Method of using an on-line device for measuring yield and tensile strength of low carbon steel 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 low carbon steel cold-rolled sheet yield and tensile strength on-line measuring device.A use method of low carbon steel cold-rolled sheet yield and tensile strength on-line measuring device, characterized by: sequentially implementing the following steps: ① determining parameters: ② on-line prediction.The present application is convenient to use, accurate in measurement, high in representativeness and strong in adaptability.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of measuring methods and devices specially applicable to metal rolling mills, in particular to a method for using a device for on-line measuring of the yield and tensile strength of low-carbon cold-rolled sheet steel. BACKGROUND

[0002] In the production of cold-rolled strip steel, low-carbon steel is an important raw material for downstream users such as automobile and home appliance industries due to its easy workability. The downstream users have strict requirements on various indicators of the base material performance of the strip steel. Among them, the size and stability of the strength indicators of the cold-rolled sheet steel, including the yield strength and the tensile strength, are important standards for measuring the quality of the product and are the main basis for product design and material selection. Providing the strip steel with accurate and qualified strength indicators to the users is one of the prerequisites for steel plants to improve their market competitiveness.

[0003] Currently, there are mainly two methods for detecting the strength characteristics of cold-rolled sheet steel, namely, the off-line tensile method and the residual magnetism measurement method.

[0004] The off-line tensile method: This is the method widely used at present. That is, samples are cut from a certain part of a roll of strip steel, such as the head and tail, and then sent to a tensile testing machine for tensile testing to obtain the strength values of the samples, from which the strength values of a section of the strip steel are inferred. This method has the advantages of simplicity, direct results, and high precision. However, this method has the following disadvantages: First, the data time lag is large, and the measured values are often obtained after several hours or even a day, which is of limited help to the production process and is even more impossible for on-line control. Second, the data are incomplete, and only the physical property data values of the head and tail of a roll of strip steel can be reflected, which causes low user satisfaction. Third, there is waste in cutting. When the rolling mill is stopped or produces at low speed due to some reason, in order to maintain the experience judgment that "if the head and tail are qualified, then the middle part is also qualified", a section of "suspected unqualified" strip steel is usually cut off. There is no judgment standard for how much to cut, and only as much as possible is cut, which obviously causes waste. Fourth, it is necessary to have someone working at the mill all day long, which is high in labor intensity and labor cost.

[0005] Residual magnetism measurement method: the main principle is to use pulse magnetic field excitation, the strip steel runs through two groups of rollers, the upper and lower of the strip steel in this area are arranged with measuring devices, this measuring method essentially uses the magnetization of electromagnetic materials, and then arranges residual magnetism signal acquisition coil downstream. From the electromagnetic detection, this detection method is single detection, that is, only one electromagnetic model, which is commonly known as IMPOC value in the industry. The model input variables of this method have four items, which are residual magnetism parameter (IM), strip steel thickness (TH), production line skin pass mill elongation (SKD) and production line stretch leveling machine elongation (TLD), of which only one is an electrical measurement parameter value, that is, the residual magnetism parameter (IM), and the other three are characteristic values transmitted by the mill computer to the detection computer. This detection method has simple equipment and easy-to-construct detection model. The disadvantage is that the detection principle only uses one electromagnetic characteristic value, the model depends on the process parameters of the strip steel and the production line, and the detection accuracy is easily affected by these parameters. In addition, it must be installed after the skin pass mill and the stretch leveling machine, and for some occasions without these equipment or without investment, the model detection only depends on the thickness and residual magnetism value of the strip steel, and the applicability of the measurement is relatively weak. SUMMARY

[0006] In order to overcome the defects of the prior art, a metal material performance test method which is convenient to use, accurate in measurement, high in representativeness and strong in adaptability is provided, and a use method of a low-carbon steel cold-rolled sheet yield and tensile strength on-line measuring device is disclosed.

[0007] The application achieves the purpose of the application through the following technical scheme:

[0008] A use method of a low-carbon steel cold-rolled sheet yield and tensile strength on-line measuring device, the low-carbon steel cold-rolled sheet yield and tensile strength on-line measuring device comprises rollers, a strip steel, a base, vertical rails, flat rails, slide rods, lifting drive cylinders, an electromagnetic detection unit, distance sensors and a controller,

[0009] Each roller is arranged parallel to each other and on the same horizontal plane, the strip steel is arranged on each roller, and the rollers rotate in the same direction to push the strip steel to move;

[0010] The base is arranged between two adjacent rollers, four vertical rails are vertically arranged on the base, the connecting line of the top ends of the four vertical rails forms a rectangle with the parallel rollers, the two ends of each slide rod are movably embedded in one vertical rail, the three of the roller, the vertical rail and the slide rod are perpendicular to each other in pairs, the two ends of the flat rail are fixed on the middle part of one slide rod, the cylinder body of the lifting drive cylinder is fixed on the base, and the moving end of the piston rod of the lifting drive cylinder is connected to the middle part of the flat rail;

