WO2011158711A1 - 欠陥判別装置 - Google Patents
欠陥判別装置 Download PDFInfo
- Publication number
- WO2011158711A1 WO2011158711A1 PCT/JP2011/063119 JP2011063119W WO2011158711A1 WO 2011158711 A1 WO2011158711 A1 WO 2011158711A1 JP 2011063119 W JP2011063119 W JP 2011063119W WO 2011158711 A1 WO2011158711 A1 WO 2011158711A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- defect
- defect type
- value
- function
- feature
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B19/00—Programme-control systems
- G05B19/02—Programme-control systems electric
- G05B19/18—Numerical control [NC], i.e. automatically operating machines, in particular machine tools, e.g. in a manufacturing environment, so as to execute positioning, movement or co-ordinated operations by means of programme data in numerical form
- G05B19/406—Numerical control [NC], i.e. automatically operating machines, in particular machine tools, e.g. in a manufacturing environment, so as to execute positioning, movement or co-ordinated operations by means of programme data in numerical form characterised by monitoring or safety
- G05B19/4065—Monitoring tool breakage, life or condition
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/84—Systems specially adapted for particular applications
- G01N21/88—Investigating the presence of flaws or contamination
- G01N21/89—Investigating the presence of flaws or contamination in moving material, e.g. running paper or textiles
- G01N21/892—Investigating the presence of flaws or contamination in moving material, e.g. running paper or textiles characterised by the flaw, defect or object feature examined
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/84—Systems specially adapted for particular applications
- G01N21/88—Investigating the presence of flaws or contamination
- G01N21/95—Investigating the presence of flaws or contamination characterised by the material or shape of the object to be examined
- G01N21/952—Inspecting the exterior surface of cylindrical bodies or wires
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/84—Systems specially adapted for particular applications
- G01N21/88—Investigating the presence of flaws or contamination
- G01N21/89—Investigating the presence of flaws or contamination in moving material, e.g. running paper or textiles
- G01N21/8914—Investigating the presence of flaws or contamination in moving material, e.g. running paper or textiles characterised by the material examined
- G01N2021/8918—Metal
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30108—Industrial image inspection
- G06T2207/30136—Metal
Definitions
- the present invention relates to a defect discriminating apparatus that discriminates defect types of defects.
- the present invention relates to a defect discriminating apparatus suitably used for discriminating defect types of defects generated in a material to be rolled (wire material, pipe material, plate material, etc.) made of a metal material (for example, steel material).
- a defect discriminating apparatus that performs image processing on a captured image of a wire taken by a camera arranged in the production line and discriminates a defect type of a defect such as a flaw generated in the wire is used. ing.
- the defect discriminating apparatus identifies a defect area corresponding to the defect generated in the wire from the captured image, and discriminates the defect type of the defect from the feature amount (for example, size / area) of the defect area.
- the defect discriminating apparatus disclosed in Patent Document 1 includes the number of feature amounts constituting feature information (vector) having a plurality of feature amounts (for example, defect dimensions and areas) indicating attributes of a discrimination target defect whose defect type is unknown.
- a mapping space having a higher number of dimensions is divided into two defect species regions by the discrimination boundary.
- This discrimination boundary is created in advance by the defect discriminating apparatus of Patent Document 1 using the feature information of the learning defect of the two defect types whose defect types have been discriminated by the user.
- the defect discriminating apparatus of Patent Document 1 maps data points (points at the tip of a vector) indicating feature information of a discrimination target defect to the mapping space, and maps the defect type of the discrimination target defect (hereinafter, referred to as a data point). It is determined that the defect type corresponds to the region where the “mapped point” is located.
- the defect discriminating apparatus disclosed in Patent Document 1 is a method similar to the method of discriminating the defect type of a defect to be discriminated (a method of mapping a data point indicating feature information into a mapping space and discriminating the defect type according to an area where the mapping point is located.
- a determination boundary is created so that the defect type of each learning defect is correctly determined (as determined by the user).
- Some of the learning defects used to create the discriminant boundary may have a characteristic value that is a unique value, and the same kind of learning defect that has a characteristic value that is not a unique value. is there.
- the defect determination apparatus of Patent Document 1 creates a determination boundary so that the defect type of each learning defect is correctly determined.
- a discriminant boundary that is excessively adapted to the feature information of the learning defect used to create the discriminant boundary is created, and overlearning, a phenomenon in which the ability to deal with discriminating target defects whose defect type is unknown, occurs.
- over-learning occurs when there is a learning defect with a unique feature value, the dimension is excessively high so that the defect type can be accurately determined even for a learning defect with a unique feature value.
- the discriminant boundary is created. When such overlearning occurs, the defect type of the determination target defect whose defect type is unknown may not be accurately determined.
- an object of the present invention is to provide a defect discriminating apparatus that suppresses overlearning and accurately discriminates defect types of defects.
- the present invention provides a mapping space having a higher number of dimensions than the number of feature quantities constituting the feature information, by representing data points indicating feature information including a plurality of feature quantities indicating the attributes of the discrimination target defect whose defect type is unknown. It is determined in which of the two defect type areas formed by bisecting the mapping space the mapping point mapped to is, and the defect type of the defect to be determined is determined to be the mapping point
- a defect discriminating apparatus that discriminates a defect type corresponding to a region, an acquisition unit for acquiring the feature information, a determination unit that determines a discrimination function that divides the mapping space into two, and the determination unit
- the discriminant function is a kernel
- the kernel function k (x, x ′) Is a kernel function in which the matrix K given by the element k (x, x') is a semi-definite value, x is feature information of the defect for learning of the one defect type, and x ' , Feature information of the defect for learning of the other defect type, and the determination unit
- Each of the characteristic information has a positive correlation with the discriminant error that
- the feature information is configured with respect to a predetermined regularization parameter so as to minimize the value of an error function that is variable according to the weight of the feature amount and is a sum of the regularization term multiplied by the regularization parameter.
- the weight of each feature is specified and the error function is When the weight of each feature amount constituting the feature information specified so as to minimize the value of the number is temporarily adopted as the weight of each feature amount constituting the discriminant function, for learning of the one defect type
- the one defect type is greater than the absolute value of the difference between the output value of the discriminant function when the defect feature information is input to the kernel function k (x, x ′) and the value corresponding to the one defect type.
- the absolute value of the difference between the output value of the discriminant function and the value corresponding to the other defect type when the learning defect feature information is input to the kernel function k (x, x ′) is smaller.
- the output value of the discriminant function and the other when the number of learning defects of the one defect type and the feature information of the learning defect of the other defect type are input to the kernel function k (x, x ′).
- the absolute value of the difference from the value corresponding to the defect type of the other The absolute value of the difference between the output value of the discriminant function and the value corresponding to the one defect type when the characteristic information of the defect for learning of the defect type is input to the kernel function k (x, x ′).
- the regularization parameter is adjusted so that the value of the error function is minimized again.
- the weight of each feature amount constituting the feature information is specified, and when the misclassification number is less than a predetermined value, each feature constituting the feature information specified so as to minimize the value of the error function.
- the defect discriminating apparatus is such that a mapping point to a high-dimensional number mapping space of a defect to be discriminated whose defect type is unknown is located in any of the two defect type regions formed by bisecting the mapping space. Determine whether to do.
- the defect discriminating apparatus discriminates the defect type of the defect to be discriminated to be a defect type corresponding to an area in which the mapping point of the discriminating target defect is located in the area of the two defect types. To do.
