Improved particle swarm algorithm-based plate shape identification method and device, and storage medium

Abstract

The invention discloses a plate shape recognition method and device based on an improved particle swarm algorithm and a storage medium, and relates to the technical field of plate shape recognition, and the method comprises the steps: collecting the plate shape measurement value of each measurement segment in the width direction of a target plate strip; inputting the strip shape measurement value into a trained feature extraction network for feature extraction so as to output strip shape feature parameters; performing linear combination on the strip shape characteristic parameters to generate strip shape linear characteristics; and inputting the plate shape linear features into a trained I PSO-E lman neural network to carry out plate shape identification so as to output a plate shape defect prediction value of the target plate strip. According to the invention, the strip shape quality can be effectively monitored, and the objectivity, accuracy and real-time performance of strip shape quality anomaly detection are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of plate shape recognition, and in particular to a plate shape recognition method, device and storage medium based on an improved particle swarm algorithm. Background Art

[0002] Plate and strip materials are important products in my country's modern steel industry and are important raw materials for national industrial products. With the rapid development of my country's economy and the continuous improvement of science and technology, the domestic industry's demand for plate and strip products is growing, and the requirements for the quality and variety of plate and strip products are also increasing. Therefore, the control accuracy and speed in the rolling process of plate and strip production have become key issues that need to be urgently solved in today's plate and strip product production control. In the rolling production of plate and strip materials, plate shape is one of the key technical indicators to measure the quality of plate and strip materials. With the increasing requirements for plate shape accuracy in the production of modern industrial products, plate shape quality has become an important issue that needs to be solved urgently.

[0003] At present, many scholars at home and abroad have conducted in-depth research on the control theory and technology of plate shape, including taking different methods to study the mathematical model of plate shape control. Some scholars use artificial intelligence optimization algorithms such as neural networks to identify plate shape. Among them, the pattern recognition link of plate shape defects is a crucial step in the plate shape control process and plays a guiding role in the control of plate shape. Accordingly, by analyzing the residual stress signal of the plate shape and comparing and classifying it with the set defect type pattern, the plate shape defect type pattern of the detected plate strip section can be determined, and the precise control amount can be provided to the plate shape control execution device, so as to obtain excellent plate shape quality.

[0004] Chinese invention patent CN106862284A discloses a method for cold-rolled strip signal pattern recognition, which applies an improved quantum neural network with a multi-layer excitation function optimized by a genetic algorithm to the plate shape pattern recognition technology. However, in the early neural network plate shape pattern recognition model, only the basic type of plate shape to be identified is given qualitatively, rather than quantitatively outputting the actual deviation value of each plate shape pattern from the target plate shape, that is, the plate shape membership problem, and it is impossible to provide quantitative control data to the control mechanism of the rolling mill; therefore, the plate shape pattern recognition method based on the back propagation (BP) network is commonly used, but its convergence speed is slow and it is easy to fall into the dilemma of local optimal solution. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a plate shape recognition method, device and storage medium based on an improved particle swarm algorithm, which can effectively monitor the plate shape quality of plate and strip materials and improve the objectivity, accuracy and real-time performance of abnormal detection of plate shape quality of plate and strip materials.

[0006] In order to solve the above technical problems, the present invention provides a plate shape recognition method based on an improved particle swarm algorithm, comprising: collecting plate shape measurement values ​​of each measurement section in the width direction of the target plate and strip; inputting the plate shape measurement values ​​into a trained feature extraction network for feature extraction to output plate shape feature parameters; linearly combining the plate shape feature parameters to generate plate shape linear features; inputting the plate shape linear features into a trained IPSO-Elman neural network for plate shape recognition to output plate shape defect prediction values ​​of the target plate and strip.

[0007] As an improvement of the above solution, the plate shape linear characteristic is expressed as:

[0008] Δσ(x)=a1p1(x)+a2p2(x)+a3p3(x)+a4p4(x)

[0009] Among them, a1, a2, a3, a4 are plate shape characteristic parameters, p1(x), p2(x), p3(x), p4(x) are plate shape defect types, and Δσ(x) is the plate shape linear characteristic.

