Flameproof transformer shell corrugation bending forming process and bending equipment

Abstract

The invention relates to the technical field of transformers, in particular to an explosion-proof transformer shell corrugation bending forming process and bending equipment. The process comprises the following steps: acquiring a plate material and target ripple shape information; inputting the process control parameter sequence into a pre-trained deep learning model to generate an optimized process control parameter sequence; the control equipment executes bending; in the process, process monitoring data are collected in real time and fed back to the model, and the model carries out real-time optimization and adjustment. Correspondingly, the equipment comprises a bending execution mechanism, a data acquisition module and an intelligent processing and control module integrated with a deep learning model. According to the method, the deep learning model is used for replacing artificial experience to carry out process decision making and online self-adaptive adjustment, the problems that a traditional process depends on experience, is poor in consistency and cannot adapt to dynamic working conditions are effectively solved, and high-quality, high-consistency and high-efficiency intelligent production of corrugations of the flame-proof transformer shell is achieved.

Description

Technical Field

[0001] This invention relates to the field of transformer technology, specifically to a corrugated bending forming process and bending equipment for explosion-proof transformer shells. Background Technology

[0002] As a critical industrial piece of equipment, the mechanical strength and sealing performance of the explosion-proof transformer shell directly affect the safety and reliability of the equipment in flammable and explosive environments. The corrugated structure on the shell surface plays an important role in enhancing overall rigidity and expanding heat dissipation area; therefore, the forming quality of the shell corrugations is a core aspect of the manufacturing process. Currently, the industry mainly uses CNC bending machines with specialized molds to bend sheet metal segment by segment to form corrugations. However, existing technical solutions still largely rely on the experience of operators for setting process parameters and debugging molds. Specifically, key parameters in the bending process, such as the target position of the lower mold slide, applied pressure, feed speed curve, and holding time, are usually preset based on the experience of process personnel or basic formulas. This method, which relies on manual experience, has significant limitations. First, it is difficult to guarantee the consistency of corrugation shape, dimensional accuracy, and angle under mass production, resulting in insufficient product quality stability. Second, when dealing with sheet metal with performance fluctuations between different batches, or molds whose dimensions change due to long-term use, the preset fixed process parameters cannot be adaptively adjusted, easily leading to forming defects such as inconsistent springback, overstretching, or wrinkling, making it difficult to further improve the product qualification rate. Furthermore, developing process solutions for new specifications of corrugated shells often requires repeated trial-and-error physical experiments to determine feasible parameter combinations. This process is time-consuming, costly, and makes it difficult to find the optimal process. Although relevant utility model patents have been applied for in the prior art, their protection focuses mainly on the mold structure or equipment configuration itself. Effective solutions are still lacking for how to achieve intelligent, adaptive, and highly consistent corrugated bending processes. Therefore, there is an urgent need in this field for an innovative method and equipment that can overcome the above-mentioned shortcomings, achieve intelligent decision-making and online dynamic optimization of process parameters, and thus ensure high-quality, high-efficiency, and highly consistent corrugated shell forming of explosion-proof transformers.

[0003] Therefore, the existing technology still needs further development. Summary of the Invention

[0004] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide a corrugated bending forming process and bending equipment for explosion-proof transformer shells, so as to solve the problems existing in the prior art.

[0005] To achieve the above-mentioned technical objectives, according to a first aspect of the present invention, the present invention provides a corrugated bending forming process for an explosion-proof transformer housing, comprising: S1. Obtain the material information of the sheet material to be bent and the three-dimensional shape information of the target ripple; S2. Input the material information and the three-dimensional shape information into a pre-trained deep learning model; S3. The deep learning model outputs a corresponding sequence of corrugated bending process control parameters based on the input information. The sequence of process control parameters includes control instructions for the bending equipment in multiple bending steps. S4. Control the bending equipment to perform corrugated bending operation on the sheet material to be bent according to the process control parameter sequence; S5. During the bending operation, process monitoring data reflecting the forming state is collected in real time, and the process monitoring data is fed back to the deep learning model. S6. The deep learning model optimizes and adjusts the currently executed process control parameter sequence in real time based on the feedback process monitoring data.

[0006] Specifically, in step S1, the material information includes at least one of the plate grade, thickness, and hardness; the three-dimensional shape information is obtained through the product's CAD model and includes at least one of the ripple depth, spacing, and radius of curvature.

[0007] Specifically, in step S2, the information input to the deep learning model also includes the identification information of the bending die, and the deep learning model associates the corresponding die feature parameters according to the identification information of the bending die.

[0008] Specifically, in step S3, the process control parameter sequence includes the target position of the lower die slider, the applied pressure value, the feed speed curve, and the holding time in each bending step.

[0009] Specifically, the process monitoring data mentioned in step S5 includes: At least one of the following: real-time bending force acquired by a force sensor, real-time displacement of the mold or sheet material acquired by a displacement sensor, and surface image information of the sheet material acquired by a vision sensor.

[0010] Specifically, the real-time optimization and adjustment method described in step S6 is as follows: The deep learning model compares the real-time collected process monitoring data with the expected monitoring data range of the current step. If the deviation exceeds the allowable threshold, it immediately generates an adjusted control command for subsequent steps and re-plans the process control parameters for the remaining steps.