- the defect discriminating apparatus determines a discriminant function indicating a discriminant boundary that bisects the mapping space as follows.
- the defect discriminating apparatus first provides feature information on a predetermined regularization parameter so as to minimize the value of an error function consisting of a sum of a discrimination error and a regularization term multiplied by the regularization parameter.
- the weight of each feature quantity constituting is specified.
- the defect discriminating apparatus tentatively adopts the weight of each feature amount constituting the feature information specified to minimize the value of the error function as the weight of each feature amount constituting the discriminant function.
- the number of misclassifications is less than a predetermined value, it is determined that the specified weight is adopted as the weight of each feature quantity constituting the discrimination function, and the discrimination function is determined.
- the regularization term In order to reduce the value of the error function consisting of the sum of the discrimination error and the regularization term multiplied by the regularization parameter, at least one of the discrimination error and the regularization term needs to be reduced.
- the regularization term when the regularization parameter is large, the regularization term has a large influence on the value of the error function.
- the regularization term varies according to the weight of each feature quantity constituting the feature information. For this reason, when the regularization parameter is large, the weight that makes the regularization term sufficiently small is specified as the weight that minimizes the value of the error function.
- the number of dimensions of the discriminant function and the regularization term have a positive correlation.
- a weight that makes the regularization term sufficiently small is specified as a weight that minimizes the value of the error function, and that the weight is adopted as the weight of each feature quantity constituting the discriminant function.
- the defect according to the present invention adjusts the regularization parameter and again specifies the weight that minimizes the value of the error function. If the regularization parameter is reduced by adjusting the regularization parameter described above, the influence of the regularization term on the value of the error function is reduced, while the influence of the discrimination error on the value of the error function is increased. For this reason, when the regularization parameter is adjusted to be small, the weight that makes the discrimination error smaller than before the adjustment can be specified as the weight that minimizes the value of the error function.
- the discriminant error is an output value of the discriminant function when the feature information of the defect for learning of one defect type is input to the kernel function k (x, x ′) (hereinafter, “output of discriminant function corresponding to one defect type”).
- Discriminant function output when the difference between the value corresponding to one defect type and the feature information of the defect for learning of the other defect type are input to the kernel function k (x, x ′). It is defined by the difference between the value (hereinafter referred to as “the output value of the discriminant function corresponding to the other defect type”) and the value corresponding to the other defect type.
- the discrimination error includes the difference between the output value of the discriminant function corresponding to one defect type and the value corresponding to the one defect type, and the output value of the discriminant function corresponding to the other defect type and the other defect.
- the absolute value of any of the differences from the value corresponding to the seed decreases, it decreases, and when it increases, it increases.
- the discriminant error becomes small, the absolute value of the difference between the output value of the discriminant function corresponding to one defect type and the value corresponding to one defect type, or the output value of the discriminant function corresponding to the other defect type And the absolute value of the difference between the value corresponding to the other defect type becomes small.
- the output value of the discriminant function corresponding to one defect type is smaller than the absolute value.
- the absolute value of the difference from the value corresponding to the other defect type increases, and the number of learning defects of one defect type decreases.
- the discriminant function corresponding to the other defect type is more than the absolute value.
- the number of learning defects of the other defect type decreases, where the absolute value of the difference between the output value and the value corresponding to one defect type increases.
- the determination error when the determination error is reduced, the number of erroneous determinations is reduced. Therefore, even if the number of misclassifications when the weight specified before adjusting the regularization parameter is adopted as the weight of each feature quantity constituting the discrimination function is a predetermined value or more, the regularization parameter is small. By adjusting so that the weight that the misclassification number is less than the predetermined value can be identified, it is determined that the identified weight is adopted as the weight of each feature quantity constituting the discriminant function, and the discriminant function is determined. it can.
- the defect discriminating apparatus determines that the specified weight is adopted as the weight of each feature quantity constituting the discriminant function, and determines the discriminant function. For this reason, the defect determination apparatus according to the present invention can accurately determine the defect type.
- the possibility of overlearning increases as the regularization term increases.
- the regularization parameter is increased in the initial stage, and when the weight at which the number of misclassifications is less than a predetermined value cannot be specified, the number of misclassifications can be reduced by adjusting the regularization parameter to be gradually reduced. It is preferable to specify a weight that is less than a predetermined value.
- the discriminant error described above is, for example, the difference between the output value of the discriminant function corresponding to one defect type and the value corresponding to one defect type, and the output value of the discriminant function corresponding to the other defect type. It is a value having a positive correlation with the square sum of the difference from the value corresponding to the other defect type.
- the value having a positive correlation with the square sum is, for example, the square root of the square sum.
- the discriminant function is composed of a kernel function k (x, x ') and the weight of each feature quantity, and has no mapping function. For this reason, it is not necessary to calculate a mapping function in order to determine a discriminant function. The calculation amount of the mapping function is enormous. Therefore, the defect discriminating apparatus according to the present invention that does not need to calculate the mapping function for determining the discriminant function can determine the discriminant function with a small amount of calculation.
- the concept of “defects” in the present invention includes fake patterns such as dirt attached to the material to be rolled in addition to flaws.
- the “value corresponding to the defect type” in the present invention is a value determined in advance so that one defect type and the other defect type can be distinguished, and one defect type and the other defect type. And different values.
- the defect discriminating apparatus is configured such that at least the acquisition unit is arranged on a rolling line of the material to be rolled and discriminates the defect type of the defect to be discriminated that occurs in the material to be rolled.
- a hot rolling line for example, hot rolling of a wire rod or steel bar
- a hot rolling line can be performed quickly.
- Line and cold rolling line rolling mill setting adjustment and the like.
- a continuous hot rolling line is normally comprised from a rough rolling mill row
- the defect discriminating apparatus having the preferred configuration can be arranged between two rolling mill rows, but in order to efficiently discriminate on-line the defect type of the discrimination target defect generated in the material to be rolled, a finish rolling mill row It is most preferable to arrange it on the downstream side (downstream in the rolling direction of the material to be rolled).
- a cooling device for cooling the material to be rolled is usually installed on the downstream side of the finishing rolling mill row.
- the defect discriminating device is arranged on the downstream side of the cooling device, the difference in luminance value between the defect area corresponding to the defect to be discriminated in the captured image and the other area is increased by cooling the material to be rolled. In terms of ease, it can be expected that the feature quantity of the defect to be determined is obtained with high accuracy.
- the defect discriminating device is arranged on the downstream side of the cooling device, the guide, which is normally arranged on the downstream side of the cooling device, and the material to be rolled come into contact with each other to simulate the dirt attached to the material to be rolled. There is a risk that the pattern will appear in the captured image and the accuracy of discrimination between the flaw and the pseudo pattern will be reduced.
- the defect determination device upstream of the cooling device. Actually, it is only necessary to evaluate in advance a test or the like whether the defect discriminating device is arranged on the upstream side or the downstream side of the cooling device, and determine the arrangement position of the defect discriminating device.
- the defect determination device at least an imaging device for the material to be rolled as the acquisition unit is arranged along with the eddy current flaw detection device in the rolling line, and image processing is performed on the captured image of the material to be rolled imaged by the imaging device. It is set as the structure which discriminate
- a minute defect is detected by an eddy current flaw detector (for example, a differential type penetrating eddy current flaw detector), and a defect extending in the longitudinal direction of the material to be rolled is detected by a defect discriminating device (defect).
- a defect discriminating device defect discriminating device It is possible to increase the defect detection accuracy of the material to be rolled by performing the role sharing of two defect types determined by the determination device as defects extending in the longitudinal direction on the same rolling line.