[0010] As an improvement of the above scheme, the training steps of the IPSO-Elman neural network include: constructing a particle swarm; constructing an Elman neural network; initializing the particle swarm, mapping the particles in the particle swarm with the weights and thresholds of the Elman neural network, and setting the initial values ​​of the velocity vector, position vector, global optimal value and individual optimal value of the particle; using the output error of the Elman neural network as the fitness function of IPSO, calculating the fitness of the particle, and updating the individual optimal value and the global optimal value according to the fitness; updating the velocity vector and the position vector according to the corrected inertia weight and the learning factor; when the output error is less than the expected error or the current number of iterations reaches the preset maximum number of iterations, extracting the optimal weight and optimal threshold corresponding to the current global optimal solution, otherwise reinitializing the particle swarm; reinitializing the Elman neural network to use the optimal weight as the initial weight of the Elman neural network, and the optimal threshold as the initial threshold of the Elman neural network; training the Elman neural network after the reinitialization until the output error reaches the expected error.

[0011] As an improvement of the above scheme, the Elman neural network includes an input layer, a hidden layer, a receiving layer and an output layer. The nonlinear spatial state expression of the Elman neural network is:

[0012] y(k)=g(w 1 x(k)

[0013] x(k)=f(w 3 xc (k)+w 2 (u(k-1)))

[0014] x c (k) = x(k-1)

[0015] Among them, y(k) is the output vector at time k, x(k) is the output vector of the hidden layer at time k, and x c (k) is the output vector of the receiving layer at time k, f(*) is the activation function of the hidden layer, g(*) is the activation function of the output layer, u(k-1) is the input vector of the input layer at time k-1, w 1 is the connection weight from the hidden layer to the output layer, w 2 is the connection weight from the input layer to the hidden layer, w 3 is the connection weight from the network receiving layer to the hidden layer.

[0016] As an improvement of the above scheme, the output error of the Elman neural network is:

[0017]

[0018] Where E is the output error of the Elman neural network, y d (k) is the expected output value of the Elman neural network, and y(k) is the actual output value of the Elman neural network.

[0019] As an improvement of the above scheme, the step of updating the individual optimal value and the global optimal value according to the fitness includes: comparing the fitness with the individual optimal value, if the fitness is better than the individual optimal value, then using the fitness as the new individual optimal value; comparing the fitness with the global optimal value, if the fitness is better than the global optimal value, then using the fitness as the new global optimal value.

[0020] As an improvement of the above solution, the steps of modifying the inertia weight and the learning factor include:

[0021] Apply a linear correction to the inertia weight:

[0022] ω=(ω1-ω2)[(t max -t) / t max ]+ω2

[0023] The trigonometric function method is introduced to perform nonlinear correction on the learning factor:

[0024] c1=2tan(πt / 4t max )+0.5

[0025] c2=2.5-c1

[0026] Among them, ω is the corrected inertia weight, c1 and c2 are the corrected learning factors, ω1 is the maximum value of the inertia weight, ω2 is the minimum value of the inertia weight, t max is the maximum number of iterations, and t is the current number of iterations.

[0027] As an improvement of the above solution, the steps of calculating the velocity vector and the position vector include:

[0028] Calculate the velocity vector:

[0029] v id (t+1)=ωv id (t)+c1r1(p id -x id (t))+c2r2(p gd -x id (t))

[0030] Calculate the position vector:

[0031] x id (t+1)=x id (t)+v id (t+1)

[0032] Among them, ω is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers, v id (t) is the velocity vector of the ith particle in d dimension at the tth iteration, x id (t) is the position vector of the ith particle in d dimension at the tth iteration, p id is the optimal position information of the individual searched by the population particles, p gd is the optimal position information of all particles in the population.

[0033] Correspondingly, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein the processor implements the steps of the above-mentioned plate shape recognition method based on the improved particle swarm algorithm when executing the computer program.

[0034] Correspondingly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-mentioned plate shape recognition method based on the improved particle swarm algorithm.

