Micro-control rolling process time series planning method based on inverse neural network
By using a time series planning method for micro-controlled rolling processes based on inverse neural networks, the problems of resource waste and insufficient generalization ability in determining the timing of controlled rolling in traditional micro-controlled rolling technology are solved, and precise controlled rolling time control is achieved in welding and additive manufacturing.
Patent Information
- Application Number
- CN202610992775.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-06
- Publication Date
- 2026-08-25
AI Technical Summary
Traditional micro-controlled rolling technology relies on trial and error to determine the timing of controlled rolling in welding and additive manufacturing, resulting in resource waste and insufficient generalization ability, making it difficult to achieve precise control.
A time series planning method for micro-controlled rolling process based on inverse neural network is adopted. By designing single-pass welding process, determining the heat source parameters of temperature field distribution function, constructing inverse neural network, training and deploying inverse function to accurately control rolling time.
It enables precise control of rolling time in welding and additive manufacturing processes, reduces experimental resource consumption, and improves the flexibility and generalization of rolling timing.
Smart Images
Figure CN122632626A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent welding and additive manufacturing technology, specifically to a time series planning method for micro-controlled rolling processes based on inverse neural networks. Background Technology
[0002] With the increasing application of welding and additive manufacturing in critical technology fields such as aerospace, bridge and shipbuilding, energy and transportation, the traditional reliance on alloying to improve the performance and lightweighting of metal parts and large structural components can no longer meet the requirements of the engineering field. In addition, the high requirements for the precision, cost and manufacturing cycle of welding technology have led to a shift from extensive production to a model of digital design, precise process control and intelligent manufacturing.
[0003] Traditional welding and additive manufacturing are highly complex processes involving multiple physical effects (including heat, force, light, electricity, and magnetism generated by electric arcs or lasers) and complex metallurgical processes (covering the melting of welding wire or powder, flow of the molten pool, solidification of the molten zone, and subsequent solid-state phase transformation). During this process, strong coupling exists between multiple physical fields such as temperature field, flow field, phase field, and stress field. The metallic material rapidly transforms between solid, liquid, and gas phases, resulting in extremely complex heat and mass transfer mechanisms. These complex mechanisms ultimately shape the corresponding microstructure and stress distribution, which are crucial factors affecting weld quality.
[0004] As can be seen from the above, welding or additive manufacturing process parameters directly affect the forming quality and structural strength and toughness. However, no matter how these parameters are optimized, defects such as porosity, cracks, and lack of fusion, as well as the tensile stress distribution after cooling, cannot be completely avoided, posing challenges to achieving an efficient and stable welding process and optimizing component performance.
[0005] Micro-controlled rolling technology is an effective method to avoid the aforementioned defects and reduce tensile stress. Its principle is that after the molten pool solidifies, an impact force is applied by a controlled rolling device at an appropriate temperature to compact the cooled material and reduce residual tensile stress. Applying the rolling impact force too early will cause molten pool splashing, resulting in more processing defects. Applying controlled rolling too late will cause the material to be subjected to impact force after it has already hardened, also causing additional processing defects.
[0006] Therefore, finding the optimal timing for controlled rolling is key to the efficient and stable application of micro-controlled rolling technology in welding and additive manufacturing.
[0007] Traditional methods for determining the controlled rolling process primarily rely on trial and error. This method often requires repeated experiments, consuming significant experimental resources and time. Under constantly changing thermal conditions, the timing of controlled rolling also varies throughout the welding or additive manufacturing process. Therefore, the trial and error method can only determine the optimal timing under specific conditions. Furthermore, re-calibration through trial and error is necessary after changing the manufacturing material.
[0008] This method relies heavily on experimental data and suffers from problems such as low prediction position, lack of flexibility in use, weak generalization ability, and limited functionality. Therefore, it greatly limits the application of micro-controlled rolling technology in welding or additive manufacturing process design. Summary of the Invention
[0009] One of the technical problems to be solved by this application is to overcome the shortcomings of the above-mentioned related technologies and provide a time series planning method for micro-controlled rolling process based on inverse neural network for achieving precise control of controlled rolling.
