CN116976192A - JS-BP model-based die forging defect accurate repair process parameter decision method - Google Patents
JS-BP model-based die forging defect accurate repair process parameter decision method Download PDFInfo
- Publication number
- CN116976192A CN116976192A CN202310743956.XA CN202310743956A CN116976192A CN 116976192 A CN116976192 A CN 116976192A CN 202310743956 A CN202310743956 A CN 202310743956A CN 116976192 A CN116976192 A CN 116976192A
- Authority
- CN
- China
- Prior art keywords
- jellyfish
- neural network
- layer
- weight
- output
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 230000007547 defect Effects 0.000 title claims abstract description 66
- 238000005242 forging Methods 0.000 title claims abstract description 55
- 238000000034 method Methods 0.000 title claims abstract description 50
- 230000008439 repair process Effects 0.000 title claims abstract description 45
- 238000013528 artificial neural network Methods 0.000 claims abstract description 83
- 230000008569 process Effects 0.000 claims abstract description 30
- 238000005457 optimization Methods 0.000 claims abstract description 24
- 238000005498 polishing Methods 0.000 claims abstract description 19
- 239000013598 vector Substances 0.000 claims abstract description 10
- 241000242583 Scyphozoa Species 0.000 claims description 138
- 210000002569 neuron Anatomy 0.000 claims description 80
- 230000006870 function Effects 0.000 claims description 68
- 230000033001 locomotion Effects 0.000 claims description 59
- 238000012986 modification Methods 0.000 claims description 24
- 230000004048 modification Effects 0.000 claims description 24
- 238000003062 neural network model Methods 0.000 claims description 22
- 238000013507 mapping Methods 0.000 claims description 19
- 238000012549 training Methods 0.000 claims description 12
- 238000012546 transfer Methods 0.000 claims description 12
- 238000010845 search algorithm Methods 0.000 claims description 8
- 230000007246 mechanism Effects 0.000 claims description 7
- 230000008859 change Effects 0.000 claims description 4
- 230000000739 chaotic effect Effects 0.000 claims description 4
- 238000012937 correction Methods 0.000 claims description 3
- 238000011478 gradient descent method Methods 0.000 claims description 3
- 230000001105 regulatory effect Effects 0.000 claims description 3
- 230000004913 activation Effects 0.000 claims description 2
- 230000001276 controlling effect Effects 0.000 claims description 2
- 238000005192 partition Methods 0.000 claims description 2
- 230000000694 effects Effects 0.000 abstract description 5
- 238000004886 process control Methods 0.000 abstract 2
- 238000004519 manufacturing process Methods 0.000 description 11
- 238000005516 engineering process Methods 0.000 description 4
- 238000011161 development Methods 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000009776 industrial production Methods 0.000 description 1
- 238000002715 modification method Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/17—Mechanical parametric or variational design
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Geometry (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Computer Hardware Design (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Data Mining & Analysis (AREA)
- Biophysics (AREA)
- Mathematical Optimization (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Analysis (AREA)
- Computational Mathematics (AREA)
- Medical Informatics (AREA)
- Feedback Control In General (AREA)
Abstract
The invention discloses a method for accurately repairing die forging defects based on a JS-BP model, which comprises the steps of optimizing initial weight and threshold of a BP neural network by utilizing a JS algorithm, predicting a repairing feed speed, an operating angle and a force control size by utilizing three process control parameters of crack width, crack depth and polishing disc rotating speed in a die forging defect repairing process by utilizing the BP neural network, and finally, optimizing a process control parameter input vector by the repairing feed speed, the operating angle and the force control size in the die forging defect repairing process by utilizing the JS algorithm. The method can make a decision on the die forging defect repair process parameters, and can provide reliable data support for the field of automatic die forging defect repair. The BP neural network after JS optimization can not only improve the convergence rate of the BP neural network and prevent the condition of local optimum, but also has accurate decision-making effect and self-learning capability.
Description
Technical Field
The invention relates to the field of die forging modification, in particular to a JS-BP-based die forging defect accurate modification process parameter decision method.
Background
Forging is widely used as a traditional production process in mechanical, metallurgical, shipbuilding, aerospace, weapon and other industrial production departments, and plays a very important role in national economy. With the high-speed development of economy, net shape and near net shape manufacturing with the purposes of improving quality and reducing cost has become a new development trend of the current manufacturing industry, and various advanced manufacturing technologies are also emerging continuously. Under the new environment, the traditional production process-forging production with long history has the advantages of good mechanical property, high production efficiency, material saving and the like, not only maintains the irreplaceable position, but also receives importance of people, and has quite wide application prospect. However, during the manufacture of the die forging, defects such as cracks, inclusions and the like often occur due to material and process differences, and the defects seriously affect the quality and performance of the die forging. For these drawbacks, appropriate corrective measures are required to eliminate or mitigate the effects thereof, thereby ensuring the quality of the die forging. However, most of the conventional optimization decisions of the modification process parameters are based on experience and technical level of the process personnel, the optimal process parameters cannot be accurately judged, and the modification efficiency is low, so that the intelligent decision technology for researching the modification process parameters is particularly critical.
The traditional modification method is often based on experience or trial and error, has low efficiency and is easy to generate errors, and in the die forging modification process, the part information and the process parameters are in complex nonlinear relations, so that the quantitative description is difficult. The design of the technological parameters of the current aviation forging is mainly based on human experience and a trial and error method. Because the technological parameters of the die forging are formulated, repeated tests are required before production, and time waste and cost rise are caused. In recent years, with the development of artificial intelligence technology, defect modification by using artificial intelligence technology has become a new research direction. Among them, neural network-based methods have been widely used for die forging defect recognition and modification.