[0011] Specifically, the deep learning model is trained in the following way: Collect multiple sets of data samples corresponding to historical successful molding cases. Each set of data samples includes an input feature set and a corresponding label process control parameter sequence. The input feature set includes material information, three-dimensional shape information, mold identification information, and corresponding process monitoring data. Use the data samples to supervise the training of the initial deep learning network until the error between its output prediction parameter sequence and the label parameter sequence meets the requirements.

[0012] Specifically, the training process of the deep learning model adopts a reinforcement learning framework. The deep learning model acts as an agent, and its output action is the sequence of process control parameters. The environmental state includes real-time process monitoring data and the completed bending shape. The reward function is constructed based on the dimensional accuracy, surface quality and forming efficiency of the shell after molding.

[0013] Specifically, before step S2, the following steps are also included: The deep learning model is used to perform virtual simulation evaluation on multiple sets of candidate process control parameter sequences, predict the molding results corresponding to each set of parameters, and select the optimal one from multiple sets of candidate parameter sequences as the initial process control parameter sequence for actual execution based on the prediction results.

[0014] According to a second aspect of the present invention, a bending device is provided, comprising: A bending actuator, including an upper die, a lower die slider, and a drive device, is used to perform bending operations; The data acquisition module is used to acquire material information, the three-dimensional shape information of the target ripples, and to collect process monitoring data in real time during the bending process; The intelligent processing and control module integrates the pre-trained deep learning model. The intelligent processing and control module is configured to: receive information acquired by the data acquisition module; run the deep learning model to generate and output the process control parameter sequence; control the bending actuator according to the process control parameter sequence; receive real-time feedback of process monitoring data, and call the deep learning model for real-time optimization and adjustment, generate adjustment instructions and send them to the bending actuator.

[0015] Beneficial effects: The deep learning-based corrugated bending forming process and equipment for explosion-proof transformer shells provided by this invention brings many significant benefits compared to existing technologies, mainly reflected in the dimensions of intelligent process, adaptive quality control, and improved development efficiency.

[0016] First, this invention fundamentally changes the traditional parameter setting mode that relies on human experience by introducing a pre-trained deep learning model as the core of process decision-making. This model can integrate multi-source information such as material properties, the three-dimensional geometry of the target corrugation, and mold features. Through its internally learned complex nonlinear mapping relationships, it automatically generates a precise and optimized sequence of process control parameters. This greatly reduces the reliance on the personal experience of highly skilled technicians, realizes the digitization and standardization of the process, and fundamentally ensures a significant improvement in the forming accuracy and consistency of products across different batches and on different equipment.

[0017] Secondly, this invention constructs a complete real-time closed loop of "perception-decision-execution". During the bending process, the system collects process data in real time through force sensors, displacement sensors, etc., and feeds it back to the deep learning model. The model can evaluate the current process status in real time and compare it with the expected status. Once a deviation caused by interference factors such as material fluctuations or mold wear is detected, the model can immediately make online predictive adjustments and replanning for subsequent unexecuted process instructions. This adaptive capability gives the process system strong anti-interference robustness, can automatically compensate for various uncertainties, significantly reduce scrap and rework, and improve the stability and first-pass yield of the production process.

[0018] Furthermore, the technical solution of this invention organically combines offline global optimization with online local tuning. Before production, multiple candidate process schemes can be virtually simulated and evaluated using a model to quickly select the optimal initial process parameters. This replaces traditional physical trial and error, significantly shortening the process development and debugging cycle of new products and reducing R&D costs. During production, real-time fine-tuning is achieved through online closed-loop control. This two-stage optimization mode not only improves efficiency but also taps into process potential, achieving an optimal balance between quality and efficiency.

[0019] Finally, this invention also solidifies the intelligent process into dedicated equipment, providing users with a ready-to-use intelligent solution by integrating data acquisition, intelligent processing, and control modules. This equipment can automatically complete the entire process from information input, intelligent decision-making, precise execution to quality assurance, greatly lowering the barrier for users to apply advanced technologies. This facilitates the rapid promotion and industrial application of this intelligent technology, driving the upgrading of the entire industry's manufacturing level. Attached Figure Description

[0020] Figure 1 This is a schematic flowchart of the corrugated bending forming process of the explosion-proof transformer shell provided in a specific embodiment of the present invention; Figure 2 This is a schematic diagram of the bending device provided in a specific embodiment of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments in this application, other similar embodiments obtained by those skilled in the art without creative effort should all fall within the scope of protection of this application. Furthermore, directional terms mentioned in the following embodiments, such as "up," "down," "left," and "right," are only for reference to the directions in the accompanying drawings; therefore, the directional terms used are for illustrative purposes and not for limiting the invention.

[0022] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0023] Please see Figure 1 This invention provides a corrugated bending forming process for an explosion-proof transformer housing, comprising: S1. Obtain the material information of the sheet material to be bent and the three-dimensional shape information of the target ripple; It should be further explained that this invention constructs a complete intelligent process closed loop. Specifically, in S1, the operator manually inputs or scans the material code on the board through the equipment's HMI (Human-Machine Interface) to obtain material information; at the same time, the system imports the three-dimensional CAD model file (such as STEP or IGES format) of the shell to be processed through a communication interface (such as USB or Ethernet) and automatically parses out the shape features of the target ripples.