[0010] The technical solution adopted by this invention to solve the technical problem is: a time series planning method for micro-controlled rolling processes based on inverse neural networks, specifically including the following steps: Design a single-pass welding process: Design the corresponding single-pass welding process based on welding or additive manufacturing; Determine the heat source parameters of the temperature field distribution function: Based on the experimental data of the single-pass welding process test, determine the heat source parameters of the temperature field distribution function; Simplify the temperature field distribution function and generate a training dataset: Calculate the stable molten pool size based on the temperature field distribution function of the determined heat source parameters; analyze the heat source scanning path of welding or additive manufacturing, and extract the feature path based on the molten pool size, stabilization distance, and control rolling spacing; simplify the temperature field distribution function and generate a training dataset by using the control rolling point positions on the feature path. Constructing an inverse neural network: Construct an inverse neural network that takes temperature as input and applied time as output; Obtain the inverse function of the temperature field distribution function: Train and evaluate the inverse neural network using the training dataset to obtain the inverse function of the temperature field distribution function; Deploying the inverse function to the micro-controlled rolling system: Substitute the controlled rolling temperature into the trained inverse neural network to obtain the controlled rolling time at the corresponding controlled rolling position; construct inverse neural networks for other controlled rolling points in sequence and obtain the controlled rolling time to form a controlled rolling time sequence; deploy the obtained controlled rolling time sequence to the actual welding or additive manufacturing controlled rolling process.
[0011] Compared with related technologies, the advantages of this invention are: As a preferred option, the design of a single-pass welding process specifically involves designing a corresponding single-pass welding process based on the welding or additive manufacturing equipment, materials, structural form, and connection method.
[0012] As a preferred option, based on the experimental data from single-pass welding process testing, the heat source parameters of the temperature field distribution function are determined as follows: Experimental data for single-pass welding process testing were obtained using a synchronized infrared camera. The experimental data included equipment and material grades, process parameters, real-time temperature at the location behind the molten pool measured by the infrared camera, and weld cross-sectional profile. The heat source parameters in the temperature field distribution function are determined by using the weld cross-sectional profile and the real-time temperature at a point behind the molten pool measured by an infrared camera. The formula is: ; in, It is the input power of the heat source. It is the efficiency of the heat source, It is the density of metallic materials, It refers to the specific heat of the corresponding material; These are position coordinates. These are the coordinates of the location of the heat source; The corresponding heat source radius The parameters.
[0013] Preferably, the heat source parameters include the efficiency of the heat source. and the radii of the three heat sources , .
[0014] As a preferred method for determining heat source parameters, an iterative method is adopted: heat source radius and heat source radius The estimated value is half the width of the molten pool cross-section, and the radius of the heat source. The estimated value is the depth of the molten pool cross-section and the efficiency of the heat source. For empirical estimates, an iteration step size is set, iterating from 150% to 50% of the estimated radius of each heat source. By comparing the width and length of the cross-section with the real-time temperature of the location behind the molten pool captured by the infrared camera, reasonable values for each heat source parameter are found.
[0015] Preferably, the inverse neural network consists of an input layer, a hidden layer, and an output layer.
[0016] As a preferred embodiment, the filler metal for the single-pass welding process test is S13800 stainless steel welding wire, the wire welding current is 10A, the welding speed is 60m / min, the wire feed speed is 70m / min, the shielding gas is pure argon with a flow rate of 20L / min, and the infrared camera is placed 40mm behind the welding torch; the reverse neural network is a deep neural network, including an input layer, two hidden layers and an output layer.
[0017] Preferably, the depth of the stable molten pool is 6 mm and the width is 15 mm.
[0018] Preferably, the length of the stable molten pool is 20 mm.
[0019] Preferably, the simplified formula for the temperature field distribution function is: ; in, , and The initial time and temperature are known. Attached Figure Description Figure 1 This is a flowchart of the present invention.
[0020] Figure 2 This is a schematic diagram of the single-pass welding experiment and simulation of the present invention.
[0021] Figure 3 This is a cross-section and a schematic diagram of the heat source used in this invention for calibrating heat source parameters.