- the defect discriminating apparatus having such a preferable configuration is also applied to the downstream side of the finish rolling mill row of the hot rolling line, it can be provided on either the upstream side or the downstream side of the cooling apparatus. It is also possible to arrange.
- the defect discriminating apparatus it is also possible to discriminate flaws and pseudo patterns such as dirt adhering to the material to be rolled, which are difficult to discriminate by a known discrimination method. That is, in the defect discriminating apparatus, the two defect types can be made into a fake pattern such as a flaw and a dirt adhering to the material to be rolled. In addition, even if it uses a well-known discrimination
- the defect discriminating apparatus is applied only to the defect discriminated (extracted) by this known discriminating method, it is possible to discriminate flaws and pseudo patterns efficiently.
- the feature amount used in the defect determination apparatus is not particularly limited.
- the acquisition unit includes an imaging device for the material to be rolled, an eddy current flaw detection device that performs eddy current flaw detection on the material to be rolled, and ultrasonic flaw detection that performs ultrasonic flaw detection on the material to be rolled.
- the present invention can provide a defect discriminating apparatus that suppresses overlearning and accurately discriminates defect types of defects.
- a target for which a discrimination target defect occurs is a wire that is a material to be rolled, and a feature amount obtained by performing image processing on a captured image of the wire is used.
- a case will be described as an example.
- a parameter shown in bold italics means a vector.
- FIG. 1A is a schematic configuration diagram illustrating an example of the defect determination apparatus 1 according to the present embodiment.
- the defect determination device 1 includes an imaging device (camera 2 and a light source (not shown)) as an acquisition unit for acquiring a feature value, a determination unit 3, and a determination unit 4.
- a plurality of (for example, four) cameras 2 are arranged in the circumferential direction of the wire 21 conveyed along the hot rolling line, and images the wire 21.
- a light source (not shown) for illuminating the wire 21 is disposed around the camera 2.
- the camera 2 and the light source are arranged on the downstream side of the finishing rolling mill row 5 installed in the hot rolling line (downstream side in the rolling direction (conveying direction) of the wire 21).
- FIG.1 (b) is a schematic block diagram which shows the other example of the defect determination apparatus 1 of this embodiment.
- a differential type penetrating eddy current flaw detector 6 is disposed downstream of the finishing rolling mill row 5, and the camera 2 and the light source are disposed downstream of the eddy current flaw detector 6. Is different from the example shown in FIG.
- the determination unit 3 determines a discriminant function indicating a discrimination boundary for discriminating the defect type of the discrimination target defect whose defect type is unknown.
- This discrimination boundary is obtained by mapping a mapping space having a number of dimensions higher than the number of feature amounts constituting feature information (vector) having a plurality of feature amounts indicating the attributes of the defect to be identified as two defect types (hereinafter, 2 One of the two defect types is divided into two regions, ie, “defect type A” and the other defect type is called “defect type B”.
- the defect type A and the defect type B are different defect types set in advance by a user of the defect determination device 1 or the like.
- the defect type A and the defect type B may be defect types having different degrees of seriousness (influence on the quality of the wire 21).
- the defect type A and the defect type B can be, for example, defect types with different causes of occurrence.
- the feature amount a dimension, an area, a luminance value, and the like of a defect area corresponding to a defect obtained by performing image processing on a captured image captured by the camera 2 can be used.
- the number of feature amounts constituting the feature information is not limited as long as it is plural.
- the determination unit 3 determines the discriminant function using the feature information of the learning defect that is known to be the defect type A or the defect type B.
- the feature information of the learning defect input to the determination unit 3 is obtained by using, for example, an image processing function of the determination unit 4 described later. That is, the learning defect feature information is obtained by inputting the captured image of the learning defect captured by the camera 2 to the determination unit 4 and performing image processing by the determination unit 4. Then, the obtained feature information of the learning defect is input to the determination unit 3.
- the determination unit 3 itself has an image processing function
- a captured image of the learning defect captured by the camera 2 is input to the determination unit 3, and image processing is performed by the determination unit 3. It is also possible to obtain feature information.
- the feature information of the defect for learning of the defect type A, the feature information of the defect for learning of the defect type B, and the feature information of the aforementioned defect to be discriminated are composed of the same type of feature quantity.
- the learning defect is a defect whose defect type is known (the defect type is identified by the user).
- each feature amount constituting the feature information of each learning defect of the defect type A and each feature amount constituting the feature information of each learning defect of the defect type B are the learning defects.
- the identifier and the defect type of the learning defect are associated with each other and stored in the determination unit 3.
- the value of each feature amount constituting the feature information of each learning defect is normalized so as to be within a range of 0 to 1.
- the discriminant function f (x) determined by the determining unit 3 is expressed by the following formula (1).
- “w” indicates weight information (vector) whose component is the weight of each feature quantity constituting the feature information.
- X in the equation (1) indicates feature information (vector) of the defect for learning of the defect type A or the defect type B.
- ⁇ ( ⁇ ) represents a mapping function that maps data points (points at the tips of vectors) indicating feature information in the mapping space and has positive definiteness.
- a mapping function having positive definiteness includes a Gaussian distribution function.
- the discriminant function f (x) expressed by the following equation (2) is used as the discriminant function so that the discriminant function f (x) can be determined with a small amount of calculation.
- the discriminant function f (x) means the discriminant function f (x) expressed by the following formula (2).
- ⁇ indicates the weight of each feature amount constituting the feature information.
- k (x, x ′) represents a kernel function in which a matrix K whose elements are given by k (x, x ′) is a semi-definite value.
- X described in the equation (2) and after indicates the feature information (vector) of the defect for learning of the defect type A.
- x ′ indicates feature information (vector) of the defect for learning of the defect type B.
- the matrix K whose elements are given by k (x, x ′) is the output value of the kernel function obtained when the feature information x of the defect for learning of the defect type A is input to the kernel function k (x, x ′).
- the kernel function k (x, x ′) is a matrix whose elements are output values of the kernel function obtained when the learning defect feature information x ′ of the defect type B is input.
- kernel functions k (x, x ′) are examples of kernel functions k (x, x ′) whose elements are given by kernel functions k (x, x ′) and are positive semidefinite. . Further, as other examples of the kernel function k (x, x ′) in which the matrix K given by the kernel function k (x, x ′) is a semi-definite value, there are a sigmoid function and a Gaussian function represented by the following equations. .
- f (•) indicates an arbitrary function
- q (•) indicates a non-negative coefficient polynomial
- k a (•, •) and k b (•, •) denote arbitrary kernel functions
- subscripts a and b denote identifiers for learning defects
- ⁇ denotes sigmoid function gain
- ⁇ indicates dispersion.
- the above equation (2) is derived as follows.
- the following formula (4) is derived.
- d indicates the number of feature quantities constituting the feature information. If the number of feature amounts constituting the feature information is sufficiently increased, the following equation (5) is derived from the above equation (1).
- the weight information w is expressed by the following equation (6).
- the above formula (2) is derived from the above formula (1) using the above formula (6).
- the discriminant function f (x) in the above equation (2) is a function in which the number of dimensions is influenced by the number of feature amounts constituting the feature information of the learning defect.
- the determination unit 3 first determines whether or not the number of erroneous determinations is less than a predetermined value (step S1 in FIG. 3).