[0035] The implementation of the present invention has the following beneficial effects:

[0036] The present invention collects the width data set of the plate and strip material to form the width characteristic parameters characterizing the plate and strip material, and linearly combines the plate shape characteristic parameters to generate the plate shape linear characteristics, and then uses the plate shape linear characteristics as input and the plate shape quality label of the plate and strip material as output, so as to effectively monitor the plate shape quality of the plate and strip material and improve the objectivity, accuracy and real-time performance of the plate shape quality abnormality detection of the plate and strip material;

[0037] Furthermore, the present invention improves the particle swarm algorithm and combines it with the Elman neural network to form a unique IPSO-Elman neural network, and applies the IPSO-Elman neural network to the plate shape pattern recognition technology; effectively utilizes the advantages of the Elman neural network, such as strong real-time performance, fast learning speed, memory with the particle swarm optimization algorithm, few adjustment parameters, and fast convergence speed, to provide a new way for efficient and intelligent identification of plate shape defects and a reliable basis for plate shape control; thereby significantly improving the training efficiency of the network, effectively solving the problems encountered in traditional plate shape recognition methods, such as unsatisfactory accuracy and real-time performance, complex network structure and long training time, poor stability and robustness, etc., to provide a reliable basis for plate shape control, and has strong practicality and industrial value. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flow chart of an embodiment of a plate shape recognition method based on an improved particle swarm algorithm of the present invention;

[0039] Figure 2 It is a flow chart of an embodiment of the training method of the IPSO-Elman neural network in the present invention;

[0040] Figure 3 It is a schematic diagram of the structure of the IPSO-Elman neural network in the present invention. DETAILED DESCRIPTION

[0041] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings.

[0042] See also Figure 1 , Figure 1 The flowchart of an embodiment of the plate shape recognition method based on the improved particle swarm algorithm of the present invention is shown, which includes:

[0043] S101, collecting the plate shape measurement values ​​of each measuring section in the width direction of the target plate and strip;

[0044] In practical applications, the shape meter can be used to measure the shape values ​​of each measuring section in the width direction of the target plate and strip (e.g., cold-rolled strip steel) online;

[0045] Correspondingly, the number of measurement segments is m, and the flatness measurement value of the i-th measurement segment is Fi For example, the number of measurement segments can be set to 20, but it is not limited to this and can be adjusted according to actual conditions.

[0046] S102, inputting the plate shape measurement value into the trained feature extraction network for feature extraction to output plate shape feature parameters;

[0047] The plate shape measurement value is input into an n-layer feature extraction network as the feature extraction layer. The feature extraction network is trained to automatically extract features to eliminate artificial traces.

[0048] Since the board shape mainly includes eight defect modes, namely left wave, right wave, middle wave, double wave, right three-point wave, left three-point wave, four-point wave and side middle wave, the output of the feature extraction network can be a i ,i=1,2,3...m, where m is the number of outputs. For example, m=8, then the calculation formula for each layer of the feature extraction network is:

[0049]

[0050] in,

[0051] I is the Ith layer of the feature extraction network;

[0052] W I is the network weight of layer I;

[0053] x I is the input of layer I;

[0054] b I Bias the neurons in the network;

[0055] f(*) is the activation function, which is used to introduce nonlinear factors so that the network can learn complex features;

[0056] represents the output node value of the neural network.

[0057] In the actual production process, in order to identify the defect type of plate shape, a series of plate shape stress signals (i.e., plate shape measurement values) measured online by the plate shape meter are usually converted into plate shape characteristic parameters in a specific range to determine the defect type of the plate shape.

[0058] S103, performing linear combination on the plate shape characteristic parameters to generate a plate shape linear characteristic;

[0059] The plate shape in actual production is a linear combination of the above basic plate shape characteristic parameters:

[0060] Δσ(x)=a1p1(x)+a2p2(x)+a3p3(x)+a4p4(x)

[0061] p1(x)=x

[0062] p2(x)=3x 2 / 2-1 / 2

[0063] p3(x)=(35x 4 -30x 2 +3) / 8

[0064] in,

[0065] a1, a2, a3, and a4 are plate shape characteristic parameters;

[0066] p1(x) is a plate shape defect (left wave / right wave);

[0067] p2(x) is the secondary shape defect (middle wave / double wave);

[0068] p3(x) is the fourth plate shape defect (quarter wave / side middle wave);

[0069] p4(x) is a high-order plate shape defect, which is mainly eliminated by segmented cooling of the rolls and is not considered in the present invention;

[0070] Δσ(x) is the linear characteristic of the plate shape.

[0071] S104, inputting the plate shape linear features into the trained IPSO-Elman neural network for plate shape recognition to output a plate shape defect prediction value of the target plate and strip.

