Method for using an on-line detection device for the yield elongation of cold-rolled sheet

Abstract

The present application relates to the field of measuring method and device specially used for metal rolling mill, in particular to a use method of yield elongation online detection device for cold-rolled sheet. The use method of yield elongation online detection device for cold-rolled sheet is characterized by: sequentially implementing the following steps: ① determining parameters; ② online prediction. The present application is convenient to use, accurate in measurement, high in representativeness and strong in adaptability.

Description

Technical Field

[0001] This invention relates to the field of measurement methods and apparatus specifically applicable to metal rolling mills, and more specifically to a method of using an online detection device for the yield elongation of cold-rolled thin plates. Background Technology

[0002] The yield elongation of strip steel refers to the percentage of the elongation of the extensometer gauge length to the extensometer gauge length between the start of yielding and the start of uniform work hardening in a metallic material exhibiting obvious yielding.

[0003] Currently, domestic steel companies widely adopt the offline tensile testing method for testing the yield elongation (YPel) of cold-rolled thin strip steel. This involves cutting samples from certain parts of a coil of strip steel, such as the beginning and end, and then sending them to the laboratory for offline tensile testing to obtain the yield elongation of the samples, thereby inferring the yield elongation of the entire coil of strip steel.

[0004] The offline tensile testing procedure for the specimen is as follows: Figure 1 As shown, for materials with discontinuous yielding, the yield elongation Ae is obtained by subtracting the elongation corresponding to the upper yield strength ReL from the elongation at the beginning of uniform work hardening on the force-elongation diagram. The elongation at the beginning of uniform work hardening is obtained as follows: on the curve, draw a horizontal line through the final minimum value of the discontinuous yielding stage, or a regression line through the yield range before uniform work hardening; the intersection point with the highest slope line of the curve at the beginning of uniform work hardening determines the yield elongation. Dividing the yield point elongation by the extensometer gauge length Le yields the yield elongation.

[0005] While the offline tensile testing method using cut samples is simple to operate and provides direct results, it has the following drawbacks: 1. Large data lag, offering limited assistance to the production process and failing to support online control. 2. Incomplete data, only reflecting the values ​​of the beginning and end of a single coil of strip. 3. Waste caused by shearing: During production, if the unit stops or operates at low speed for some reason, to maintain the empirical judgment that "if the beginning and end are qualified, then the middle is also qualified," a section of "suspected unqualified" strip is usually cut off. There is no standard for the amount cut, leading to the tendency to cut as much as possible and wasting steel. 4. Requires 24 / 7 human intervention at the machine, resulting in high labor intensity and high labor costs. Summary of the Invention

[0006] In order to overcome the shortcomings of the existing technology and provide a metal rolling mill testing device that is easy to use, accurate in measurement, highly representative, and highly adaptable, this invention discloses a method for using an online detection device for yield elongation of cold-rolled thin plates.