- the number of misidentifications is the output value of the discriminant function f (x) when the learning defect feature information x of the defect type A is input to the kernel function k (x, x ′) of the discriminant function f (x) (
- the output value of the discriminant function corresponding to the defect type A and the defect type B rather than the absolute value of the difference between “the output value of the discriminant function corresponding to the defect type A”
- the difference between the output value of the discriminant function f (x) (hereinafter referred to as “the output value of the discriminant function corresponding to the defect type B”) when input to x, x ′) and the value corresponding to the defect type B
- the defect type B having a smaller absolute value of the difference between the output value of the discriminant function corresponding to the defect type B and the value corresponding to the defect type A than the absolute value of Is a number obtained by adding the number of the learning defect.
- the weight ⁇ of each feature amount constituting the feature information of the discriminant function f (x) is set to an arbitrary value (for example, 1).
- the value corresponding to the defect type A is 1 and the value corresponding to the defect type B is -1.
- the value corresponding to the defect type A and the value corresponding to the defect type B are determined in advance by the user of the defect determination apparatus 1 or the like, and are different from each other so that the defect type A and the defect type B can be determined. .
- the determination unit 3 determines that the number of misclassifications is equal to or greater than a predetermined value, the determination unit 3 minimizes the value of an error function that is the sum of the determination error and the regularization term ⁇ T K ⁇ multiplied by the regularization parameter ⁇ .
- the weight ⁇ of each feature amount constituting the feature information to be specified is specified (step S2 in FIG. 3).
- the minimum value of the error function is expressed by the following formula (7).
- Superscript (i) indicates an identifier of a learning defect.
- the regularization parameter ⁇ takes a value in the range of 0 to 1.
- the discriminant error is the difference between the output value of the discriminant function f (x) when the feature information of the defect for learning of the defect type A is input to the kernel function k (x, x ′) and the value corresponding to the defect type A.
- the difference between the output value of the discriminant function f (x) when the feature information of the defect for learning of the defect type B is input to the kernel function k (x, x ′) and the value corresponding to the defect type B It is defined and becomes smaller when the absolute value of one of the two differences becomes smaller, and becomes larger when it becomes larger.
- the ⁇ cost of the equation (7) is expressed by the following equation (8).
- the above equation (8) is a convex function approximating the following equation (9).
- y represents a vector whose component is the weight of each feature amount.
- the regularization term ⁇ T K ⁇ is expressed by the following equation (11).
- the subscript i indicates an identifier representing the type of feature quantity constituting the feature information of the defect for learning of the defect type A.
- the subscript j indicates an identifier indicating the type of feature quantity constituting the feature information of the defect for learning of the defect type B.
- the regularization term ⁇ T K ⁇ has a positive correlation with the weight ⁇ of each feature quantity.
- the regularization term ⁇ T K ⁇ is derived as follows.
- the linear sum w 0 of each feature quantity of the defect for learning of the defect type A is expressed by the following equation (12). Since the weight information w is obtained by adding a ⁇ component orthogonal to the mapping point ⁇ (x (i) ) obtained by mapping the data point indicating the feature information of the learning defect to the linear sum w 0 , It is expressed by (13).
- Equation (5) f (x) in the above-described equation (5) is It is expressed by the following formula (14). That is, it can be seen that ⁇ cost on the left side of Equation (8) does not depend on the value of the ⁇ component. Further, the following equation (15) can be derived from the orthogonality between the linear sum w 0 and the ⁇ component. From equation (15), it is clear that ⁇
- the equation (11) can be derived from the equation (15).
- the determination unit 3 inputs arbitrary values (for example, 1) to the weight ⁇ i and the weight ⁇ j of the above equation (11), and the learning type of the defect type A is input to x (i) of the above equation (11).
- the feature information of the defect is input, the feature information of the defect for learning of the defect type B is input to x (j) , and the regularization term ⁇ T K ⁇ is calculated (step S21 in FIG. 3).
- ⁇ cost of the left side of the above equation (8) is input to y (i) of the above equation (7), and each defect type A or defect type B is used for x (i) of the above equation (7). enter the characteristic information of the defect, and inputs a value of the regularization term alpha T K [alpha calculated at step S21 to the regularization term alpha T K [alpha in the formula (7) (step S22 in FIG. 3).
- the regularization parameter of the above equation (7) at this time is an initial value, and the initial value is 1 here.
- Characteristic information x (i) if the output of the determination error for the values y (i) xi] i, the minimum value of the output xi] i is two inequalities (17), the minimum value defined by (18) Become.
- the output ⁇ i at the time of the minimum value is called a slack variable, and by introducing the output ⁇ i at the time of the minimum value into the above equation (7), the above equations (17) and (18) are used as constraints.
- the above equation (7) is converted into the above equation (16).
- the above equation (16) is in the form of a convex quadratic programming problem relating to the output ⁇ and the weight ⁇ of each feature quantity constituting the feature information.
- the solution of the convex quadratic programming problem of the above equation (16) will be shown.
- Equation (16) is solved using Lagrange's undetermined multiplier method.
- the following formula (19) is defined as Lagrangian. Domain: ⁇ R n R n represents the entire real number.
- the quadratic programming problem can be converted into a dual problem with simpler constraints.
- the determination unit 3 sets the weight ⁇ of each feature amount constituting the feature information of the learning defect identified as described above to the weight ⁇ of each feature amount constituting the feature information of the learning defect of the discriminant function f (x). Temporarily adopted. Then, in the same manner as in step S1 of FIG. 3, the determination unit 3 tentatively adopts the weight ⁇ of each identified feature quantity as the weight ⁇ of each feature quantity constituting the discrimination function f (x). Is calculated. If the calculated misclassification number is equal to or greater than the predetermined value, the determination unit 3 adjusts the regularization parameter ⁇ to be small, and again, as described above, each feature constituting the feature information that minimizes the error function. The quantity weight ⁇ is specified (step S2 in FIG. 3).
- the weight ⁇ i and the weight ⁇ j in the above-described equation (11) are set to the respective values specified in the previous step S25.
- the feature amount weight ⁇ is input.
- the weight ⁇ of each specified feature quantity is adopted as the weight ⁇ of each feature quantity constituting the discrimination function f (x).
- the function f (x) is determined (step S3 in FIG. 3).
- the determination unit 3 of the present embodiment sets the initial value of the regularization parameter ⁇ as the maximum value of the regularization parameter ⁇ , and reduces the regularization parameter ⁇ when the number of misclassifications is less than a predetermined value. adjust.
- the regularization parameter ⁇ is large, the regularization term ⁇ T K ⁇ has a large influence on the value of the error function. Therefore, when the regularization parameter ⁇ is large, the weight ⁇ of each feature amount that makes the regularization term ⁇ T K ⁇ sufficiently small is specified as the weight ⁇ of each feature amount that minimizes the value of the error function.
- the number of dimensions of the discriminant function f (x) and the regularization term ⁇ T K ⁇ have a positive correlation.
- the weight ⁇ of each feature amount that makes the regularization term ⁇ T K ⁇ sufficiently small is specified as the weight ⁇ of each feature amount that minimizes the value of the error function, and the weight ⁇ of each feature amount is discriminated.
- the discriminant function is determined by determining that it is adopted as the weight ⁇ of each feature quantity constituting the function f (x)
- the higher-order discriminant function discriminant boundary
- the weight ⁇ of each feature amount is used as the weight ⁇ of each feature amount constituting the discriminant function f (x).
- the regularization parameter ⁇ is adjusted to be small, and the weight ⁇ of each feature amount that minimizes the value of the error function is specified again.