However, the model modification parameters are not sufficiently accurate and reliable only by adopting the neural network model, and the optimal model modification parameter combination is determined by combining other optimization algorithms, so that the process parameter decision of accurate model modification of the die forging defects is finally realized, and the quality and performance of the die forging are ensured.
The artificial jellyfish searching algorithm (JS) is an optimization algorithm with global searching and high efficiency, and can effectively search an optimal solution, so that the algorithm can be used for optimizing the modification process parameter decision problem of the neural network algorithm.
Disclosure of Invention
Aiming at the defects and the demands, the invention provides a precise repair process parameter decision method aiming at the defects of a die forging, which comprises the following steps: the method is based on a BP neural network model (JS-BP) optimized by an artificial jellyfish search algorithm to carry out a repair process parameter decision on the die forging defects. Establishing a BP neural network model with input quantity of crack width, crack depth and polishing disc rotating speed, optimizing the weight and threshold of the model by using an artificial jellyfish searching algorithm, and training the optimized model by using a good data set to finally obtain a stable and accurate repair process parameter decision model; inputting the values of crack width, crack depth and polishing disc rotation speed randomly in a certain range into a trained BP neural network model, outputting the values of repair feeding speed, operation angle and force control size by using the BP neural network model, initializing jellyfish population by using chaotic logic mapping, continuously updating jellyfish positions based on a time control mechanism and boundary conditions and matching with the current optimal positions until the positions are matched with the optimal positions, and continuously adjusting the input vector of the BP neural network by repeating the jellyfish optimal position matching process until the error fitness of the BP neural network is converged to the minimum, thereby determining the optimal technological parameters.
The invention discloses a die forging defect accurate repair process parameter decision method based on a JS-BP algorithm, which comprises the following steps of:
step one: three technological parameters with the largest influence factors on the modification result in the modification process are used as the input quantity H of the BP neural network k ＝(x 1 ,x 2 ,x 3), wherein ,x1 ,x 2 ,x 3 The crack width, the crack depth and the polishing disc rotating speed are respectively, k is any one of n groups of experimental data sets, k=1, 2, … and n, and the repair feeding speed, the running angle and the force control size corresponding to the parameters are taken as output quantity T k ＝(y 1 ,y 2 ,y 3 ) And establishing a mapping relation among the rotation speed of the polishing disc, the defect width, the defect depth, the repair feeding speed, the operation angle and the force control size. And the reliable data in the actual modification is recorded into a database by utilizing the mapping relation to form a reliable data set for BP neural network training:
step two: and (3) establishing a BP neural network with a three-layer topological structure through the die forging defect repair process parameter mapping relation established in the step one, wherein the BP neural network comprises an input layer neuron, a hidden neuron and an output layer neuron. The number of the neurons of the input layer is 3, the number of the neurons of the hidden layer is 7, and the number of the neurons of the output layer is 3. Assuming that net is entered on the ith neuron node of the hidden layer i The functional relationship is:
wherein M is the number of neurons of the j th layer, j is the node of the neurons of the input layer, and w ij For inputting the weight value between the jth neuron node of the layer and the ith neuron node of the hidden layer, x j For input of the jth neuron, θ i Is the bias of the ith neuron.
The output of the ith neuron node of the hidden layer has the following functional relationship:
where f is the activation function of the neuron.
The logsig function is selected as a transfer function of a neural network hidden layer, and the function relation is as follows:
wherein e is a natural constant, s j Refers to the input of the j hidden layer node, and the function relation is as follows:
where j=1, 2 … 7. Wherein w is ij For inputting the weight value between the jth neuron node of the layer and the ith neuron node of the hidden layer, theta j Is the threshold value of the j node of the hidden layer, x i For crack width, crack depth and sanding disc rotational speed.
The pureline type function is selected as a transfer function of the neural network output layer, and the function relation is as follows:
y＝x
input net of the kth neuron node of the output layer k Is a function of:
wherein ak Output o of the kth neuron node of the output layer is the threshold of the kth neuron node of the output layer k The functional relation is as follows:
step three: and (3) using the neural network in the second step, taking the input quantity in the first step as an input variable of the BP neural network, and taking the output quantity as an output variable of the BP neural network. And establishing a BP neural network prediction model according to the mapping relation of the rotation speed, defect width, defect depth, repair feeding speed, operation angle and force control size of the polishing disc.
Step four: and initializing a weight and a threshold between the connection of the input layer and the output layer of the BP neural network model by using the input variable data and the output variable data of the experimental data set in the step one. Optimizing the weight and threshold value of the BP neural network initialization according to the JS artificial jellyfish search algorithm, and finally giving the optimal initial threshold value weight to the BP neural network model for prediction:
step 4-1: initializing the weight w of the neural network model constructed in the steps by using chaotic logic mapping 0 and θ0 The threshold value is used as an initial jellyfish population of the JS algorithm, the maximum iteration number is set to be 50 generations, and the population number is 30;
step 4-2: the movement type of jellyfish in the ocean can be changed continuously along with the time, and the movement mode of the time-controlled population can be changed from passive movement to active movement, and the movement of jellyfish to ocean current is controlled. The initial weight w of the BP neural network is regulated by a time control mechanism 0 And an initial threshold value theta 0 The time control function is:
where t is the designated time of the iteration number, max_iter is the maximum iteration number, and is an initial parameter of the algorithm.
