the surface image of the material to be leveled may be acquired by a camera disposed on a control system of the screed.
In the control system 100 of the leveling machine, the image dividing unit 120 is configured to divide the surface image of the material to be leveled into a two-dimensional image block sequence. That is, the image of the surface of the material to be leveled is diced to obtain non-overlapping and fixed-size image blocks, wherein each image block represents the surface state of each local area of the material to be leveled. In this way, provision is made for feature extraction of the medium-long distance dependent image of the material image to be flattened.
Specifically, in the embodiment of the present application, the image dividing unit 120 is further configured to uniformly divide the surface image of the material to be leveled into the two-dimensional image block sequence.
In the control system 100 of the above-mentioned leveling machine, the image block vectorization unit 130 is configured to pass each two-dimensional image block in the two-dimensional image block sequence through a linear embedding layer to obtain a sequence of image block feature vectors. That is, each two-dimensional image block in the two-dimensional image block sequence is respectively passed through a linear embedding layer to convert each two-dimensional image block into a one-dimensional embedding vector to obtain a sequence of image block feature vectors, wherein the linear embedding layer linearly projects each two-dimensional image block in the two-dimensional image block sequence into a one-dimensional embedding vector through a learning embedding matrix, and the one-dimensional embedding vector represents an embedding vector representation of the surface state of each local area of the material to be leveled. In other words, in the technical scheme of the application, the linear embedded layer is located at the front end of the network, the blocking operation is performed on the defect image, so that the non-overlapped image blocks with fixed sizes can be obtained, mapped into the embedded vectors, and then the category vectors and the position codes are added.
Specifically, in the embodiment of the present application, the image block vectorization unit 130 is further configured to use the linear embedding layer to perform linear projection on each two-dimensional image block in the two-dimensional image block sequence with a learning embedding matrix to obtain the sequence of image block feature vectors.
In the control system 100 of the above-mentioned leveling machine, the first scale context encoding unit 140 is configured to pass the sequence of image block feature vectors through a context encoder based on a converter to obtain a plurality of image block semantic feature vectors. That is, based on the transformer concept, the converter is used to capture the long-range context dependent characteristic, and the global context-based semantic coding is performed on each image block feature vector in the sequence of image block feature vectors to obtain a context semantic association feature representation with the overall semantic association of the sequence of image block feature vectors as a context, i.e. the plurality of image block semantic feature vectors. It should be appreciated that in the context of the present application, a long-range contextual semantic association feature representation of the surface state features of the respective localized areas of the material to be leveled relative to the overall surface state features of the material to be leveled may be captured by the transducer-based context encoder.
Specifically, in the embodiment of the present application, the first scale context encoding unit 140 is further configured to: arranging the sequence of the image block feature vectors into an input vector; converting the input vector into a query vector and a key vector through a learning embedding matrix respectively; calculating the product between the query vector and the transpose vector of the key vector to obtain a self-attention correlation matrix; carrying out standardization processing on the self-attention association matrix to obtain a standardized self-attention association matrix; inputting the standardized self-attention association matrix into a Softmax activation function to activate so as to obtain a self-attention feature matrix; and multiplying the self-attention feature matrix with each image block feature vector in the sequence of image block feature vectors to obtain the plurality of image block semantic feature vectors.
In the control system 100 of the above-mentioned leveling machine, the cascade unit 150 is configured to cascade the plurality of image block semantic feature vectors to obtain a long-distance dependent image feature vector. That is, the plurality of image semantic feature vectors are concatenated to integrate long-range contextual semantic-related feature representations of the surface state features of the respective localized regions relative to the overall surface state features of the material to be leveled.
In the control system 100 of the above-mentioned leveling machine, the second scale context encoding unit 160 is configured to pass the sequence of image block feature vectors through a two-way long-short term memory neural network model to obtain the medium-distance dependent image feature vector. That is, the two-way long and short term memory neural network model is utilized to capture mid-range context dependent characteristics. It should be understood that in the technical solution of the present application, the mid-distance context semantic association feature representation of the surface state features of each local area of the material to be leveled relative to the surface state features of other local areas of the material to be leveled may be captured by the two-way long-short term memory neural network model. The two-way long-short-term memory neural network (BiLSTM) is formed by adopting a two-way connection mode for cell structures of the long-short-term memory neural network (LSTM), so that the gradient disappearance problem of the traditional circulating neural network (RNN) is solved, and the integrity of the extracted information is ensured.