- the regularization parameter ⁇ is reduced, the influence of the regularization term ⁇ T K ⁇ on the value of the error function is reduced, while the influence of the discrimination error on the value of the error function is increased.
- the discriminant error is the difference between the output value of the discriminant function corresponding to the defect type A and the value corresponding to the defect type A, and the output value of the discriminant function corresponding to the defect type B and the value corresponding to the defect type B. It is defined by the difference. Further, the discriminant error corresponds to the difference between the output value of the discriminant function corresponding to the defect type A and the value corresponding to the defect type A, and the output value of the discriminant function corresponding to the defect type B and the defect type B.
- the absolute value of any of the differences from the value becomes smaller, it becomes smaller, and when any of the differences becomes larger, it becomes larger. That is, when the discrimination error is reduced, the absolute value of the difference between the output value of the discriminant function corresponding to the defect type A and the value corresponding to the defect type A, or the output value of the discriminant function corresponding to the defect type B and the defect type
- the absolute value of the difference from the value corresponding to B becomes smaller.
- the absolute value of the difference between the output value of the discriminant function corresponding to the defect type A and the value corresponding to the defect type A becomes smaller, the output value of the discriminant function corresponding to the defect type A and the defect type B are smaller than the absolute value.
- the number of defects for learning of the defect type A in which the absolute value of the difference from the value corresponding to is increased is reduced.
- the output value of the discriminant function corresponding to the defect type B becomes smaller than the absolute value.
- the number of defects for learning of the defect type B in which the absolute value of the difference from the value corresponding to the defect type B is large is reduced. Therefore, when the determination error is reduced, the number of erroneous determinations is reduced.
- the discriminant function f (x) expressed by the above-described equation (2) has a kernel function k (x, x ′) and a weight ⁇ of each feature quantity constituting the feature information, and has a mapping function. Not done. For this reason, it is not necessary to calculate a mapping function when calculating the number of misclassifications. In other words, it is not necessary to calculate a mapping function to determine the discriminant function. The calculation amount of the mapping function is enormous. Therefore, the defect discriminating apparatus 1 that does not need to calculate the mapping function for determining the discriminant function f (x) can determine the discriminant function f (x) with a small amount of calculation.
- the discriminating unit 4 discriminates whether the defect type of the discrimination target defect generated in the wire 21 imaged by the camera 2 is the defect type A or the defect type B.
- the determination unit 4 has an image processing function, specifies a defect area corresponding to a determination target defect of the wire 21 from a captured image of the wire 21 captured by the camera 2 by a known image processing method, and determines the determination target.
- the feature amount of the defect is calculated.
- the discriminating unit 4 inputs the feature information composed of the calculated feature quantities into the kernel function k (x, x ′) of the discriminant function f (x) determined by the determining unit 3, and the discriminant function f (x) , That is, a mapping point obtained by mapping a data point indicating the feature information in the mapping space.
- the determination unit 4 determines that the defect type of the defect to be determined is a defect type corresponding to an area determined to have a mapping point located in one of the two areas. Specifically, the discriminating unit 4 determines the absolute value of the difference between the output value of the discriminant function f (x) when the feature information of the discrimination target defect is input and the value corresponding to the defect type A, and the discrimination. The absolute value of the difference between the output value of the function f (x) and the value corresponding to the defect type B is compared. If the former is smaller, it is determined that the defect to be determined is the defect type A, and the latter is If it is smaller, it is determined that the determination target defect is the defect type B.
- the determination unit 3 specifies the weight ⁇ of each feature amount constituting the feature information that minimizes the error function.
- the error function is a function composed of the sum of the discrimination error and the regularization term ⁇ T K ⁇ .
- the discriminant error is the difference between the output value of the discriminant function corresponding to the defect type A and the value corresponding to the defect type A, and the output value of the discriminant function corresponding to the defect type B and the value corresponding to the defect type B.
- the discrimination error varies depending on the output value of the discrimination function f (x).
- the discriminant function f (x) varies according to the weight of the feature amount from the above-described equation (2). For this reason, the variation amount of the discrimination error when the weight of the feature amount that hardly affects the output value of the discrimination function f (x) or does not affect the output value is small.
- the regularization term ⁇ T K ⁇ has a positive correlation with the weight ⁇ of each feature quantity.
- the value of the error function consisting of the sum of the discrimination error and the regularization term is likely to be reduced by minimizing the weight of the feature quantity having a small change amount of the discrimination error (that is, 0). Therefore, the weight of the feature quantity that hardly affects the output value of the discriminant function f (x) or does not affect the output value is highly likely to be specified as 0 by the determination unit 3.
- the feature information input to the kernel function k (x, x ′) of the discriminant function f (x) determined by the determining unit 3 in order for the discriminating unit 4 to discriminate the defect type of the discrimination target defect is as described above. It is good also as feature information comprised from the feature-values other than the feature-value specified as the weight 0.
- feature information comprised from the feature-values other than the feature-value specified as the weight 0.
- the feature amount identified as having a weight of 0 is the discriminant function. It is not input to the kernel function k (x, x ′) of f (x), and the amount of calculation required for determining the defect type of the determination target defect is reduced accordingly.
- the defect type of the determination target defect can be determined at high speed.
- the weight of the feature amount that hardly affects the output value of the discriminant function f (x) or does not affect it at all is specified as 0. Therefore, even if the feature quantity identified as having a weight of 0 is not input to the kernel function k (x, x ′) of the discriminant function f (x), the discrimination target to be performed using the output value of the discriminant function f (x)
- the defect type of the defect can be determined with a certain accuracy or higher.
- the defect discriminating apparatus 1 has been described for discriminating defects that have occurred in the wire.
- the defects that the defect discriminating apparatus 1 discriminates are not limited to those that have occurred in the wire. It can be a defect generated in the rolled material.
- the discrimination error is expressed using ⁇ cost of the above-described equation (8) that is a convex function. Since ⁇ cost in Equation (8) described above is a convex function, the weight ⁇ that minimizes the value of the discrimination error can be obtained without falling into a local solution. Therefore, it is possible to efficiently specify the weight ⁇ that makes the discrimination error less than a predetermined value.
- defect type A and defect type B the defect type of the defect to be determined.
- the determination unit 3 and the determination unit 4 described above. By repeating this operation, it is also possible to determine which of the three or more defect types is the defect type of the determination target defect. For example, consider a case where the defect type A can be further classified into either the defect type A1 or the defect type A2. That is, consider a case where the defect type can be classified into one of the defect type A1, the defect type A2, and the defect type B.
- the determination unit 3 first determines a discriminant function indicating a discriminant boundary for discriminating whether the defect type of the defect to be discriminated is the defect type A or the defect type B by the procedure described above.
- the discriminating unit 4 discriminates whether the defect type of the defect to be discriminated is the defect type A or the defect type B using the discriminant function determined by the determining unit 3.
- a discriminant function indicating a discriminant boundary for discriminating whether the defect type of the discrimination target defect that has been discriminated to be the defect type A is the defect type A1 or the defect type A2. This is determined by the same procedure as described above.
- the discriminating unit 4 discriminates whether the defect type of the discriminating target discriminated to be the defect type A is the defect type A1 or the defect type A2 by using the discriminant function determined by the determining unit 3. As a result, it is determined which of the three defect types A1, A2, and B is the defect to be determined. By repeating a procedure similar to the procedure described above, it is possible to determine which of the four or more defect types is the defect type of the defect to be determined.
- the defect discrimination device 1 of this embodiment was arrange
- the surface temperature of the material to be rolled picked up by the camera 2 was approximately 1000 ° C.