[0007] The present invention achieves its objective through the following technical solution: A method for using an online detection device for the yield elongation of cold-rolled sheet, wherein the online detection device for the yield elongation of cold-rolled sheet includes idler rollers, strip steel, base, vertical rail, horizontal rail, slide bar, lifting drive cylinder, electromagnetic detection unit, distance sensor and controller. Each idler roller is arranged parallel to each other and on the same horizontal plane. The strip is placed on each idler roller, and the strip moves when each idler roller rotates in the same direction. The base is located between two adjacent idler rollers. Four vertical rails are vertically mounted on the base. The line connecting the tops of the four vertical rails forms two rectangles with opposite sides parallel to the idler rollers. The two ends of two slide rods are movably embedded in one of the vertical rails. The idler rollers, vertical rails, and slide rods are perpendicular to each other. The two ends of the flat rail are fixed to the middle of one slide rod. The cylinder body of the lifting drive cylinder is fixed on the base. The moving end of the piston rod of the lifting drive cylinder is connected to the middle of the flat rail. The electromagnetic detection unit includes a housing, a servo drive motor, rollers, a tangential magnetic field harmonic analysis module, a Barkhausen noise detection module, and an incremental permeability detection module. The tangential magnetic field harmonic analysis module, the Barkhausen noise detection module, and the incremental permeability detection module are housed inside the housing. The servo drive motor is fixed to the bottom of the housing, and the rollers are rotatably located at the bottom of the housing. The output shaft of the servo drive motor is connected to the rollers, and the electromagnetic detection unit is movably mounted on a flat rail via the rollers. The distance sensor is fixed to the top of the electromagnetic detection unit housing, and the distance sensor is located directly below the strip steel; The lifting drive cylinder, servo drive motor, tangential magnetic field harmonic analysis module, Barkhausen noise detection module, incremental permeability detection module, and distance sensor are all connected to the controller via signal lines. The top of the vertical rail is equipped with a limit block, which is located at the extreme position of the slide bar when it moves upward; The controller can be a microcomputer, a single-chip microcomputer, or a programmable controller; Its characteristics are: it is implemented in the following steps in sequence: ① Determine the parameters: The tangential magnetic field harmonic analysis module of the electromagnetic detection unit detects and analyzes to obtain the excitation magnetic field EM1~EM1. 11 The eleven parameters, their parameter codes, parameter names, parameter units, and parameter meanings are as follows: ; The Barkhausen noise detection module of the electromagnetic detection unit detects and analyzes the excitation magnetic field to obtain the EM. 12 ~EM 18 The seven parameters, their parameter codes, parameter names, parameter units, and parameter meanings are as follows: ; The incremental permeability detection module of the electromagnetic detection unit detects and analyzes the EMF of the excitation magnetic field. 19 ~EM 25 The seven parameters, their parameter codes, parameter names, parameter units, and parameter meanings are as follows: ; The multi-frequency eddy current electromagnetic detection module of the electromagnetic detection unit detects and analyzes the EM of the excitation magnetic field. 26 ~EM 41 The sixteen parameters, their parameter codes, names, units, and meanings are as follows: ; Each raw detection electromagnetic signal EMi obtained from the detection analysis corresponds to an extended electromagnetic signal NM. i NM i = , i=1,2,…,41; Through data experiments and analysis, the following 17 electromagnetic parameters X1 to X2 were obtained from 41 electromagnetic parameters and their extended terms. 17 Used to calculate the elongation at break (Ae) value of cold-rolled sheet: ; Mathematical model for calculating Ae: Ae=A ae + +B ae (G real -4)——(a), In formula (a): Ae is the elongation after fracture of the cold-rolled sheet, A ae and B ae Both are constant terms, X i For the required electromagnetic parameter signal, C i For the corresponding regression coefficient, A ae B ae and C i All were obtained through data experiments; G real The distance between the lower surface of the strip and the probe surface is a key parameter for judgment. Due to external factors such as vibration during strip operation and the inherent plate-like properties of the strip, the distance between the strip and the distance sensor fixed on the electromagnetic detection unit fluctuates slightly. The distance G is measured in real time by the distance sensor. real G real As an input to the mathematical model for testing, G real The target value is 5mm, with an allowable error of ±1mm, when 4mm≤G real When the thickness is ≤6mm, the measurement is valid, and the test result is corrected. When Greal >6mm or G real When the thickness is less than 4mm, the system is in an abnormal state, the detection conditions are not met, and the detection is invalid.

[0008] ② Online prediction: The online prediction of strip yield elongation is achieved using a backpropagation (BP) neural network, a type of artificial neural network. The specific algorithm steps are as follows: i. Determine the BP neural network structure and network parameter settings: Before training the network, it is necessary to construct the designed network structure and set appropriate network parameters to ensure that the trained network can achieve the expected requirements; ii. Initialize weights and thresholds: BP networks that do not use a specified algorithm to set weights and thresholds are randomly assigned a set of small non-zero values; iii. Process the training sample data and feed it into the network for training: Before feeding the data into the network for training, it is necessary to normalize the training sample data of different orders of magnitude, and then feed the data into the network to start training. iv. Forward propagation phase: Calculate the output layer by layer sequentially, and obtain the final output Y at the output layer. m ; v. Compare the actual output with the expected output: If the actual output Y m With expected output O m If they match, training ends; if the actual output Y is consistent, training ends. m With expected output O m If there is a discrepancy, the backpropagation phase begins; vi. Backpropagation phase: Calculate the actual output Y m The total error E between the output Om and the expected output is distributed among all layers and nodes, and the weights and thresholds of each layer and node are adjusted accordingly. vii. The updated network is trained again: The network with adjusted weights and thresholds is used as a new model and trained again. If the training result meets the expected requirements, the training ends; otherwise, the error backpropagation stage begins again. This process is repeated until the preset number of iterations is reached or the performance function value is less than the preset error accuracy, at which point the training process stops.

[0009] The method of using the online yield elongation detection device for cold-rolled thin plates is characterized by: Step ② shall be performed as follows: A backpropagation (BP) neural network is a feedforward neural network where the signal propagates forward and the error propagates backward from the output layer. The basic principle is to use sample data, mapping and transforming it layer by layer from the input layer through hidden layers, finally obtaining the output data at the output layer. The obtained output data is compared with the true value, and the error between the two is distributed to each node in each layer to modify the weights and thresholds of each node. This forward and backward propagation process is repeated continuously until the network becomes a mapping network model that achieves the desired result.