In the control system 100 of the above-mentioned leveling machine, the multi-scale fusion unit 170 is configured to fuse the long-distance dependent image feature vector and the medium-distance dependent image feature vector to obtain a classification feature vector. In this way, the long-distance dependent image feature vector and the middle-distance dependent image feature vector are fused to obtain the overall feature distribution information of the surface state of the material to be leveled, so as to obtain the classification feature vector.
Specifically, in the present embodiment, the multi-scale fusion unit 170 is further configured to: fusing the long-distance dependent image feature vector and the medium-distance dependent image feature vector to obtain the classification feature vector by the following formula; wherein, the formula is:wherein->Representing the long-range dependent image feature vector, < >>Representing the mid-range dependent image feature vector, < >>Representing a cascading function.
When the long-distance-dependent image feature vector and the middle-distance-dependent image feature vector are fused to obtain the classification feature vector, in order to make full use of the long-distance-dependent information and the middle-distance-dependent information of the image feature vector, the classification feature vector is preferably obtained by directly concatenating the long-distance-dependent image feature vector and the middle-distance-dependent image feature vector, but in the case of direct concatenation, the explicit difference of the feature distributions of the long-distance-dependent image feature vector and the middle-distance-dependent image feature vector still causes the problem of discretization of the overall feature distribution of the classification feature vector, so that the classification feature vector has poor convergence with respect to a predetermined target class probability when classified regression is performed through a classifier, and thus the training speed of the classifier and the accuracy of the classification result are affected.
In the control system 100 of the above-mentioned leveling machine, the vector correction unit 180 is configured to perform geometric constraint re-parameterization of the normal-localization space on the classification feature vector to obtain a corrected classification feature vector.
Specifically, in the embodiment of the present application, the vector correction unit 180 is further configured to: performing geometric constraint re-parameterization of positive definite excipient space on the classification feature vector by the following formula to obtain a corrected classification feature vector; wherein, the correction formula is:wherein->Is the classification feature vector,/->And->Is feature set +.>Mean and variance of>Is a transpose of the classification feature vector,representing the square of the two norms of the vector, +.>Frobenius norms of the matrix are represented, < >>Is the classification feature vector->Is>Characteristic value of individual position->Is the +.f of the corrected classification feature vector>Characteristic values of the individual positions.
Here, the classification feature vectorThe geometric constrained repartitioning of the forward-defined excipient space of (2) may be based on a projection modulo length relation of the Bessel inequality by projecting the square of the vector norm expressed in the form of an inner product within the associated set space of vectors themselves such that the set of distributions of vectors has modulo length constraints within the geometric metric subspace of the forward-defined excipient space to repartitionize the distribution space to a bounded forward-defined excipient space having a closed subspace based on the geometric constraints of the feature distribution. Thus, the classification feature vector +. >The overall characteristic distribution has convergence relative to the target class probability value in the classification domain, so that the training speed of the classifier and the accuracy of the classification result are improved.
In the control system 100 of the above-mentioned leveling machine, the leveling control result generating unit 190 is configured to pass the corrected classification feature vector through a classifier to obtain a classification result, where the classification result is used to represent a type tag of the roll shaft. That is, the classification processing is performed with the classification feature vector, and a classification result for representing the type tag of the roll shaft can be obtained.
Specifically, in the embodiment of the present application, the leveling control result generating unit 190 is further configured to: processing the corrected classification feature vector using the classifier in the following formula to generate the classification result; wherein, the formula is:wherein->For outputting result vector, ++>And->Respectively +.>Weights and bias vectors corresponding to the respective classifications, +.>An exponential operation representing a vector that represents a natural exponential function value that is a power of a eigenvalue of each position in the vector.
In the control system 100 of the above-described screed, a training module 200 for training the linear embedded layer, the converter-based context encoder, the two-way long and short term memory neural network model, and the classifier is further included.