- the defect type A is a stain on the surface of the wire
- the defect type B is a flaw generated on the surface of the wire.
- the defect discriminating apparatus 1 of the present embodiment discriminates 45 defects from the defect type B, that is, a flaw among the defects generated in the wire having a wire diameter of 20 mm, and among the defects generated in the wire having a wire diameter of 13 mm, Twenty-three defects were identified as defect type B, that is, flaws.
- defect discriminating apparatus 1 of the present embodiment one overdetection occurred in both of the two wire rods having the wire diameters of 20 mm and 13 mm, but all of them online including the flaws that could not be detected by the penetrating eddy current flaw detector 6.
- the flaws could be determined as defect type B. From the above, it has become clear that the defect determination apparatus 1 of the present embodiment can accurately determine the defect type of a defect.
- an imaging device (camera 2 and a light source) is provided as an acquisition unit for acquiring a feature quantity, and a defect type is obtained using a feature quantity obtained by performing image processing on a captured image of a material to be rolled.
- the defect discriminating apparatus is not limited to this, and may be configured as shown in FIG.
- FIG. 5 is a schematic configuration diagram showing a defect determination apparatus 1A according to another embodiment of the present invention. As illustrated in FIG. 5, the defect determination device 1 ⁇ / b> A according to the present embodiment includes an acquisition unit 7, a determination unit 3, and a determination unit 4.
- the acquisition unit 7 includes at least one of an imaging device for a material to be rolled, an eddy current flaw detection device that performs eddy current flaw detection on the material to be rolled, and an ultrasonic flaw detection device that performs ultrasonic flaw detection on the material to be rolled.
- the determination unit 4 is a feature amount obtained by performing image processing on a captured image of a material to be rolled imaged by the imaging device, a feature amount obtained by performing eddy current flaw detection on the material to be rolled by the eddy current flaw detection device, and the It is configured to discriminate the defect type of the discrimination target defect generated in the material to be rolled using at least one type of feature value among the feature values obtained by performing ultrasonic flaw detection on the material to be rolled by the ultrasonic flaw detector. ing.
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Biochemistry (AREA)
- Pathology (AREA)
- Immunology (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Chemical & Material Sciences (AREA)
- Analytical Chemistry (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Quality & Reliability (AREA)
- Textile Engineering (AREA)
- Software Systems (AREA)
- Automation & Control Theory (AREA)
- Manufacturing & Machinery (AREA)
- Human Computer Interaction (AREA)
- General Engineering & Computer Science (AREA)
- Medical Informatics (AREA)
- Computing Systems (AREA)
- Evolutionary Computation (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Investigating Materials By The Use Of Optical Means Adapted For Particular Applications (AREA)
- Image Analysis (AREA)
- Investigating Or Analyzing Materials By The Use Of Magnetic Means (AREA)
- Image Processing (AREA)
- Length Measuring Devices With Unspecified Measuring Means (AREA)
- Investigating Or Analyzing Materials By The Use Of Ultrasonic Waves (AREA)
Abstract
Description
また、本発明における「欠陥種に対応する値」とは、一方の欠陥種と他方の欠陥種とを判別し得るように予め決定された値であって、一方の欠陥種と他方の欠陥種とで異なる値とされている。
なお、連続式の熱間圧延ラインは、通常、粗圧延機列及び仕上圧延機列から構成されるか、或いは、粗圧延機列、中間圧延機列及び仕上圧延機列から構成される。前記好ましい構成の欠陥判別装置は、二つの圧延機列間に配置することも可能であるが、被圧延材に生じる判別対象欠陥の欠陥種をオンラインで効率よく判別するには、仕上圧延機列の下流側(被圧延材の圧延方向下流側)に配置することが最も好ましい。
また、仕上圧延機列の下流側には、通常、被圧延材を冷却するための冷却装置が設置されている。前記好ましい構成の欠陥判別装置が、被圧延材の撮像画像に画像処理を施すことによって得られる特徴量を用いる場合、前記欠陥判別装置は、前記冷却装置の上流側(被圧延材の圧延方向上流側)及び下流側の何れに配置することも可能である。前記欠陥判別装置を前記冷却装置の下流側に配置すれば、被圧延材が冷却されることにより、撮像画像における判別対象欠陥に対応する欠陥領域とその他の領域との輝度値の差が大きくなり易い点で、判別対象欠陥の特徴量を精度良く得ることが期待できる。この反面、前記欠陥判別装置を前記冷却装置の下流側に配置すれば、冷却装置の下流側に通常配置されるガイドと被圧延材とが接触することによって被圧延材に付着した汚れ等の疑似模様が撮像画像に写り込んでしまい、きずと疑似模様との判別の精度が低下してしまうおそれがある。このおそれを低減する点では、前記欠陥判別装置を前記冷却装置の上流側に配置することが好ましい。実際には、前記欠陥判別装置を前記冷却装置の上流側に配置する場合と下流側に配置する場合との得失を予め試験等によって評価し、前記欠陥判別装置の配置位置を確定すればよい。
なお、公知の判別手法を用いても、被圧延材におけるきず及び疑似模様が混合した欠陥と、被圧延材の健全部位とを判別することは可能である。例えば、被圧延材の撮像画像に公知の画像処理を施すことにより、上記の健全部位に対応する領域を除外して上記の欠陥に対応する領域のみを抽出することが可能である。そして、この公知の判別手法によって判別(抽出)された欠陥に対してのみ、前記欠陥判別装置を適用することにすれば、きずと疑似模様とを効率良く判別することが可能である。
図1(b)は、本実施形態の欠陥判別装置1の他の例を示す概略構成図である。図1(b)に示す例では、仕上圧延機列5の下流側に差動方式の貫通型渦流探傷装置6が配置されており、この渦流探傷装置6よりも下流側にカメラ2及び前記光源が配置されている点が、図1(a)に示す例と異なる。
凸二次計画問題
定義域:Ω⊆Rn
1-βi-γi≠0…(34)
1-βi-γi=0…(35)
このように、ラグランジュ関数が1次式となっている変数はその係数は0なので、双対問題は出力ξiと無関係である。ゆえに、重みαiを上記式(33)で置換し、上記式(35)の制約条件のもとで、下記式(36)のラグランジュ関数を最大化する。
0≦γi≦1…(37)
最急降下法や内点法といった公知の最適解探索手法を用いて、上記式(36)からγiを算出し(図3のステップS24)、該算出したγiを式(33)に代入すれば、誤差関数を最小とする特徴情報を構成する各特徴量の重みαが特定される(図3のステップS25)。