[0072] The plate shape defect prediction values ​​are u1, u2, u3, and u4, which are used to represent the degree of membership of the plate shape, where:

[0073] If u1>0, it means the board has a left wave, if u1<0, it means the board has a right wave;

[0074] If u2>0, it means the board has a middle wave, if u2<0, it means the board has a double-sided wave;

[0075] If u3>0, it means the board has three waves on the left, and if u3<0, it means the board has three waves on the right;

[0076] If u4>0, it means that the board shape has quarter waves. If u4<0, it means that the board shape has side and middle waves.

[0077] Therefore, the present invention collects the width data set of the plate and strip to form the width characteristic parameters that characterize the plate and strip, and linearly combines the plate shape characteristic parameters to generate plate shape linear features, and then uses the plate shape linear features as input and the plate shape quality label of the plate and strip as output, so as to effectively monitor the plate shape quality of the plate and strip and improve the objectivity, accuracy and real-time performance of abnormal detection of plate shape quality of the plate and strip.

[0078] See also Figure 2 , Figure 2 The training method of the IPSO-Elman neural network in the present invention is shown, which includes:

[0079] S201, constructing a particle swarm;

[0080] Set the particle swarm population size m and the maximum inertia weight ω max , minimum inertia weight ω min , the maximum number of iterations of the algorithm t max , the particle velocity variation range [V min ,V max ] and position change range [X min ,X max ].

[0081] S202, constructing Elman neural network;

[0082] like Figure 3 As shown in the figure, the Elman neural network includes an input layer, a hidden layer, a receiving layer and an output layer. When constructing the Elman neural network, it is necessary to set the number of neuron nodes in the input layer, the hidden layer, the receiving layer and the output layer.

[0083] At the same time, when constructing the sample data set, it is necessary to divide the sample data set into two parts: training sample data and test sample data in proportion, so as to train and test the Elman neural network respectively.

[0084] Correspondingly, the nonlinear spatial state expression of the Elman neural network is:

[0085] y(k)=g(w 1 x(k)

[0086] x(k)=f(w 3 x c (k)+w 2 (u(k-1)))

[0087] x c (k) = x(k-1)

[0088] in,

[0089] y is the m-dimensional output node vector, y(k) is the output vector at time k;

[0090] x is the n-dimensional hidden layer node vector, x(k) is the output vector of the hidden layer at time k;

[0091] x c is the n-dimensional feedback state vector, x c(k) is the output vector of the receiving layer at time k;

[0092] u is the r-dimensional input vector, u(k-1) is the input vector of the input layer at time k-1;

[0093] f(*) is the activation function of the hidden layer, and f(*) is generally a nonlinear function, that is: f(x) = 1 / (1+e -x );

[0094] g(*) is the activation function of the output layer, and g(*) is generally a linear function, that is: y(k) = w 1 x(k);

[0095] w 1 is the connection weight from the hidden layer to the output layer;

[0096] w 2 is the connection weight from the input layer to the hidden layer;

[0097] w 3 is the connection weight from the network receiving layer to the hidden layer.

[0098] S203, initializing the particle swarm, mapping the particles in the particle swarm with the weights and thresholds of the Elman neural network, and setting the initial values ​​of the particle's velocity vector, position vector, global optimal value, and individual optimal value;

[0099] Initialize the entire particle swarm, map each particle to the initial weight and initial threshold of the Elman neural network one by one, and then randomly generate the initial velocity vector V of each particle i and the initial position vector X i , and get the initial global optimal value p G and the initial individual optimal value p i .

[0100] S204, using the output error of the Elman neural network as the fitness function of IPSO, calculating the fitness of the particle, and updating the individual optimal value and the global optimal value according to the fitness;

[0101] The output error of the Elman neural network is:

[0102]

[0103] in,

[0104] E is the output error of the Elman neural network;

[0105] y d (k) is the expected output value of the Elman neural network;

[0106] y(k) is the actual output value of the Elman neural network.

[0107] Specifically, the steps of updating the individual optimal value and the global optimal value according to the fitness include:

[0108] (1) Compare the fitness with the individual optimal value. If the fitness is better than the individual optimal value, the fitness is used as the new individual optimal value.

[0109] That is, the current fitness is compared with the current individual optimal value p i Compare, if the current fitness is better than the current individual optimal value p i , then the individual optimal value p i Update with the current fitness value.