FIG. 3 illustrates a block diagram of a training module in a control system of a screed according to an embodiment of the present application. As shown in fig. 3, the training module 200 includes: a training surface image acquisition unit 210 for acquiring training data including a training surface image of the material to be leveled and a true value of a type tag of the roller shaft; a training image dividing unit 220 for dividing the training surface image of the material to be leveled into a training two-dimensional image block sequence; a training image block vectorization unit 230, configured to pass each training two-dimensional image block in the training two-dimensional image block sequence through the linear embedding layer to obtain a sequence of training image block feature vectors; a training first scale context encoding unit 240, configured to pass the sequence of training image block feature vectors through the converter-based context encoder to obtain a plurality of training image block semantic feature vectors; a training concatenation unit 250, configured to concatenate the plurality of training image block semantic feature vectors to obtain a training long-distance dependent image feature vector; a training second scale context encoding unit 260, configured to pass the sequence of training image block feature vectors through the two-way long-short-term memory neural network model to obtain a training mid-distance dependent image feature vector; a training multi-scale fusion unit 270, configured to fuse the training long-distance dependent image feature vector and the training medium-distance dependent image feature vector to obtain a training classification feature vector; a classification loss function value calculation unit 280, configured to pass the training classification feature vector through the classifier to obtain a classification loss function value; a classification mode cancellation suppression loss function value calculation unit 290 configured to calculate a classification mode cancellation suppression loss function value of the training long-distance dependent image feature vector and the training medium-distance dependent image feature vector; and a training unit 300 for training the linear embedding layer, the converter-based context encoder, the two-way long and short term memory neural network model, and the classifier as a weighted sum of the classification loss function value and the suppression loss function value in the classification mode.
In particular, in the technical solution of the present application, since the classification feature vector is obtained by fusing the long-distance dependent image feature vector and the intermediate-distance dependent image feature vector, when the gradient of the loss function is calculated and counter-propagated from the classifier to the model in the model training, the gradient may pass through the context encoder based on the converter and the two-way long-short-term memory neural network model, respectively, and at this time, the resolution of the feature pattern extracted by the context encoder based on the converter and the two-way long-short-term memory neural network model may be caused due to abnormal gradient branching. Thus, in addition to the classification loss function, a classification mode resolution suppression loss function is further introduced to address resolution of the extracted feature mode.
Specifically, in the embodiment of the present application, the classification mode resolution suppression loss function value calculation unit 290 is further configured to: calculating the classification mode digestion inhibition loss function values of the training long-distance dependent image feature vector and the training medium-distance dependent image feature vector according to the following formula; wherein, the formula is: Wherein->And->The training long-distance dependent image feature vector and the training medium-distance dependent image feature vector are respectively +.>And->The classifier is a weight matrix of the training long-distance dependent image feature vector and the training medium-distance dependent image feature vector respectively, < >>Representation ofSquare of two norms of vector, +.>Frobenius norms of the matrix are represented, < >>Representing subtraction by position +.>Representing an exponential operation.
That is, by introducing a classification mode digestion suppression loss function, the pseudo-difference of classifier weights is pushed to the true feature distribution difference of the long-distance dependent image feature vector and the medium-distance dependent image feature vector, so that the directional derivative in the process of gradient back propagation is guaranteed to be regularized near a gradient branching point, that is, the gradient is subjected to over-weighting between the context encoder based on the converter and the two-way long-short-term memory neural network model, so that the classification mode digestion of the features is suppressed, and the extraction capacity of classification features of the context encoder based on the converter and the two-way long-short-term memory neural network model is improved, so that the accuracy of classification results of the classification feature vectors is correspondingly improved. In this way, the roller shaft having an appropriate diameter can be adaptively selected based on the actual condition of the surface of the operated object to improve the leveling effect.