欠陥種Aは、線材の表面の汚れであり、欠陥種Bは、線材の表面に生じたきずである。本実施形態の欠陥判別装置1は、線径が20mmの線材に生じた欠陥のうち、45個の欠陥を欠陥種Bすなわちきずと判別し、線径が13mmの線材に生じた欠陥のうち、23個の欠陥を欠陥種Bすなわちきずと判別した。一方、貫通型渦流探傷装置6では、線径が20mmの線材では31個のきず、線径が13mmの線材では15個のきずしか検出できなかった。
圧延後に巻き取られた線径が20mm及び13mmの2本の線材コイルを、各々50mmの長さに切断し、オフラインで磁粉探傷によりきずの確認を行った。その結果、線径が20mmの線材にはきずが44個存在しており、線径が13mmの線材にはきずが22個存在していた。
本実施形態の欠陥判別装置1では、線径が20mm及び13mmの2本の線材とも、過検出が1個発生したが、貫通型渦流探傷装置6では検出できなかったきずを含め、オンラインにおいて全てのきずを欠陥種Bと判定することができた。
以上のことから、本実施形態の欠陥判別装置1は、精度良く欠陥の欠陥種を判別できることが明らかとなった。
図5は、本発明の他の実施形態に係る欠陥判別装置1Aを示す概略構成図である。図5に示すように、本実施形態に係る欠陥判別装置1Aは、取得部7と、決定部3と、判別部4とを備える。取得部7は、被圧延材の撮像装置、被圧延材に渦流探傷を施す渦流探傷装置及び被圧延材に超音波探傷を施す超音波探傷装置のうちの少なくとも一つの装置から構成されている。判別部4は、前記撮像装置で撮像した被圧延材の撮像画像に画像処理を施すことによって得られる特徴量、前記渦流探傷装置で被圧延材に渦流探傷を施すことによって得られる特徴量及び前記超音波探傷装置で被圧延材に超音波探傷を施すことによって得られる特徴量のうち、少なくとも一種の特徴量を用いて、被圧延材に生じる判別対象欠陥の欠陥種を判別するように構成されている。
Claims (4)
- 欠陥種が未知の判別対象欠陥の属性を示す複数の特徴量を成分とする特徴情報を示すデータ点を、前記特徴情報を構成する特徴量の数よりも高い次元数の写像空間に写像した写像点が、前記写像空間を二分することによって形成された2つの欠陥種の領域の何れに位置するかを判別し、前記判別対象欠陥の欠陥種を前記写像点が位置すると判別した領域に対応する欠陥種であると判別する欠陥判別装置において、
前記特徴量を取得するための取得部と、前記写像空間を二分する判別境界を示す判別関数を決定する決定部と、該決定部によって決定された判別関数に前記判別対象欠陥の特徴情報を入力したときの該判別関数の出力値に基づいて、前記判別対象欠陥の欠陥種を判別する判別部とを備え、
前記2つの欠陥種のそれぞれは、予め設定された互いに異なる欠陥種であり、
前記決定部は、前記2つの欠陥種の何れであるかが既知である学習用欠陥の前記特徴情報を用いて前記判別関数を決定し、
前記判別関数は、前記2つの欠陥種のうち一方又は他方の欠陥種の前記学習用欠陥の特徴情報が入力されると、前記特徴情報が入力された学習用欠陥の写像点を出力するカーネル関数k(x,x’)と、前記カーネル関数k(x,x’)に付され、前記特徴情報を構成する各特徴量の重みとから構成された関数であり、
前記カーネル関数k(x,x’)は、要素がk(x,x’)で与えられる行列Kが半正定値であるカーネル関数であり、xは、前記一方の欠陥種の学習用欠陥の特徴情報であり、x’は、前記他方の欠陥種の学習用欠陥の特徴情報であり、
前記決定部は、
前記一方の欠陥種の学習用欠陥の特徴情報を前記カーネル関数k(x,x’)に入力したときの前記判別関数の出力値と前記一方の欠陥種に対応する値との差、及び、前記他方の欠陥種の学習用欠陥の特徴情報を前記カーネル関数k(x,x’)に入力したときの前記判別関数の出力値と前記他方の欠陥種に対応する値との差で規定され、前記2つの差の何れかの絶対値が小さくなれば小さくなり、大きくなれば大きくなる判別誤差と、前記判別関数の次元数に対し正の相関を有し、前記特徴情報を構成する各特徴量の重みに応じて変動すると共に、正則化パラメータが乗じられた正則化項との和からなる誤差関数の値を最小にするように、所定の正則化パラメータについて、前記特徴情報を構成する各特徴量の重みを特定し、
前記誤差関数の値を最小にするように特定した前記特徴情報を構成する各特徴量の重みを、前記判別関数を構成する各特徴量の重みとして仮に採用したときにおいて、前記一方の欠陥種の学習用欠陥の特徴情報を前記カーネル関数k(x,x’)に入力したときの前記判別関数の出力値と前記一方の欠陥種に対応する値との差の絶対値よりも、前記一方の欠陥種の学習用欠陥の特徴情報を前記カーネル関数k(x,x’)に入力したときの前記判別関数の出力値と前記他方の欠陥種に対応する値との差の絶対値の方が小さくなる前記一方の欠陥種の学習用欠陥の数と、前記他方の欠陥種の学習用欠陥の特徴情報を前記カーネル関数k(x,x’)に入力したときの前記判別関数の出力値と前記他方の欠陥種に対応する値との差の絶対値よりも、前記他方の欠陥種の学習用欠陥の特徴情報を前記カーネル関数k(x,x’)に入力したときの前記判別関数の出力値と前記一方の欠陥種に対応する値との差の絶対値の方が小さくなる前記他方の欠陥種の学習用欠陥の数とを加算した誤判別個数が、所定値以上の場合は、前記正則化パラメータを調整して、再度、前記誤差関数の値を最小にするように、前記特徴情報を構成する各特徴量の重みを特定し、
前記誤判別個数が所定値未満の場合は、前記誤差関数の値を最小にするように特定した前記特徴情報を構成する各特徴量の重みを、前記判別関数を構成する各特徴量の重みとして採用することを確定して、前記判別関数を決定することを特徴とする欠陥判別装置。 - 少なくとも前記取得部が被圧延材の圧延ラインに配置され、前記被圧延材に生じる判別対象欠陥の欠陥種を判別することを特徴とする請求項1に記載の欠陥判別装置。
- 少なくとも前記取得部としての被圧延材の撮像装置が前記圧延ラインに渦流探傷装置と共に配置され、
前記撮像装置で撮像した前記被圧延材の撮像画像に画像処理を施すことによって得られる特徴量を用いて、前記被圧延材に生じる判別対象欠陥の欠陥種を判別することを特徴とする請求項2に記載の欠陥判別装置。 - 前記取得部が、前記被圧延材の撮像装置、前記被圧延材に渦流探傷を施す渦流探傷装置及び前記被圧延材に超音波探傷を施す超音波探傷装置のうちの少なくとも一つの装置から構成され、
前記撮像装置で撮像した前記被圧延材の撮像画像に画像処理を施すことによって得られる特徴量、前記渦流探傷装置で前記被圧延材に渦流探傷を施すことによって得られる特徴量及び前記超音波探傷装置で前記被圧延材に超音波探傷を施すことによって得られる特徴量のうち、少なくとも一種の特徴量を用いて、前記被圧延材に生じる判別対象欠陥の欠陥種を判別することを特徴とする請求項2に記載の欠陥判別装置。
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2012520387A JP5505818B2 (ja) | 2010-06-14 | 2011-06-08 | 欠陥判別装置 |
US13/703,536 US8977580B2 (en) | 2010-06-14 | 2011-06-08 | Defect classification apparatus |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2010-135218 | 2010-06-14 | ||
JP2010135218 | 2010-06-14 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2011158711A1 true WO2011158711A1 (ja) | 2011-12-22 |
Family
ID=45348113
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2011/063119 WO2011158711A1 (ja) | 2010-06-14 | 2011-06-08 | 欠陥判別装置 |
Country Status (3)
Country | Link |
---|---|
US (1) | US8977580B2 (ja) |
JP (1) | JP5505818B2 (ja) |
WO (1) | WO2011158711A1 (ja) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110147325A (zh) * | 2019-05-22 | 2019-08-20 | 电信科学技术第十研究所有限公司 | 一种基于自动化测试的数据生成方法及装置 |
JP2021004738A (ja) * | 2019-06-25 | 2021-01-14 | 神鋼検査サービス株式会社 | 超音波探傷用機械学習装置、該方法および該プログラムならびに超音波探傷装置 |
Families Citing this family (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9589246B2 (en) * | 2014-06-26 | 2017-03-07 | Ford Global Technologies, Llc | Marking the surface of metal coils with material property data |
CN108431584B (zh) * | 2015-12-25 | 2021-04-27 | 富士胶片株式会社 | 损伤信息处理装置及损伤信息处理方法 |
KR102021944B1 (ko) * | 2017-09-20 | 2019-09-17 | 주식회사 에이치엔에스휴먼시스템 | 제철소 철강제품 품질관리를 위한 지능형 결함 제어 방법 및 시스템 |
WO2020254259A1 (en) * | 2019-06-18 | 2020-12-24 | Tetra Laval Holdings & Finance S.A. | Detection of deviations in packaging containers for liquid food |
CN110766628B (zh) * | 2019-10-16 | 2020-12-11 | 哈尔滨工程大学 | 一种基于多波段自适应正则化迭代的目标边缘反演方法 |
CN111060520B (zh) * | 2019-12-30 | 2021-10-29 | 歌尔股份有限公司 | 一种产品缺陷检测方法、装置与系统 |
CN111239263B (zh) * | 2020-01-19 | 2022-10-28 | 国网宁夏电力有限公司电力科学研究院 | 一种gis设备内部的异物缺陷的检测方法及系统 |
CN111754505B (zh) * | 2020-06-30 | 2024-03-15 | 创新奇智(成都)科技有限公司 | 辅料检测方法、装置、电子设备及存储介质 |
CN112485325B (zh) * | 2020-11-20 | 2024-04-09 | 浙江树人学院(浙江树人大学) | 一种基于pec/ut数据融合的亚表面缺陷深度检测方法及系统 |