[0022] Figure 4 This is a schematic diagram of the dimensions of the molten pool at different stages in an embodiment of the present invention.
[0023] Figure 5 This is a simplified process diagram of the heat source scanning path in an embodiment of the present invention.
[0024] Figure 6 This is a diagram of the deep neural network structure in an embodiment of the present invention. Detailed Implementation
[0025] First, those skilled in the art should understand that these embodiments are merely used to explain the technical principles of the embodiments of this application and are not intended to limit the scope of protection of the embodiments of this application. Those skilled in the art can make adjustments as needed to adapt to specific application scenarios.
[0026] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0027] This invention discloses, as follows Figure 1 The time series planning method for micro-controlled rolling processes shown includes the following specific steps: Design a single-pass welding process: Design the corresponding single-pass welding process based on welding or additive manufacturing; Determine the heat source parameters of the temperature field distribution function: Based on the experimental data of the single-pass welding process test, determine the heat source parameters of the temperature field distribution function; Simplify the temperature field distribution function and generate a training dataset: Calculate the stable molten pool size based on the temperature field distribution function (analytical temperature field function) of the determined heat source parameters; analyze the heat source scanning path of welding or additive manufacturing, and extract the feature path based on the molten pool size, stabilization distance, and control rolling spacing; simplify the temperature field distribution function and generate a training dataset by using the control rolling point positions on the feature path. Constructing an inverse neural network: Constructing an inverse neural network (also known as an inverse deep neural network) with controlled rolling temperature as input and controlled rolling application time as output. Obtain the inverse function of the temperature field distribution function: Train and evaluate the inverse neural network using the training dataset to obtain the inverse function of the temperature field distribution function; that is, use the data generated by the forward analytical temperature field function to train and evaluate the neural network to obtain the micro-controlled rolling time series prediction model. Deploying the inverse function to the micro-controlled rolling system: Substituting the controlled rolling temperature into a trained inverse neural network to obtain the controlled rolling time at the corresponding controlled rolling position; subsequently constructing inverse neural networks for other controlled rolling points and obtaining the controlled rolling time to form a controlled rolling time series; deploying the obtained controlled rolling time series to the actual welding or additive manufacturing controlled rolling process. In other words, the micro-controlled rolling time series prediction model is deployed in the manufacturing process to achieve precise control of the micro-controlled rolling impact force at the corresponding time points.
[0028] As a preferred option, the design of a single-pass welding process specifically involves designing a corresponding single-pass welding process based on the welding or additive manufacturing equipment, materials, structural form, and connection method.
[0029] As a preferred option, based on the experimental data from single-pass welding process testing, the heat source parameters of the temperature field distribution function are determined as follows: Using a synchronized infrared camera, corresponding single-pass process experiments were designed based on the welding / additive manufacturing process, and experimental data were obtained. The experimental data included equipment and material grades, process parameters, real-time temperature at the location behind the molten pool measured by the infrared camera, and weld cross-sectional profile. The heat source parameters in the temperature field distribution function are determined by using the weld cross-sectional profile and the real-time temperature at a point behind the molten pool measured by an infrared camera. The formula is: ; in, It is the input power of the heat source. It is the efficiency of the heat source, It is the density of metallic materials, It refers to the specific heat of the corresponding material; These are position coordinates. These are the coordinates of the location of the heat source. and The initial time and temperature are known; The corresponding heat source radius The parameters of the heat source include its efficiency. and the radius of the heat source in three dimensions , .
[0030] As a preferred method for determining heat source parameters, an iterative method is adopted: heat source radius and heat source radius The estimated value is half the width of the molten pool cross-section, and the radius of the heat source. The estimated value is the depth of the molten pool cross-section and the efficiency of the heat source. For empirical estimates, an iteration step size is set, iterating from 150% to 50% of the estimated radius of each heat source. By comparing the width and length of the cross-section with the real-time temperature of the location behind the molten pool captured by the infrared camera, reasonable values for each heat source parameter are found.
[0031] Preferably, the inverse neural network consists of an input layer, a hidden layer, and an output layer.