[0110] (2) Compare the fitness with the global optimal value. If the fitness is better than the global optimal value, the fitness is used as the new global optimal value.

[0111] That is, the current fitness is compared with the current global optimal value p G Compare, if the current fitness is better than the current global optimal value p G , then the global optimal value p G Update with the current fitness value.

[0112] S205, updating the velocity vector and the position vector according to the corrected inertia weight and the learning factor;

[0113] In different stages of the optimization process, in order to effectively balance the global search ability and local search ability of particles, the modified inertia weight and learning factor can be substituted into the velocity vector and position vector calculation formula of the particle swarm, and then the velocity vector and position vector of the particle are updated and calculated. Specifically:

[0114] The mathematical formula for linear inertia weight reduction is:

[0115] ω=(ω1-ω2)[(t max -t) / t max ]+ω2

[0116] in,

[0117] ω is the corrected inertia weight;

[0118] ω1 is the maximum value of the inertia weight;

[0119] ω2 is the minimum value of inertia weight;

[0120] t max is the maximum number of iterations;

[0121] t is the current iteration number.

[0122] The trigonometric function method is introduced to perform nonlinear correction on the values ​​of the two learning factors in the algorithm iteration:

[0123] c1=2tan(πt / 4t max )+0.5

[0124]

[0125] in,

[0126] c1 and c2 are the corrected learning factors. The value range of c1 and c2 is [0.5, 2.5], and the value is generally 2;

[0127] t max is the maximum number of iterations;

[0128] t is the current iteration number.

[0129] Accordingly, the velocity vector can be updated according to the following formula:

[0130] v id (t+1)=ωv id (t)+c1r1(p id -x id (t))+c2r2(p gd -x id (t))

[0131] in,

[0132] i=1,2,…,m;

[0133] d=1,2,…,D;

[0134] ω is the corrected inertia weight;

[0135] c1 and c2 are the corrected learning factors;

[0136] r1 and r2 are random numbers, and the value range of r1 and r2 is [0,1];

[0137] v id (t) is the velocity vector of the ith particle in d dimension at the tth iteration;

[0138] x id (t) is the position vector of the ith particle in d dimension at the tth iteration;

[0139] p id The optimal position information of individuals searched for population particles;

[0140] p gd is the optimal position information of all particles in the population.

[0141] At the same time, the position vector can be updated according to the following formula:

[0142] x id (t+1)=x id (t)+v id (t+1)

[0143] in,

[0144] v id (t+1) is the velocity vector of the ith particle in d dimension at the t+1th iteration;

[0145] x id (t+1) is the position vector of the ith particle in d dimension at the next iteration (i.e., the t+1th iteration);

[0146] x id (t) is the position vector of the ith particle in d dimension at the tth iteration.

[0147] Specifically, the revised IPSO parameter values ​​are shown in Table 1 below:

[0148] Table 1

[0149] parameter index <![CDATA[Maximum number of iterations t max > 50 <![CDATA[Maximum inertia weight ω max > 0.9 <![CDATA[Minimum inertia weight ω min > 0.4 <![CDATA[Learning factor c1]]> [0.5,2.5] <![CDATA[Learning factor c2]]> [0,2] Particle speed range [-2.2] Particle position variation range [-5,5] <![CDATA[Random numbers r1 and r2]]> [0,1]

[0150] S206, when the output error is less than the expected error or the current number of iterations reaches the preset maximum number of iterations, extract the optimal weight and optimal threshold corresponding to the current global optimal solution, otherwise reinitialize the particle swarm;

[0151] It should be noted that when the output error is less than the expected error (i.e. the global optimal value p G The fitness value reaches the requirement), or the number of iterations reaches the maximum number of iterations t max When , the optimization process is terminated and the global optimal value p is output G (i.e., the optimal weight and the optimal threshold), otherwise return to step S203.

[0152] S207, re-initializing the Elman neural network to use the optimal weight as the initial weight of the Elman neural network, and use the optimal threshold as the initial threshold of the Elman neural network;

[0153] The optimal weight is used as the initial weight of the Elman neural network, and the optimal threshold is used as the initial threshold of the Elman neural network, that is, the secondary initialization of the Elman neural network is completed.