The time control mechanism includes a time control function C (t) and a constant C (0). The time control function is a random value that fluctuates from 0 to 1 over time. When C (t) exceeds C (0), jellyfish follows the ocean current. Jellyfish moves in the population when it is less than C (0). Since C (0) is not exact, the average value is set to C (0) =0.5. If C (t) is more than or equal to 0.5, the jellyfish moves along with ocean currents, otherwise, the jellyfish moves in the population; when moving within the jellyfish population, if rand (0, 1) > (1-C (t)), it appears as passive movement, otherwise it appears as active movement;
step 4-3: simulating the real state of the ocean, setting a boundary condition, and returning to the opposite boundary when the jellyfish moves out of the boundary, namely, when the boundary condition is exceeded, changing the weight w and the threshold value theta by a boundary condition function, wherein the boundary condition function is as follows:
wherein ,Xi,d Is the position of the ith jellyfish in the d dimension; x'. i,d To check the updated position after boundary constraint: u (U) b,d 、L b,d The upper and lower bounds of the search space are d dimensions.
Step 4-4: according to the different motion states of jellyfish mentioned in the step 4-2, firstly simulating the motion relation of jellyfish flowing down along ocean, wherein the direction of ocean current is determined by the average value of all vectors from each jellyfish in ocean to the jellyfish at the optimal position at present, and the average position of jellyfish:
wherein ,the weight w is the optimal value of the threshold value theta, which is the optimal vector of the jellyfish at the optimal position, and n is the number of jellyfishes (weight w threshold value theta).
A single jellyfish (weight w, threshold value θ) and the whole population jellyfish (w 1 ，w 2 …，w n And theta 1 ，θ 2 …θ n ) The average position difference of (2) is:
Δu＝eu
where e is a factor controlling attractive force. e is expressed mathematically as follows:
e＝β*rand
where β=3, which is the partition coefficient.
Trend of change in position for each jellyfish:
wherein ,X* Is the optimal position of jellyfish population.
The update formula for the position (weight w, threshold θ) of each jellyfish is:
X i (t+1)＝X i (t)+rand*(X * -β*rand*u)
step 4-5: the motion relation of jellyfish motions in jellyfish groups is simulated, and in the groups, the motions of jellyfish are divided into two types: active motion and passive motion. The passive motion is that jellyfish moves around the position of the jellyfish, and the formula of the position update (weight w, threshold value theta) is as follows:
X i (t+1)＝X i (t)+γ*rand(0，1)*(U b -L b )
wherein ,Ub 、L b The upper and lower boundaries of the exploration space are respectively; gamma is a motion coefficient, and γ > 0, and γ=1 is set in relation to the motion length around the jellyfish position.
The active motion of jellyfish is simulated, jellyfish (j) other than the jellyfish of interest is randomly selected, and the vector from the jellyfish (i) of interest to the selected jellyfish (j) is used to determine the direction of motion. When the quantity of food at the selected jellyfish (j) position exceeds the quantity of food at the jellyfish (i) position of interest, the latter is moved to the former; if the selected jellyfish (j) has a lower amount of food available than the jellyfish (i) of interest, then the jellyfish (i) leaves the jellyfish (j) directly. Thus, each jellyfish moves in a better direction, looking for food in the population:
in addition:
wherein f is the objective function of position X;is the movement direction of jellyfish.
The latest jellyfish position (weight w and threshold θ) is:
step 4-6, establishing a JS artificial jellyfish optimization algorithm model of step 4-1-4.5, wherein the threshold value theta 'of the BP neural network is to be calculated' m And weight w' z Optimizing to finally obtain the optimal initial weight and threshold value of the BP neural network; using the die forging defect repair process parameter data set in the first step to input an input quantity H k Substituting the optimal initial threshold value theta 'into a BP neural network prediction model' m And weight w' z Predicting by BP neural network model to obtain predicted output O of die forging defect repair process parameters k . Output O k And the modeling reality value Z k There is an error between them;
step five: and constructing a BP neural network error back propagation model, subtracting actual values from values of all neurons on an output layer to obtain error values, and continuously adjusting weights and thresholds through back propagation. And correcting the direction and the magnitude of the partial derivative of the neuron weight and the threshold value of each layer through the obtained error until the error reaches a preset condition.
Step 5-1, constructing a desired output quantity T of the die forging defect repair process parameter k (crack width, crack depth, and polished disk rotational speed) and predicted output O of BP neural network prediction model k The sum of squares error function of the difference (modified feed speed, angle of operation, force control magnitude), the average error of this function being
Wherein k is the number of output nodes. The process of minimizing the error function is the back propagation of the error of the BP neural network;
in step 5-2, since there are multiple data in the training data set, there are p training data sets in the sample, and the total error function is the objective function:
step 5-3, calculating the error function in step 5-1 by using a gradient descent method to ensure that the optimal initial threshold value theta 'in step 4-6' m And weight w' z Further optimization is obtained, along with the increase of iterative optimization times, when the error value of the error function meets the error precision requirement, the iterative optimization process is ended:
according to the modified connection parameter value, the weight Deltaw between the system output layer and the hidden layer is obtained in a sorting way ki The connection parameters have the following functional relationship:
threshold Δa on output layer neuron node k The function relationship is as follows:
Δw of weights between hidden layer and input layer ij The change amount has the following functional relationship:
the threshold variation of each neuron node of the hidden layer has the following functional relationship:
wherein ：
the weight and the threshold value after the correction corresponding to the input layer and the output layer can be obtained by arranging the formulas:
in the formula ：wij The weight between the jth neuron node of the input layer and the ith neuron node of the hidden layer is obtained; x and y are respectively the input quantity of the node and the output quantity of the node; w (w) kj The weight between the ith neuron node of the hidden layer and the kth neuron node of the output layer is obtained; x is x j An input for a j-th neuron node of the input layer; i is the threshold of the hidden layer i-th neuron node; f (x) is the transfer function of the hidden layer; a, a k A threshold value for the kth neuron node of the output layer; g (x) is the output layer transfer function; o (o) k Is the output of the kth neuron node of the output layer.