[0032] The scan path analysis and feature extraction step is particularly important in the entire micro-controlled rolling time series prediction model. This is due to the temperature field distribution function... It is a multivariate function. In order to obtain its inverse function through an inverse neural network, the temperature field distribution function needs to be further simplified into a single-element function.
[0033] Typically, controlled rolling units apply controlled rolling impact forces at equal intervals. Therefore, once the molten pool has reached a stable state, the applied frequency is fixed. This is the first case for predicting the controlled rolling position, such as... Figure 5 As shown between the two vertical white lines. In this case, it is only necessary to solve for the time difference between adjacent controlled rolling positions. The scanning path in this case can be rotated to a horizontal position to simplify the calculation process. For a horizontal scanning path, on the specimen surface... and With the value zero, the only unknown input to the temperature field distribution function is the location term. and Therefore, only the starting position is needed. The temperature field distribution function can be expressed as This becomes a mapping of a one-to-one correspondence.
[0034] The second scenario involves an unstable molten pool, which often occurs at the beginning or end of the scan path. In this case, the situation becomes more complex. First, the scan path is also rotated to a horizontal position, making... and The value is zero. The scan start position is recorded as zero. At this point, the temperature field distribution function can be expressed as: Substitute these other known positions in sequence ( , , A series of simplified temperature field distribution functions can be obtained from these simplified temperature field distribution functions. The corresponding controlled rolling time can be calculated from the inverse functions of these simplified temperature field distribution functions. , , (...). Due to the limited scanning distance in unstable cases, the number of these simplified temperature field distribution functions is also limited.
[0035] Depending on the specific circumstances, training datasets are generated from simplified temperature field distribution functions. Each simplified temperature field distribution function produces a corresponding... This is used to train the corresponding inverse neural network to solve its inverse function.
[0036] Inverse neural networks can be trained to generate the inverse function of a function. The method to implement an inverse function is to train a pre-built inverse neural network such that when the output of the inverse neural network is used as input, the output of the function is equal to the input of the inverse neural network. This makes the output of the inverse neural network the inverse function of the original function. or .
[0037] Training process: using temperature field distribution function For example, the corresponding neural network is constructed as follows: Training points The input used for this inverse neural network, i.e. Then the output of the inverse neural network is applied to the input of the inverse function to be found, i.e. The inverse neural network is trained using a loss function, with the aim of ensuring that the loss function satisfies the following relationship. .
[0038] After training, the inverse neural network's expression of the inverse function is obtained, i.e. .
[0039] Typically, inverse neural networks are trained using known input-output data pairs. The training input data is fed into the inverse neural network, and then a loss function is used to compare the resulting output with the training output data. The loss value returned by this loss function is used through the backpropagation algorithm to adjust the network's weights to reduce the loss. Inverse neural networks use a custom loss function to constrain the neural network, causing it to produce an output that conforms to an inverse function. The response.
[0040] For the current temperature field distribution function This formula has no closed-form solution. Therefore, it is also necessary to utilize an integral neural network. This is used to represent the function, which is used as part of the loss function.
[0041] The loss function of this inverse neural network consists of two parts, such as Figure 6 As shown. Loss function two includes forcing the output of the inverse neural network to conform to the inverse function. The loss function includes a conditional function for a single data point, used to stabilize the training process. After training, the inverse function is obtained. .
[0042] The following specific embodiment further illustrates the time series planning method for micro-controlled rolling processes based on inverse neural networks of the present invention.
[0043] Example: This embodiment constructs a prediction system for weld formation of S13800 stainless steel welding wire based on deep neural network. The welding equipment adopted is Jinrui MFR series pulse gas shielded welding machine, the arc welding robot model is OTC FD-V6L, the welding wire material is S13800 steel, the base material material is C250, and the test plate size is 200mm×100mm×6mm.
[0044] The specimen was constructed of flat plates stacked together, welded using a surfacing method. The welding torch was placed vertically above the specimen plate, without a bevel. The arc type was single-wire weld pool welding, the filler metal was S13800 welding wire with a diameter of 2mm, and the arc voltage was 30V. Figure 2 As shown; the current wire welding current is 10A; the welding speed is 60m / min, the wire feed speed is 70m / min, and the shielding gas is pure argon with a flow rate of 20L / min.