[0024] S2. Input the material information and the three-dimensional shape information into a pre-trained deep learning model; It should be further explained that in S2, this structured information is encapsulated into a feature vector, which is then input into a deep learning model inference engine deployed in a device controller (such as an industrial PC). This model is a multi-layer neural network trained on a large amount of data.

[0025] S3. The deep learning model outputs a corresponding sequence of corrugated bending process control parameters based on the input information. The sequence of process control parameters includes control instructions for the bending equipment in multiple bending steps. It should be further explained that in S3, the model calculates the feature vector and outputs a structured JSON or XML process file, which precisely specifies the specific action parameters of the equipment in each bending step from the first ripple to the last ripple.

[0026] S4. Control the bending equipment to perform corrugated bending operation on the sheet material to be bent according to the process control parameter sequence; It should be further explained that in S4, the equipment's numerical control system (CNC) reads and parses the process document, converts it into control signals for the servo drive and hydraulic valve group, and drives the bending machine slide, back gauge and other mechanisms to move in sequence.

[0027] S5. During the bending operation, process monitoring data reflecting the forming state is collected in real time, and the process monitoring data is fed back to the deep learning model. It should be further explained that in S5, sensors installed at key locations on the equipment (such as pressure sensors installed on the hydraulic cylinder oil circuit, grating rulers that move synchronously with the slider, and industrial cameras facing the bending area) synchronously collect data at a sampling frequency of not less than 100Hz and transmit it to the controller in real time via fieldbus (such as EtherCAT).

[0028] S6. The deep learning model optimizes and adjusts the currently executed process control parameter sequence in real time based on the feedback process monitoring data. It's worth further explaining that in S6, the model inference engine within the controller not only makes one-time predictions in the preceding steps but also functions as a continuously running "observation-decision" loop module. It receives real-time data streams and compares them at high speed with the expected data for the current step (computation cycle <10ms). Once an adjustment is determined, it immediately calculates a new set of optimized parameters for the currently incomplete part and all subsequent steps, and updates subsequent motion commands online through the CNC system. This process is millisecond-level, ensuring that the adjustment takes effect in the next control cycle. Understandably, this closed loop transforms the traditional open-loop, experience-driven process into a data-driven, adaptive intelligent process, fundamentally solving the problem of poor molding quality consistency caused by uncertainties such as material property fluctuations, mold wear, and environmental temperature changes, significantly improving the dimensional accuracy and yield of the explosion-proof housing corrugations.

[0029] Specifically, in step S1, the material information includes at least one of the following: the grade of the sheet material, its thickness, and its hardness. The three-dimensional shape information is obtained through the product's CAD model and includes at least one of the following: ripple depth, spacing, and radius of curvature.

[0030] It should be further explained that this invention specifies the specific physical quantities of the model input features: ①Material Information: 1) Grade, such as Q235B, SUS304, this information is associated with a set of default material mechanical property parameters (such as elastic modulus, yield strength). 2) Thickness t, in millimeters (mm), is the core parameter for calculating bending force. It is usually measured by vernier calipers and then manually entered, with an accuracy of 0.1mm. 3) Hardness, preferably Brinell hardness (HB), for example, the typical hardness range of Q235B sheet is 120-160 HB. Hardness information can be obtained by measuring the edges and corners of the sheet with a portable hardness tester. It is a key indicator for evaluating the material's resistance to plastic deformation and its springback tendency. ② 3D Shape Information: The system automatically identifies and extracts ripple features from the imported 3D model by calling the API of the CAD kernel (such as ACIS, Parasolid) or using a dedicated parsing library. For a typical U-shaped ripple, the extracted parameters include: ripple depth H (vertical distance from trough to crest), ripple spacing P (center distance between adjacent crests), and inner fillet radius R (inner radius of the ripple bend). For example, a specific ripple might be parsed as H=12mm, P=50mm, and R=3mm. These parameters are the basic basis for calculating the number of bending steps, the target position of each step, and the selection of the lower die opening. Using these specific and measurable parameters as model input allows the model's predictions to be based on deterministic physical geometry and material properties, significantly improving the accuracy and reliability of the predictions.

[0031] Understandably, the beneficial effect of the above scheme is that it provides high-information-density and unambiguous input for deep learning models, ensuring the physical interpretability and repeatability of process decisions.

[0032] Specifically, in step S2, the information input to the deep learning model also includes the identification information of the bending die, and the deep learning model associates the corresponding die feature parameters according to the identification information of the bending die.

[0033] It should be further noted that this invention emphasizes the crucial role of molds in intelligent process systems. In specific implementation, each upper and lower mold has a unique identification information carrier installed at its tail or side, preferably an oil-resistant and impact-resistant RFID tag (such as a high-frequency HF 13.56MHz tag). An RFID reader is installed on the equipment's workbench. Once the mold is in place, the reader automatically reads the mold ID. Based on this ID, the system retrieves the corresponding mold characteristic parameters from a local or cloud database. This parameter set includes at least: 1) The radius of the tip of the upper die (punch) is R_punch, such as R2, R4, R6, etc.; 2) The V-groove angle α_die of the lower die (concave die) is usually 30°, 45°, 88°, etc.; 3) Lower die opening width V_die, for example, 12mm, 20mm, 32mm, etc.; 4) The mold wear compensation coefficient C_w, initially set to 1.0, is updated after each mold maintenance or inspection based on the deviation between the measured and nominal dimensions (e.g., if the opening width increases after wear, C_w is adjusted to 1.02). In step S2, these mold feature parameters, along with material and shape information, are encoded into feature vectors and input into the model. For example, a feature vector might be: [Material grade code, thickness t=2.0, hardness HB=140, corrugation depth H=12, spacing P=50, radius of curvature R=3, upper mold radius R_punch=4, lower mold angle α_die=88°, lower mold opening V_die=20, wear coefficient C_w=1.0]. The deep learning model establishes a complex mapping relationship between these features and optimal process parameters during the learning phase.