[0011] The electromagnetic detection unit comprises a shell, a servo drive motor, a roller, a tangent magnetic field harmonic analysis module, a Barkhausen noise detection module, an incremental magnetic permeability detection module, a multi-frequency eddy current electromagnetic detection module and an electromagnetic ultrasonic nondestructive detection module, the shell is internally provided with the tangent magnetic field harmonic analysis module, the Barkhausen noise detection module, the incremental magnetic permeability detection module, the multi-frequency eddy current electromagnetic detection module and the electromagnetic ultrasonic nondestructive detection module, the servo drive motor is fixed at the bottom of the shell, the roller is rotatably arranged at the bottom of the shell, the output shaft of the servo drive motor is connected with the roller, and the electromagnetic detection unit is movably arranged on the flat rail through the roller;

[0012] The distance sensor is fixed at the top of the electromagnetic detection unit shell and arranged directly below the strip steel;

[0013] The lifting drive cylinder, the servo drive motor, the tangent magnetic field harmonic analysis module, the Barkhausen noise detection module, the incremental magnetic permeability detection module and the distance sensor are connected with the controller through signal lines;

[0014] The top of the vertical rail is provided with a limiting block, and the limiting block is arranged at the limit position of upward movement of the sliding rod;

[0015] The controller is selected from a microcomputer, a single-chip microcomputer or a programmable controller;

[0016] The features are that the following steps are sequentially implemented:

[0017] ① Determining parameters:

[0018] The tangent magnetic field harmonic analysis module of the electromagnetic detection unit detects and analyzes eleven parameters EM1-EM11 of the excitation magnetic field, and the parameter code, parameter name, parameter unit and parameter meaning of the eleven parameters are as follows:

[0019] Parameter code Parameter name Parameter unit Parameter meaning EM1 A3 A / cm Amplitude of the third harmonic EM2 A5 A / cm Amplitude of the fifth harmonic EM3 A7 A / cm Amplitude of the seventh harmonic EM4 P3 Rad Phase of the third harmonic EM5 P5 Rad Phase of the fifth harmonic EM6 P7 Rad Phase of the seventh harmonic EM7 UHS A / cm Sum of the amplitudes of the third, fifth, seventh and ninth harmonics EM8 K ％ Deformation coefficient EM9 Hco A / cm Coercive field strength EM10 Hro A / cm Harmonic amplitude at the zero point of the hysteresis loop EM11 Vmag V Steady voltage of the electromagnetic coil ;

[0020] The Barkhausen noise detection module of the electromagnetic detection unit detects and analyzes seven parameters EM12-EM18 of the excitation magnetic field, and the parameter code, parameter name, parameter unit and parameter meaning of the seven parameters are as follows:

[0021] Parameter code Parameter name Parameter unit Parameter meaning EM12 MMAX V Maximum amplitude of the excitation field EM13 MMEAN V Average value of the amplitude of the excitation field over one field period EM14 MR V Remanence point amplitude EM15 HCM A / cm Coercive field strength for M = MMAX EM16 DH25M A / cm Width of the Barkhausen curve for M = 25% MMAX EM17 DH50M A / cm Width of the Barkhausen curve for M = 50% MMAX EM18 DH75M A / cm Width of the Barkhausen curve for M = 75% MMAX ;

[0022] The incremental magnetic permeability detection module of the electromagnetic detection unit detects and analyzes seven parameters EM19-EM25 of the excitation magnetic field, and the parameter code, parameter name, parameter unit and parameter meaning of the seven parameters are as follows:

[0023] Parameter code Parameter name Parameter unit Parameter meaning EM19 UMAX V Maximum amplitude of the excitation field EM20 UMEAN V Average value of the amplitude of the excitation field over one field period EM21 UR V Remanence point amplitude EM22 HCU A / cm Coercive field strength for U = UMAX EM23 DH25U A / cm Width of the permeability curve for U = 25% UMAX EM24 DH50U A / cm Width of the permeability curve for U = 50% UMAX EM25 DH75U A / cm Width of the permeability curve for U = 75% UMAX ;

[0024] The multi-frequency eddy current electromagnetic detection module of the electromagnetic detection unit detects and analyzes sixteen parameters of the excitation magnetic field EM26-EM41, and the parameter code, parameter name, parameter unit and parameter meaning of the sixteen parameters are as follows:

[0025]

[0026] The electromagnetic ultrasonic nondestructive testing module of the electromagnetic detection unit detects and analyzes three parameters of the excitation magnetic field EM42-EM44, and the parameter code, parameter name, parameter unit and parameter meaning of the three parameters are as follows:

[0027] Parameter code Parameter name Parameter unit Parameter meaning EM42 UEMAT_MAX V Peak value of the butterfly diagram EM43 UEMAT_MIN V Valley value of the butterfly diagram EM44 URatio V Ratio of the two peak values

[0028] Each of the original detection electromagnetic signals EM i corresponds to an extended electromagnetic signal NM i ,

[0029] ② Online prediction:

[0030] The online prediction of the yield elongation of the strip steel is realized based on the BP neural network in the artificial neural network, and the specific algorithm steps are as follows:

[0031] i. Determine the BP neural network structure and network parameter setting: before network training, the designed network structure needs to be constructed, and appropriate network parameters need to be set to ensure that the trained network can meet the expected requirements;