US20220318667A1 (en) * | 2021-03-30 | 2022-10-06 | Accenture Global Solutions Limited | Intelligent real-time defect prediction, detection, and ai driven automated correction solution |
CN113219053B (zh) * | 2021-04-21 | 2022-05-13 | 大连理工大学 | 一种涂层表界面完整性参数的灵敏度矩阵超声反演方法 |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2009186243A (ja) * | 2008-02-04 | 2009-08-20 | Nippon Steel Corp | 判別装置、判別方法及びプログラム |
JP2009281742A (ja) * | 2008-05-19 | 2009-12-03 | Nippon Steel Corp | 判別方法、判別装置及びプログラム |
Family Cites Families (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6246787B1 (en) * | 1996-05-31 | 2001-06-12 | Texas Instruments Incorporated | System and method for knowledgebase generation and management |
JP4253522B2 (ja) * | 2003-03-28 | 2009-04-15 | 株式会社日立ハイテクノロジーズ | 欠陥分類方法及び装置 |
US20050177040A1 (en) * | 2004-02-06 | 2005-08-11 | Glenn Fung | System and method for an iterative technique to determine fisher discriminant using heterogenous kernels |
US7521917B2 (en) * | 2004-06-25 | 2009-04-21 | General Electric Company | Method and apparatus for testing material integrity |
GB0601982D0 (en) * | 2006-02-01 | 2006-03-15 | Rolls Royce Plc | Method and apparatus for examination of objects and structures |
-
2011
- 2011-06-08 US US13/703,536 patent/US8977580B2/en not_active Expired - Fee Related
- 2011-06-08 WO PCT/JP2011/063119 patent/WO2011158711A1/ja active Application Filing
- 2011-06-08 JP JP2012520387A patent/JP5505818B2/ja active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2009186243A (ja) * | 2008-02-04 | 2009-08-20 | Nippon Steel Corp | 判別装置、判別方法及びプログラム |
JP2009281742A (ja) * | 2008-05-19 | 2009-12-03 | Nippon Steel Corp | 判別方法、判別装置及びプログラム |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110147325A (zh) * | 2019-05-22 | 2019-08-20 | 电信科学技术第十研究所有限公司 | 一种基于自动化测试的数据生成方法及装置 |
CN110147325B (zh) * | 2019-05-22 | 2023-04-07 | 电信科学技术第十研究所有限公司 | 一种基于自动化测试的数据生成方法及装置 |
JP2021004738A (ja) * | 2019-06-25 | 2021-01-14 | 神鋼検査サービス株式会社 | 超音波探傷用機械学習装置、該方法および該プログラムならびに超音波探傷装置 |
Also Published As
Publication number | Publication date |
---|---|
US8977580B2 (en) | 2015-03-10 |
JP5505818B2 (ja) | 2014-05-28 |
JPWO2011158711A1 (ja) | 2013-08-19 |
US20130173508A1 (en) | 2013-07-04 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
JP5505818B2 (ja) | 欠陥判別装置 | |
Rifai et al. | Evaluation of turned and milled surfaces roughness using convolutional neural network | |
CN102292187B (zh) | 用于监控要在工件上实施的激光加工过程的方法和装置以及具有这种装置的激光加工头 | |
TW201921542A (zh) | 識別在一晶圓上偵測到之缺陷中之損害及所關注缺陷 | |
EP4027300B1 (en) | Apparatus, program, and method for anomaly detection and classification | |
CN105719291A (zh) | 品种可选择的表面缺陷图像分类系统 | |
Kunkel et al. | Quality assurance in metal powder bed fusion via deep-learning-based image classification | |
US20220044383A1 (en) | Learned model generation method, learned model, surface defect inspection method, steel manufacturing method, pass/fail determination method, grade determination method, surface defect determination program, pass/fail determination program, determination system, and steel manufacturing equipment | |
Shen et al. | Automatic classification of weld defects in radiographic images | |
JP5218084B2 (ja) | 検査方法 | |
Schmitt et al. | Machine vision system for inspecting flank wear on cutting tools | |
CN110427019B (zh) | 一种基于多变量判别分析的工业过程故障分类方法及控制装置 | |
Banda et al. | Machine vision and convolutional neural networks for tool wear identification and classification | |
TWI763451B (zh) | 利用自動地選擇演算法模組來檢驗樣本的系統、方法、和非暫時性電腦可讀媒體 | |
Tian et al. | Signal processing schemes for Eddy Current Testing of steam generator tubes of nuclear power plants | |
Kerscher et al. | Steel type determination by spark test image processing with machine learning | |
Jakubowski et al. | Roll wear prediction in strip cold rolling with physics-informed autoencoder and counterfactual explanations | |
Liu et al. | Deep learning for coating condition assessment with active perception | |
Guldur et al. | Automated classification of detected surface damage from point clouds with supervised learning | |
Agarwal et al. | Knowledge discovery in steel bar rolling mills using scheduling data and automated inspection | |
Jin et al. | Quality prediction and control in rolling processes using logistic regression | |
CN111695582A (zh) | 一种颤振纹理的检测方法及其装置 | |
Liu et al. | A deep learning approach to defect detection in additive manufacturing of titanium alloys | |
Elanangai et al. | Automated system for defect identification and character recognition using IR images of SS-plates | |
Karthikeyan et al. | DWT based LCP features for the classification of steel surface defects in SEM images with KNN classifier |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 11795618 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2012520387 Country of ref document: JP |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
WWE | Wipo information: entry into national phase |
Ref document number: 13703536 Country of ref document: US |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 11795618 Country of ref document: EP Kind code of ref document: A1 |