[0034] Understandably, the beneficial effect of the above solution is that the system can accurately sense and adapt to the tooling used in practice, automatically match and optimize the process of different mold combinations, and especially achieve adaptive compensation of mold state through the wear coefficient C_w, thereby maintaining process stability throughout the mold life cycle.

[0035] Specifically, in step S3, the process control parameter sequence includes the target position of the lower die slider, the applied pressure value, the feed speed curve, and the holding time in each bending step.

[0036] It should be further explained that this invention defines a specific instruction set for the output of the deep learning model. For a shell containing N ripples, the model outputs a sequence containing N elements. Each element (corresponding to a bend in a ripple) is a parameter structure containing: 1) Target position (L_target): The unit is mm. It is the absolute coordinate that the slider needs to move downward from the reference zero point (usually the position where the upper and lower dies just contact the sheet metal). For example, for a corrugation with a depth H=12mm, considering springback, the target position L_target may be calculated as -13.5mm (the negative sign indicates downward movement); 2) Applied pressure value (F_set): The unit is kN, which is the force that the hydraulic system or servo motor needs to reach and maintain during the bending process. Its value is determined by the properties of the sheet material, thickness, bending length, and lower die opening. For example, for a 2mm thick and 1-meter long Q235B sheet, bending it 90° in a V20 lower die, F_set may be 300kN; 3) Feed Rate Curve: This is a polygonal line defined by multiple points (speed, position). For example, three segments can be defined: rapid descent (from 0mm to -2mm, speed 100mm / s), feed (from -2mm to -12mm, speed 5mm / s), and precision positioning (from -12mm to -13.5mm, speed 1mm / s). This curve controls the dynamic characteristics of the forming process; 4) Holding Time (T_hold): Measured in seconds, this is the duration for which the pressure F_set is maintained after the slider reaches L_target. The preferred holding time is 1.0 to 2.0 seconds, with 1.5 seconds being the optimal value. The rationale for this value is as follows: Too short a time (<1.0s) results in insufficient plastic deformation of the material, inadequate release of internal stress, and increased and unstable springback; too long a time (>2.0s) reduces production efficiency and diminishing marginal returns to improving springback. 1.5 seconds is a proven value that achieves the best balance between springback control and production cycle time. This parameter sequence output by the model is sent to the bending machine CNC via a device communication protocol (such as MODBUSTCP), where the CNC parses it into specific axis control commands.

[0037] Understandably, the beneficial effect of the above scheme is that the control commands output by the model are extremely detailed and executable, directly driving the equipment to complete precise and flexible movements, and realizing a seamless transition from intelligent decision-making to physical action.

[0038] Specifically, the process monitoring data mentioned in step S5 includes at least one of the following: real-time bending force collected by a force sensor, real-time displacement of the mold or sheet material collected by a displacement sensor, and sheet material surface image information collected by a vision sensor.

[0039] It should be further explained that this invention specifies the "sensory aspect" of the process system. Specific implementation methods are as follows: ① Real-time bending force: A high-precision strain gauge pressure sensor is installed in the oil inlet circuit of the main hydraulic cylinder of the bending machine or on the force-bearing plate of the slide block. The range covers the maximum tonnage of the equipment (e.g., 1000kN), and the nonlinearity is better than ±0.1%FS. This sensor measures the actual force value F_actual(t) during the bending process in real time, with a sampling frequency of at least 500Hz to capture instantaneous changes in force; ② Real-time displacement: A high-resolution magnetic or optical scale with a resolution of 0.001 mm is installed on the side of the slider. It measures the absolute position L_actual(t) of the slider in real time. ③ Surface image information of the sheet material: An industrial area CCD camera is installed diagonally above the bending area, and images I(x, y, t) of the bending area are captured at a rate of more than 25 frames per second, using coaxial light or a strip light source. These data are synchronously acquired and timestamped via a high-speed data acquisition card. Preferably, force F_actual(t) and displacement L_actual(t) are the core variables that must be monitored, as they constitute the most basic and direct observation pair describing the mechanical state of the bending process. Images I(x, y, t) serve as a supplement to detect subtle slippage or early signs of surface wrinkling that are difficult to observe with the naked eye. In S5, these time-series data are packaged in real time and sent as a data frame (e.g., one frame every 10ms, containing the instantaneous values ​​of force and displacement and the image index at that moment) to the real-time processing thread of the deep learning model.

[0040] Understandably, the beneficial effect of the above scheme is that, through the fusion of multi-source heterogeneous sensors, the system obtains comprehensive, synchronous, and high-precision perception of force-displacement-morphology in the molding process, providing the model with rich information input comparable to or even surpassing the senses of skilled workers for real-time decision-making.