[0032] ii. Initialize the weight and threshold: the BP network with unspecified algorithm to set the weight and threshold adopts a set of small non-zero values randomly;

[0033] iii. Process the training sample data and send it to the network for training. Before sending the data to the network for training, the training sample data of different orders of magnitude needs to be normalized, and then the data is sent to the network for training;

[0034] iv. Forward propagation stage: calculate the output of each layer in turn, and get the final output Ym in the output layer;

[0035] v. Compare the actual output with the expected output. If the actual output Ym is consistent with the expected output Om, the training is ended; if the actual output Ym is not consistent with the expected output Om, the back propagation stage is started;

[0036] vi. Back propagation stage: find the total error E between the actual output Ym and the expected output Om, and allocate the total error E to each layer and each node to adjust the weight and threshold of each layer and each node;

[0037] vii. The updated network is trained next time: the network with adjusted weights and thresholds is taken as a new model for training again, if the result after training is consistent with the expected requirement, the training is ended, if it is not consistent, it enters the error back propagation stage again, and the cycle is repeated, when the iteration number reaches the preset number or the performance function value is less than the preset error precision, the training process is stopped.

[0038] The use method of the low-carbon steel cold-rolled sheet yield and tensile strength online measurement device is characterized in that:

[0039] Step 2 is implemented as follows:

[0040] The BP (Back Propagation) neural network algorithm is a multi-layer feedforward network learning and training according to the error back propagation algorithm, and the network structure is composed of three parts of input layer, hidden layer and output layer, wherein the input layer and the output layer each have only one, and the hidden layer can have multiple, and each layer is composed of a plurality of neurons. The nodes in the layer are not connected to each other, and the nodes between the layers adopt a full connection mode, that is, any one neuron node of the input layer is connected to all nodes of the hidden layer, and any one neuron node of the hidden layer is connected to all nodes of the output layer.

[0041] The execution process of the BP neural network learning algorithm can be summarized as two main parts: forward propagation of signals and backward propagation of errors. For each training sample, during forward propagation, the input vector is transmitted from the input layer to the output layer layer by layer; during backward propagation, the error is transmitted from the hidden layer to the input layer layer by layer in a certain form. The two processes are repeated until the set termination condition is met.

[0042] I. Determine the network structure:

[0043] The input layer and the output layer are fixed to have only one, and the number of neuron nodes of the input layer and the output layer is determined according to the actual problem to be solved and the data representation method, so the key of the network structure design lies in the design of the hidden layer structure, including the selection of the number of layers and the number of nodes. For the design of the number of hidden layers, usually one hidden layer is first considered, and when a hidden layer cannot meet the related performance requirements, the number of hidden layers is increased. Experimental research shows that when two hidden layers are used, if the number of neuron nodes of the first hidden layer is more than that of the second hidden layer, the performance of the BP network will be improved to a certain extent. However, for some practical problems, the number of neuron nodes required by the double hidden layer may be less than that of the single hidden layer. The more the number of hidden layers, the longer the training and learning time of the BP network, and the local minimum error will also increase, that is, the probability of the network falling into a local optimal solution will increase, so only when the increase of the number of hidden layer nodes cannot significantly improve the network performance, the number of hidden layers is considered to be increased.

[0044] The number of nodes of the hidden layer neurons is determined according to the following empirical estimation formula:

[0045]

[0046] m=log2n-(k),

[0047]

[0048] wherein m represents the number of nodes of the hidden layer neurons, n represents the number of nodes of the input layer, l represents the number of nodes of the output layer, and a is a constant between 1 and 10.

[0049] The number of nodes of the hidden layer neurons is the key to the success of network design. If the number of nodes of the hidden layer is too small, the learning ability of the network from the training samples is poor, and the learned rules are not enough to reflect the rules of the entire sample set, so the prediction error of the test sample is increased. If the number of nodes of the hidden layer is too large, the training time is greatly prolonged, and the network may also learn the non-rules in the sample (such as noise), and the problem of "overfitting" occurs, which reduces the generalization ability of the network.

[0050] II. Selecting a transfer function:

[0051] The selection of the parameters of the BP neural network includes the transfer function, the learning algorithm, the initial weight and threshold, and the iteration number, the learning rate, the training target error and the like. Among them, the transfer function is set according to the network requirements and the relationship between the input and the output. Common transfer functions include the logarithmic S-type transfer function (logsid), the hyperbolic tangent S-type transfer function (tangsig), and the linear transfer function, etc. The most commonly used S (sigmoid) type function is selected in the present application, and the function expression is:

[0052]

[0053] The S-type function has the characteristics of nonlinearity and everywhere derivability, and has a good gain control on the signal: when the value of |x| is small, f(x) has a large gain; when the value of |x| is large, f(x) has a small gain, which can prevent the network from entering the saturation state to a certain extent, so the function has been widely used.

