Search

CN-122020032-A - Forging dimension online prediction method and system in forging process

CN122020032ACN 122020032 ACN122020032 ACN 122020032ACN-122020032-A

Abstract

The invention provides a forging size online prediction method and system in a forging process, and relates to the technical field of intelligent manufacturing, wherein the method comprises the steps of acquiring multi-element sensing data in the forging process, wherein the multi-element sensing data comprises forging pressure data, slider running speed data and blank temperature data; the method comprises the steps of preprocessing multi-element sensing data, constructing a time sequence database containing the multi-element sensing data and corresponding forging sizes, constructing a forging size prediction model which is a mixed model fused with a one-dimensional convolutional neural network, a gate control circulation unit and a time sequence attention mechanism, training the forging size prediction model through the time sequence database, establishing a mapping relation between the context vector and a forging size prediction value, inputting real-time acquired forging process time sequence data into the trained forging size prediction model, and outputting the current forging size prediction value.

Inventors

  • HU ZHILI
  • MAO YAWEI
  • HUA LIN

Assignees

  • 武汉理工大学

Dates

Publication Date
20260512
Application Date
20260121

Claims (10)

  1. 1. An online prediction method for the size of a forging piece in a forging process is characterized by comprising the following steps: Acquiring multi-element sensing data in the forging process, wherein the multi-element sensing data comprises forging pressure data, slider running speed data and blank temperature data; Preprocessing the multi-element sensing data to construct a time sequence database containing the multi-element sensing data and corresponding forging sizes; Building a forge piece size prediction model, wherein the forge piece size prediction model is a mixed model fused with a one-dimensional convolutional neural network, a gating circulating unit and a time sequence attention mechanism, and the time sequence attention mechanism is used for generating a context vector focused on a key process time step; Training the forge piece size prediction model through the time sequence database, and establishing a mapping relation between the context vector and the forge piece size prediction value; And inputting the time sequence data of the forging process acquired in real time into a trained forging size prediction model, and outputting the size prediction value of the current forging.
  2. 2. The method for online predicting the size of a forging in a forging process according to claim 1, wherein the forging size prediction model specifically comprises: the one-dimensional convolutional neural network 1DCNN module is used for extracting local time sequence characteristics from the multi-element sensing data; The gate control circulating unit GRU layer is used for learning the time sequence dependency relationship of the local time sequence characteristics and outputting a hidden state sequence; the time sequence attention layer is used for weighting the hidden state sequence to generate a context vector focused on a key process window; And the output layer is used for generating a size predicted value according to the context vector.
  3. 3. The online prediction method of forging size in forging process according to claim 2, wherein the one-dimensional convolutional neural network 1DCNN module comprises two layers of convolutional layers, the number of convolution kernels is 64 and 128 respectively, the convolution kernels are used for extracting local process features layer by layer from forging time sequence data, the convolution kernels are 3 and the step length is 1 so as to capture short-time dependency in forging fluctuation, and each convolutional layer is followed by a largest pooling layer with the window size of 2 so as to enhance feature robustness and reduce noise sensitivity; The gate control circulation unit GRU layer unit number is 128, and is used for learning a long time sequence dependency relationship among process parameters in the forging process, and setting a dropout rate to be 0.2 so as to inhibit overfitting caused by material batch fluctuation or sensor noise; The time sequence attention layer is used for carrying out dynamic weight distribution on the hidden state output by the GRU layer so as to focus on a key process window with obvious influence on the size of the forging; the output layer is a full-connection output layer and is used for mapping the weighted context vector to the forge piece size predicted value.
  4. 4. The method of claim 1, wherein the step of generating a context vector focused on critical process time steps specifically comprises: based on the trainable parameter matrix and the bias vector, carrying out nonlinear transformation on the hidden state of each time step, and calculating an importance score; Normalizing the importance scores of all the time steps by using a normalization index function to obtain a weight coefficient of each time step; Multiplying the hidden state of each time step with the corresponding weight coefficient, and then summing to obtain the context vector.
  5. 5. The method of claim 4, wherein the formula for generating a context vector focused on critical process time steps is: ; ; ; wherein: Is an importance score; is the first Hidden state of each time step; And Is a trainable parameter matrix and bias vector; Is the weight coefficient, representing the first The contribution degree of the hidden states of the time steps to the final prediction task; is a trainable background vector for scoring with importance Dot product is performed to evaluate the importance; As a context vector, a context vector is used, Is the total number of time steps.
  6. 6. The method for online predicting the size of a forging in a forging process according to claim 1, wherein the key process time steps specifically comprise: in the initial preheating stage of the forging process, the temperature of the blank reaches the lower limit of the set process range to a window with the temperature change rate which tends to be gentle; A stable deformation stage in the middle of the forging process, wherein the deviation between the forging pressure and the actual and set values of the speed of the sliding block is continuously in a window within a tolerance range; the final forming stage of the forging process begins with the first occurrence of significant jitter in the forging pressure exceeding 15% of the average or abrupt change in the slider speed exceeding 10% of the set point, until the end of the forging stroke.
  7. 7. The method for online prediction of forging dimensions in a forging process according to claim 1, wherein the input of the forging dimension prediction model is a time sequence matrix with dimension t×3, wherein T is a time step covering a complete forging cycle, and 3 characteristic dimensions correspond to forging pressure, a running speed of a slide block and a blank temperature respectively so as to completely represent a time sequence state of pressure, a running speed and temperature coupling in the forging process; the output dimension of the forging size prediction model is 2, and the forging size prediction model corresponds to the geometric dimension of the forging in the radial direction and the axial direction respectively, so that double-index synchronous prediction of the forming integrity of the forging is realized.
  8. 8. An on-line prediction system for forging dimension in forging process, comprising: the acquisition module is used for acquiring multi-element sensing data in the forging process, wherein the multi-element sensing data comprise forging pressure data, slider running speed data and blank temperature data; the database construction module is used for preprocessing the multi-element sensing data and constructing a time sequence database containing the multi-element sensing data and corresponding forging sizes; The prediction model construction module is used for constructing a forge piece size prediction model, wherein the forge piece size prediction model is a mixed model fused with a one-dimensional convolutional neural network, a gating circulation unit and a time sequence attention mechanism, and the time sequence attention mechanism is used for generating a context vector focused on a key process time step; The training module is used for training the forge piece size prediction model through the time sequence database and establishing a mapping relation between the context vector and the forge piece size prediction value; the output module is used for inputting the real-time acquired time sequence data of the forging process into the trained forging size prediction model and outputting the size prediction value of the current forging.
  9. 9. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor, when executing the computer program, implements the steps of the forging size online prediction method in a forging process as claimed in any one of claims 1 to 7.
  10. 10. A non-transitory computer readable storage medium having stored thereon a computer program, wherein the computer program when executed by a processor implements the steps of the forging dimension online prediction method in a forging process according to any one of claims 1 to 7.

Description

Forging dimension online prediction method and system in forging process Technical Field The invention relates to the technical field of intelligent manufacturing, in particular to an online forging size prediction method and system in a forging process. Background In the current forging production, real-time monitoring and prediction of the critical geometric dimensions of the forging are core problems for guaranteeing the manufacturing quality of high-performance components. The prior art scheme mainly relies on an off-line and spot check type size detection method, namely, after the forging batch is completed, the size of the sampled workpiece is measured by a coordinate measuring machine. The method is seriously dependent on a preset selective inspection plan, and obvious space-time disjoint exists between the detection result and the real-time state of the forging process, so that quality information cannot be embedded into the production takt, and the method essentially still belongs to a passive quality control mode of post detection. However, the forging process is essentially a complex physical process of highly dynamic, pressure-motion velocity-temperature strongly coupled multi-physical fields, with significant time-varying and nonlinear characteristics of the process parameters. In actual production, factors such as equipment vibration, material batch difference, external environment fluctuation and the like often cause instantaneous deviation of forming pressure, sliding block speed and blank temperature, so that metal flow stress change and dynamic load mismatch are caused, and uneven filling and size fluctuation are caused. The inherent data feedback lag of the existing offline sampling inspection scheme makes the production system unable to sense and respond to the process disturbance in real time, and unable to perform early warning and intervention on the dimension deviation in the forging process, thereby severely restricting the online closed-loop optimization of the process parameters and remarkably increasing the risk of batch quality defects. Disclosure of Invention The invention aims to provide an online prediction method and an online prediction system for the size of a forging in the forging process, so as to solve the problems that the prior art scheme mainly depends on an offline and spot check type size detection method, cannot sense and respond to the process disturbance in real time, cannot early warn and intervene the size deviation in the forging process, and remarkably increases the batch quality defect. The method comprises the steps of obtaining multi-element sensing data in the forging process, wherein the multi-element sensing data comprise forging pressure data, slider running speed data and blank temperature data, preprocessing the multi-element sensing data to construct a time sequence database containing the multi-element sensing data and corresponding forging sizes, constructing a forging size prediction model which is a mixed model integrating a one-dimensional convolutional neural network, a gate control circulation unit and a time sequence attention mechanism, wherein the time sequence attention mechanism is used for generating a context vector focused on a key process time step, training the forging size prediction model through the time sequence database, establishing a mapping relation between the context vector and a forging size prediction value, inputting the real-time acquired forging process time sequence data into the trained forging size prediction model, and outputting the current forging size prediction value. The forge piece size prediction model specifically comprises a one-dimensional convolutional neural network 1DCNN module, a gating circulation unit GRU layer, a time sequence attention layer and an output layer, wherein the one-dimensional convolutional neural network 1DCNN module is used for extracting local time sequence features from multi-element sensing data, the gating circulation unit GRU layer is used for learning time sequence dependency relations of the local time sequence features and outputting a hidden state sequence, the time sequence attention layer is used for weighting the hidden state sequence to generate a context vector focused on a key process window, and the output layer is used for generating a size prediction value according to the context vector. Optionally, the one-dimensional convolutional neural network 1DCNN module includes two convolutional layers, the number of the convolutional kernels is 64 and 128, the convolutional kernels are used for extracting local process features from forging time sequence data layer by layer, the size of the convolutional kernels is 3, the step size is 1, short-time dependency in forging fluctuation is captured, each convolutional layer is connected with a largest pooling layer with the window size of 2, the robustness of the features is enhanced, the sensitivity to noise is reduced, th