[0010] Let x0, x1, ..., x n-1 Let y0 be the input vector, and n be the number of nodes in the input layer; y0, y1, ..., y2 are the input vectors. m-1 The output vector is a0, a1, ..., a1, ..., a2. l-1 The threshold of the hidden layer neurons is denoted by , and l is the number of nodes in the hidden layer; b0, b1, ..., b m-1 is the threshold of the output layer neuron; v(0,0), v(0,1), ..., v(n-1,l-1) are the weights from the input layer to the hidden layer; w(0,0), w(0,1), ..., w(l-1,m-1) are the weights from the hidden layer to the output layer.

[0011] The learning process of a BP neural network consists of two stages: the forward propagation stage of the input signal and the backpropagation stage of the error. The input data is transmitted to the output end through forward propagation. Then, the output data is compared with the actual data. If the output data does not match the expected data, the error is distributed to each layer and each node. The network is updated by modifying the threshold of each node and the weights between nodes in each layer. The BP neural network gradually approaches the expected model through such continuous iterative updates. The following uses a single hidden layer as an example to illustrate the working process of the BP neural network.

[0012] Each data point in the input vector has a different physical meaning, with varying values ​​and dimensions. This can easily lead to significant differences in magnitude during data processing, resulting in errors in the training model. Furthermore, excessively large datasets can slow down network training and hinder convergence. Therefore, the input data needs to be normalized before the forward propagation process.

[0013] The key reason why backpropagation (BP) neural networks can achieve nonlinear mapping lies in the role of the transfer functions between network layers. Commonly used transfer functions include the sigmoid function, tanh function, ReLU function, and softmax function, etc. Different transfer functions produce different mapping results. BP neural networks use the sigmoid function f(x) as the transfer function. The expression for the sigmoid function is: f(x) = —(1); The forward propagation process consists of the normalized input vectors x0, x1, ..., x2. n-1 The input layer passes through the weights v(0,0), v(0,1), ..., v(n-1,l-1) between the input layer and the hidden layer, and the thresholds a0, a1, ..., a1 of the hidden layer. l-1 The mapping relationship formed yields a hidden layer output H. l The formula is: H l =f1( )——(2), In equation (2): x i Let v(i,l) be the input vector, v(i,l) be the weights from the input layer to the hidden layer, and a be the weights from the input layer to the hidden layer. l is the threshold of the hidden layer, n is the dimension of the input vector, l is the number of nodes in the hidden layer, and f1 is the transfer function from the input layer to the hidden layer; The output Hl of the hidden layer is then processed by the weights w(0,0), w(0,1), ..., w(l-1,m-1) between the hidden layer and the output layer, and by the thresholds b0, b1, ..., b of the output layer. m-1 The mapping relationship formed yields the final output Y. m The formula is: Y m =f2(a i=0 (w(i,m)H i +b m )——(3), In equation (3), H i Let w(i,m) be the output vector of the hidden layer, and let b be the weights from the hidden layer to the output layer. m f1 is the threshold of the output layer, m is the dimension of the output vector, and f2 is the transfer function from the hidden layer to the output layer. When the output data obtained by the output layer does not match the expected data, the error backpropagation stage begins. The error backpropagation stage distributes the error between the output data of the output layer and the real data to each layer and each node in a certain way, thereby adjusting the weights between nodes and the thresholds of nodes, and completing the learning of the network.

[0014] The purpose of backpropagation is to continuously reduce the error between the output data and the true data. Therefore, the weights and thresholds should be adjusted in the direction of the negative gradient of the error. Let the desired output be O. k The total error is represented by the sum of squared residuals, E, and the formula is: E = ——(4), In equation (4), m is the dimension of the output vector. The purpose of dividing by 2 is that E is squared, and when differentiating the sum of squared errors, a doubling relationship will occur, so it is necessary to divide by 2 to cancel it out. The weights and thresholds should be adjusted in the direction of the negative gradient, and the amount of change should satisfy equations (5) to (8): Δv(i,l)=-α —(5), Δw(l,m)=-α —(6), Δal=-α ——(7), Δbm=-α ——(8), In equations (5) to (8): α is the learning rate, which is a constant in the interval (0,1); By continuously adjusting the weights and thresholds between nodes through forward and backward propagation, the desired network model is finally obtained, which outputs the elongation at break of the cold-rolled sheet, and then the yield elongation is calculated.

[0015] This invention develops a method for online measurement of the yield elongation of cold-rolled thin strip. By applying comprehensive electromagnetic detection to the running strip, multiple electromagnetic signals are acquired in real time. The electromagnetic signals are expanded, and the spacing affecting the electromagnetic parameters is corrected, as well as the influence of the strip thickness, are taken into account. The developed method does not rely on the real-time process parameters of the unit, and achieves the purpose of online accurate measurement of the yield elongation of the strip.