Step 5-4, the optimal weight w obtained by repeated iterative optimization in the step 5-3 is obtained 0 And a threshold value theta 0 Weight w as final BP neural network 0 And a threshold value theta 0 Substituting the input variable of the die forging defect repair process parameter data set into the BP neural network prediction model again to be used as the input quantity of the input of the BP neural network input layer, and outputting the numerical values of the repair feed speed, the operation angle and the force control size to be the optimal predicted quantity through the mapping relation between the polishing disc rotating speed, the defect width, the defect depth, the repair feed speed, the operation angle and the force control size, which are established in the first step.
Compared with the prior art, the invention has the beneficial effects that:
aiming at the nonlinear problem established by the die forging production process, a neural network is adopted, and the decision model is obtained by training production actual data. The mapping between the part information and the technological parameters is established, a JS (artificial jellyfish search) optimized BP neural network is adopted to establish an aviation forging defect treatment technological parameter decision method, and the decision can be made on the die forging defect repair technological parameters (repair feeding speed, operation angle and force control size), so that reliable data support can be provided for the field of automatic die forging defect repair. The BP neural network after JS optimization can not only improve the convergence rate of the BP neural network and prevent the condition of local optimum, but also has accurate decision-making effect and self-learning capability.
Drawings
FIG. 1 is a flowchart of the optimization of BP neural network model by using an artificial jellyfish search algorithm in the invention;
FIG. 2 is a flowchart of the optimization of the artificial jellyfish search algorithm employed in the present invention;
FIG. 3 is a diagram showing the comparison between the decision-making errors of the process parameters of the JS-BP neural network and the actual processing parameters of the modified feed speed in the invention;
FIG. 4 is a chart showing the comparison of the decision-making error of the process parameters of the JS-BP neural network and the actual processing parameters of the polishing operation angle of the repairing equipment;
FIG. 5 is a chart showing the comparison of the decision-making error of the process parameters of the JS-BP neural network and the actual processing parameters of the force control size of the repairing equipment;
Detailed Description
Other advantages and effects of the present invention will become apparent to those skilled in the art from the following disclosure, which is to be read in light of the accompanying examples. The invention may be practiced or carried out in other embodiments that depart from the specific details, and the details of the present description may be modified or varied from the spirit and scope of the present invention.
The invention relates to a die forging defect repairing method, which is characterized in that an aviation blade die forging is used as a machined part, the die forging is polished for crack defects on blades, a force control polisher is used for polishing the cracks of the die forging until the width-depth ratio of the defect position meets the requirement, and experimental data sources of the method are repairing process parameter data sets formed by a large number of experiments and manual experience.
The artificial neural network learning process is a process of performing continuous iterative optimization on data, and the weights and the thresholds among neurons of an input layer, a hidden layer and an output layer are continuously corrected, so that the weights and the thresholds are finally converged within an error range set in advance.
The basic steps of the artificial jellyfish searching algorithm are as follows: initializing population, calculating fitness function, evaluating fitness, updating time parameters and jellyfish positions, and evaluating fitness again.
Because the traditional BP neural network model has low convergence rate and is easy to fall into the local optimum condition, the BP neural network with a three-layer structure is adopted, and the initial weight and the threshold value of the BP neural network are optimized by using an artificial jellyfish search algorithm. Establishing a JS-BP neural network optimization model, carrying out network training on a die forging defect repair process parameter data set by combining with the actual die forging defect repair situation, and finally obtaining the die forging defect repair process parameter JS-BP neural network model which can meet the expected prediction result through repeated iterative training.
The method is to establish a BP neural network model (JS-BP) optimized by an artificial jellyfish search algorithm for repairing the defects of the die forging. Setting the optimized input quantity of the JS-BP neural network model as crack width, crack depth and polishing disc rotating speed, and setting the output quantity as feeding speed, operating angle and force control size; the weight and the threshold of the model are optimized by using the artificial jellyfish searching algorithm, then the optimized model is trained by using a large number of data sets obtained by real experiments and artificial experiences, and finally the stable and accurate modification process parameter decision model is obtained.