In summary, the control system 100 of the trowel according to embodiments of the present application is illustrated that adaptively selects a roller of an appropriate diameter to enhance the flattening effect based on the topography of the material to be flattened. Specifically, in the technical solution of the present application, considering that this is essentially a classification problem, that is, for example, for a three-axis roll leveler, an artificial intelligence algorithm based on deep learning is adopted to extract long-distance context semantic association information and mid-distance context semantic association information from a surface image of a material to be leveled, and the feature information is fused to perform type tag classification of the three-axis roll. In this way, the roller shaft having an appropriate diameter can be adaptively selected based on the actual condition of the surface of the operated object to improve the leveling effect.
An exemplary method is: fig. 4 illustrates a flowchart of a control method of the screed according to an embodiment of the present application. As shown in fig. 4, the control method of the leveling machine according to the embodiment of the application includes the steps of: s110, acquiring a surface image of a material to be leveled; s120, dividing the surface image of the material to be leveled into a two-dimensional image block sequence; s130, respectively passing each two-dimensional image block in the two-dimensional image block sequence through a linear embedding layer to obtain a sequence of image block feature vectors; s140, passing the sequence of image block feature vectors through a context encoder based on a converter to obtain a plurality of image block semantic feature vectors; s150, cascading the plurality of image block semantic feature vectors to obtain a long-distance dependent image feature vector; s160, passing the sequence of the image block feature vectors through a two-way long-short-term memory neural network model to obtain a medium-distance dependent image feature vector; s170, fusing the long-distance dependent image feature vector and the medium-distance dependent image feature vector to obtain a classification feature vector; s180, performing geometric constraint re-parameterization of a positive definite excipient space on the classification feature vector to obtain a corrected classification feature vector; and S190, passing the corrected classification feature vector through a classifier to obtain a classification result, wherein the classification result is used for representing the type label of the roll shaft.
In one example, in the control method of the leveling machine, the dividing the surface image of the material to be leveled into a two-dimensional image block sequence includes: and uniformly dividing the surface image of the material to be leveled into the two-dimensional image block sequence.
In an example, in the control method of the leveling machine, the step of passing each two-dimensional image block in the two-dimensional image block sequence through a linear embedding layer to obtain a sequence of image block feature vectors includes: and linearly projecting each two-dimensional image block in the two-dimensional image block sequence by using the linear embedding layer in a learning embedding matrix to obtain the sequence of the image block feature vectors.
In one example, in the control method of the above leveling machine, the step of passing the sequence of image block feature vectors through a context encoder based on a converter to obtain a plurality of image block semantic feature vectors includes: arranging the sequence of the image block feature vectors into an input vector; converting the input vector into a query vector and a key vector through a learning embedding matrix respectively; calculating the product between the query vector and the transpose vector of the key vector to obtain a self-attention correlation matrix; carrying out standardization processing on the self-attention association matrix to obtain a standardized self-attention association matrix; inputting the standardized self-attention association matrix into a Softmax activation function to activate so as to obtain a self-attention feature matrix; and multiplying the self-attention feature matrix with each image block feature vector in the sequence of image block feature vectors to obtain the plurality of image block semantic feature vectors.
In one example, in the control method of the above leveling machine, the fusing the long-distance dependent image feature vector and the medium-distance dependent image feature vector to obtain the classification feature vector includes: fusing the long-distance dependent image feature vector and the medium-distance dependent image feature vector to obtain the classification feature vector by the following formula; wherein, the formula is:wherein->Representing the long-range dependent image feature vector, < >>Representing the mid-range dependent image feature vector, < >>Representing a cascading function.
In one example, in the control method of the above leveling machine, performing geometric constraint re-parameterization of the classification feature vector in a positive excipient space to obtain a corrected classification feature vector includes: the classification characteristic is oriented by the following formulaPerforming geometric constraint re-parameterization of the normal vector space to obtain corrected classification feature vectors; wherein, the correction formula is:wherein->Is the classification feature vector,/->Andis feature set +.>Mean and variance of>Is a transpose of the classification feature vector, < >>Representing the square of the two norms of the vector, +. >Frobenius norms of the matrix are represented, < >>Is the classification feature vector->Is>Characteristic value of individual position->Is the +.f of the corrected classification feature vector>Characteristic values of the individual positions.