[0054] Taking x as the input vector, x=(x0, x1, …, x n－1 ), x i represents the input of the i-th neuron of the input layer, wherein i=(0, 1, …, n-1), b j represents the output of the j-th neuron of the hidden layer, wherein j=(0, 1, …, l-1), and y kwhere v ij represents the weight from the i-th neuron node in the input layer to the j-th node in the hidden layer, w jk represents the weight from the j-th node in the hidden layer to the k-th node in the output layer, then:

[0055] The output of each neuron node in the hidden layer is:

[0056] The output of each neuron node in the output layer is:

[0057] where θ j , respectively represent the threshold value of the hidden layer node and the output layer node; f1 and f2 respectively represent the transfer function of the hidden layer and the output layer.

[0058] Let the expected output result of the output layer node be O k , then the total error E is calculated as follows:

[0059]

[0060] The purpose of error back propagation is to continuously reduce the total error by adjusting the weight, so the weight should be adjusted in the direction of the negative gradient of the error, that is, the change of the weight should satisfy the following formula:

[0061]

[0062] In the above formulas (d) and (e), η is a constant, called learning rate, usually taking 0<η<1.

[0063] Combining formulas (a)~(e), the weight adjustment formulas of the hidden layer and the output layer are respectively:

[0064]

[0065]

[0066] Similarly, the threshold value adjustment formulas of the hidden layer and the output layer are respectively:

[0067]

[0068]

[0069] After calculating the new weight and threshold value according to the weight and threshold value adjustment formulas, enter a new round of forward propagation process.

[0070] The derivation process is carried out for a single training sample based on the BP neural network with a single hidden layer, and the basic idea is also applicable to the BP neural network with multiple hidden layers, and when calculating the adjustment amount of the weight and the threshold value, the output layer is recursively propagated to the first hidden layer through the intermediate hidden layers, and the error calculation mode of multiple training samples is the cumulative sum of the single sample error.

[0071] Thus, by continuously adjusting the weights between nodes and node thresholds through continuous forward and backward propagation, the desired network model is ultimately obtained, i.e., the yield and tensile strength of the low-carbon steel cold-rolled thin plate.

[0072] The application develops a method for online measurement of the yield elongation of a cold-rolled thin strip, which obtains multiple electromagnetic signals in real time by applying comprehensive electromagnetic detection to the running strip steel, and extends the electromagnetic signals, considers the correction of the interval affecting the electromagnetic parameters, and considers the influence of the thickness of the strip steel.

[0073] The application can be used for online detection of the bake hardening value of the online running strip steel, and obtains high-density data values of the full length of the strip steel, and the sample qualified rate is more than 90% within the relative error accuracy range of 10%.

[0074] The application is applied to the online detection system of the mechanical properties of the cold-rolled strip steel, and realizes real-time online detection of the tensile strength and yield strength of the cold-rolled strip steel, realizes continuous detection, classification and recording of the production quality of the steel plate, and plays a very positive role in improving the production efficiency, product quality and product competitiveness.

[0075] The application has the advantages of convenient use, accurate measurement, high representativeness and strong adaptability. DETAILED DESCRIPTION

[0076] Figure 1 is a schematic view of the front view direction of the application,

[0077] Figure 2 is a schematic view of the top view direction of the application,

[0078] Figure 3 is a schematic view of the electromagnetic detection unit in the application,

[0079] Figure 4 is a topological structure diagram of the BP neural network,

[0080] Figure 5 is a flowchart of the BP neural network learning algorithm,

[0081] Figure 6 is a function image of the S (sigmoid) type function,

[0082] Figure 7 is a measurement result of a yield strength Rp value of a coil of strip steel measured using the present application,

[0083] Figure 8 is a measurement result of a tensile strength Rm value of a coil of strip steel measured using the present application. DETAILED DESCRIPTION

[0084] The present application is further illustrated by the following specific examples.

[0085] Example 1

[0086] A yield and tensile strength on-line measuring device for a low-carbon steel cold-rolled sheet, comprising a plurality of rollers 11, a strip steel 12, a base 2, a plurality of vertical rails 31, a plurality of horizontal rails 32, a plurality of slide rods 4, a plurality of lifting drive cylinders 5, an electromagnetic detection unit 6, a distance sensor 7 and a controller 8, as shown in Figures 1 to 3 the specific structure is as follows:

[0087] The plurality of rollers 11 are arranged parallel to each other and on the same horizontal plane, the strip steel 12 is arranged on the plurality of rollers 11, the plurality of rollers 11 rotate in the same direction to push the strip steel 12 to move, the base 2 is arranged between two adjacent rollers 11, the base 2 is vertically provided with four vertical rails 31, the top ends of the four vertical rails 31 form two opposite sides of a rectangle parallel to the rollers 11, the two ends of each of the plurality of slide rods 4 are movably embedded in one vertical rail 31, the three of the roller 11, the vertical rail 31 and the slide rod 4 are perpendicular to each other in pairs, the two ends of the horizontal rail 32 are fixed to the middle part of one slide rod 4, the cylinder body of the lifting drive cylinder 5 is fixed to the base 2, and the moving end of the piston rod of the lifting drive cylinder 5 is connected to the middle part of the horizontal rail 32.