[0045] The infrared camera is positioned 40mm behind the welding torch, and the temperature it measures is located behind the welding torch.
[0046] The length of a single weld is 120mm. After welding, the plate is cut open from the middle to measure the width and depth of the weld pool. Figure 3 As shown. Measurements show that the depth of the molten pool is 6 mm and the width is 15 mm. Furthermore, analysis of the temperature near the center point using infrared imaging suggests that the temperature at a location 40 mm from the arc center is 800℃.
[0047] Assign the depth of the molten pool to , and Assign half the width of the molten pool as the initial reference value. ( and the efficiency of the heat source Within the specified interval, with an iteration step size of 10, the temperature field distribution function of the three-dimensional analytical heat source is repeatedly used to simulate the single-pass welding process, and the depth, width, and temperature of the infrared camera test point corresponding to the experimental position are measured. Note that since this is plate welding, the above-mentioned analytical temperature function can be used. If the single-pass welding is performed on a pre-formed thin wall, it is necessary to use pseudo-heat sources on both sides to simulate the adiabatic boundary conditions of the thin wall.
[0048] By analyzing all simulation results, it was found that when mm, mm, mm and When the value best matches the experimentally measured value, the values of each heat source parameter can be determined.
[0049] In addition, such as Figure 4 As shown, by analyzing the size of the molten pool, it was found that the molten pool reaches stability when the heat source is 25mm from the starting point. The size of the molten pool begins to change when it is 20mm from the ending point of the heat source. The length of the stable molten pool is 20mm. Therefore, a single-pass welding model with a length of 85mm is sufficient to represent the changes in the molten pool under the current welding or additive manufacturing processing parameters. By plotting the controlled rolling position alongside this single-pass welding scan path, the required controlled rolling position can be obtained, i.e., In the current processing technology, there are 8 locations that need to be calculated. Therefore, 8 deep neural networks need to be constructed to solve the corresponding inverse functions.
[0050] By location ( Taking mm as an example, we can construct a deep neural network as an inverse neural network, such as... Figure 6 As shown. Before solving for the inverse function using the established deep neural network, an integral neural network needs to be built to calculate the integral expression of the temperature field distribution function. To simplify the calculation, the scanning path is rotated to a horizontal position and only the surface temperature is needed; therefore, the simplified equation of the temperature field distribution function is: The location of the heat source is a function of time with respect to a known scanning speed, i.e. .
[0051] To train this integral neural network First, the training set is generated using the simplified equations described above. This deep neural network consists of an input layer, two hidden layers, and an output layer. Each hidden layer contains 512 neurons. The final system parameters are batch_size 64, epochs 5000, training dataset size 100, and learning rate 0.1. An integral neural network is obtained through training.
[0052] Furthermore, a deep neural network is used to inverse the simplified equation above, yielding its inverse function. The training samples still use the training set from the previous step, but the input and output positions are interchanged. Similarly, this deep neural network contains an input layer, two hidden layers, and an output layer. Each hidden layer contains 256 neurons. The final system parameters are batch_size 32, epochs 5000, training dataset size 100, and learning rate 0.1. The loss function of this deep neural network consists of two parts, one part of which is used to force the output to conform to the inverse function, i.e. Another portion is used to stabilize the training process.
[0053] The position was obtained after training ended. The inverse function of the temperature field distribution function at (mm) is substituted into the controlled rolling temperature. The controlled rolling time can then be obtained. s.
[0054] Train using the same method. , , , , , , A deep neural network at millimeter speed yields a series of controlled rolling time series. , , , , , , s. The time difference between adjacent control rolling points when stable is s. This allows for the generalization of the controlled rolling time series to the original scan path's controlled rolling time series ( , , , , , , , , , , , , s).