In one ofIn an example, in the control method of the above leveling machine, the step of passing the corrected classification feature vector through a classifier to obtain a classification result includes: processing the corrected classification feature vector using the classifier in the following formula to generate the classification result; wherein, the formula is:wherein->For outputting result vector, ++>And->Respectively +.>Weights and bias vectors corresponding to the respective classifications, +.>An exponential operation representing a vector that represents a natural exponential function value that is a power of a eigenvalue of each position in the vector.
In one example, the control method of the leveling machine further includes a training phase for training the linear embedded layer, the context encoder based on the converter, the two-way long-short term memory neural network model and the classifier.
Fig. 5 illustrates a flowchart of a training phase in a control method of a screed according to an embodiment of the present application. As shown in fig. 5, the training phase includes the steps of: s210, acquiring training data, wherein the training data comprises training surface images of materials to be leveled and true values of type labels of the roll shafts; s220, dividing the training surface image of the material to be leveled into a training two-dimensional image block sequence; s230, respectively passing each training two-dimensional image block in the training two-dimensional image block sequence through the linear embedding layer to obtain a sequence of training image block feature vectors; s240, passing the sequence of training image block feature vectors through the context encoder based on the converter to obtain a plurality of training image block semantic feature vectors; s250, cascading the plurality of training image block semantic feature vectors to obtain training long-distance dependent image feature vectors; s260, passing the sequence of the training image block feature vectors through the two-way long-short-term memory neural network model to obtain a distance dependent image feature vector in training; s270, fusing the training long-distance dependent image feature vector and the training medium-distance dependent image feature vector to obtain a training classification feature vector; s280, passing the training classification feature vector through the classifier to obtain a classification loss function value; s290, calculating a classification mode digestion inhibition loss function value of the training long-distance dependent image feature vector and the training medium-distance dependent image feature vector; and S300, training the linear embedded layer, the context encoder based on the converter, the two-way long-short-term memory neural network model and the classifier by taking the weighted sum of the classification loss function value and the suppression loss function value as the loss function value.
In summary, a control method of the trowel according to embodiments of the present application is illustrated that adaptively selects a roll shaft of an appropriate diameter to enhance a flattening effect based on a surface condition of a material to be flattened. Specifically, in the technical solution of the present application, considering that this is essentially a classification problem, that is, for example, for a three-axis roll leveler, an artificial intelligence algorithm based on deep learning is adopted to extract long-distance context semantic association information and mid-distance context semantic association information from a surface image of a material to be leveled, and the feature information is fused to perform type tag classification of the three-axis roll. In this way, the roller shaft having an appropriate diameter can be adaptively selected based on the actual condition of the surface of the operated object to improve the leveling effect.
Claims (9)
1. A control system for a troweling machine, comprising:
the surface image acquisition unit is used for acquiring a surface image of the material to be leveled;
an image dividing unit for dividing the surface image of the material to be leveled into a two-dimensional image block sequence;
the image block vectorization unit is used for enabling each two-dimensional image block in the two-dimensional image block sequence to pass through a linear embedding layer respectively so as to obtain a sequence of image block feature vectors;
A first scale context coding unit, configured to pass the sequence of image block feature vectors through a context encoder based on a converter to obtain a plurality of image block semantic feature vectors;
the cascade unit is used for cascading the plurality of image block semantic feature vectors to obtain long-distance dependent image feature vectors;
the second scale context coding unit is used for enabling the sequence of the image block feature vectors to pass through a two-way long-short-term memory neural network model to obtain a middle-distance dependent image feature vector;
the multi-scale fusion unit is used for fusing the long-distance dependent image feature vector and the middle-distance dependent image feature vector to obtain a classification feature vector;
the vector correction unit is used for carrying out geometric constraint re-parameterization of a positive sizing space on the classification feature vector so as to obtain a corrected classification feature vector; and
the leveling control result generation unit is used for enabling the corrected classification feature vector to pass through a classifier to obtain a classification result, wherein the classification result is used for representing the type label of the roll shaft;
wherein the vector correction unit is further configured to: performing geometric constraint re-parameterization of positive definite excipient space on the classification feature vector by using the following correction formula to obtain a corrected classification feature vector;
Wherein, the correction formula is:
wherein,,is the classification feature vector,/->And->Is feature set +.>Mean and variance of>Is a transpose of the classification feature vector, < >>Representing the square of the two norms of the vector, +.>Frobenius norms of the matrix are represented, < >>Is the classification feature vector->Is>Characteristic value of individual position->Is the +.f of the corrected classification feature vector>Characteristic values of the individual positions.