[0016] This invention is applied to an online detection system for the mechanical properties and quality of cold-rolled strip steel. It enables real-time online detection of indicators such as yield elongation of cold-rolled strip steel, realizing continuous detection, classification, and recording of steel plate production quality. This will play a very positive role in improving production efficiency, product quality, and product competitiveness.

[0017] The present invention has the following advantages: it is easy to use, accurate in measurement, highly representative, and highly adaptable. Attached Figure Description

[0018] Figure 1 This is a graph showing the yield elongation test. Figure 2 This is a schematic diagram of the main view direction of the online yield elongation detection device for cold-rolled thin plates. Figure 3 This is a top-view schematic diagram of the online yield elongation testing device for cold-rolled thin plates. Figure 4 This is a schematic diagram of the electromagnetic detection unit in an online yield elongation testing device for cold-rolled thin plates. Figure 5 This is the neural network topology diagram used in this invention. Figure 6 It is the graph of the Sigmoid function. Figure 7This is a graph showing the test results of the yield elongation of the strip along its entire length using this invention. Detailed Implementation

[0019] The present invention will be further illustrated below through specific embodiments. Example

[0020] An online detection device for the yield elongation of cold-rolled sheet includes a roller 11, a strip 12, a base 2, a vertical rail 31, a horizontal rail 32, a slide bar 4, a lifting drive cylinder 5, an electromagnetic detection unit 6, a distance sensor 7, and a controller 8. Figures 2-4 As shown, the specific structure is: Each idler roller 11 is arranged parallel to each other and on the same horizontal plane. The strip steel 12 is placed on each idler roller 11. When each idler roller 11 rotates in the same direction, it pushes the strip steel 12 to move. The base 2 is placed between two adjacent idler rollers 11. Four vertical rails 31 are vertically arranged on the base 2. The line connecting the top of the four vertical rails 31 forms two rectangles with opposite sides parallel to the idler rollers 11. The two ends of the two slide rods 4 are movably embedded in one of the vertical rails 31. The idler rollers 11, vertical rails 31 and slide rods 4 are perpendicular to each other. The two ends of the flat rail 32 are fixed to the middle of one slide rod 4. The cylinder body of the lifting drive cylinder 5 is fixed on the base 2. The moving end of the piston rod of the lifting drive cylinder 5 is connected to the middle of the flat rail 32. The electromagnetic detection unit 6 includes a housing 61, a servo drive motor 62, a roller 63, a tangential magnetic field harmonic analysis module, a Barkhausen noise detection module, and an incremental permeability detection module. The housing 61 houses the tangential magnetic field harmonic analysis module, the Barkhausen noise detection module, and the incremental permeability detection module. The servo drive motor 62 is fixed to the bottom of the housing 61, and the roller 63 is rotatably disposed at the bottom of the housing 61. The output shaft of the servo drive motor 62 is connected to the roller 63. The electromagnetic detection unit 6 is movably disposed on the flat rail 32 via the roller 63. The distance sensor 7 is fixed on the top of the housing 61 of the electromagnetic detection unit 6, and the distance sensor 7 is located directly below the strip steel 12; The lifting drive cylinder 5, the servo drive motor 62, the tangential magnetic field harmonic analysis module, the Barkhausen noise detection module, and the distance sensor 7 are all connected to the controller 8 via signal lines.

[0021] In this embodiment: a limiting block 311 is provided at the top of the vertical rail 31, and the limiting block 311 is located at the extreme position of the slide bar 4 when it moves upward.

[0022] In this embodiment, the controller 8 is selected from microcomputers, single-chip microcomputers or programmable controllers.