The invention relates to a JS-BP algorithm-based die forging defect repair process parameter decision method which comprises the following steps of
Step one: using 300 reliable experimental sample data for modification, taking three technological parameters (crack width, crack depth and polishing disc rotating speed) with the largest influence factors on modification results in the modification process as input quantity of a BP neural network, taking modification feeding speed, operation angle and force control size corresponding to the parameters as output quantity, and establishing a mapping relation between polishing disc rotating speed, defect width, defect depth and modification feeding speed, operation angle and force control size in the 300 groups of data;
step two: and (3) establishing a BP neural network with a three-layer topological structure through the die forging defect repair process parameter mapping relation established in the step one, wherein the BP neural network comprises an input layer neuron, a hidden neuron and an output layer neuron. The number of the neurons of the input layer is 3, the number of the neurons of the hidden layer is 7, and the number of the neurons of the output layer is 3. Selecting logsig function as transfer function of hidden layer of neural network, its functional relationship is The pureline type function is selected as a transfer function of the neural network output layer, and the function relation is as follows: y=x.
Step three: and (3) using the neural network in the second step, taking the input quantity in the first step as an input variable of the BP neural network, and taking the output quantity as an output variable of the BP neural network. And establishing a BP neural network prediction model according to the mapping relation of the rotation speed, defect width, defect depth, repair feeding speed, operation angle and force control size of the polishing disc.
Step four: and initializing a weight and a threshold between the connection of the input layer and the output layer of the BP neural network model by using the input variable data and the output variable data of the experimental data set in the step one. Optimizing the weight and threshold value of the BP neural network initialization according to the JS artificial jellyfish search algorithm, and finally giving the optimal initial threshold value weight to the BP neural network model for prediction:
step 4-1: initializing the weight and the threshold of the neural network model constructed in the steps by using chaotic logic mapping as an initial jellyfish population of a JS algorithm, setting the maximum iteration number as 50 generations and setting the population number as 30;
step 4-2: as the movement type of jellyfish in the ocean can be changed continuously along with the time, the movement mode of the time-controlled population can be changed from passive movement to active movement, and the water is controlledMovement of the parent to the ocean currents. The initial weight w of the BP neural network is regulated by a time control mechanism 0 And an initial threshold value theta 0 The time control function is:
where t is the specified time of the iteration number and max_iter is the maximum iteration number.
The time control mechanism includes a time control function C (t) and a constant C (0), and an average value is set to C (0) =0.5. If C (t) is more than or equal to 0.5, the jellyfish moves along with ocean currents, otherwise, the jellyfish moves in the population; when moving within the jellyfish population, if rand (0, 1) > (1-C (t)), it appears as passive movement, otherwise it appears as active movement;
step 4-3: simulating the real state of the ocean, setting boundary conditions, and returning to the opposite boundary when jellyfish moves out of the boundary, wherein the boundary condition functions are as follows:
wherein ,Xi，d Is the position of the ith jellyfish in the d dimension; x'. i，d To check the updated position after boundary constraint: u (U) b，d 、L b，d The upper and lower bounds of the search space are d dimensions.
Step 4-4: according to different movement states of jellyfish mentioned in the step 4-2, firstly simulating the movement relation of jellyfish flowing down along with the ocean, and the average position of jellyfish:
wherein ,all vectors of jellyfish which are the best positions, namely weight w and threshold valueThe optimal value of θ, n is the jellyfish (weight w threshold θ) number.
The location update formula for each jellyfish is:
X i (t+1)＝X i (t)+rand*(X * -β*rand*u)
step 4-5: the motion relation of jellyfish motions in jellyfish groups is simulated, and in the groups, the motions of jellyfish are divided into two types: active motion and passive motion. The passive motion is that jellyfish moves around the position of the jellyfish, and the update formula of the position (weight w, threshold value theta) is as follows:
X i (t+1)＝X i (t)+γ*rand(0，1)*(U b -L b )
wherein ,Ub 、L b The upper and lower boundaries of the exploration space are respectively; gamma is a motion coefficient, and γ > 0, and γ=1 is set in relation to the motion length around the jellyfish position.
The active motion of jellyfish is simulated, jellyfish (j) other than the jellyfish of interest is randomly selected, and the vector from the jellyfish (i) of interest to the selected jellyfish (j) is used to determine the direction of motion. When the quantity of food at the selected jellyfish (j) position exceeds the quantity of food at the jellyfish (i) position of interest, the latter is moved to the former; if the selected jellyfish (j) has a lower amount of food available than the jellyfish (i) of interest, then the jellyfish (i) leaves the jellyfish (j) directly. Thus, each jellyfish moves in a better direction, looking for food in the population:
in addition:
wherein f is the objective function of position X；Is the movement direction of jellyfish.
The latest jellyfish position (weight w and threshold θ) is:
step 4-6, establishing a JS artificial jellyfish optimization algorithm model of step 4-1-4.5, wherein the threshold value theta 'of the BP neural network is to be calculated' m And weight w' z Optimizing to finally obtain the optimal initial weight and threshold value of the BP neural network; using the die forging defect repair process parameter data set in the first step to input an input quantity H k Substituting the optimal initial threshold value theta 'into a BP neural network prediction model' m And weight w' z Predicting by BP neural network model to obtain predicted output O of die forging defect repair process parameters k . Output O k And the modeling reality value Z k There is an error between them;
step five: and constructing a BP neural network error back propagation model, subtracting actual values from values of all neurons on an output layer to obtain error values, and continuously adjusting weights and thresholds through back propagation. And correcting the direction and the magnitude of the partial derivative of the neuron weight and the threshold value of each layer through the obtained error until the error reaches the preset condition.