[0154] S208, training the Elman neural network after the secondary initialization until the output error reaches the expected error.

[0155] The Elman neural network with secondary initialization is trained. When the output error reaches the expected error, the training is completed and the IPSO-Elman neural network can be constructed.

[0156] Therefore, the present invention utilizes the advantages of the Elman neural network, such as strong real-time performance and fast learning speed, and the particle swarm optimization algorithm, such as memory, few adjustment parameters and fast convergence speed, to provide a new way for efficient and intelligent identification of plate shape defects; at the same time, the present invention improves the particle swarm algorithm and combines it with the Elman neural network to form the IPSO-Elman neural network, thereby effectively improving the accuracy of plate shape recognition.

[0157] In order to more fully compare the recognition performance of the existing Elman neural network with the IPSO-Elman neural network in the present invention, the following four evaluation indicators, namely, mean absolute error, mean square error, root mean square error, and mean absolute percentage error, are used to evaluate the recognition accuracy of the plate shape. Specifically:

[0158] Mean absolute error E mae for:

[0159]

[0160] Mean square error E mse for:

[0161]

[0162] Root mean square error E rmse for:

[0163]

[0164] Mean absolute percentage error E mape for:

[0165]

[0166] Among them, y j is the true value of the plate shape, y j ′ is the predicted value of IPSO-Elman neural network, and N is the number of test samples.

[0167] Accordingly, the various error operation results obtained by using the Elman neural network and IPSO-Elman neural network prediction methods are shown in Table 2 below:

[0168] Table 2

[0169] parameter index Elman mean absolute error 0.13453 IPSO-Elman mean absolute error 0.09107 Elman mean square error 0.03607 IPSO-Elman mean square error 0.01021 Elman RMSE 0.18991 IPSO-Elman RMSE 0.10104 Elman Percent Error 23.9869％ PSO-Elman Percent Error 3.2868％

[0170] From the comparison of the error parameters in Table 2 above, it can be seen that the IPSO-Elman neural network adopted in the present invention has significantly improved prediction accuracy and performs well in flatness pattern recognition.

[0171] Correspondingly, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein the processor implements the steps of the above-mentioned plate shape recognition method based on the improved particle swarm algorithm when executing the computer program.

[0172] At the same time, the present invention also discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, implements the steps of the above-mentioned plate shape recognition method based on the improved particle swarm algorithm.

[0173] In summary, the plate shape recognition method based on the improved particle swarm algorithm of the present invention has the following beneficial effects:

[0174] (1) The present invention collects a width data set of plate and strip materials to form width characteristic parameters that characterize the plate and strip materials, and linearly combines the plate shape characteristic parameters to generate plate shape linear features. The plate shape linear features are then used as input and the plate shape quality labels of the plate and strip materials are used as output. A plate shape quality anomaly recognition model based on the particle swarm algorithm and the optimized Elman neural network is constructed using a training sample set. The plate shape quality of the plate and strip materials can be effectively monitored, and the objectivity, accuracy and real-time performance of plate shape quality anomaly detection of the plate and strip materials can be improved.

[0175] (2) The present invention utilizes the advantages of Elman neural network, such as strong real-time performance, fast learning speed, memory, few adjustment parameters and fast convergence speed, to provide a new way for efficient and intelligent identification of plate shape defects and a reliable basis for plate shape control.

[0176] (3) The present invention improves the particle swarm algorithm to optimize the Elman neural network to form a unique IPSO-Elman neural network, and applies the IPSO-Elman neural network to the plate shape pattern recognition technology, which significantly improves the training efficiency of the network and effectively solves the problems encountered in the traditional plate shape recognition method, such as the accuracy and real-time performance are not ideal, the network structure is complex and the training time is long, and the stability and robustness are poor. It provides a reliable basis for plate shape control and has strong practicality and industrial value.

[0177] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A plate shape recognition method based on improved particle swarm algorithm, characterized in that: include: Collect the plate shape measurement values ​​of each measuring section in the width direction of the target plate and strip; Inputting the plate shape measurement value into a trained feature extraction network for feature extraction to output plate shape feature parameters; Linearly combining the plate shape characteristic parameters to generate a plate shape linear characteristic; The plate shape linear features are input into the trained IPSO-Elman neural network for plate shape recognition to output the plate shape defect prediction value of the target plate and strip.