[0088] The electromagnetic detection unit 6 comprises a shell 61, a servo drive motor 62, a roller 63, a tangent magnetic field harmonic analysis module, a Barkhausen noise detection module, an incremental magnetic permeability detection module, a multi-frequency eddy current electromagnetic detection module and an electromagnetic ultrasonic nondestructive detection module, the shell 61 is provided with the tangent magnetic field harmonic analysis module, the Barkhausen noise detection module, the incremental magnetic permeability detection module, the multi-frequency eddy current electromagnetic detection module and the electromagnetic ultrasonic nondestructive detection module, the servo drive motor 62 is fixed to the bottom of the shell 61, the roller 63 is rotatably arranged at the bottom of the shell 61, the output shaft of the servo drive motor 62 is connected to the roller 63, and the electromagnetic detection unit 6 is movably arranged on the horizontal rail 32 through the roller 63.

[0089] The distance sensor 7 is fixed to the top of the shell 61 of the electromagnetic detection unit 6, and the distance sensor 7 is arranged directly below the strip steel 12.

[0090] The lifting driving cylinder 5, the servo driving motor 62, the tangent magnetic field harmonic analysis module, the Barkhausen noise detection module, the incremental permeability detection module and the distance sensor 7 are connected to the controller 8 through signal lines;

[0091] The top of the vertical rail 31 is provided with a limiting block 311, which is arranged at the limit position of the upward movement of the slide rod 4;

[0092] In the embodiment, the controller 8 is selected from a microcomputer, a single-chip microcomputer or a programmable controller;

[0093] When the embodiment is used, the following steps are sequentially implemented:

[0094] ①Determination of parameters:

[0095] The tangent magnetic field harmonic analysis module of the electromagnetic detection unit 6 detects and analyzes eleven parameters EM1-EM11 of the excitation magnetic field, and the parameter code, parameter name, parameter unit and parameter meaning of the eleven parameters are as follows:

[0096] Parameter code Parameter name Parameter unit Parameter meaning EM1 A3 A / cm Amplitude of the third harmonic EM2 A5 A / cm Amplitude of the fifth harmonic EM3 A7 A / cm Amplitude of the seventh harmonic EM4 P3 Rad Phase of the third harmonic EM5 P5 Rad Phase of the fifth harmonic EM6 P7 Rad Phase of the seventh harmonic EM7 UHS A / cm Sum of the amplitudes of the third, fifth, seventh and ninth harmonics EM8 K ％ Deformation coefficient EM9 Hco A / cm Coercive field strength EM10 Hro A / cm Harmonic amplitude at the zero point of the hysteresis loop EM11 Vmag V Steady voltage of the electromagnetic coil ;

[0097] The Barkhausen noise detection module of the electromagnetic detection unit 6 detects and analyzes seven parameters EM12-EM18 of the excitation magnetic field, and the parameter code, parameter name, parameter unit and parameter meaning of the seven parameters are as follows:

[0098] Parameter code Parameter name Parameter unit Parameter meaning EM12 MMAX V Maximum amplitude of the excitation field EM13 MMEAN V Average value of the amplitude of the excitation field over one field period EM14 MR V Remanence point amplitude EM15 HCM A / cm Coercive field strength for M = MMAX EM16 DH25M A / cm Width of the Barkhausen curve for M = 25% MMAX EM17 DH50M A / cm Width of the Barkhausen curve for M = 50% MMAX M is the width of the Barkhausen curve at 50% MMAX EM18 DH75M A / cm M is the width of the Barkhausen curve at 75% MMAX ;

[0099] The incremental permeability detection module of the electromagnetic detection unit 6 detects and analyzes seven parameters EM19-EM25 of the excitation magnetic field, and the parameter code, parameter name, parameter unit and parameter meaning of the seven parameters are as follows:

[0100]

[0101]

[0102] The multi-frequency eddy current electromagnetic detection module of the electromagnetic detection unit 6 detects and analyzes sixteen parameters EM26-EM41 of the excitation magnetic field, and the parameter code, parameter name, parameter unit and parameter meaning of the sixteen parameters are as follows:

[0103]

[0104] The electromagnetic ultrasonic nondestructive testing module of the electromagnetic detection unit 6 detects and analyzes three parameters EM42-EM44 of the excitation magnetic field, and the parameter code, parameter name, parameter unit and parameter meaning of the three parameters are as follows:

[0105] Parameter Code Parameter Name Parameter Unit Parameter Meaning EM42 UEMAT_MAX V Peak value of the butterfly diagram EM43 UEMAT_MIN V Valley value of the butterfly diagram EM44 URatio V Ratio of two peak points ;

[0106] Each detection analysis obtains a raw detection electromagnetic signal EM i corresponding to an extended electromagnetic signal NM i ,

[0107] ② Online prediction:

[0108] The online prediction of the yield elongation of the strip steel is realized based on the BP neural network in the artificial neural network, and the specific algorithm steps are as follows:

[0109] i. Determine the BP neural network structure and network parameter setting: Before network training, the designed network structure needs to be constructed, and appropriate network parameters need to be set to ensure that the trained network can meet the expected requirements;

[0110] ii. Initialize the weight and threshold: The BP network whose weight and threshold are not set by the specified algorithm randomly assigns a set of small non-zero values to it;