[0023] When using this embodiment, follow these steps in sequence: ① Determine the parameters: The tangential magnetic field harmonic analysis module of electromagnetic detection unit 6 detects and analyzes to obtain the excitation magnetic field EM1~EM1. 11 The eleven parameters, their parameter codes, parameter names, parameter units, and parameter meanings are as follows: ; The Barkhausen noise detection module of electromagnetic detection unit 6 detects and analyzes seven parameters of the excitation magnetic field EM12 to EM18. The parameter codes, parameter names, parameter units, and parameter meanings of the seven parameters are as follows: ; The incremental permeability detection module of electromagnetic detection unit 6 detects and analyzes to obtain seven parameters of the excitation magnetic field EM19 to EM25. The parameter codes, parameter names, parameter units, and parameter meanings of the seven parameters are as follows: ; The multi-frequency eddy current electromagnetic detection module of electromagnetic detection unit 6 detects and analyzes to obtain sixteen parameters of the excitation magnetic field EM26~EM41. The parameter codes, parameter names, parameter units, and parameter meanings of the sixteen parameters are as follows: ; The raw detection electromagnetic signal EM obtained from each detection analysis i Each corresponds to an extended electromagnetic signal NM i NM i = , i=1,2,…,41; Through data experiments and analysis, the following 17 electromagnetic parameters X1 to X2 were obtained from 41 electromagnetic parameters and their extended terms. 17 Used to calculate the elongation at break (Ae) value of cold-rolled sheet: ; Mathematical model for calculating Ae: Ae=A ae + +B ae (G real -4)——(a), In formula (a): Ae is the elongation after fracture of the cold-rolled sheet, A ae and B ae Both are constant terms, X i For the required electromagnetic parameter signal, C i For the corresponding regression coefficient, A ae B ae and C i All were obtained through data experiments; G realThe distance between the lower surface of the strip 12 and the probe surface is a key parameter for judgment. Due to external factors such as vibration during the operation of the strip 12 and the inherent plate-like motion of the strip 12, the distance between the strip 12 and the distance sensor 7 fixed on the electromagnetic detection unit 6 fluctuates slightly. The distance G is measured in real time by the distance sensor 7. real G real As an input to the mathematical model for testing, G real The target value is 5mm, with an allowable error of ±1mm, when 4mm≤G real When the thickness is ≤6mm, the measurement is valid, and the test result is corrected. When G real >6mm or G real When the thickness is less than 4mm, the system is in an abnormal state, the detection conditions are not met, and the detection is invalid.

[0024] ② Online prediction: The online prediction of strip yield elongation is achieved using a backpropagation (BP) neural network, a type of artificial neural network. The specific algorithm steps are as follows: i. Determine the BP neural network structure and network parameter settings: Before training the network, it is necessary to construct the designed network structure and set appropriate network parameters to ensure that the trained network can achieve the expected requirements; ii. Initialize weights and thresholds: BP networks that do not use a specified algorithm to set weights and thresholds are randomly assigned a set of small non-zero values; iii. Process the training sample data and feed it into the network for training: Before feeding the data into the network for training, it is necessary to normalize the training sample data of different orders of magnitude, and then feed the data into the network to start training. iv. Forward propagation phase: Calculate the output layer by layer sequentially, and obtain the final output Y at the output layer. m ; v. Compare the actual output with the expected output: If the actual output Y m With expected output O m If they match, training ends; if the actual output Y is consistent, training ends. m With expected output O m If there is a discrepancy, the backpropagation phase begins; vi. Backpropagation phase: Calculate the actual output Y m With expected output O m The total error E between layers is distributed to each node, and the weights and thresholds of each node in each layer are adjusted. vii. The updated network is trained again: The network with adjusted weights and thresholds is used as a new model and trained again. If the training result meets the expected requirements, the training ends; otherwise, the error backpropagation stage is entered again. This process is repeated until the number of iterations reaches the preset number or the performance function value is less than the preset error accuracy, at which point the training process stops.

[0025] Step ② shall be performed as follows: A backpropagation (BP) neural network is a feedforward neural network where the signal propagates forward and the error propagates backward from the output layer. The basic principle is to use sample data, mapping and transforming it layer by layer from the input layer through hidden layers, finally obtaining the output data at the output layer. The obtained output data is compared with the true value, and the error between the two is distributed to each node in each layer to modify the weights and thresholds of each node. This forward and backward propagation process is repeated continuously until the network becomes a mapping network model that achieves the desired result. Its topology is as follows: Figure 5 As shown.

[0026] Figure 5 In the middle, take x0, x1, ..., x n-1 Let y0 be the input vector, and n be the number of nodes in the input layer; y0, y1, ..., y2 are the input vectors. m-1 The output vector is a0, a1, ..., a1, ..., a2. l-1 The threshold of the hidden layer neurons is denoted by , and l is the number of nodes in the hidden layer; b0, b1, ..., b m-1 is the threshold of the output layer neuron; v(0,0), v(0,1), ..., v(n-1,l-1) are the weights from the input layer to the hidden layer; w(0,0), w(0,1), ..., w(l-1,m-1) are the weights from the hidden layer to the output layer. The learning process of a BP neural network consists of two stages: the forward propagation stage of the input signal and the backpropagation stage of the error. The input data is transmitted to the output end through forward propagation. Then, the output data is compared with the actual data. If the output data does not match the expected data, the error is distributed to each layer and each node. The network is updated by modifying the threshold of each node and the weights between nodes in each layer. The BP neural network gradually approaches the expected model through such continuous iterative updates. The following uses a single hidden layer as an example to illustrate the working process of the BP neural network.