2. The plate shape recognition method based on improved particle swarm algorithm according to claim 1, characterized in that: The plate shape linear characteristic is expressed as: Δσ(x)=a1p1(x)+a2p2(x)+a3p3(x)+a4p4(x) Among them, a1, a2, a3, a4 are plate shape characteristic parameters, p1(x), p2(x), p3(x), p4(x) are plate shape defect types, and Δσ(x) is the plate shape linear characteristic.

3. The plate shape recognition method based on improved particle swarm algorithm according to claim 1, characterized in that: The training steps of the IPSO-Elman neural network include: Construct particle swarm; Construct Elman neural network; Initializing the particle swarm, mapping the particles in the particle swarm with the weights and thresholds of the Elman neural network, and setting the initial values ​​of the velocity vector, position vector, global optimal value and individual optimal value of the particle; The output error of the Elman neural network is used as the fitness function of IPSO, the fitness of the particle is calculated, and the individual optimal value and the global optimal value are updated according to the fitness; Update the velocity vector and the position vector according to the corrected inertia weight and the learning factor; When the output error is less than the expected error or the current iteration number reaches the preset maximum iteration number, the optimal weight and the optimal threshold corresponding to the current global optimal solution are extracted, otherwise the particle swarm is reinitialized; Performing secondary initialization on the Elman neural network, so as to use the optimal weight as the initial weight of the Elman neural network, and use the optimal threshold as the initial threshold of the Elman neural network; The Elman neural network after secondary initialization is trained until the output error reaches the expected error.

4. The plate shape recognition method based on improved particle swarm algorithm as claimed in claim 3, characterized in that: The Elman neural network includes an input layer, a hidden layer, a receiving layer and an output layer. The nonlinear spatial state expression of the Elman neural network is: x c (k)=x(k-1) Among them, y(k) is the output vector at time k, x(k) is the output vector of the hidden layer at time k, and x c (k) is the output vector of the receiving layer at time k, f(*) is the activation function of the hidden layer, g(*) is the activation function of the output layer, u(k-1) is the input vector of the input layer at time k-1, w 1 is the connection weight from the hidden layer to the output layer, w 2 is the connection weight from the input layer to the hidden layer, w 3 is the connection weight from the network receiving layer to the hidden layer.

5. The plate shape recognition method based on improved particle swarm algorithm as claimed in claim 3, characterized in that: The output error of the Elman neural network is: Where E is the output error of the Elman neural network, y d (k) is the expected output value of the Elman neural network, and y(k) is the actual output value of the Elman neural network.

6. The plate shape recognition method based on improved particle swarm algorithm as claimed in claim 3, characterized in that: The step of updating the individual optimal value and the global optimal value according to the fitness comprises: Comparing the fitness with the individual optimal value, if the fitness is better than the individual optimal value, taking the fitness as a new individual optimal value; The fitness is compared with the global optimal value, and if the fitness is better than the global optimal value, the fitness is used as a new global optimal value.

7. The plate shape recognition method based on improved particle swarm algorithm according to claim 1, characterized in that: The steps of modifying the inertia weight and the learning factor include: Apply a linear correction to the inertia weight: ω=(ω1-ω2)[(t max -t) / t max ]+ω2 The trigonometric function method is introduced to perform nonlinear correction on the learning factor: c1=2tan(πt / 4t max )+0.5 c2=2.5-c1 Among them, ω is the corrected inertia weight, c1 and c2 are the corrected learning factors, ω1 is the maximum value of the inertia weight, ω2 is the minimum value of the inertia weight, t max is the maximum number of iterations, and t is the current number of iterations.

8. The plate shape recognition method based on improved particle swarm algorithm according to claim 7, characterized in that: The calculation steps of the velocity vector and the position vector include: Calculate the velocity vector: v id (t+1)=ωv id (t)+c1r1(p id -x id (t))+c2r2(p gd -x id (t)) Calculate the position vector: x id (t+1)=x id (t)+v id (t+1) Among them, ω is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers, v id (t) is the velocity vector of the ith particle in d dimension at the tth iteration, x id (t) is the position vector of the ith particle in d dimension at the tth iteration, p id is the optimal position information of the individual searched by the population particles, p gd is the optimal position information of all particles in the population.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images