[0111] iii. Process the training sample data and send it to the network for training. Before sending the network for training, the training sample data of different orders of magnitude needs to be normalized, and then the data is sent to the network for training;

[0112] iv. Forward propagation stage: Calculate the output of each layer in turn, and get the final output Ym in the output layer;

[0113] v. Compare the actual output with the expected output. If the actual output Ym is consistent with the expected output Om, the training is ended; if the actual output Ym is not consistent with the expected output Om, the back propagation stage is started;

[0114] vi. Back propagation stage: find the total error E between the actual output Ym and the expected output Om, and distribute the total error E to each layer and each node, and adjust the weight and threshold of each layer and each node;

[0115] vii. The updated network is trained next time: the network with adjusted weight and threshold is used as a new model for retraining, if the training result meets the expected requirement, the training is ended, if it does not meet the requirement, the error back propagation stage is entered again, and the cycle is repeated, when the iteration times reach the preset times or the performance function value is less than the preset error precision, the training process is stopped.

[0116] The BP (Back Propagation) neural network algorithm is a multi-layer feedforward network that learns and trains according to the error back propagation algorithm, and its network topology structure is as follows: Figure 4As shown, the network structure is composed of input layer, hidden layer and output layer, in which the input layer and the output layer each has only one, while the hidden layer can have multiple, and each layer is composed of a number of neurons. The nodes of the neurons in the layer are not connected to each other, while the nodes of the neurons between layers are connected in a full connection manner, i.e. any neuron node of the input layer is connected to all nodes of the hidden layer, and any neuron node of the hidden layer is connected to all nodes of the output layer.

[0117] The flow chart of the BP neural network learning algorithm is shown in Fig. 1. Figure 5 As shown, the execution process can be summarized as two main parts: forward propagation of signals and backward propagation of errors. For each training sample, in the forward propagation, the input vector is transmitted from the input layer to the output layer layer by layer; in the backward propagation, the error is transmitted from the hidden layer to the input layer layer by layer in a certain form. The two processes are repeated until the set termination condition is met.

[0118] The specific steps are as follows:

[0119] I. Determine the network structure:

[0120] The input layer and the output layer are fixed to have only one, and the number of neuron nodes of the input layer and the output layer is determined according to the actual problem to be solved and the data representation method, so the key of the network structure design lies in the design of the hidden layer structure, including the selection of the number of layers and the number of nodes. For the design of the number of hidden layers, usually one hidden layer is first considered, and the number of hidden layers is increased only when the one hidden layer cannot meet the related performance requirements. Experimental research shows that when two hidden layers are used, if the number of neuron nodes of the first hidden layer is more than that of the second hidden layer, the performance of the BP network will be improved to a certain extent. However, for some practical problems, the number of neuron nodes required by the double hidden layer may be less than that of the single hidden layer. The more the number of hidden layers, the longer the training learning time of the BP network, and the local minimum error will also increase, i.e. the probability of the network falling into a local optimal solution will increase, so only when the number of hidden layer nodes cannot significantly improve the network performance, the number of hidden layers is considered to be increased.

[0121] The number of neuron nodes of the hidden layer is determined by referring to the following several empirical estimation formulas:

[0122]

[0123] m = log2n — (k),

[0124]

[0125] wherein m represents the number of neuron nodes of the hidden layer, n represents the number of nodes of the input layer, I represents the number of nodes of the output layer, and a is a constant between 1 and 10.

[0126] The selection of the number of nodes of the hidden layer neurons is the key to the success of network design. If the number of nodes of the hidden layer is too small, the learning ability of the network from the training samples is poor, and the learned rules are not enough to reflect the rules of the entire sample set, so the prediction error of the test sample will increase. If too many nodes are set, the training time will be greatly prolonged, and the network may also learn the non-rules in the sample (such as noise), and the problem of "overfitting" occurs, which reduces the generalization ability of the network.

[0127] II. Selection of transfer function:

[0128] The selection of the parameters of the BP neural network includes the transfer function, learning algorithm, initial weight and threshold, and the number of iterations, learning rate, training target error and other parameters. Among them, the transfer function is set according to the network requirements and the relationship between the input and output. Common transfer functions include log-sigmoid transfer function (logsig), hyperbolic tangent sigmoid transfer function (tangsig), and linear transfer function. The most commonly used S (sigmoid) type function is selected, and the function expression is:

[0129]

[0130] The image of the S-type function is shown in Figure 6 .

[0131] The S-type function has the characteristics of nonlinearity and everywhere derivability, and has a good gain control on the signal: when the value of |x| is small, f(x) has a large gain; when the value of |x| is large, f(x) has a small gain, which can prevent the network from entering the saturation state to a certain extent, so the function has been widely used.