[0027] Each data point in the input vector has a different physical meaning, with varying values ​​and dimensions. This can easily lead to significant differences in magnitude during data processing, resulting in errors in the training model. Furthermore, excessively large datasets can slow down network training and hinder convergence. Therefore, the input data needs to be normalized before the forward propagation process.

[0028] The key reason why backpropagation (BP) neural networks can achieve nonlinear mapping lies in the role of the transfer functions between network layers. Commonly used transfer functions include the sigmoid function, tanh function, ReLU function, and softmax function, etc. Different transfer functions produce different mapping results. BP neural networks use the sigmoid function f(x) as the transfer function. The expression for the sigmoid function is: f(x) = —(1), the graph of the function is as follows Figure 6 As shown.

[0029] The forward propagation process consists of the normalized input vectors x0, x1, ..., x2. n-1 The input layer passes through the weights v(0,0), v(0,1), ..., v(n-1,l-1) between the input layer and the hidden layer, and the thresholds a0, a1, ..., a1 of the hidden layer. l-1 The mapping relationship formed yields a hidden layer output H. l The formula is: H l =f1( )——(2), In equation (2): x i Let v(i,l) be the input vector, v(i,l) be the weights from the input layer to the hidden layer, and a be the weights from the input layer to the hidden layer. l is the threshold of the hidden layer, n is the dimension of the input vector, l is the number of nodes in the hidden layer, and f1 is the transfer function from the input layer to the hidden layer; The output Hl of the hidden layer is then processed by the weights w(0,0), w(0,1), ..., w(l-1,m-1) between the hidden layer and the output layer, and by the thresholds b0, b1, ..., b of the output layer. m-1 The mapping relationship formed yields the final output Y. m The formula is: Y m =f2(a i=0 (w(i,m)H i +b m )——(3), In equation (3), H i Let w(i,m) be the output vector of the hidden layer, and let b be the weights from the hidden layer to the output layer. m f1 is the threshold of the output layer, m is the dimension of the output vector, and f2 is the transfer function from the hidden layer to the output layer. When the output data obtained by the output layer does not match the expected data, the error backpropagation stage begins. The error backpropagation stage distributes the error between the output data and the actual data across all layers and nodes in a specific way, thereby adjusting the weights and thresholds of the nodes and completing the network's learning process. The purpose of backpropagation is to continuously reduce the error between the output data and the true data. Therefore, the weights and thresholds should be adjusted in the direction of the negative gradient of the error. Let the desired output be O. k The total error is represented by the sum of squared residuals, E, and the formula is: E = ——(4), In equation (4), m is the dimension of the output vector. The purpose of dividing by 2 is that E is squared, and when differentiating the sum of squared errors, a doubling relationship will occur, so it is necessary to divide by 2 to cancel it out. The weights and thresholds should be adjusted in the direction of the negative gradient, and the amount of change should satisfy equations (5) to (8): Δv(i,l)=-α —(5), Δw(l,m)=-α —(6), Δa l =-α ——(7), Δb m =-α ——(8), In equations (5) to (8): α is the learning rate, which is a constant in the interval (0,1); By continuously adjusting the weights and thresholds between nodes through forward and backward propagation, the desired network model is finally obtained, which outputs the elongation at break of the cold-rolled sheet, and then the yield elongation is calculated.

[0030] Through a data experiment of a certain scale, A was obtained. ae =-8227.6, B ae =2.392, the electromagnetic parameter set X and the corresponding coefficient C are shown in the table below: .

[0031] Substituting Aae, Bae, and coefficient C, along with the input parameters X and Greal obtained from real-time detection, the calculation results are shown in the table below: , Continued: .

[0032] The model in this embodiment was used for real-time detection of a coil of steel strip. The detection results of the yield elongation Ae along the entire length are as follows: Figure 7 As shown in the figure. Compared to existing technologies that can only test by shearing the sample, the amount of data and real-time performance are greatly improved.

[0033] This embodiment was used to measure the elongation after fracture of 100 coils of SEDDQ strip steel on a certain production line online. The same sample was taken from the beginning and end, and the elongation after fracture was obtained by offline tensile testing. A total of 2000 sets of results were obtained. The results were compared with the corresponding position values ​​of online measurement, and the reliability was 92%.