[0055] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A time series planning method for micro-controlled rolling processes based on inverse neural networks, characterized in that, Specifically, the steps include the following: Design a single-pass welding process: Design the corresponding single-pass welding process based on welding or additive manufacturing; Determine the heat source parameters of the temperature field distribution function: Based on the experimental data of the single-pass welding process test, determine the heat source parameters of the temperature field distribution function; Simplify the temperature field distribution function and generate a training dataset: Calculate the stable molten pool size based on the temperature field distribution function of the determined heat source parameters; analyze the heat source scanning path for welding or additive manufacturing, and extract the feature path based on the molten pool size, stabilization distance, and controlled rolling spacing; By simplifying the temperature field distribution function through the location of the control rolling point on the feature path, a training dataset is generated. Constructing an inverse neural network: Construct an inverse neural network that takes temperature as input and applied time as output; Obtain the inverse function of the temperature field distribution function: Train and evaluate the inverse neural network using the training dataset to obtain the inverse function of the temperature field distribution function; Deploying the inverse function to the micro-controlled rolling system: Substitute the controlled rolling temperature into the trained inverse neural network to obtain the controlled rolling time at the controlled rolling position; construct inverse neural networks for other controlled rolling points in sequence and obtain the controlled rolling time, obtain the controlled rolling time difference between adjacent controlled rolling points when stable, and form a controlled rolling time sequence; deploy the obtained controlled rolling time sequence to the actual welding or additive manufacturing controlled rolling process.
2. The time series planning method for micro-controlled rolling process based on inverse neural network according to claim 1, characterized in that, The design of a single-pass welding process specifically involves designing a corresponding single-pass welding process based on the welding or additive manufacturing equipment, materials, structural form, and connection method.
3. The time series planning method for micro-controlled rolling process based on inverse neural network according to claim 2, characterized in that, Based on the experimental data from the single-pass welding process test, the specific heat source parameters of the temperature field distribution function are determined as follows: Experimental data for single-pass welding process testing were obtained using a synchronized infrared camera. The experimental data included equipment and material grades, process parameters, real-time temperature at the location behind the molten pool measured by the infrared camera, and weld cross-sectional profile. The heat source parameters in the temperature field distribution function are determined by using the weld cross-sectional profile and the real-time temperature at a point behind the molten pool measured by an infrared camera. The formula is: ； in, It is the input power of the heat source. It is the efficiency of the heat source, It is the density of metallic materials, It refers to the specific heat of the corresponding material; These are position coordinates. These are the coordinates of the location of the heat source; The corresponding heat source radius The parameters.
4. The time series planning method for micro-controlled rolling process based on inverse neural network according to claim 3, characterized in that, The heat source parameters include the efficiency of the heat source. and the radii of the three heat sources , .
5. The time series planning method for micro-controlled rolling process based on inverse neural network according to claim 4, characterized in that, The method for determining heat source parameters is an iterative method: heat source radius and heat source radius The estimated value is half the width of the molten pool cross-section, and the radius of the heat source. The estimated value is the depth of the molten pool cross-section and the efficiency of the heat source. For empirical estimates, an iteration step size is set, iterating from 150% to 50% of the estimated radius of each heat source. By comparing the width and length of the cross-section with the real-time temperature of the location behind the molten pool captured by the infrared camera, reasonable values for each heat source parameter are found.
6. The time series planning method for micro-controlled rolling process based on inverse neural network according to claim 5, characterized in that, The inverse neural network consists of an input layer, a hidden layer, and an output layer.
7. The time series planning method for micro-controlled rolling processes based on inverse neural networks according to claim 6, characterized in that, The filler metal for the single-pass welding process test was S13800 stainless steel welding wire, and the wire welding current was 10A. The welding speed is 60 m / min, the wire feed speed is 70 m / min, the shielding gas is pure argon with a flow rate of 20 L / min, and the infrared camera is placed 40 mm behind the welding torch; the reverse neural network is a deep neural network, including an input layer, two hidden layers and an output layer.
8. The time series planning method for micro-controlled rolling process based on inverse neural network according to claim 7, characterized in that, The depth of the stabilized molten pool is 6 mm and the width is 15 mm.
9. The time series planning method for micro-controlled rolling process based on inverse neural network according to claim 8, characterized in that, The length of the stabilized molten pool is 20 mm.
10. The time series planning method for micro-controlled rolling process based on inverse neural network according to claim 9, characterized in that, The simplified formula for the temperature field distribution function is: ； in, , and The initial time and temperature are known.