Step 5-1, constructing a desired output quantity T of the die forging defect repair process parameter k (crack width, crack depth, and polished disk rotational speed) and predicted output O of BP neural network prediction model k The square sum error function of the difference (correction feed speed, running angle, force control size), because there are multiple data in the training data set, so there are p training data sets in the sample, the total error function, its objective function is:
wherein q is the number of output nodes. The process of minimizing the error function is the back propagation of the error of the BP neural network;
step 5-2, calculating the error function in step 5-1 by using a gradient descent method to ensure that the optimal initial threshold value theta 'in step 4-6' m And weight w' z Further optimization is obtained, along with the increase of iterative optimization times, when the error value of the error function meets the error precision requirement, the iterative optimization process is ended:
according to the modified connection parameter values, the corrected weight and threshold corresponding to the input layer and the output layer can be obtained:
in the formula ：wij The weight between the jth neuron node of the input layer and the ith neuron node of the hidden layer is obtained; x and y are respectively the input quantity of the node and the output quantity of the node; w (w) kj The weight between the ith neuron node of the hidden layer and the kth neuron node of the output layer is obtained; x is x j An input for a j-th neuron node of the input layer; i is the threshold of the hidden layer i-th neuron node; f (x) is the transfer function of the hidden layer; a, a k A threshold value for the kth neuron node of the output layer; g (x) is the output layer transfer function; o (o) k Is the output of the kth neuron node of the output layer.
Step six: and (3) taking the input variable of the die forging defect repair process parameter data set as the input quantity of the input layer of the JS-optimized BP neural network, wherein the numerical values of the repair feed speed, the operation angle and the force control size output by the JS-optimized BP neural network are the optimal predicted values through the mapping relation of the rotation speed, the defect width, the defect depth, the repair feed speed, the operation angle and the force control size of the polishing disc established in the first step.
In the accurate die forging defect repairing process, the method combines the strong nonlinear fitting capacity of the BP neural network with the high convergence speed and the strong optimizing capacity of the artificial jellyfish searching algorithm, the combined JS-BP optimizing model can well complement the defects of the respective algorithm, the BP neural network after JS optimization can not only improve the convergence speed of the BP neural network, prevent the situation of being in local optimum, and have accurate decision making effect and self-learning capacity.
Claims (3)
1. The method for accurately repairing the die forging defect by using the parameter decision based on the JS-BP model is characterized by comprising the following steps of:
step one: three technological parameters with the largest influence factors on the modification result in the modification process are used as the input quantity H of the BP neural network k ＝(x 1 ,x 2 ,x 3), wherein ,x1 ,x 2 ,x 3 The crack width, the crack depth and the polishing disc rotating speed are respectively, k is any one of n groups of experimental data sets, k=1, 2, … and n, and the repair feeding speed, the running angle and the force control size corresponding to the parameters are taken as output quantity T k ＝(y 1 ,y 2 ,y 3 ) Establishing a mapping relation between the rotation speed of the polishing disc, the defect width, the defect depth, the repair feeding speed, the operation angle and the force control size; and the reliable data in the actual modification is recorded into a database by utilizing the mapping relation to form a reliable data set for BP neural network training:
step two: through the die forging defect repair process parameter mapping relation established in the first step, a BP neural network with a three-layer topological structure is established, wherein the BP neural network comprises an input layer neuron, a hidden neuron and an output layer neuralMenstruation; the number of neurons of the input layer is 3, the number of neurons of the hidden layer is 7, and the number of neurons of the output layer is 3; assuming that net is entered on the ith neuron node of the hidden layer i The functional relationship is:
wherein M is the number of neurons of the j th layer, j is the node of the neurons of the input layer, and w ij For inputting the weight value between the jth neuron node of the layer and the ith neuron node of the hidden layer, x j For input of the jth neuron, θ i Bias for the ith neuron;
the output of the ith neuron node of the hidden layer has the following functional relationship:
wherein f is the activation function of the neuron;
the logsig function is selected as a transfer function of a neural network hidden layer, and the function relation is as follows:
wherein e is a natural constant, wherein s j Refers to the input of the j hidden layer node, and the function relation is as follows:
wherein j=1, 2 … 7; wherein w is ij For inputting the weight value between the jth neuron node of the layer and the ith neuron node of the hidden layer, theta j Is the threshold value of the j node of the hidden layer, x i The crack width, the crack depth and the polishing disc rotating speed are adopted;
the pureline type function is selected as a transfer function of the neural network output layer, and the function relation is as follows:
y＝x
input net of the kth neuron node of the output layer k Is a function of:
wherein ak Output o of the kth neuron node of the output layer is the threshold of the kth neuron node of the output layer k The functional relation is as follows:
step three: using the neural network in the second step, taking the input quantity in the first step as the input variable of the BP neural network, and taking the output quantity as the output variable of the BP neural network; establishing a BP neural network prediction model according to the mapping relation of the rotation speed, defect width, defect depth, repair feeding speed, operation angle and force control size of the polishing disc;
step four: initializing a weight and a threshold between the connection of the input layer and the output layer of the BP neural network model by using the input variable data and the output variable data of the experimental data set in the first step; optimizing the weight and the threshold value of the BP neural network initialization according to the JS artificial jellyfish search algorithm, and finally giving the optimal initial threshold value weight to the BP neural network model for prediction;
step five: constructing a BP neural network error back propagation model, subtracting actual values from values of all neurons on an output layer to obtain error values, and continuously adjusting weights and thresholds through back propagation; and correcting the direction and the magnitude of the partial derivative of the neuron weight and the threshold value of each layer through the obtained error until the error reaches a preset condition.