[0132] Figure 4 In the formula, x is the input vector, x=(x0, x1, …, x n－1 ), x i represents the input of the i-th neuron of the input layer, where i=(0, 1, …, n-1), b j is the output of the hidden layer neuron node, where j=(0, 1, …, l-1), y k is the output of the output layer node, k=(0, 1, …, m-1), n, l and m represent the number of neuron nodes of the input layer, the hidden layer and the output layer respectively, v ij represents the weight of the i-th neuron node of the input layer to the j-th node of the hidden layer, w jk represents the weight of the j-th node of the hidden layer to the k-th node of the output layer, then:

[0133] The output of each neuron node of the hidden layer is:

[0134] The output of each neuron node of the output layer is:

[0135] where θ j , respectively represent the threshold value of the hidden layer node and the output layer node; f1 and f2 respectively represent the transfer function of the hidden layer and the output layer.

[0136] Let the expected output result of the output layer node be O k , then the total error E is calculated as follows:

[0137]

[0138] The purpose of error back propagation is to continuously reduce the total error by adjusting the weight, so the weight should be adjusted in the negative gradient direction of the error, that is, the change of the weight should satisfy the following formula:

[0139]

[0140] In the above formulas (d) and (e), η is a constant, called learning rate, usually 0<η<1;

[0141] Combining formulas (a) to (e), the weight adjustment formulas of the hidden layer and the output layer are respectively:

[0142]

[0143]

[0144] Similarly, the threshold value adjustment formulas of the hidden layer and the output layer are respectively:

[0145]

[0146]

[0147] After calculating the new weight and threshold value according to the weight and threshold value adjustment formulas, enter a new round of forward propagation process.

[0148] The above derivation process is for a single training sample based on BP neural network with a single hidden layer, and the basic idea is also applicable to BP neural network with multiple hidden layers. When calculating the adjustment amount of the weight and the threshold value, the first hidden layer is recursively propagated from the output layer through the intermediate hidden layers, and the error calculation method for multiple training samples is the cumulative sum of the error of each single sample.

[0149] In this way, by continuously adjusting the weight between nodes and the node threshold value through continuous forward and backward propagation, the network model required by the expected requirement, i.e. the yield and tensile strength of low carbon steel cold rolled sheet, is finally obtained.

[0150] The present embodiment is applied in a production line, and main parameters of the artificial neural network are as follows:

[0151] Model: BP multilayer feedforward neural network

[0152] Model structure: three-layer structure: input layer, hidden layer, and output layer, wherein:

[0153] Model input layer: 43 parameters, including: 41 electromagnetic parameters EM1-EM41; Gap, the distance between the probe and the lower surface of the strip, satisfying the condition 4≤Gap≤6; and strip thickness THK

[0154] Hidden layer: 12 layers, which are optimized according to data test during model training

[0155] Output layer: one parameter: yield strength or tensile strength

[0156] The transfer function between model layers is a Sigmoid function.

[0157] First, the model training stage, with low carbon steel produced in the past period of time in the production line as the training sample, the specific input parameters and output parameters of the model are shown in Table 1:

[0158] Table 1:

[0159]

[0160]

[0161] The developed model is used for model verification, which is shown in Table 2. The model calculation values of five low carbon steel samples are compared with the test values of the offline samples, and the qualified rate is 100% according to the requirement of relative error of 10%. It shows that the model can meet the technical requirements of online detection of yield strength (Rp) and tensile strength (Rm).

[0162] Table 2:

[0163]

[0164]

[0165]

[0166] The model input parameters (part) of the yield strength (Rp) training stage are shown in Table 3:

[0167] Table 3:

[0168] Test 1 Test 2 Test 3 Test 4 Test 5 Model calculated value (MPa) 161 155 148 163 152 Sample tensile value (MPa) 155 146 149 158 160 Error (MPa) 6 9 -1 5 -8 Relative error (%) 3.726 5.803 0.675 3.067 5.2635

[0169] The model input parameters (partial) for the tensile strength (Rm) training phase are shown in Table 4:

[0170] Table 4:

[0171] Test 1 Test 2 Test 3 Test 4 Test 5 Model calculated value (MPa) 309 294 295 298 305 Sample tensile value (MPa) 301 302 297 299 303 Error (MPa) 8 -8 -2 -1 2 Relative error (%) 2.588 -2.721 0.677 0.335 0.6557 .

[0172] This embodiment is used for online measurement of Rp and Rm of 500 coils of SEDDQ strip steel on a certain production line. One sample was taken from each end, and the strength values ​​were obtained using an offline tensile testing method. A total of 1000 sets of samples were obtained for each method. The test results for the yield strength (Rp) and tensile strength (Rm) along the entire length are as follows: Figure 7 and Figure 8 As shown, the results obtained are compared with the corresponding position values ​​measured online, and the confidence levels are Rp: 97.8% and Rm: 97.1%. Compared with existing technologies that can only test by shearing the sample, the data volume and real-time performance are greatly improved.

Claims

1. A method for using a low carbon steel cold-rolled sheet yield and tensile strength on-line measuring device, the low carbon steel cold-rolled sheet yield and tensile strength on-line measuring device comprising a roller (11), a strip steel (12), a base (2), a vertical rail (31), a flat rail (32), a sliding rod (4), a lifting driving cylinder (5), an electromagnetic detection unit (6), a distance sensor (7) and a controller (8), each roller (11) is arranged in parallel and on the same horizontal plane, the strip steel (12) is arranged on each roller (11), and each roller (11) drives the strip steel (12) to move when rotating in the same direction; the base (2) is arranged between two adjacent rollers (11), four vertical rails (31) are vertically arranged on the base (2), the top ends of the four vertical rails (31) form two opposite sides of a rectangle parallel to the rollers (11), the two ends of each sliding rod (4) are movably embedded in one vertical rail (31), the roller (11), the vertical rail (31) and the sliding rod (4) are perpendicular to each other in pairs, the two ends of the flat rail (32) are fixed on the middle part of one sliding rod (4), the cylinder body of the lifting driving cylinder (5) is fixed on the base (2), and the moving end of the piston rod of the lifting driving cylinder (5) is connected to the middle part of the flat rail (32); the electromagnetic detection unit (6) comprises a shell (61), a servo driving motor (62), a roller (63), a tangent magnetic field harmonic analysis module, a Barkhausen noise detection module, an incremental magnetic permeability detection module, a multi-frequency eddy current electromagnetic detection module and an electromagnetic ultrasonic nondestructive detection module, the tangent magnetic field harmonic analysis module, the Barkhausen noise detection module, the incremental magnetic permeability detection module, the multi-frequency eddy current electromagnetic detection module and the electromagnetic ultrasonic nondestructive detection module are arranged in the shell (61), the servo driving motor (62) is fixed on the bottom of the shell (61), the roller (63) is rotatably arranged on the bottom of the shell (61), the output shaft of the servo driving motor (62) is connected to the roller (63), and the electromagnetic detection unit (6) is movably arranged on the flat rail (32) through the roller (63); the distance sensor (7) is fixed on the top of the shell (61) of the electromagnetic detection unit (6) and arranged directly below the strip steel (12); the lifting driving cylinder (5), the servo driving motor (62), the tangent magnetic field harmonic analysis module, the Barkhausen noise detection module, the incremental magnetic permeability detection module and the distance sensor (7) are connected to the controller (8) through signal lines; the top of the vertical rail (31) is provided with a limiting block (311), and the limiting block (311) is arranged at the limit position of the upward movement of the sliding rod (4); the controller (8) is selected from a microcomputer, a single-chip microcomputer or a programmable controller; and the following steps are sequentially implemented: ① determining parameters: the tangent magnetic field harmonic analysis module of the electromagnetic detection unit (6) detects and analyzes eleven parameters EM1-EM11 of the excitation magnetic field, and the parameter code, parameter name, parameter unit and parameter meaning of the eleven parameters are as follows: ​ ​ ​ ​ ​ ​ ​ ​ The application is characterized in that ​ ​ ​ ； The Barkhausen noise detection module of the electromagnetic detection unit (6) detects and analyzes seven parameters of EM12-EM18 of the excitation magnetic field, and the parameter code, parameter name, parameter unit and parameter meaning of the seven parameters are as follows: ； The incremental permeability detection module of the electromagnetic detection unit (6) detects and analyzes seven parameters of EM19-EM25 of the excitation magnetic field, and the parameter code, parameter name, parameter unit and parameter meaning of the seven parameters are as follows: ； The multi-frequency eddy current electromagnetic detection module of the electromagnetic detection unit (6) detects and analyzes sixteen parameters of EM26-EM41 of the excitation magnetic field, and the parameter code, parameter name, parameter unit and parameter meaning of the sixteen parameters are as follows: ； The electromagnetic ultrasonic nondestructive testing module of the electromagnetic detection unit (6) detects and analyzes three parameters of EM42-EM44 of the excitation magnetic field, and the parameter code, parameter name, parameter unit and parameter meaning of the three parameters are as follows: ； Each of the raw detection electromagnetic signals EM obtained by the detection analysis i corresponds to an extended electromagnetic signal NM i= , i = 1, 2,..., 44; ② Online prediction: The online prediction of the yield elongation of the strip steel is realized based on the BP neural network in the artificial neural network, and the specific algorithm steps are as follows: i. Determine the BP neural network structure and network parameter setting: before network training, the designed network structure needs to be constructed, and appropriate network parameters need to be set to ensure that the trained network can meet the expected requirements; ii. Initialize the weight and threshold: the BP network with unspecified algorithm to set the weight and threshold adopts a set of small non-zero values randomly; iii. Process the training sample data and send it to the network for training: before sending the network for training, the training sample data of different orders of magnitude needs to be normalized, and then the data is sent to the network for training; iv. Forward propagation stage: calculate the output of each layer in turn, and get the final output Ym in the output layer; v. Compare the actual output with the expected output: if the actual output Ym is consistent with the expected output Om, the training is ended; if the actual output Ym is not consistent with the expected output Om, the back propagation stage is started; vi. Back propagation stage: the total error E between the actual output Ym and the expected output Om is calculated, and the total error E is allocated to each layer and each node, and the weight and threshold of each layer and each node are adjusted; vii. The updated network is trained next time: the network with adjusted weight and threshold is used as a new model for further training, if the training result meets the expected requirements, the training is ended, if it does not meet the requirements, the error back propagation stage is entered again, and the cycle is repeated, when the iteration times reach the preset times or the performance function value is less than the preset error precision, the training process is stopped.

Images