Claims

1. A method of using an online detection device for the yield elongation of cold-rolled sheet, the online detection device for the yield elongation of cold-rolled sheet comprising a roller (11), a strip (12), a base (2), a vertical rail (31), a horizontal rail (32), a slide bar (4), a lifting drive cylinder (5), an electromagnetic detection unit (6), a distance sensor (7), and a controller (8). Each idler roller (11) is arranged parallel to each other and on the same horizontal plane. The strip steel (12) is placed on each idler roller (11). When each idler roller (11) rotates in the same direction, it pushes the strip steel (12) to move. The base (2) is located between two adjacent rollers (11). Four vertical rails (31) are vertically arranged on the base (2). The line connecting the tops of the four vertical rails (31) forms two rectangles with opposite sides parallel to the rollers (11). The two ends of the two slide rods (4) are movably embedded in one of the vertical rails (31). The rollers (11), the vertical rails (31) and the slide rods (4) are perpendicular to each other. The two ends of the flat rail (32) are fixed to the middle of one slide rod (4). The cylinder body of the lifting drive cylinder (5) is fixed on the base (2). The moving end of the piston rod of the lifting drive cylinder (5) is connected to the middle of the flat rail (32). The electromagnetic detection unit (6) includes a housing (61), a servo drive motor (62), a roller (63), a tangential magnetic field harmonic analysis module, a Barkhausen noise detection module, an incremental permeability detection module, and a multi-frequency eddy current electromagnetic detection module. The housing (61) is equipped with the tangential magnetic field harmonic analysis module, the Barkhausen noise detection module, and the incremental permeability detection module. The servo drive motor (62) is fixed at the bottom of the housing (61), and the roller (63) is rotatably disposed at the bottom of the housing (61). The output shaft of the servo drive motor (62) is connected to the roller (63). The electromagnetic detection unit (6) is movably disposed on the flat rail (32) via the roller (63). The distance sensor (7) is fixed on the top of the housing (61) of the electromagnetic detection unit (6), and the distance sensor (7) is located directly below the strip (12); The lifting drive cylinder (5), the servo drive motor (62), the tangential magnetic field harmonic analysis module, the Barkhausen noise detection module, the incremental permeability detection module and the distance sensor (7) are all connected to the controller (8) via signal lines; The top of the vertical rail (31) is provided with a limiting block (311), which is located at the limit position of the slide bar (4) when it moves upward. The controller (8) is a microcomputer, single-chip microcomputer or programmable controller; Its characteristics are: Follow these steps in sequence: ① Determine the parameters: The tangential magnetic field harmonic analysis module of the electromagnetic detection unit (6) detects and analyzes to obtain the excitation magnetic field EM1~ ... 11 The eleven parameters, their parameter codes, parameter names, parameter units, and parameter meanings are as follows: ； The Barkhausen noise detection module of the electromagnetic detection unit (6) detects and analyzes the excitation magnetic field to obtain the EM. 12 ~EM 18 The seven parameters, their parameter codes, parameter names, parameter units, and parameter meanings are as follows: ； The incremental permeability detection module of the electromagnetic detection unit (6) detects and analyzes the EM of the excitation magnetic field to obtain the excitation magnetic field. 19 ~EM 25 The seven parameters, their parameter codes, parameter names, parameter units, and parameter meanings are as follows: ； The multi-frequency eddy current electromagnetic detection module of the electromagnetic detection unit (6) detects and analyzes the excitation magnetic field to obtain the EM. 26 ~EM 41 The sixteen parameters, their parameter codes, names, units, and meanings are as follows: ； The raw detection electromagnetic signal EM obtained from each detection analysis i Each corresponds to an extended electromagnetic signal NM i NM i = , i=1,2,…,41; Through data experiments and analysis, the following 17 electromagnetic parameters X1 to X2 were obtained from 41 electromagnetic parameters and their extended terms. 17 Used to calculate the elongation at break (Ae) value of cold-rolled sheet: ； Mathematical model for calculating Ae: Ae=A ae + +B ae (G real -4)——(a), In formula (a): Ae is the elongation after fracture of the cold-rolled sheet, A ae and B ae Both are constant terms, X i For the required electromagnetic parameter signal, C i For the corresponding regression coefficient, A ae B ae and C i All were obtained through data experiments; G real The distance between the lower surface of the strip (12) and the probe surface is a key parameter for judgment. Due to the vibration of the strip (12) during operation and the influence of external factors such as the inherent plate nature of the strip (12), the distance between the strip (12) and the distance sensor (7) fixed on the electromagnetic detection unit (6) fluctuates slightly. The distance G is measured in real time by the distance sensor (7). real G real As an input to the mathematical model for testing, G real The target value is 5mm, with an allowable error of ±1mm, when 4mm≤G real When the thickness is ≤6mm, the measurement is valid, and the test result is corrected. When G real >6mm or G real When the thickness is less than 4mm, the system is in an abnormal state, the detection conditions are not met, and the detection is invalid. ② Online prediction: The online prediction of strip yield elongation is achieved using a backpropagation (BP) neural network, a type of artificial neural network. The specific algorithm steps are as follows: i. Determine the BP neural network structure and network parameter settings: Before training the network, it is necessary to construct the designed network structure and set appropriate network parameters to ensure that the trained network can achieve the expected requirements; ii. Initialize weights and thresholds: BP networks that do not use a specified algorithm to set weights and thresholds are randomly assigned a set of small non-zero values; iii. Process the training sample data and feed it into the network for training: Before feeding the data into the network for training, it is necessary to normalize the training sample data of different orders of magnitude, and then feed the data into the network to start training. iv. Forward propagation phase: Calculate the output layer by layer sequentially, and obtain the final output Y at the output layer. m ; v. Compare the actual output with the expected output: If the actual output Y m With expected output O m If they match, training ends; if the actual output Y is consistent, training ends. m With expected output O m If there is a discrepancy, the backpropagation phase begins; vi. Backpropagation phase: Calculate the actual output Y m With expected output O m The total error E between layers is distributed to each node, and the weights and thresholds of each node in each layer are adjusted. vii. The updated network is trained again: The network with adjusted weights and thresholds is used as a new model and trained again. If the training result meets the expected requirements, the training ends; otherwise, the error backpropagation stage is entered again. This process is repeated until the number of iterations reaches the preset number or the performance function value is less than the preset error accuracy, at which point the training process stops.

2. The method of using the online yield elongation detection device for cold-rolled thin plates as described in claim 1, characterized in that: Let x0, x1, ..., x n-1 Let y0 be the input vector, and n be the number of nodes in the input layer; y0, y1, ..., y2 are the input vectors. m-1 The output vector is a0, a1, ..., a1, ..., a2. l-1 The threshold of the hidden layer neurons is denoted by , and l is the number of nodes in the hidden layer; b0, b1, ..., b m-1 is the threshold of the output layer neuron; v(0,0), v(0,1), ..., v(n-1,l-1) are the weights from the input layer to the hidden layer; w(0,0), w(0,1), ..., w(l-1,m-1) are the weights from the hidden layer to the output layer. The learning process of a BP neural network consists of two stages: the forward propagation stage of the input signal and the backpropagation stage of the error. The input data is propagated to the output through forward propagation. The output data is then compared with the actual data. If the output data does not match the expected data, the error is distributed across all layers and nodes. The network is updated by modifying the thresholds of each node and the weights between nodes in each layer. Using the sigmoid function f(x) as the transfer function, the expression for the sigmoid function is: f(x) = —(1); The forward propagation process consists of the normalized input vectors x0, x1, ..., x2. n-1 The input layer passes through the weights v(0,0), v(0,1), ..., v(n-1,l-1) between the input layer and the hidden layer, and the thresholds a0, a1, ..., a1 of the hidden layer. l-1 The mapping relationship formed yields a hidden layer output H. l The formula is: H l =f1( )——(2), In equation (2): x i Let v(i,l) be the input vector, v(i,l) be the weights from the input layer to the hidden layer, and a be the weights from the input layer to the hidden layer. l is the threshold of the hidden layer, n is the dimension of the input vector, l is the number of nodes in the hidden layer, and f1 is the transfer function from the input layer to the hidden layer; The output H of the hidden layer l Then, the weights w(0,0), w(0,1), ..., w(l-1,m-1) between the hidden layer and the output layer, and the thresholds b0, b1, ..., b of the output layer are used. m-1 The mapping relationship formed yields the final output Y. m The formula is: Y m =f2(a i=0 (w(i,m)H i +b m )——(3), In equation (3), H i Let w(i,m) be the output vector of the hidden layer, and let b be the weights from the hidden layer to the output layer. m f1 is the threshold of the output layer, m is the dimension of the output vector, and f2 is the transfer function from the hidden layer to the output layer. Let the expected output be O k The total error is represented by the sum of squared residuals, E, and the formula is: E = ——(4), In equation (4), m is the dimension of the output vector. The purpose of dividing by 2 is that E is squared, and when differentiating the sum of squared errors, a doubling relationship will occur, so it is necessary to divide by 2 to cancel it out. The weights and thresholds should be adjusted in the direction of the negative gradient, and the amount of change should satisfy equations (5) to (8): Δv(i,l)=-α ——(5), Δw(l,m)=-α ——(6), Δal=-a ——(7), Δbm=-a ——(8), In equations (5) to (8): α is the learning rate, which is a constant in the interval (0,1); By continuously adjusting the weights and thresholds between nodes through forward and backward propagation, the desired network model is finally obtained, which outputs the elongation at break of the cold-rolled sheet, and then the yield elongation is calculated.

Images