[0041] Specifically, the real-time optimization and adjustment method in step S6 is as follows: the deep learning model compares the real-time collected process monitoring data with the expected monitoring data range of the current step. If the deviation exceeds the allowable threshold, it immediately generates an adjusted control command for subsequent steps and re-plans the process control parameters for the remaining steps.

[0042] It should be further explained that this invention reveals the core algorithm logic for online adaptive adjustment of the model. The specific steps are as follows: In each control cycle (e.g., 10ms), the model receives the latest process monitoring data frame. The model internally maintains the expected monitoring data range for the currently executed bending step i, which is calculated together when the original process parameters are generated in step S3. For example, for step i, it is expected that when the displacement reaches L_i = -8mm, the bending force should be within the interval [F_low_i, F_high_i] = [280kN, 320kN]. The model compares the real-time collected force F_actual with the expected force range corresponding to the current displacement point.

[0043] Furthermore, regarding the deviation determination scheme: calculate the relative deviation δ = |(F_actual - F_expected) / F_expected|, where F_expected is the median of the expected range (300kN in this example). Set an allowable threshold Th. For bending force, the preferred value for the threshold Th is ±7%. The reason for choosing this value is: based on statistical analysis of a large amount of stable molding process data, normal fluctuations are usually within ±5%. When the deviation exceeds ±7%, there is a greater than 95% probability that it is caused by non-random interference factors (such as significantly higher / lower material hardness, a sharp increase in the coefficient of friction due to poor lubrication, or severe local wear of the mold). This threshold can capture meaningful process anomalies in a timely manner without being overly sensitive (avoiding false triggering due to noise).

[0044] Furthermore, regarding the adjustment triggering and execution scheme: if δ > 7%, the model immediately triggers the adjustment procedure. First, the model diagnoses the possible causes of the deviation based on current and historical monitoring data (such as the overall shape of the force-displacement curve) (e.g., determining that the material is too hard). Then, based on the diagnostic results, the model performs two operations: 1) Immediately generate adjusted control commands: For example, increase the target pressure F_set from 300kN to 330kN during the remaining stroke of the current step; 2) Replanning the remaining steps: The model takes the current state as a new starting point, combines the updated material hardness estimate, and runs a fast inference again to calculate a completely new and optimized sequence of process control parameters for all subsequent steps (i+1, i+2, ..., N), and immediately updates the subsequent instructions via CNC. This adjustment is completed in milliseconds.

[0045] Understandably, the beneficial effect of the above scheme is that it achieves true "feedforward-feedback" composite control, enabling predictive compensation before defects actually occur, and greatly enhancing the robustness of the process and its ability to suppress interference.

[0046] Specifically, the deep learning model is trained in the following way: multiple sets of data samples corresponding to historical successful molding cases are collected. Each set of data samples includes an input feature set and a corresponding label process control parameter sequence. The initial deep learning network is trained under supervision using the data samples until the error between its output prediction parameter sequence and the label parameter sequence meets the requirements.

[0047] It should be further noted that this invention details a specific method for training a model through supervised learning, specifically including: ① Data Sample Construction: Export M historical success cases from the factory's MES / SCADA system. The data sample for each case j includes: 1) Input feature set X_j: a vector containing all defined features, such as X_j=[grade code, thickness, hardness, corrugation depth, spacing, radius of curvature, upper die radius, lower die angle, lower die opening, wear coefficient], and perform standardization processing (such as Z-score standardization). 2) Label Y_j: A vector, which is the sequence of process parameters actually used and successfully verified in case j. It is expanded into a one-dimensional vector, for example, Y_j=[L_target1, F_set1, T_hold1, speed curve coefficient 1, ..., L_targetN, F_setN, T_holdN, speed curve coefficient N].

[0048] ② Network Structure: The initial deep learning network is preferably a deep feedforward neural network, with the following structure: The number of nodes in the input layer equals the dimension of X_j (e.g., 12-dimensional). Three fully connected hidden layers, each containing 128 neurons. The hidden layer activation function is ReLU. The number of nodes in the output layer equals the dimension of Y_j, using a linear activation function. The reasons for choosing 3 layers with 128 nodes and ReLU are: a three-layer network is sufficient to fit the complex nonlinear relationships in the process; 128 neurons provide sufficient model capacity, while reducing the risk of overfitting and computational cost compared to larger networks (e.g., 256 nodes), making it suitable for deployment on industrial controllers and real-time inference; the ReLU activation function is computationally efficient, effectively mitigating the gradient vanishing problem in deep networks and accelerating training convergence.

[0049] ③ Training Process: Mean squared error is used as the loss function. The optimizer Adam is used, with its parameters set to default values ​​(β1=0.9, β2=0.999). The initial learning rate is set to 0.001, and a strategy of decaying to 0.9 times the original value every 100 epochs is adopted. The reason for choosing 0.001 is that this is a robust initial learning rate widely used in deep learning, achieving a good balance between large gradient updates (which may lead to oscillations) and small updates (which lead to slow convergence). Combined with the decay strategy, it can more precisely approximate the optimal solution. The batch size is set to 32. The reason for choosing 32 is that this is a common batch size. Compared to larger batches (e.g., 128), it allows for deeper networks with limited GPU memory; compared to smaller batches (e.g., 8), it provides more stable gradient estimates, resulting in a smoother training process. The training cycle is set to 2000 epochs. The reason for choosing 2000 epochs is that by observing the loss curves of the training and validation sets, the loss usually converges sufficiently and tends to stabilize around 2000 epochs. Continuing training yields little benefit and may lead to overfitting. After training is complete, save the network weights.

[0050] Understandably, the beneficial effect of the above solution is that it provides a clear, operable model training path based on actual industrial data, and the resulting model can accurately learn and reproduce the expert experience contained in historically successful processes.

[0051] 8. The corrugated bending forming process for explosion-proof transformer housing according to claim 7, characterized in that the training process of the deep learning model adopts a reinforcement learning framework, the deep learning model acts as an agent, and its output action is the process control parameter sequence, the environmental state includes real-time process monitoring data and the completed bending shape, and the reward function is comprehensively constructed based on the dimensional accuracy, surface quality and forming efficiency of the formed housing.

[0052] It should be further explained that this invention provides a more advanced model training method that autonomously explores optimal strategies through interaction with the environment. The specific implementation is as follows: ① Framework configuration: The proximal policy optimization algorithm is used for training. The agent is a deep neural network with a policy network π_θ(a|s) parameter of θ, input state s, and probability distribution of output action a; ②State s_t: At the t-th step of training (corresponding to a bending step decision point), the state includes 1) real-time process monitoring data of the current step (such as current force and displacement), 2) the shape of the completed bending part (which can be simplified to a depth sequence of a formed ripple), 3) the geometric information of the remaining ripples to be formed, and 4) global features such as material and mold. ③Action a_t: These are the process control parameters for this step (target position, pressure, speed curve, and holding time). ④ Environment: A high-fidelity finite element analysis simulation environment, such as a shell-fitting bending dynamics simulation model built on Abaqus or Ansys, capable of receiving the action a_t, calculating the result after execution, and outputting the next state s_{t+1} and the reward r_t. The design of the reward function R is crucial, and its formal definition is as follows: in: The total reward earned in a single complete bend (one round); Dimensional accuracy score. The calculation formula is as follows: ,in This is the simulation result for the i-th ripple depth. Where N is the target depth and N is the total number of ripples. The score is 1 when the dimensions are perfectly matched; the larger the error, the lower the score, down to 0. Surface quality score. This score is evaluated by analyzing the stress concentration and deformation of the sheet material in the simulation results. If there are no signs of wrinkling or cracking, the score is 1; otherwise, it is reduced to 0.5 or 0 depending on the severity. Efficiency score. The calculation formula is as follows: ,in It is a baseline molding time. This is the total formation time resulting from the agent's decision-making. Scores encourage quick completion. : Failure penalty. If unacceptable defects such as sheet metal tearing or excessive stretching occur during simulation, then... Otherwise, it is 0; Weighting coefficient. Preferred value: Reason for selection: This weighting setting clearly defines the principle of "quality first", assigning the largest weight (0.5) to dimensional accuracy, placing surface quality (0.3) in an important position, giving appropriate encouragement to efficiency (0.1), and imposing a large fixed penalty (-1.0) on any failure, so as to guide the agent to resolutely avoid generating waste; ⑤ Training Process: The agent explores millions of rounds in a simulation environment. In each round, based on the current state s_t, the agent samples an action a_t according to the probability distribution output by the policy network π_θ and executes it. The environment returns a reward r_t and a new state s_{t+1}. These experiences (s_t, a_t, r_t, s_{t+1}) are stored in the experience replay pool. During training, a batch of data is sampled from the pool, and the policy network parameters θ are updated using the pruning objective function of the PPO algorithm to maximize the expected cumulative discount reward. PPO algorithm parameters: discount factor γ = 0.99, pruning range ε ​​= 0.2, optimizer learning rate α = 3e-4. After sufficient training, the agent's policy network becomes a deep learning model capable of outputting high-quality, highly robust process parameters.

[0053] Understandably, the beneficial effect of the above approach is that it does not rely on a large amount of historical "perfect" data, but allows AI to learn through trial and error in a virtual environment. It can not only learn to imitate, but also discover better bending process strategies that surpass human experience, which is especially suitable for scenarios with new materials and new structures where historical data is lacking.

[0054] 9. The corrugated bending forming process for explosion-proof transformer housing according to any one of claims 1-8, characterized in that, before step S2, it further includes: using the deep learning model to perform virtual simulation evaluation on multiple sets of candidate process control parameter sequences, predicting the forming results corresponding to each set of parameters, and selecting the optimal one from multiple sets of candidate parameter sequences based on the prediction results as the initial process control parameter sequence to be actually executed.

[0055] It should be further explained that this invention adds an offline process pre-planning and optimization step. The specific implementation is as follows: Before formal production, the system enters a "virtual trial molding" mode. First, based on baseline process parameters (which can be obtained from empirical formulas or simple case analogies), multiple sets of candidate process control parameter sequences are generated. An efficient generation method is "Gaussian sampling": using the baseline parameter sequence P_baseline as the mean, a standard deviation σ is set (e.g., target position standard deviation 0.5mm, pressure standard deviation 10kN), and K sets (e.g., K=50) of candidate sequences {P_candidate_k} fluctuating around the baseline are randomly generated, k=1, 2, ..., K. Then, for each set of candidate parameters P_candidate_k, the system inputs it along with fixed input features X (material, shape, mold) into the deep learning model. However, at this time, the model does not perform conventional inference, but instead runs its built-in or associated fast result prediction subnetwork. This sub-network, a branch of the main model, is specially trained to directly predict the molding result Y_pred based on input features X and process parameters P. This includes the predicted final dimensions, predicted maximum stress, and predicted force-displacement curve shape. Next, the system scores the prediction results for each set of candidate parameters according to a pre-defined comprehensive evaluation function. The evaluation function is as follows: in: The overall score of the k-th candidate parameter group; the higher the score, the better. : The weighted average error between the predicted size and the target size; The predicted peak equivalent stress, after normalization; The predicted total molding time corresponds to this set of parameters; Weights, for example, set to 0.6, 0.3, 0.1, emphasize accuracy and safety (low stress). The system calculates the score {S_k} of all K groups of candidate parameters, selects the candidate sequence with the highest score as the optimal initial process control parameter sequence, and passes it to step S2 and subsequent online closed-loop control processes. The reason for choosing K=50 groups is to achieve a balance between computation time (usually several seconds to tens of seconds) and search space coverage. Too few groups (e.g., 10) may not find a solution significantly better than the baseline; too many groups (e.g., 200) will significantly increase computation time, but the incremental benefit decreases. 50 groups is an empirically good number for exploring the parameter space with an acceptable response time. Understandably, the beneficial effect of the above scheme is that it realizes "offline global optimization" of process parameters, which can intelligently select the optimal or near-optimal starting point from many possible solutions before production begins. Combined with online "local real-time adjustment", it forms a complete two-stage intelligent process optimization system, which greatly improves the first-piece success rate and taps the process potential.

[0056] Please see Figure 2 The present invention provides another embodiment, which provides a bending device, the bending device comprising: The bending actuator 100 includes an upper die, a lower die slider, and a driving device, and is used to perform bending operations. The data acquisition module 200 is used to acquire material information, the three-dimensional shape information of the target ripple, and to acquire process monitoring data in real time during the bending process; The intelligent processing and control module 300 integrates the pre-trained deep learning model. The intelligent processing and control module 300 is configured to: receive information acquired by the data acquisition module 200; run the deep learning model to generate and output the process control parameter sequence; control the bending actuator 100 according to the process control parameter sequence; receive real-time feedback of process monitoring data, and call the deep learning model for real-time optimization and adjustment, generate adjustment instructions and send them to the bending actuator 100.

[0057] It should be further noted that this invention protects the physical equipment for implementing the aforementioned intelligent process. The bending actuator 100 is the equipment body, including a frame, a worktable, an upper die mounting base, a lower die slider driven by a servo motor or hydraulic cylinder, and high-precision guide rails, etc. The data acquisition module 200 is the equipment's sensing system, specifically including: 1) Information input unit: Touch screen HMI for manual input of material grade, thickness, and hardness; Ethernet port for importing CAD files; RFID reader for reading mold ID; 2) Process Sensing Unit: This unit includes a piezoelectric force sensor installed in the hydraulic lines; a magnetic scale coupled to the slider; and an industrial camera (e.g., a 5-megapixel CMOS camera) installed inside the protective window. The Intelligent Processing and Control Module 300 is the brain of the equipment. Its hardware consists of an industrial-grade embedded computer (e.g., based on an Intel i7 processor), and its software layer includes a real-time operating system (or a real-time Linux kernel), a CNC system kernel, and a deep learning model inference engine deployed within it (e.g., accelerated using Tensor RT or ONNX Runtime). The module's configuration is achieved through software logic: it receives all data from the HMI, documents, RFID, and sensors; calls the inference engine to calculate the model; outputs process parameter sequences to the NC (numerical control) kernel; the NC kernel converts these into servo drive instructions; simultaneously, it receives real-time data streams from sensors, calls the model's real-time evaluation and adjustment routines, and updates the calculated adjustment instructions online to the NC kernel's execution queue. The entire equipment constitutes an integrated intelligent entity. Users only need to perform simple operations such as loading, mold selection, and program invocation, and the equipment can automatically complete the entire process from intelligent process planning and adaptive execution to online quality assurance.

[0058] Understandably, the beneficial effect of the above solution is that it solidifies the innovative process into a mass-producible and operable intelligent equipment, enabling advanced deep learning-driven bending technology to be delivered to users in the form of a standard industrial product. This significantly reduces the threshold for users to apply the technology and the complexity of integration, and has extremely high industrial promotion value.

[0059] In a preferred embodiment, this application also provides an electronic device, the electronic device comprising: The computer device includes a memory and a processor, wherein the memory stores computer-readable instructions that, when executed by the processor, implement the described corrugated bending forming process for the explosion-proof transformer housing. The computer device can be broadly categorized as a server, terminal, or any other electronic device with the necessary computing and / or processing capabilities. In one embodiment, the computer device may include a processor, memory, network interface, communication interface, etc., connected via a system bus. The processor of the computer device can be used to provide the necessary computing, processing, and / or control capabilities. The memory of the computer device may include a non-volatile storage medium and internal memory. The non-volatile storage medium may store an operating system, computer programs, etc. The internal memory can provide an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface and communication interface of the computer device can be used to connect and communicate with external devices via a network. When the computer program is executed by the processor, it performs the steps of the method of the present invention.

[0060] This invention can be implemented as a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the steps of the methods of embodiments of the invention to be performed. In one embodiment, the computer program is distributed across multiple network-coupled computer devices or processors, such that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, may be executed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations may be executed by one or more computer devices or processors, and one or more other method steps / operations may be executed by one or more other computer devices or processors. One or more computer devices or processors may execute a single method step / operation, or execute two or more method steps / operations.

[0061] Those skilled in the art will understand that the method steps of this invention can be performed by a computer program instructing related hardware, such as a computer device or processor, to perform the steps of this invention when executed. Depending on the context, any references herein to memory, storage, databases, or other media may include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0062] The technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification, provided that such combination does not contain contradictions.

[0063] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A deep learning-based process for bending and forming corrugations of an explosion-proof transformer housing, characterized in that, Includes the following steps: S1. Obtain the material information of the sheet material to be bent and the three-dimensional shape information of the target ripple; S2. Input the material information and the three-dimensional shape information into a pre-trained deep learning model; S3. The deep learning model outputs a corresponding sequence of corrugated bending process control parameters based on the input information. The sequence of process control parameters includes control instructions for the bending equipment in multiple bending steps. S4. Control the bending equipment to perform corrugated bending operation on the sheet material to be bent according to the process control parameter sequence; S5. During the bending operation, process monitoring data reflecting the forming state is collected in real time, and the process monitoring data is fed back to the deep learning model. S6. The deep learning model optimizes and adjusts the currently executed process control parameter sequence in real time based on the feedback process monitoring data.

2. The corrugated bending forming process for the explosion-proof transformer shell according to claim 1, characterized in that, In step S1, the material information includes at least one of the plate grade, thickness, and hardness; the three-dimensional shape information is obtained through the product's CAD model and includes at least one of the ripple depth, spacing, and radius of curvature.

3. The corrugated bending forming process for the explosion-proof transformer shell according to claim 2, characterized in that, In step S2, the information input to the deep learning model also includes the identification information of the bending die, and the deep learning model associates the corresponding die feature parameters according to the identification information of the bending die.

4. The corrugated bending forming process for the explosion-proof transformer shell according to claim 3, characterized in that, In step S3, the process control parameter sequence includes the target position of the lower die slider, the applied pressure value, the feed speed curve, and the holding time in each bending step.

5. The corrugated bending forming process for the explosion-proof transformer shell according to claim 4, characterized in that, The process monitoring data mentioned in step S5 includes: At least one of the following: real-time bending force acquired by a force sensor, real-time displacement of the mold or sheet material acquired by a displacement sensor, and surface image information of the sheet material acquired by a vision sensor.

6. The corrugated bending forming process for the explosion-proof transformer shell according to claim 5, characterized in that, The specific method for real-time optimization and adjustment in step S6 is as follows: The deep learning model compares the real-time collected process monitoring data with the expected monitoring data range of the current step. If the deviation exceeds the allowable threshold, it immediately generates an adjusted control command for subsequent steps and re-plans the process control parameters for the remaining steps.

7. The corrugated bending forming process for the explosion-proof transformer shell according to claim 6, characterized in that, The deep learning model is trained in the following way: Collect multiple sets of data samples corresponding to historical successful molding cases. Each set of data samples includes an input feature set and a corresponding label process control parameter sequence. The input feature set includes material information, three-dimensional shape information, mold identification information, and corresponding process monitoring data. Use the data samples to supervise the training of the initial deep learning network until the error between its output prediction parameter sequence and the label parameter sequence meets the requirements.

8. The corrugated bending forming process for the explosion-proof transformer shell according to claim 7, characterized in that, The training process of the deep learning model adopts a reinforcement learning framework. The deep learning model acts as an agent, and its output action is the sequence of process control parameters. The environmental state includes real-time process monitoring data and the completed bending shape. The reward function is constructed based on the dimensional accuracy, surface quality and forming efficiency of the shell after molding.

9. The corrugated bending forming process for the explosion-proof transformer shell according to any one of claims 1-8, characterized in that, Before step S2, the following is also included: The deep learning model is used to perform virtual simulation evaluation on multiple sets of candidate process control parameter sequences, predict the molding results corresponding to each set of parameters, and select the optimal one from multiple sets of candidate parameter sequences as the initial process control parameter sequence for actual execution based on the prediction results.

10. A bending device, characterized in that, The process for implementing the corrugated bending forming process of the explosion-proof transformer housing according to any one of claims 1-9 includes: A bending actuator, including an upper die, a lower die slider, and a drive device, is used to perform bending operations; The data acquisition module is used to acquire material information, the three-dimensional shape information of the target ripples, and to collect process monitoring data in real time during the bending process; The intelligent processing and control module integrates the pre-trained deep learning model. The intelligent processing and control module is configured to: receive information acquired by the data acquisition module; run the deep learning model to generate and output the process control parameter sequence; control the bending actuator according to the process control parameter sequence; receive real-time feedback of process monitoring data, and call the deep learning model for real-time optimization and adjustment, generate adjustment instructions and send them to the bending actuator.

Images