2. The method for accurately repairing die forging defects according to claim 1, wherein the fourth step comprises the following steps: initializing the weight and the threshold of the neural network model constructed in the steps by using chaotic logic mapping as an initial jellyfish population of a JS algorithm, setting the maximum iteration number as 50 generations and setting the population number as 30;
step 4-2: the movement type of jellyfish in the ocean can be changed continuously along with the time, so that the movement mode of the time-controlled population can be changed from passive movement to active movement, and the movement of jellyfish to ocean current is controlled; the initial weight w of the BP neural network is regulated by a time control mechanism 0 And an initial threshold value theta 0 The time control function is:
wherein t is the appointed time of iteration times, max_iter is the maximum iteration times and is an initial parameter of the algorithm;
the time control mechanism comprises a time control function C (t) and a constant C (0); the time control function is a random value that fluctuates from 0 to 1 over time; when C (t) exceeds C (0), jellyfish will follow the ocean current; jellyfish moves in the population when it is less than C (0); since C (0) is not exact, the average value is set to C (0) =0.5; if C (t) is more than or equal to 0.5, the jellyfish moves along with ocean currents, otherwise, the jellyfish moves in the population; when moving within the jellyfish population, if rand (0, 1) > (1-C (t)), it appears as passive movement, otherwise it appears as active movement;
step 4-3: simulating the real state of the ocean, setting a boundary condition, and returning to the opposite boundary when the jellyfish moves out of the boundary, namely, when the boundary condition is exceeded, changing the weight w and the threshold value theta by a boundary condition function, wherein the boundary condition function is as follows:
wherein ,Xi,d For the position of the ith jellyfish in the d dimensionPlacing; x'. i,d To check the updated position after boundary constraint: u (U) b,d 、L b,d Upper and lower bounds for the d-dimensional search space;
step 4-4: according to the different motion states of jellyfish mentioned in the step 4-2, firstly simulating the motion relation of jellyfish flowing down along ocean, wherein the direction of ocean current is determined by the average value of all vectors from each jellyfish in ocean to the jellyfish at the optimal position at present, and the average position of jellyfish:
wherein ,the optimal value of the weight w and the threshold value theta is the all vectors of the jellyfish at the optimal position, and n is the number of jellyfish (the weight w threshold value theta);
single jellyfish weight w threshold value theta and whole population jellyfish w 1 ,w 2 …,w n And theta 1 ,θ 2 …θ n The average position difference of (2) is:
Δu＝eu
wherein e is a factor controlling attractive force; e is expressed mathematically as follows:
e＝β*rand
wherein β=3, which is the partition coefficient;
trend of change in position for each jellyfish:
wherein ,X* Is the optimal position of jellyfish population;
the update formula for the position (weight w, threshold θ) of each jellyfish is:
X i (t+1)＝X i (t)+rand*(X * -β*rand*u)
step 4-5: the motion relation of jellyfish motions in jellyfish groups is simulated, and in the groups, the motions of jellyfish are divided into two types: active motion and passive motion; the passive motion is that jellyfish moves around the position of the jellyfish, and the position updating formula is as follows:
X i (t+1)＝X i (t)+γ*rand(0,1)*(U b -L b )
wherein ,Ub 、L b The upper and lower boundaries of the exploration space are respectively; gamma is the motion coefficient, gamma>0, regarding the length of movement around the jellyfish position, let γ=1;
simulating the active movement of jellyfish, randomly selecting jellyfish (j) except the jellyfish of interest, and determining the movement direction by using the vector from the jellyfish (i) of interest to the selected jellyfish (j); when the quantity of food at the selected jellyfish (j) position exceeds the quantity of food at the jellyfish (i) position of interest, the latter is moved to the former; if the amount of food available to the selected jellyfish (j) is lower than the amount of food available to the jellyfish (i) of interest, then jellyfish (i) leaves jellyfish (j) directly; thus, each jellyfish moves in a better direction, looking for food in the population:
in addition:
wherein f is the objective function of position X;is the movement direction of jellyfish;
the latest jellyfish position (weight w, threshold θ) is:
step 4-6, establishing a JS artificial jellyfish optimization algorithm model of step 4-1-4.5, wherein the threshold value theta 'of the BP neural network is to be calculated' m And weight w' z Optimizing to finally obtain the optimal initial weight and threshold value of the BP neural network; using the die forging defect repair process parameter data set in the first step to input an input quantity H k Substituting the optimal initial threshold value theta 'into a BP neural network prediction model' m And weight w' z Predicting by BP neural network model to obtain predicted output O of die forging defect repair process parameters k The method comprises the steps of carrying out a first treatment on the surface of the Output O k And the modeling reality value Z k There is an error between them.