2. The control system of a screed according to claim 1, wherein the image dividing unit is further configured to uniformly divide the surface image of the material to be screed into the two-dimensional image block sequence.
3. The control system of claim 2, wherein the image block vectorization unit is further configured to linearly project each two-dimensional image block in the sequence of two-dimensional image blocks with a learnable embedding matrix using the linear embedding layer to obtain the sequence of image block feature vectors.
4. The control system of a trowel according to claim 3, wherein the first scale context encoding unit is further configured to:
arranging the sequence of the image block feature vectors into an input vector;
Converting the input vector into a query vector and a key vector through a learning embedding matrix respectively;
calculating the product between the query vector and the transpose vector of the key vector to obtain a self-attention correlation matrix;
carrying out standardization processing on the self-attention association matrix to obtain a standardized self-attention association matrix;
inputting the standardized self-attention association matrix into a Softmax activation function to activate so as to obtain a self-attention feature matrix; and
and multiplying the self-attention feature matrix with each image block feature vector in the sequence of image block feature vectors respectively to obtain a plurality of image block semantic feature vectors.
5. The control system of a trowel as in claim 4, wherein the trowel control result generation unit is further configured to: processing the corrected classification feature vector using the classifier in the following formula to generate the classification result;
wherein, the formula is:
wherein the method comprises the steps ofFor outputting result vector, ++>And->Respectively +.>The weight and bias vector corresponding to each category,an exponential operation representing a vector that represents a natural exponential function value that is a power of a eigenvalue of each position in the vector.
6. The control system of a trowel according to claim 1, further comprising a training module for training the linear embedded layer, the transducer-based context encoder, the two-way long and short term memory neural network model, and the classifier.
7. The control system of a trowel as in claim 6, wherein said training module comprises:
the training device comprises a training surface image acquisition unit for training a material to be leveled, and a training data acquisition unit for acquiring training data, wherein the training data comprise training surface images of the material to be leveled and the true value of the type label of the roll shaft;
the training image dividing unit is used for dividing the training surface image of the material to be leveled into a training two-dimensional image block sequence;
the training image block vectorization unit is used for enabling each training two-dimensional image block in the training two-dimensional image block sequence to pass through the linear embedding layer respectively so as to obtain a sequence of training image block feature vectors;
a training first scale context coding unit, configured to pass the sequence of training image block feature vectors through the converter-based context encoder to obtain a plurality of training image block semantic feature vectors;
The training cascade unit is used for cascading the plurality of training image block semantic feature vectors to obtain training long-distance dependent image feature vectors;
the training second scale context coding unit is used for enabling the sequence of the training image block feature vectors to pass through the two-way long-short-term memory neural network model to obtain the training middle-distance dependent image feature vectors;
the training multi-scale fusion unit is used for fusing the training long-distance dependent image feature vector and the training medium-distance dependent image feature vector to obtain a training classification feature vector;
a classification loss function value calculation unit, configured to pass the training classification feature vector through the classifier to obtain a classification loss function value;
the classification mode digestion inhibition loss function value calculation unit is used for calculating the classification mode digestion inhibition loss function value of the training long-distance dependent image feature vector and the training medium-distance dependent image feature vector; and
and the training unit is used for training the linear embedded layer, the context encoder based on the converter, the two-way long-short-term memory neural network model and the classifier by taking the weighted sum of the classification loss function value and the suppression loss function value as the loss function value.
8. The control system of a trowel according to claim 7, wherein the classification mode resolution suppression loss function value calculation unit is further configured to:
calculating the classification mode digestion inhibition loss function values of the training long-distance dependent image feature vector and the training medium-distance dependent image feature vector according to the following formula;
wherein, the formula is:
wherein the method comprises the steps ofAnd->The training long-distance dependent image feature vector and the training medium-distance dependent image feature vector are respectively +.>And->The classifier is a weight matrix of the training long-distance dependent image feature vector and the training medium-distance dependent image feature vector respectively, < >>Representing the square of the two norms of the vector, +.>Frobenius norms of the matrix are represented, < >>Representing subtraction by position +.>Representing an exponential operation.
9. A control method of a leveling machine, comprising:
acquiring a surface image of a material to be leveled;
dividing the surface image of the material to be leveled into a two-dimensional image block sequence;
respectively passing each two-dimensional image block in the two-dimensional image block sequence through a linear embedding layer to obtain a sequence of image block feature vectors;
Passing the sequence of image block feature vectors through a context encoder based on a converter to obtain a plurality of image block semantic feature vectors;
cascading the plurality of image block semantic feature vectors to obtain a long-distance dependent image feature vector;
the sequence of the image block feature vectors is processed through a two-way long-short-term memory neural network model to obtain a medium-distance dependent image feature vector;
fusing the long-distance dependent image feature vector and the medium-distance dependent image feature vector to obtain a classification feature vector;
performing geometric constraint re-parameterization of the positive assignment norm space on the classification feature vector to obtain a corrected classification feature vector; and
the corrected classification feature vector passes through a classifier to obtain a classification result, wherein the classification result is used for representing the type label of the roll shaft;
wherein, carry on the geometric constraint repartitioning of the normal definite form space to the said classification characteristic vector, in order to get the classification characteristic vector after correcting, include: performing geometric constraint re-parameterization of positive definite excipient space on the classification feature vector by using the following correction formula to obtain a corrected classification feature vector;
wherein, the correction formula is:
Wherein,,is the classification feature vector,/->And->Is feature set +.>Mean and variance of>Is a transpose of the classification feature vector, < >>Representing the square of the two norms of the vector, +.>Frobenius norms of the matrix are represented, < >>Is the classification feature vector->Is>Characteristic value of individual position->Is the +.f of the corrected classification feature vector>Characteristic values of the individual positions.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310389735.7A CN116109879B (en) | 2023-04-13 | 2023-04-13 | Control system and control method of leveling machine |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310389735.7A CN116109879B (en) | 2023-04-13 | 2023-04-13 | Control system and control method of leveling machine |
Publications (2)
Publication Number | Publication Date |
---|---|
CN116109879A CN116109879A (en) | 2023-05-12 |
CN116109879B true CN116109879B (en) | 2023-07-04 |
Family
ID=86265926
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202310389735.7A Active CN116109879B (en) | 2023-04-13 | 2023-04-13 | Control system and control method of leveling machine |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN116109879B (en) |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108996179A (en) * | 2017-06-07 | 2018-12-14 | 鸿劲精密股份有限公司 | The flattening mechanism of electronic component carrier and its job class equipment of application |
Family Cites Families (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
GB149343A (en) * | 1919-07-25 | 1921-07-07 | Nordiska Kullager Ab | An improved method and apparatus for gauging and sorting rollers |
ES2032881T3 (en) * | 1988-02-08 | 1993-03-01 | Thomas Josef Heimbach Gmbh & Co. | HOLLIN FILTER. |
US7058221B1 (en) * | 2000-07-07 | 2006-06-06 | Tani Electronics Industry Co., Ltd. | Method of recognizing object based on pattern matching and medium for recording computer program having same |
JP3787082B2 (en) * | 2000-11-30 | 2006-06-21 | 株式会社リコー | Classification device and image forming apparatus |
US7551768B2 (en) * | 2003-01-09 | 2009-06-23 | Panasonic Corporation | Image recognition apparatus and method for surface discrimination using reflected light |
JP2005062285A (en) * | 2003-08-20 | 2005-03-10 | Fuji Photo Film Co Ltd | Image forming apparatus, setup system for image forming apparatus and setup method for image forming apparatus |
FR2893520B1 (en) * | 2005-11-22 | 2009-05-15 | Vai Clecim Soc Par Actions Sim | METHOD FOR PLACING A FLAT PRODUCT IN THE FORM OF A STRIP OR A TELE IN A PLANER MILLING MACHINE WITH IMBRIC ROLLERS AND A PLANAR INSTALLATION FOR IMPLEMENTING THE METHOD |
PL2313215T3 (en) * | 2008-07-10 | 2012-04-30 | Arku Maschb Gmbh | Method for straightening parts in a roller straightening machine |
CN208976540U (en) * | 2018-11-08 | 2019-06-14 | 中山市钜沣精密机械制造有限公司 | A kind of novel leveling device convenient for stablizing speed regulation |
US10776651B2 (en) * | 2019-01-18 | 2020-09-15 | Intelligrated Headquarters, Llc | Material handling method, apparatus, and system for identification of a region-of-interest |
JP7428197B2 (en) * | 2021-05-25 | 2024-02-06 | Jfeスチール株式会社 | Steel plate shape discrimination method, shape measurement method, shape control method, manufacturing method, shape discrimination model generation method, and shape discrimination device |
CN217438742U (en) * | 2022-06-06 | 2022-09-16 | 中交第二航务工程局有限公司 | Cast-in-place roof beam top plate concrete leveling device that vibrates |
CN114882021B (en) * | 2022-07-07 | 2022-09-09 | 江苏中清先进电池制造有限公司 | Efficient processing method and system for battery lithium film |
CN115951584B (en) * | 2023-02-09 | 2024-03-15 | 浙江上洋机械股份有限公司 | Temperature control system and method for roller fixation machine |
-
2023
- 2023-04-13 CN CN202310389735.7A patent/CN116109879B/en active Active
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108996179A (en) * | 2017-06-07 | 2018-12-14 | 鸿劲精密股份有限公司 | The flattening mechanism of electronic component carrier and its job class equipment of application |
Non-Patent Citations (1)
Title |
---|
水下抛石基床整平机械;王中林,张鸿祥;水运工程(09);全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN116109879A (en) | 2023-05-12 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110569901A (en) | Channel selection-based countermeasure elimination weak supervision target detection method | |
CN115586755A (en) | Production management control system and method for preparing lithium hexafluorophosphate | |
CN116167668B (en) | BIM-based green energy-saving building construction quality evaluation method and system | |
CN110472695B (en) | Abnormal working condition detection and classification method in industrial production process | |
CN116765925A (en) | CNC-based die cutting machining control system and method thereof | |
CN115357065B (en) | Remote intelligent dehumidification control system and method for offshore wind turbine | |
CN115239946B (en) | Small sample transfer learning training and target detection method, device, equipment and medium | |
Heidrich et al. | Forecasting energy time series with profile neural networks | |
CN116030018A (en) | Incoming material qualification inspection system and method for door processing | |
CN117227005A (en) | Production control system and method for concrete raw material processing | |
CN116109879B (en) | Control system and control method of leveling machine | |
CN107527064A (en) | A kind of application of manifold learning in fault diagnosis data extraction | |
CN106997599A (en) | A kind of video moving object subdivision method of light sensitive | |
CN116894180B (en) | Product manufacturing quality prediction method based on different composition attention network | |
Peng et al. | Research on intelligent recognition method for self‐blast state of glass insulator based on mixed data augmentation | |
CN117076983A (en) | Transmission outer line resource identification detection method, device, equipment and storage medium | |
CN115409776A (en) | Power transmission line hardware fitting graph convolution detection method and system based on orientation reasoning | |
CN114723877A (en) | Countermeasure sample generation method for three-dimensional sparse convolution network | |
Seon et al. | GraphSAGE with contrastive encoder for efficient fault diagnosis in industrial IoT systems | |
CN116821745B (en) | Control method and system of intelligent linear cutting slow wire-moving equipment | |
Joshi et al. | Forged Image Detection using SOTA Image Classification Deep Learning Methods for Image Forensics with Error Level Analysis | |
Jana et al. | Evaluation of visualization algorithms for commsense system | |
CN115061427B (en) | Material layer uniformity control system of blow molding machine and control method thereof | |
CN116825217B (en) | Method for preparing high-purity phosphorus pentafluoride | |
CN117325432A (en) | Temperature control system and method for blow molding machine |
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 |