3. The method for determining parameters of a precision repair process for a die forging defect based on a JS-BP model as recited in claim 1, wherein the fifth step comprises the steps of 5-1, constructing a desired output quantity T of the die forging defect repair process parameters k (crack width, crack depth, and polished disk rotational speed) and predicted output O of BP neural network prediction model k The sum of squares error function of the difference (modified feed speed, angle of operation, force control magnitude), the average error of this function being
Wherein k is the number of output nodes; the process of minimizing the error function is the back propagation of the error of the BP neural network;
in step 5-2, since there are multiple data in the training data set, there are p training data sets in the sample, and the total error function is the objective function:
step 5-3, calculating the error function in step 5-1 by using a gradient descent method to ensure that the optimal initial threshold value theta 'in step 4-6' m And weight w' z Further optimization is obtained, along with the increase of iterative optimization times, when the error value of the error function meets the error precision requirement, the iterative optimization process is ended:
according to the modified connection parameter value, the weight Deltaw between the system output layer and the hidden layer is obtained in a sorting way ki The connection parameters have the following functional relationship:
threshold Δa on output layer neuron node k The function relationship is as follows:
Δw of weights between hidden layer and input layer ij The change amount has the following functional relationship:
the threshold variation of each neuron node of the hidden layer has the following functional relationship:
wherein ：
the weight and the threshold value after the correction corresponding to the input layer and the output layer can be obtained by arranging the formulas:
in the formula ：wij The weight between the jth neuron node of the input layer and the ith neuron node of the hidden layer is obtained; x and y are respectively the input quantity of the node and the output quantity of the node; w (w) kj The weight between the ith neuron node of the hidden layer and the kth neuron node of the output layer is obtained; x is x j For input ofAn input of a layer j neuron node; i is the threshold of the hidden layer i-th neuron node; f (x) is the transfer function of the hidden layer; a, a k A threshold value for the kth neuron node of the output layer; g (x) is the output layer transfer function; o (o) k The output of the kth neuron node of the output layer;
step 5-4, the optimal weight w obtained by repeated iterative optimization in the step 5-3 is obtained 0 And a threshold value theta 0 Weight w as final BP neural network 0 And a threshold value theta 0 Substituting the input variable of the die forging defect repair process parameter data set into the BP neural network prediction model again to be used as the input quantity of the input of the BP neural network input layer, and outputting the numerical values of the repair feed speed, the operation angle and the force control size to be the optimal predicted quantity through the mapping relation between the polishing disc rotating speed, the defect width, the defect depth, the repair feed speed, the operation angle and the force control size, which are established in the first step.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310743956.XA CN116976192A (en) | 2023-06-24 | 2023-06-24 | JS-BP model-based die forging defect accurate repair process parameter decision method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310743956.XA CN116976192A (en) | 2023-06-24 | 2023-06-24 | JS-BP model-based die forging defect accurate repair process parameter decision method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116976192A true CN116976192A (en) | 2023-10-31 |
Family
ID=88473989
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310743956.XA Pending CN116976192A (en) | 2023-06-24 | 2023-06-24 | JS-BP model-based die forging defect accurate repair process parameter decision method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116976192A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117313555A (en) * | 2023-11-28 | 2023-12-29 | 南京信息工程大学 | Distributed storage-based adaptive OATS-AJSA improved GRU humidity prediction method |
-
2023
- 2023-06-24 CN CN202310743956.XA patent/CN116976192A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117313555A (en) * | 2023-11-28 | 2023-12-29 | 南京信息工程大学 | Distributed storage-based adaptive OATS-AJSA improved GRU humidity prediction method |
| CN117313555B (en) * | 2023-11-28 | 2024-03-08 | 南京信息工程大学 | GRU humidity prediction method based on self-adaptive OATS-AJSA |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108959728B (en) | Radio frequency device parameter optimization method based on deep learning | |
| Rangwala et al. | Learning and optimization of machining operations using computing abilities of neural networks | |
| CN109241291A (en) | Knowledge mapping optimal path inquiry system and method based on deeply study | |
| Razfar et al. | Optimum surface roughness prediction in face milling by using neural network and harmony search algorithm | |
| Mahdavinejad et al. | Optimization of milling parameters using artificial neural network and artificial immune system | |
| Lin | A GA-based neural fuzzy system for temperature control | |
| CN108764540A (en) | Water supply network pressure prediction method based on parallel LSTM series connection DNN | |
| CN115099519B (en) | Oil well yield prediction method based on multi-machine learning model fusion | |
| CN116976192A (en) | JS-BP model-based die forging defect accurate repair process parameter decision method | |
| CN114357852A (en) | Layered water injection optimization method based on long-short term memory neural network and particle swarm optimization algorithm | |
| CN103778482A (en) | Aquaculture dissolved oxygen short-term prediction method based on multi-scale analysis | |
| CN113138555A (en) | GRNN electric spindle thermal error modeling method based on genetic algorithm optimization | |
| CN113128124B (en) | Multi-grade C-Mn steel mechanical property prediction method based on improved neural network | |
| CN115689070B (en) | Energy prediction method for optimizing BP neural network model based on monarch butterfly algorithm | |
| CN113408195A (en) | Method for predicting wear state of cutter in process | |
| Wang et al. | An intelligent process parameters determination method based on multi-algorithm fusion: a case study in five-axis milling | |
| CN115146529A (en) | Method for predicting cold shot blasting strengthening residual stress of automotive suspension spring | |
| Zhang et al. | Intelligent STEP-NC-compliant setup planning method | |
| CN107273971B (en) | Feed-forward neural network structure self-organization method based on neuron significance | |
| CN108614417B (en) | Optimized control and simulation test method for non-Poisson workpiece flow CSPS system | |
| CN111563614A (en) | Load prediction method based on adaptive neural network and TLBO algorithm | |
| CN108319146B (en) | Discrete particle swarm training-based method for radial basis function neural network | |
| CN116402209A (en) | Three gorges reservoir dam front drift prediction method based on improved neural network model | |
| CN113807005B (en) | Bearing residual life prediction method based on improved FPA-DBN | |
| CN115438842A (en) | Load prediction method based on adaptive improved dayflies and BP neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |