Joint assembly integrated forming and welding device for railway freight car and welding process of joint assembly integrated forming and welding device

Abstract

The integrated forming welding device comprises a fixed base, a top plate is arranged at the top end of the fixed base, a first vertical plate is fixedly installed on one side of the top end of the fixed base, and a second vertical plate is fixedly installed on the other side of the top end of the fixed base; a lifting mechanism is fixedly mounted at the top end of the fixed base, a connecting table is mounted at the top end of the lifting mechanism, a clamping mechanism is fixedly mounted at the top end of the connecting table, and welding mechanisms are fixedly mounted at the top end of one side of the first vertical plate and the top end of one side of the second vertical plate correspondingly. When products placed on the connecting table are different in thickness, the lifting air cylinder pushes the connecting table from the bottom, the placing height of the products is adjusted, a welding head makes contact with the surfaces of the products conveniently for welding machining, the traditional mode that welding is conducted after the height of the welding head is adjusted is replaced, and the welding efficiency of the welding device is improved.

Description

Technical Field

[0001] The invention relates to the technical field of welding devices, and in particular to an integrated forming welding device for a joint assembly for a railway freight car and a welding process thereof. Background Art

[0002] The main function of the railway freight car joint is to realize the connection between the locomotive and the vehicle or between the vehicle and the vehicle, transmit the traction and impact force, and keep a certain distance between the vehicles. The coupler is an important part of the railway freight car joint. It is divided into two types according to the opening method: the upper action type and the lower action type. The upper action type coupler is opened by the lifting mechanism on the upper part of the coupler hook head. Most freight cars generally adopt this type; while the lower action type coupler is opened by the action of the push lever at the lower part of the hook head. Passenger cars usually adopt this form. In addition, the coupler is further subdivided into spiral couplers, close-fitting automatic couplers, automatic couplers and rotating couplers according to its structural type to meet different usage requirements. These different types of couplers jointly ensure the safety and efficiency of freight cars during transportation, realize the flexible combination of train formations through connection, transmit traction and impact force at the same time, and maintain the necessary distance between vehicles, thereby ensuring the stability of train operation and the safety of cargo transportation. During the preparation process of the railway freight car joint, welding equipment needs to be used to weld and assemble the original railway freight car joint with the original railway freight car joint.

[0003] However, the conventional welding device has the following disadvantages:

[0004] Traditional welding devices weld products by adjusting the position of the welding gun, but the surface flatness of the products is uneven, and the user needs to adjust the height of the welding gun multiple times. Such welding operations are very cumbersome and reduce the welding efficiency of the welding device. Summary of the invention

[0005] The object of the present invention is to provide an integrated welding device for a joint assembly for a railway freight car and a welding process thereof, so as to solve the problem that the traditional welding device proposed in the above background technology welds the product by adjusting the position of the welding gun, but the surface flatness of the product is uneven, and the user needs to adjust the height of the welding gun multiple times, so the welding operation is very cumbersome and reduces the welding efficiency of the welding device.

[0006] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: an integrated forming welding device for a joint assembly for a railway freight car, comprising a fixed base, a top plate is provided at the top of the fixed base, a first vertical plate is fixedly installed on one side of the top of the fixed base, a second vertical plate is fixedly installed on the other side of the top of the fixed base, a lifting mechanism is fixedly installed at the top of the fixed base, a connecting platform is installed at the top of the lifting mechanism, a clamping mechanism is fixedly installed at the top of the connecting platform, a welding mechanism is fixedly installed at the top of one side of the first vertical plate and the top of one side of the second vertical plate, a control panel is fixedly installed at the middle of one side of the first vertical plate, the welding mechanism comprises a pushing cylinder and a sliding plate, the movable end of the pushing cylinder is fixedly connected to one side of the sliding plate, the clamping mechanism comprises two limit plates and a screw rod, the two ends of the screw rod are respectively rotatably connected to the opposite side of the two limit plates, the lifting mechanism comprises a positioning plate and a lifting cylinder, the middle part of the top of the positioning plate is fixedly connected to the fixed end of the lifting cylinder, and positioning seats are provided on both sides of the lifting cylinder.

[0007] As a preferred technical solution of the present invention, the bottom end of the slide board is rotatably connected to a first welding arm, one end of the first welding arm is rotatably connected to a second welding arm, the bottom end of the second welding arm is rotatably connected to a length rod, a welding table is fixedly installed at the bottom end of the length rod, a welding head is fixedly installed at the bottom end of the welding table, and the welding head welds the product.

[0008] As a preferred technical solution of the present invention, the middle part of one side of the sliding plate is rotatably connected with a first angle cylinder, and the movable end of the first angle cylinder is rotatably connected to the side opposite to the first welding arm, and the middle part of the first welding arm is rotatably connected with a second angle cylinder, and the movable end of the second angle cylinder is rotatably connected to the side opposite to the second welding arm, the middle parts of the two pushing cylinders are respectively fixedly connected to the first vertical plate and the second vertical plate, and the two sliding plates are fixedly connected to the top plate, the first angle cylinder performs telescopic and deflection movements, and the first angle cylinder pushes the first welding arm to deflect at an angle along the sliding plate, and the second angle cylinder performs telescopic and deflection movements, and the second angle cylinder pushes the second welding arm to deflect at an angle along the first welding arm, so as to adjust the position of the welding head.

[0009] As a preferred technical solution of the present invention, the surface of the screw rod is threadedly connected to two movable blocks, and the two movable blocks are slidably connected to the connecting platform, and the top ends of the two movable blocks are fixedly installed with clamping plates.

[0010] As a preferred technical solution of the present invention, a stepper motor for driving the screw to rotate is fixedly installed on one side of one of the limit plates, and the bottom ends of the two limit plates are fixedly connected to the connecting platform. The stepper motor starts after being energized, and drives the screw to rotate. The screw has two threads in opposite directions with its own center line as the dividing line. The two movable blocks match the threads on the surface of the screw. The two movable blocks are limited by the connecting platform, so the two movable blocks slide relative to each other along the screw. The movable blocks drive the clamps to move synchronously, and the two clamps clamp and fix the product from both sides.

[0011] As a preferred technical solution of the present invention, displacement grooves are provided on both sides of the top of the positioning plate, and displacement blocks are slidably connected inside the two displacement grooves. A lifting seat is fixedly installed on the surface of the lifting cylinder, and direction seats are fixedly installed on the tops of the two displacement blocks. One side of the two direction seats is rotatably connected with a first direction plate, and the ends of the two first direction plates away from the direction seats are respectively rotatably connected to the insides of the two positioning seats, and the other sides of the two direction seats are rotatably connected with a second direction plate, and the ends of the two second direction plates away from the direction seats are respectively rotatably connected to the two sides of the lifting seat, the lifting cylinder performs telescopic movement, and the lifting cylinder drives the lifting seat to move synchronously, the lifting seat drives the second direction plate to deflect at an angle along the direction seat, and the direction seat drives the displacement block to slide along the displacement groove, and the first direction plate to deflect at an angle along the positioning seat, thereby increasing the stability of the distance adjustment between the connecting platform and the positioning plate, and indirectly adjusting the height of the product placed on the connecting platform.

[0012] As a preferred technical solution of the present invention, the top ends of the two positioning seats and the movable end of the lifting cylinder are fixedly connected to the side opposite to the connecting platform, the bottom end of the positioning plate is fixedly connected to the fixed base, the top end of the first vertical plate and the top end of the second vertical plate are respectively fixedly connected to the two sides of the bottom end of the top plate, the lifting mechanism is installed on the connecting platform through the positioning seat and the lifting cylinder, and the lifting mechanism is installed on the fixed base through the positioning plate.

[0013] As a preferred technical solution of the present invention, the control panel includes a power interface circuit, a rectifier and filter circuit, an inverter circuit, a transformer module, a welding circuit, a voltage and current feedback module, a control module, a dimming mask drive circuit, an automatic dimming mask circuit and PWM control. The rectifier and filter circuit are connected to the power interface circuit, the inverter circuit is connected to the rectifier and filter circuit, the transformer module is connected to the inverter circuit, the welding circuit and the voltage and current feedback module are both connected to the transformer module, the automatic dimming mask circuit is connected to the dimming mask drive circuit, and the PWM control, the dimming mask drive circuit and the voltage and current feedback module are all connected to the control module.

[0014] The welding process of the integrated forming welding device of the joint assembly for railway freight cars of the present invention comprises the following steps:

[0015] Step 1: Product fixation: The product is placed on the connection table, and two clamps are used to clamp and fix the product from both sides;

[0016] Step 2: Product lifting: The lifting cylinder on the lifting mechanism performs telescopic movement to adjust the height of the product placed on the connecting table;

[0017] Step 3: Product welding: The welding head on the welding mechanism performs welding processing on the lifted product.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: by setting up a welding mechanism and a lifting mechanism, when the thickness of the products placed on the connecting table is different, the lifting cylinder pushes the connecting table from the bottom to adjust the height of the product placement, which is convenient for the welding head to contact the product surface for welding processing, replacing the traditional method of adjusting the height of the welding head before welding, thereby improving the welding efficiency of the welding device.

[0019] Improve welding quality: Through self-supervised contrastive learning, the model can more accurately extract important features in welding images. The contrast loss function enhances the model's ability to recognize detailed features by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs. This can effectively improve the integrity and defect-freeness of the weld and ensure the stability and consistency of welding quality. The welding process may involve different welding materials and different welding environments. Through self-supervised contrastive learning, the model can adapt to these complex environmental changes, thereby ensuring high-quality output of the welding process.

[0020] Enhanced model robustness: Meta-learning enables the model to quickly adapt to new welding tasks by training on multiple different welding tasks. This includes tasks of welding different materials (such as steel and aluminum) and different welding speeds. By training on different tasks, the model can learn how to quickly adjust parameters and adapt to new welding conditions, which enhances the robustness of the model in different tasks. In the actual welding process, the environment may change, such as fluctuations in temperature and pressure. Meta-learning enables the model to maintain efficient and stable performance in the face of these changes.

[0021] Improve welding efficiency: The optimal neural network architecture found by NAS not only performs well in welding quality, but also optimizes computing efficiency. The efficient architecture can process more data in a shorter time, thereby improving the overall efficiency of the welding process. The optimized model can process the large amount of image and sensor data generated during the welding process in real time. This real-time processing capability ensures the efficient operation of the welding process, reduces waiting and processing time, and improves welding efficiency.

[0022] Improve model stability: The controller continuously generates and optimizes candidate architectures through reinforcement learning algorithms. Performance evaluation is performed on each architecture, and the optimal architecture is selected to ensure the stability of the welding process. In this way, the model can maintain stable output under different welding conditions, improving the consistency of welding quality. The model optimized by reinforcement learning can remain stable during long-term operation and is not easily affected by external interference, thereby ensuring the continuity and stability of the welding process.

[0023] Realize intelligent welding control: Use the trained model for welding control, which can input welding image data and sensor data in real time, and output the position and welding speed of the welding head. Through real-time monitoring and adjustment, the welding process is more intelligent, reducing human intervention and improving the degree of automation. The intelligent welding control model can provide feedback and adjustments based on real-time monitoring data. For example, if it detects that the welding quality has deteriorated, the model can automatically adjust the welding parameters to improve the quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a side view of the present invention;

[0025] Figure 2 It is a connection diagram of the connection platform and the clamping mechanism of the present invention;

[0026] Figure 3 It is a connection diagram of the connection platform and the lifting mechanism of the present invention;

[0027] Figure 4 is a cross-sectional view of the lifting mechanism of the present invention;

[0028] Figure 5 It is a side view of the welding mechanism of the present invention;

[0029] Figure 6 A three-dimensional diagram of the welding mechanism of the present invention;

[0030] Figure 7 is a schematic diagram of the structure of the control panel of the present invention;

[0031] Figure 8 It is a flow chart of the present invention.

[0032] Fig. 9 :Flowchart of the intelligent algorithm introduced in the present invention

[0033] In the figure: 1, fixed base; 2, first vertical board; 3, control panel; 301, power interface circuit; 302, rectifier filter circuit; 303, inverter circuit; 304, transformer module; 305, welding circuit; 306, voltage and current feedback module; 307, control module; 308, dimming mask drive circuit; 309, automatic dimming mask circuit; 310, PWM control; 4, welding mechanism; 41, push cylinder; 42, slide board; 43, first welding arm; 44, second welding arm; 45 , length rod; 46, welding table; 47, welding head; 48, second angle cylinder; 49, first angle cylinder; 5, top plate; 6, second vertical plate; 7, clamping mechanism; 71, limit plate; 72, screw rod; 73, clamping plate; 74, stepping motor; 75, movable block; 8, lifting mechanism; 81, positioning seat; 82, first direction plate; 83, positioning plate; 84, displacement slot; 85, displacement block; 86, second direction plate; 87, lifting cylinder; 88, direction seat; 89, lifting seat; 9, connecting table. DETAILED DESCRIPTION

[0034] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0035] See also Figure 1-8 The present invention provides an integrated forming welding device for a joint assembly for a railway freight car, comprising a fixed base 1, a top plate 5 is provided at the top of the fixed base 1, a first vertical plate 2 is fixedly installed on one side of the top of the fixed base 1, a second vertical plate 6 is fixedly installed on the other side of the top of the fixed base 1, a lifting mechanism 8 is fixedly installed on the top of the fixed base 1, a connecting platform 9 is installed on the top of the lifting mechanism 8, a clamping mechanism 7 is fixedly installed on the top of the connecting platform 9, and a welding mechanism 4 is fixedly installed on the top of one side of the first vertical plate 2 and the top of one side of the second vertical plate 6 A control panel 3 is fixedly installed in the middle of one side of the first vertical plate 2. The welding mechanism 4 includes a pushing cylinder 41 and a sliding plate 42. The movable end of the pushing cylinder 41 is fixedly connected to one side of the sliding plate 42. The clamping mechanism 7 includes two limit plates 71 and a screw rod 72. The two ends of the screw rod 72 are rotatably connected to the opposite sides of the two limit plates 71 respectively. The lifting mechanism 8 includes a positioning plate 83 and a lifting cylinder 87. The middle part of the top of the positioning plate 83 is fixedly connected to the fixed end of the lifting cylinder 87. Positioning seats 81 are provided on both sides of the lifting cylinder 87.

[0036] The bottom end of the slide plate 42 is rotatably connected to a first welding arm 43, one end of the first welding arm 43 is rotatably connected to a second welding arm 44, the bottom end of the second welding arm 44 is rotatably connected to a length rod 45, a welding table 46 is fixedly installed at the bottom end of the length rod 45, a welding head 47 is fixedly installed at the bottom end of the welding table 46, and the welding head 47 is used to weld the product.

[0037] The middle part of one side of the sliding plate 42 is rotatably connected with a first angle cylinder 49, and the movable end of the first angle cylinder 49 is rotatably connected with a side opposite to the first welding arm 43. The middle part of the first welding arm 43 is rotatably connected with a second angle cylinder 48, and the movable end of the second angle cylinder 48 is rotatably connected with a side opposite to the second welding arm 44. The middle parts of the two pushing cylinders 41 are respectively fixedly connected to the first vertical plate 2 and the second vertical plate 6. Both sliding plates 42 are fixedly connected to the top plate 5. The first angle cylinder 49 performs telescopic and deflection movements, and the first angle cylinder 49 pushes the first welding arm 43 to deflect at an angle along the sliding plate 42. The second angle cylinder 48 performs telescopic and deflection movements, and the second angle cylinder 48 pushes the second welding arm 44 to deflect at an angle along the first welding arm 43, so as to adjust the position of the welding head 47.

[0038] The surface of the screw rod 72 is threadedly connected with two movable blocks 75 , and the two movable blocks 75 are slidably connected to the connecting platform 9 , and the top ends of the two movable blocks 75 are fixedly mounted with clamping plates 73 .

[0039] A stepper motor 74 for driving the screw rod 72 to rotate is fixedly installed on one side of one of the limit plates 71. The bottom ends of the two limit plates 71 are fixedly connected to the connecting platform 9. The stepper motor 74 starts after being energized, and drives the screw rod 72 to rotate. The screw rod 72 has two threads in opposite directions with its own center line as the dividing line. The two movable blocks 75 match the threads on the surface of the screw rod 72. The two movable blocks 75 are limited by the connecting platform 9, so the two movable blocks 75 slide relatively along the screw rod 72. The movable blocks 75 drive the clamping plates 73 to move synchronously, and the two clamping plates 73 clamp and fix the product from both sides.

[0040] Displacement grooves 84 are provided on both sides of the top of the positioning plate 83, and displacement blocks 85 are slidably connected inside the two displacement grooves 84. A lifting seat 89 is fixedly installed on the surface of the lifting cylinder 87, and a direction seat 88 is fixedly installed on the top of the two displacement blocks 85. One side of the two direction seats 88 is rotatably connected to the first direction plate 82, and the ends of the two first direction plates 82 away from the direction seat 88 are respectively rotatably connected to the inside of the two positioning seats 81, and the other sides of the two direction seats 88 are rotatably connected to the second direction plates 86, and the two second direction plates 88 are rotatably connected. The ends away from the direction seat 88 are rotatably connected to the two sides of the lifting seat 89, and the lifting cylinder 87 performs telescopic movement, and the lifting cylinder 87 drives the lifting seat 89 to perform synchronous movement, and the lifting seat 89 drives the second direction plate 86 to deflect along the direction seat 88, and the direction seat 88 drives the displacement block 85 to slide along the displacement groove 84, and the first direction plate 82 deflects along the positioning seat 81, so as to increase the stability of the distance adjustment between the connecting platform 9 and the positioning plate 83, and indirectly adjust the height of the product placed on the connecting platform 9.

[0041] The top ends of the two positioning seats 81 and the movable ends of the lifting cylinders 87 are fixedly connected to the side directly opposite to the connecting platform 9, the bottom end of the positioning plate 83 is fixedly connected to the fixed base 1, the top ends of the first vertical plate 2 and the second vertical plate 6 are respectively fixedly connected to the two sides of the bottom end of the top plate 5, the lifting mechanism 8 is installed on the connecting platform 9 through the positioning seats 81 and the lifting cylinders 87, and the lifting mechanism 8 is installed on the fixed base 1 through the positioning plate 83.

[0042] The control panel 3 includes a power interface circuit 301, a rectifier and filter circuit 302, an inverter circuit 303, a transformer module 304, a welding circuit 305, a voltage and current feedback module 306, a control module 307, a dimming mask drive circuit 308, an automatic dimming mask circuit 309 and a PWM control 310. The rectifier and filter circuit 302 is connected to the power interface circuit 301, the inverter circuit 303 is connected to the rectifier and filter circuit 302, the transformer module 304 is connected to the inverter circuit 303, the welding circuit 305 and the voltage and current feedback module 306 are both connected to the transformer module 304, the automatic dimming mask circuit 309 is connected to the dimming mask drive circuit 308, and the PWM control 310, the dimming mask drive circuit 308 and the voltage and current feedback module 306 are all connected to the control module 307.

[0043] The welding process of the integrated forming welding device of the joint assembly for railway freight cars of the present invention comprises the following steps:

[0044] Step 1: Product fixing: The product is placed on the connecting table 9, and two clamping plates 73 clamp and fix the product from both sides;

[0045] Step 2: Product lifting: The lifting cylinder 87 on the lifting mechanism 8 performs telescopic movement to adjust the height of the product placed on the connecting platform 9;

[0046] Step 3: Product welding: The welding head 47 on the welding mechanism 4 performs welding processing on the lifted product.

[0047] like Fig. 9 When welding the product in step 3, an algorithm based on self-supervised contrastive learning and meta-learning is introduced to optimize the process of the integrated forming welding device for the joint assembly of railway freight cars. The algorithm can improve the welding quality in actual operation, enhance the robustness and stability of the system, improve the welding efficiency, improve the generalization ability of the model, and realize intelligent welding control. The specific process is:

[0048] Step 1: Data collection and preprocessing

[0049] Camera data

[0050] The main camera is installed above or on the side of the welding head 47 to fully monitor the welding area and capture all details of the welding process; the camera is fixed to the frame of the welding device using an adjustable bracket to ensure that the camera viewing angle can be adjusted as needed;

[0051] The auxiliary camera is installed on the opposite side of the welding head 47 or other key positions to provide monitoring at different viewing angles; the camera is installed at an appropriate position through an auxiliary bracket to ensure that the blind spot of the viewing angle of the main camera is supplemented;

[0052] The use of a high frame rate camera can ensure real-time monitoring of the position of the welding head 47 and the quality of the weld during the welding process. When welding the joints of railway freight cars, the camera collects high frame rate images every second to record the movement trajectory of the welding head 47 and the process of weld formation.

[0053] The camera uses high-definition resolution to ensure image clarity, clearly showing the shape of the weld and the position of the welding head, thus helping technicians analyze welding quality and promptly discover and correct welding defects;

[0054] Sensor data:

[0055] The sensors are used to collect data at a high frequency; these data include temperature, pressure, current and voltage, which fully and real-time reflect the changes of temperature, current and other parameters during the welding process, and then provide accurate real-time data for the control system to ensure the stability of the welding process;

[0056] The temperature sensor should be installed near the welding area, as close to the weld as possible but not in direct contact with the welding point to avoid high temperature damage to the sensor; use high temperature resistant materials to fix it on the bracket of the welding mechanism to ensure that the sensor can stably monitor the temperature changes during the welding process;

[0057] During the welding process of railway freight car joints, the temperature of the welding area may quickly rise to hundreds of degrees Celsius. Temperature sensors can monitor these changes in time to prevent overheating damage; therefore, temperature sensors need to cover a wide temperature range;

[0058] The pressure sensor should be installed near the clamping mechanism of the welding equipment to measure the pressure generated during the clamping process; the pressure sensor is fixed to the key part of the clamping mechanism through a special bracket to ensure that the sensor can accurately measure the clamping force;

[0059] Clamping and fixing during the welding process may generate high pressure, and these pressures are monitored by pressure sensors to ensure that the workpiece operates within a safe range; therefore, it is necessary to be able to monitor high pressure values;

[0060] The current sensor should be installed on the output line of the welding power supply to monitor the current changes during welding; use a wire clamp or fixing strap to firmly install the current sensor near the power supply output line;

[0061] The current sensor monitors the fluctuation of current during welding, helps optimize welding parameters, and ensures the stability of the welding process and the quality of the weld. Therefore, the current sensor needs to support the measurement of large currents.

[0062] The voltage sensor should be installed between the power input and output terminals of the welding equipment to monitor the voltage changes during welding. The voltage sensor should be fixed to the power line through insulating materials and fixing brackets to ensure safety and stability;

[0063] Accurate voltage monitoring ensures the stability of the welding arc and prevents voltage fluctuations from affecting welding quality; the voltage variation range during welding is relatively small, but it needs to be accurately monitored;

[0064] Device configuration in actual operation:

[0065] The fixed base 1 provides a stable foundation for the entire device, ensuring mechanical stability during the welding process; the stability of the fixed base ensures the normal operation of the welding equipment under high temperature and high pressure environment, avoiding welding errors caused by equipment movement;

[0066] The first vertical plate 2 and the second vertical plate 6 are installed on the top of the fixed base to support the welding mechanism and the sensor; the vertical plates support the welding mechanism and the sensor to ensure the precise positioning of the equipment and the consistency of data collection;

[0067] The lifting mechanism 8 and the lifting cylinder 87 are used to adjust the position of the welding head 47 in real time to adapt to different welding tasks, ensure the relative position between the welding head and the welding workpiece is accurate, thereby improving the welding quality; the lifting mechanism adjusts the position of the welding head according to the real-time sensor data to ensure that the welding head maintains an optimal distance from the workpiece surface, thereby improving the accuracy and quality of welding;

[0068] Step 2: Data Enhancement

[0069] Positive sample pairs are generated by performing various random operations on the same image, such as cropping, rotation, and color jittering, so that the generated images maintain similar features in the feature space;

[0070] Cropping

[0071] The camera supported by the first vertical plate 2 and the second vertical plate 6 installed on the top of the fixed base 1 captures a high-definition image of the welding head 47, randomly crops a part of the image, and generates a small image block with a preset pixel value a×b, where a is a number of pixels in the vertical direction of the image, and b is b number of pixels in the horizontal direction of the image, and the key features of the image are maintained during the cropping process; these fixing devices ensure the stability and consistency of image acquisition, so that the randomly cropped image still maintains the core features of the welding head,

[0072] Perform cropping operation T on image I crop , the cropped image is generated as I crop ; i.e. I crop =T crop (I)

[0073] Where T crop (I) = I[x:x+a, y:y+b], where x and y are the starting coordinates of random cropping. The starting coordinates x and y of random cropping are integers randomly selected within the valid range of the image to ensure that the size of the cropped image block is a×b pixels; a is the number of pixels in the vertical direction of the image, and b is the number of pixels in the horizontal direction of the image.

[0074] Rotation

[0075] In the process of real-time position adjustment of the welding head 47 by the lifting cylinder 87 on the lifting mechanism 8, an angle is randomly selected to rotate the image between [-c°, c°]; c is a preset angle value, and the randomness of the rotation angle simulates the uncertainty in actual operation; this reflects the changes of the welding head at different positions and angles, and enhances the robustness of the model to these changes;

[0076] Perform rotation operation T on image I rotate , generate the rotated image I rotate , that is I rotate =T rotate (I)

[0077] Among them, T rotate (I) = Rotate(I, θ), θ∈(-c°, c°) where θ∈(-c°, c°) means that the rotation angle θ is a randomly selected value in the range [-c°, c°] to ensure that the image is not overly distorted after rotation. c is a pre-set angle value.

[0078] Color dithering

[0079] During the welding process, color jitter simulates different lighting conditions due to changes in ambient light. These changes are common in actual operations. By simulating these conditions, the model can better adapt to the real environment.

[0080] Adjust the brightness, contrast and saturation of the image so that it changes randomly between d and e; where d is the minimum value of the preset brightness, contrast and saturation change of the image, and e is the maximum value of the preset brightness, contrast and saturation change of the image

[0081] Perform color dithering operation T on image I color , generate the color-dithered image I color ;

[0082] I color =T color (I)

[0083] Where T color (I) = Adjust(I, b, c, s), b, c, s∈(d, e) and b, c, s∈(d, e) represent the color jitter parameters b, c and s represent the adjustment range of brightness, contrast and saturation respectively, all randomly selected between [d, e];

[0084] Positive sample pair generation

[0085] Through the above random operation, two different enhancement operations are performed on the same image I to generate two enhanced images I i and I j ; j represents a sample index different from i, i represents a sample index different from j; I i The image after performing one of the above three operations on image I, I j For image I, i After different operations, the enhanced image pairs are generated (I i , I j );

[0086] I i =T i (I),I j =T j (I)

[0087] Among them, T i and T j The random operation is any combination of cropping, rotation or color jittering to ensure that the generated image maintains similar features in the feature space; j represents a sample index different from i, and i represents a sample index different from j;

[0088] Generate negative sample pairs

[0089] During the welding process of railway freight car joints, images at different time points and under different conditions reflect different welding states and changes in welding head position; these images have large differences in feature space, and images I and J collected at different times or under different conditions are selected to form negative sample pairs;

[0090] Select images I and K collected at different times to generate negative sample pairs (I i , I k ), where I i and I k Images from different time points; k represents a sample index different from i, i represents a sample index different from k; the time interval is large enough to ensure that image I i and I k The differences in feature space are significant to avoid content being too similar;

[0091] During the welding process, the changes in sensor data reflect different welding conditions; select images I and K collected under different welding conditions to generate negative sample pairs (I i , I k ), where I i and I k Images from different welding states; welding parameters vary significantly, ensuring that images I i and I k The difference in feature space is significant; k represents a sample index different from i, and i represents a sample index different from k;

[0092] Step 3: Model architecture design

[0093] Neural Architecture Search (NAS) is an automated way to generate and optimize neural network architectures to find the best model for a specific task. In the integrated welding process of railway freight car joint assemblies, NAS helps optimize the control model of the welding head and improve welding quality and efficiency.

[0094] In the integrated forming welding process of the railway freight car joint assembly, the angle and position of the welding head 47 are adjusted through the cooperation of the lifting mechanism 8 and the welding mechanism 4 to ensure the accuracy and stability of the welding; the following is a detailed implementation case and parameter description;

[0095] Define the search space A: including convolutional neural network (CNN), recurrent neural network (RNN), and Transformer architecture; in the welding control task, the main focus is on the convolutional neural network (CNN), and its parameters are set as follows:

[0096] Define the number of convolutional layers in the network, with 3, 5, and 7 layers as candidates; different numbers of convolutional layers extract features of different depths; 3 layers are suitable for shallow feature extraction, while 5 and 7 layers are suitable for deeper feature extraction; the number of convolutional layers L is one of 3, 5, and 7;

[0097] Define the size of the convolution kernel, with 3×3 and 5×5 as candidates; the 3×3 convolution kernel is suitable for capturing detailed features, and the 5×5 convolution kernel is suitable for capturing features in a larger range; the convolution kernel size K is one of 3×3 and 5×5;

[0098] Define the type of pooling layer, with maximum pooling and average pooling as candidates; maximum pooling is used to extract the most significant features in the image, and average pooling is used to smooth the feature map; the pooling layer type P is one of MaxPooling and AvgPooling, MaxPooling is maximum pooling, and AvgPooling is average pooling;

[0099] The formula for maximum pooling is expressed as:

[0100]

[0101] Where: Y represents the output feature map after pooling; X represents the input feature map; i, j represent the coordinates of the output feature map; m, n represent the size of the pooling window; maximum pooling extracts the salient features of the image by taking the maximum value within the pooling window;

[0102] The formula for average pooling is expressed as:

[0103]

[0104] Where: k is the size of the pooling window; It means that the input feature map coordinates start from i and j and increase by m and n respectively, and m and n start from zero and end at k-1;

[0105] Average pooling smoothes the feature map by taking the average value within the pooling window;

[0106] The candidate architecture α is generated by the controller C. Through the reinforcement learning algorithm, the controller generates and optimizes the candidate neural network architecture, which is expressed as α = C (A). This formula means that the controller continuously generates and evaluates different architectures through exploration and utilization strategies, and selects the architecture with the best performance.

[0107] Controller C: This is a generative model that continuously generates and evaluates different architectures through exploration and exploitation strategies;

[0108] Search space A: defines all possible neural network architectures, including various numbers of convolutional layers, convolution kernel sizes, and pooling layer types;

[0109] Candidate architecture α: a specific neural network architecture generated by the controller in the search space;

[0110] Reinforcement Learning Process

[0111] Status S t : Including the number of convolution layers L at time t t , the convolution kernel size K at time t t , pooling layer type P at time t t ; S t =(L t , K t , P t ), S t represents the state at time t;

[0112] Action A t : Select the architecture parameters of the next layer, add a convolutional layer, change the convolution kernel size or pooling layer type; A t represents the action chosen at time t;

[0113] Reward R t : Rewards are given based on the performance evaluation results of the architecture, including the integrity and defect-freeness of the weld, the efficiency of the welding process, the stability of the parameters during the welding process, R t =Evaluate(α t ) where Evaluate(α t ) is the evaluation framework α t Function of performance, R t represents the reward calculated at time t;

[0114] Performance evaluation function Evaluate(α t ) is used to measure the candidate architecture α t Performance on welding tasks; this function comprehensively considers the three indicators of welding quality, welding speed and welding stability; its specific expression is as follows:

[0115] Evaluate(α t)=w q Q(α t )+w s S(α t )+w r R(α t )

[0116] Parameter meaning explanation:

[0117] Q(α t ): welding quality index function, which measures the integrity and defect-free degree of the weld;

[0118] S(α t ): welding speed index function, which measures the efficiency of the welding process;

[0119] R(α t ): welding stability index function, which measures the stability of parameters during welding;

[0120] w q : Weight coefficient of welding quality index, indicating the importance of welding quality in the comprehensive score;

[0121] w s : The weight coefficient of the welding speed index, indicating the importance of welding speed in the comprehensive score;

[0122] w r : The weight coefficient of welding stability index, indicating the importance of welding stability in the comprehensive score;

[0123] Welding quality index function Q(α t );

[0124] Weld quality measures the integrity and defect-freeness of the weld; it is expressed by the following formula:

[0125]

[0126] Among them, y i : The true quality value of the i-th weld; The predicted quality value of the i-th weld; N: the total number of welds;

[0127] Welding speed index function S(α t ): Welding speed measures the efficiency of the welding process; it is expressed by the following formula:

[0128]

[0129] L: total length of the weld;

[0130] T: total time used for welding;

[0131] Welding stability index function R(α t ): Welding stability measures the stability of parameters during welding; it is expressed by the following formula:

[0132]

[0133] Among them, σ j : Standard deviation of the jth welding parameter; μ j : the average value of the jth welding parameter; M: the total number of welding parameters;

[0134] Reinforcement learning controller optimization formula:

[0135] The goal of the controller is to maximize the cumulative reward

[0136] Where: π: strategy, the action selection strategy taken by the controller at each time step t; Indicates the maximum value among all strategies; E: indicates the expected value, which means averaging different possible candidate architectures; It means that the cumulative reward multiplied by the discount factor to the power of t from time t = 0 until t = T; γ: discount factor, used to balance the importance of current rewards and future rewards; γ t : represents the discount factor corresponding to time t; T: total time;

[0137] Use the images and sensor data from the welding process to train and evaluate the generated candidate architecture α; input the welding image data and sensor data collected at a high frequency to minimize the prediction error and improve the weld quality and the stability of the welding process;

[0138] Training Goals

[0139] When training the candidate architecture, the goal is to minimize the prediction error and improve the quality of the weld and the stability of the welding process. This process is achieved by minimizing a loss function L(θ, D). The following is the specific expression of the loss function and the meaning of the related parameters.

[0140] Formula

[0141] θ: model parameters, weights and biases of the neural network;

[0142] D: training data, including image data and sensor data during welding;

[0143] Indicates that the smallest value is selected among all model parameters.

[0144] L(θ, D): loss function, used to measure the difference between the model's predicted value and the actual value;

[0145] Specific expression of loss function

[0146] The loss function L(θ, D) includes multiple parts to comprehensively consider the weld quality, welding speed and stability of the welding process; the following is the specific expression of the loss function:

[0147] L(θ, D) = L q (θ,D)+λ s L s (θ,D)+λ r L r (θ, D)

[0148] Where: L q (θ, D): weld quality loss; L s (θ, D): welding speed loss; L r (θ, D): welding stability loss; λ s : Weight coefficient of welding speed loss; λ r : Weight coefficient of welding stability loss;

[0149] Weld quality loss L q (θ, D)

[0150] The weld quality loss measures the difference between the weld quality predicted by the model and the true quality; it is expressed using the mean square error (MSE):

[0151]

[0152] y i : The true quality value of the i-th weld; The predicted quality value of the ith weld; It means that the number of welds is accumulated from i=1 to i=N; N: the total number of welds;

[0153] Welding speed loss L s (θ, D)

[0154] The welding speed loss measures the difference between the welding speed predicted by the model and the desired speed; it is expressed as mean square error (MSE):

[0155]

[0156] t i : The actual welding time of the ith weld; Predicted welding time for the ith weld;

[0157] Welding stability loss L r (θ, D)

[0158] Welding stability loss measures the fluctuation of parameters during welding; it is expressed as the mean square error (MSE) of the parameter standard deviation:

[0159]

[0160] σ j : The true standard deviation of the jth welding parameter; The predicted standard deviation of the jth welding parameter; M: the total number of welding parameters;

[0161] Performance evaluation indicators:

[0162] Weld quality (precision): measures the integrity and defect-freeness of the weld;

[0163] Welding speed (time): the efficiency of the welding process;

[0164] Welding stability (stability coefficient): the stability of parameters during welding;

[0165] Step 4: Self-supervised contrastive learning

[0166] Self-supervised contrastive learning is an unsupervised learning method that enhances the feature extraction capability of the model by constructing positive and negative sample pairs. In the integrated forming welding process of railway freight car joint assembly, self-supervised contrastive learning is used to enhance the feature extraction capability of the welding head control model, improve the weld quality and the stability of the welding process. The following is the application process

[0167] Construct positive sample pairs and negative sample pairs:

[0168] According to step 2, two different enhancement operations T are performed on the same image I. i and T j , generate positive sample pairs (I i , I j )

[0169] Select different image pairs from images at different times or under different conditions (I i , I k ) as a negative sample pair;

[0170] Feature extraction:

[0171] Use a k-layer convolutional neural network to extract image features. The convolution kernel size is k×k, and the number of channels is 64, 128, and 256 respectively. Generate the feature vector h of image I at time i and time j. i and h j ;

[0172] Calculate similarity:

[0173] Use the cosine similarity function to calculate the similarity between feature vectors:

[0174] j represents a sample index different from i;

[0175] i represents a sample index different from j;

[0176] h i and h j : Feature representation of positive sample pairs, representing feature vectors obtained from the same image using different enhancement methods;

[0177] sim(h i ,h j 〕: Similarity function, usually using cosine similarity

[0178] h i ·h j Indicates h i and h j The dot product of

[0179] ||h i || and ||h j || means h i and h j The L2 norm of ;

[0180] Calculate contrast loss:

[0181] Using the contrastive loss function L defined contrastive Perform training, optimize feature extraction model, and define

[0182] Contrastive loss function L contrastive The contrast loss function is used to enhance the feature extraction ability of the model by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs for training;

[0183] The formula means:

[0184]

[0185] Parameter Description

[0186] τ: temperature parameter, used to adjust the smoothness of the similarity distribution;

[0187] L contrastive : Contrastive loss function, used to measure the performance of the feature extraction model, trained by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs;

[0188] h i and h j : Feature representation of positive sample pairs, feature vectors obtained from the same image through different data enhancement methods; h i and h jThe feature vector represents the same image after being rotated, cropped, and color-dithered in different enhancement methods;

[0189] sim(h i ,h j 〕: eigenvector h i and h j The similarity between

[0190] The denominator of the negative sample pairs performs exponential operations and sums the similarities of all negative sample pairs;

[0191] 1 [k≠i] : indicator function, which takes the value of 1 when k≠i and 0 otherwise, and is used to exclude the self-comparison of positive sample pairs;

[0192] j represents a sample index different from i;

[0193] i represents a sample index different from j;

[0194] The exponential operation result of the similarity of the positive sample pair enhances the contribution of the similarity of the positive sample pair in the contrast loss;

[0195] -log: Take the logarithm of the entire ratio and add a negative sign. The logarithmic transformation makes the optimization of the loss function more stable and combines the maximization of the positive sample similarity with the minimization of the negative sample similarity.

[0196] Step 5: Meta-learning optimization

[0197] Meta-learning optimization is a technique that allows models to adapt quickly on multiple tasks by training them. In welding control tasks, meta-learning helps models adapt quickly under different welding conditions, improving welding quality and efficiency. The following is a detailed implementation case and parameter description.

[0198] Define the meta-learning task T

[0199] The meta-learning tasks include a variety of different welding tasks, welding of different materials, different welding speeds, and different welding environments. By training on multiple tasks, the model learns how to quickly adapt to new welding tasks.

[0200] Meta-learning algorithm optimizes model parameters

[0201] The meta-learning algorithm optimizes multiple tasks to enable the model parameters to quickly adapt to new tasks. In welding control tasks, the meta-learning algorithm is used to allow the model to quickly adjust under different welding conditions to improve the welding effect.

[0202] The objective formula for optimizing model parameters is:

[0203]

[0204] in:

[0205] θ * : Optimal model parameters; these are the parameters obtained through the optimization process, which make the model perform best on multiple tasks;

[0206] Find the parameter θ that minimizes the objective function; this is an optimization operation, which means finding the parameter θ that minimizes the subsequent sum;

[0207] θ: model parameters; represents the parameters that need to be trained, such as the weights and biases of the neural network;

[0208] For all tasks T sampled from the task distribution p(T) i Sum up; this symbol indicates the accumulation of losses for multiple tasks;

[0209] T i : The i-th task; the specific task sampled from the task distribution;

[0210] p(T): task distribution; represents the probability distribution of task T, used to sample tasks from it;

[0211] In Task T i On, model f θ The loss function represents the error or performance index of the model on a specific task.

[0212] L: loss function, which measures the difference between the model output and the true value;

[0213] f θ : model function with parameter θ;

[0214] Perform model training in the welding control task and use the data collected in step 1 to perform meta-learning model training:

[0215] Initialize model parameters θ;

[0216] Task sampling: sampling a set of tasks from the task distribution p(T);

[0217] Task training: For each task T i , use its data to train the model and calculate the loss

[0218] Parameter update: Update the model parameters based on the sum of the losses of all tasks:

[0219]

[0220] in,

[0221] θ: model parameters; represents the weights and biases of the neural network;

[0222] ←: Assignment operator, which means updating the value on the right to the variable on the left;

[0223] α: learning rate; this is a hyperparameter that controls the step size of each update;

[0224] The gradient operator represents the gradient of the model parameter θ; the gradient points to the direction of increasing loss function, and updating parameters in the opposite direction reduces loss;

[0225] Numerical case study

[0226] Data Collection:

[0227] Assume that during a welding task, the following data is collected:

[0228] Image Data:

[0229] Collection time: 10 minutes

[0230] Frames per second: 30

[0231] Total number of frames: 10 minutes × 60 seconds / minute × 30 frames / second = 18,000 frames

[0232] Image resolution: 1920×1080 pixels

[0233] Sensor data:

[0234] Collection time: 10 minutes

[0235] Sampling frequency per second: 100 times

[0236] Total sampling number: 10 minutes × 60 seconds / minute × 100 times / second = 60,000 times

[0237] Data type: temperature (0-1000℃), pressure (0-10MPa), current (0-500A), voltage (0-100V)

[0238] Data Augmentation

[0239] Positive sample pair generation:

[0240] Two enhanced versions are generated for each frame

[0241] Total positive sample pairs: 18000 frames × 2 = 36000 pairs

[0242] Negative sample pair generation:

[0243] Select different image pairs from different time points and welding states

[0244] Total negative sample pairs: 18,000 pairs

[0245] Model architecture design (NAS)

[0246] Number of convolutional layers: Candidate values ​​are 3, 5, and 7

[0247] Convolution kernel size: Candidate values ​​are 3×3 and 5×5

[0248] Pooling layer types: max pooling and average pooling

[0249] In the search space, multiple candidate architectures are generated and evaluated through NAS, and the optimal architecture is finally selected as:

[0250] Number of convolutional layers: 5

[0251] Convolution kernel size: 3×3

[0252] Pooling layer type: max pooling

[0253] Self-supervised contrastive learning

[0254] Contrastive loss function formula parameters:

[0255] τ=0.07(temperature parameter)

[0256] Feature extraction network:

[0257] Convolutional layer: 5 layers

[0258] Convolution kernel size: 3×3

[0259] Number of channels: 64, 128, 256, 512, 1024

[0260] Meta-learning optimization

[0261] Optimized for different materials and welding speed tasks:

[0262] Welding tasks for different materials:

[0263] Steel: Current 300A, Voltage 30V, Welding Speed ​​5mm / s

[0264] Aluminum: Current 200A, Voltage 25V, Welding Speed ​​4mm / s

[0265] Welding tasks at different welding speeds:

[0266] Slow welding: welding speed 3mm / s, temperature 400℃

[0267] Fast welding: welding speed 8mm / s, temperature 600℃

[0268] Meta-learning optimization formula parameters:

[0269] Using the MAML algorithm

[0270] Learning rate: 0.001

[0271] Number of tasks: 10

[0272] Training and application of welding control model

[0273] Train the model using the optimal architecture:

[0274] Training data: including image data and sensor data

[0275] Training goal: Minimize prediction error, improve weld quality and welding process stability

[0276] Application in actual welding process:

[0277] Model input:

[0278] Image data: 224×224×3

[0279] Sensor data: 8 dimensions

[0280] Model output:

[0281] Welding head position (x, y, z coordinates)

[0282] Welding speed (mm / s)

[0283] In the present invention, the product is placed on the connecting platform 9, and the stepper motor 74 is started after being powered on. The stepper motor 74 drives the screw rod 72 to rotate. The screw rod 72 has two threads in opposite directions with its own center line as the dividing line. The two movable blocks 75 match the threads on the surface of the screw rod 72. The two movable blocks 75 are limited by the connecting platform 9, so the two movable blocks 75 slide relatively along the screw rod 72. The movable blocks 75 drive the clamping plates 73 to move synchronously. The two clamping plates 73 clamp and fix the product from both sides. The lifting cylinder 87 performs telescopic movement. The lifting cylinder 87 drives the lifting seat 89 to move synchronously. The lifting seat 89 drives the second direction plate 86 to move along the direction seat 88. The first direction plate 82 is deflected at an angle along the positioning seat 81, thereby increasing the stability of the distance adjustment between the connecting platform 9 and the positioning plate 83, and indirectly adjusting the height of the product placed on the connecting platform 9. The first angle cylinder 49 performs telescopic and deflection movements, and the first angle cylinder 49 pushes the first welding arm 43 to deflect at an angle along the sliding plate 42, and the second angle cylinder 48 performs telescopic and deflection movements, and the second angle cylinder 48 pushes the second welding arm 44 to deflect at an angle along the first welding arm 43, thereby adjusting the position of the welding head 47, and the welding head 47 performs welding processing on the product.

[0284] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An integrated forming welding device for a railway freight car joint assembly, comprising a fixed base (1), characterized in that: The top of the fixed base (1) is provided with a top plate (5), a first vertical plate (2) is fixedly mounted on one side of the top of the fixed base (1), a second vertical plate (6) is fixedly mounted on the other side of the top of the fixed base (1), a lifting mechanism (8) is fixedly mounted on the top of the fixed base (1), a connecting platform (9) is installed on the top of the lifting mechanism (8), a clamping mechanism (7) is fixedly mounted on the top of the connecting platform (9), a welding mechanism (4) is fixedly mounted on the top of one side of the first vertical plate (2) and a control mechanism (7) is fixedly mounted on the middle of one side of the first vertical plate (2). The panel (3) is provided with a welding mechanism (4) comprising a pushing cylinder (41) and a sliding plate (42), wherein the movable end of the pushing cylinder (41) is fixedly connected to one side of the sliding plate (42), the clamping mechanism (7) comprises two limit plates (71) and a screw rod (72), wherein the two ends of the screw rod (72) are respectively rotatably connected to the opposite sides of the two limit plates (71), and the lifting mechanism (8) comprises a positioning plate (83) and a lifting cylinder (87), wherein the middle part of the top end of the positioning plate (83) is fixedly connected to the fixed end of the lifting cylinder (87), and positioning seats (81) are provided on both sides of the lifting cylinder (87).

2. The integrated forming welding device for a railway freight car joint assembly according to claim 1 is characterized in that: The bottom end of the slide plate (42) is rotatably connected to a first welding arm (43), one end of the first welding arm (43) is rotatably connected to a second welding arm (44), the bottom end of the second welding arm (44) is rotatably connected to a length rod (45), a welding platform (46) is fixedly installed at the bottom end of the length rod (45), and a welding head (47) is fixedly installed at the bottom end of the welding platform (46).

3. The integrated forming welding device for a joint assembly for a railway freight car according to claim 2, characterized in that: The middle part of one side of the slide plate (42) is rotatably connected to a first angle cylinder (49), the movable end of the first angle cylinder (49) is rotatably connected to a side opposite to the first welding arm (43), the middle part of the first welding arm (43) is rotatably connected to a second angle cylinder (48), the movable end of the second angle cylinder (48) is rotatably connected to a side opposite to the second welding arm (44), the middle parts of the two pushing cylinders (41) are fixedly connected to the first vertical plate (2) and the second vertical plate (6) respectively, and the two slide plates (42) are fixedly connected to the top plate (5).

4. The integrated forming welding device for a railway freight car joint assembly according to claim 1, characterized in that: The surface of the screw rod (72) is threadedly connected to two movable blocks (75), and the two movable blocks (75) are slidably connected to the connecting platform (9), and the top ends of the two movable blocks (75) are fixedly mounted with clamping plates (73).

5. The integrated forming welding device for a railway freight car joint assembly according to claim 1, characterized in that: A stepping motor (74) for driving the screw rod (72) to rotate is fixedly mounted on one side of one of the limit plates (71), and the bottom ends of the two limit plates (71) are fixedly connected to the connecting platform (9).

6. The integrated forming welding device for a railway freight car joint assembly according to claim 1, characterized in that: Displacement grooves (84) are provided on both sides of the top of the positioning plate (83), and displacement blocks (85) are slidably connected inside the two displacement grooves (84). A lifting seat (89) is fixedly installed on the surface of the lifting cylinder (87), and a direction seat (88) is fixedly installed on the top of the two displacement blocks (85). One side of the two direction seats (88) is rotatably connected to a first direction plate (82), and one end of the two first direction plates (82) away from the direction seat (88) is rotatably connected to the inside of the two positioning seats (81), and the other side of the two direction seats (88) is rotatably connected to a second direction plate (86), and one end of the two second direction plates (86) away from the direction seat (88) is rotatably connected to the two sides of the lifting seat (89).

7. The integrated forming welding device for a joint assembly for a railway freight car according to claim 1, characterized in that: The top ends of the two positioning seats (81) and the movable end of the lifting cylinder (87) are fixedly connected to the side directly opposite to the connecting platform (9), the bottom end of the positioning plate (83) is fixedly connected to the fixed base (1), and the top end of the first vertical plate (2) and the top end of the second vertical plate (6) are respectively fixedly connected to the two sides of the bottom end of the top plate (5).

8. The integrated forming welding device for a joint assembly for a railway freight car according to claim 1, characterized in that: The control panel (3) comprises a power interface circuit (301), a rectifier filter circuit (302), an inverter circuit (303), a voltage conversion module (304), a welding circuit (305), a voltage and current feedback module (306), a control module (307), a dimming mask drive circuit (308), an automatic dimming mask circuit (309) and a PWM control (310). The rectifier filter circuit (302) and the power interface circuit (301), the inverter circuit (303) The device is connected to a rectifier filter circuit (302), the transformer module (304) is connected to an inverter circuit (303), the welding circuit (305) and the voltage and current feedback module (306) are both connected to the transformer module (304), the automatic dimming mask circuit (309) is connected to a dimming mask drive circuit (308), and the PWM control (310), the dimming mask drive circuit (308) and the voltage and current feedback module (306) are all connected to a control module (307).

9. The welding process of the integrated forming welding device for the joint assembly for railway freight cars according to any one of claims 1 to 8, characterized in that: The following steps are involved: Step 1: Product fixing: The product is placed on the connecting table (9), and two clamping plates (73) are used to clamp and fix the product from both sides; Step 2: product lifting: the lifting cylinder (87) on the lifting mechanism (8) performs telescopic movement to adjust the height of the product placed on the connecting platform (9); Step 3: Product welding: The welding head (47) on the welding mechanism (4) performs welding processing on the lifted product.

10. The welding process of the integrated forming welding device for the joint assembly for railway freight cars according to claim 9, characterized in that: Step 3: When welding products, an algorithm based on self-supervised contrastive learning combined with meta-learning is introduced to optimize the process of the integrated forming welding device for the joint assembly of railway freight cars. The algorithm can improve the welding quality in actual operation and realize intelligent welding control. The specific process is as follows: Step 1: Data collection and preprocessing Camera data The main camera is installed above or on the side of the welding head to fully monitor the welding area and capture all details of the welding process; the camera is fixed to the frame of the welding device using an adjustable bracket to ensure that the camera viewing angle can be adjusted as needed; The auxiliary camera is installed on the opposite side of the welding head or other key parts to provide monitoring at different angles. The camera is installed in an appropriate position through an auxiliary bracket to ensure that the blind spot of the main camera is supplemented. The use of high frame rate cameras can ensure real-time monitoring of the position of the welding head and the quality of the weld during the welding process. When welding railway freight car joints, the camera captures high frame rate images every second to record the movement trajectory of the welding head and the weld formation process. The camera uses high-definition resolution to ensure image clarity, clearly showing the shape of the weld and the position of the welding head, thus helping technicians analyze welding quality and promptly discover and correct welding defects; Sensor data: The sensors are used to collect data at a high frequency; these data include temperature, pressure, current and voltage, which fully and real-time reflect the changes of temperature and current parameters during the welding process, and then provide accurate real-time data for the control system to ensure the stability of the welding process; Device configuration in actual operation: The first vertical plate and the second vertical plate are installed on the top of the fixed base to support the welding mechanism and the sensor; the vertical plate supports the welding mechanism and the sensor to ensure the precise positioning of the equipment and the consistency of data collection; The lifting mechanism and lifting cylinder are used to adjust the position of the welding head in real time to adapt to different welding tasks, ensure the relative position between the welding head and the welding workpiece is accurate, and thus improve the welding quality; the lifting mechanism adjusts the position of the welding head according to the real-time sensor data to ensure that the welding head maintains the best distance from the workpiece surface; Step 2: Data Augmentation Positive sample pairs are generated by performing various random operations on the same image, such as cropping, rotation, and color jittering, so that the generated images maintain similar features in the feature space; Cropping: The camera supported by the first and second vertical plates installed on the top of the fixed base captures the high-definition image of the welding head, randomly crops a part of the image, and generates a small image block with a preset pixel value of a×b, where a is a number of pixels in the vertical direction of the image and b is b number of pixels in the horizontal direction of the image. The key features of the image are maintained during the cropping process. These fixtures ensure the stability and consistency of image acquisition, so that the randomly cropped image still maintains the core features of the welding head. Perform cropping operation T on image I crop , the cropped image is generated as I crop ; i.e. I crop =T crop (I); where T crop (I) = I[x:x+a, y:y+b], where x and y are the starting coordinates of random cropping. The starting coordinates x and y of random cropping are integers randomly selected within the valid range of the image to ensure that the size of the cropped image block is a×b pixels; a means that the image has a pixels in the vertical direction, and b means that the image has b pixels in the horizontal direction; Rotation: When the position of the welding head is adjusted in real time by the lifting cylinder on the lifting mechanism, a random angle is selected to rotate the image between [-c°, c°]; c is a preset angle value, and the randomness of the rotation angle simulates the uncertainty in actual operation; this reflects the changes of the welding head at different positions and angles, and enhances the robustness of the model to these changes; Perform rotation operation T on image I rotate , generate the rotated image I rotate , that is I rotate =T rotate (I); where T rotate (I) = Rotate(I, θ), θ(-c°, c°), where θ(-c°, c°) represents the rotation angle θ, which is a value randomly selected in the range of [-c°, c°] to ensure that the image will not be overly distorted after rotation, and c is a preset angle value; Color jitter: During the welding process, due to the changes in ambient light, color jitter simulates different lighting conditions. These changes are common in actual operations. By simulating these conditions, the model can better adapt to the real environment; Adjust the brightness, contrast and saturation of the image so that it changes randomly between d and e; where d is the preset minimum value of the brightness, contrast and saturation change of the image, and e is the preset maximum value of the brightness, contrast and saturation change of the image; Perform color dithering operation T on image I color , generate the color-dithered image I color ; I color =T color (I) Where T color (I) = Adjust(I, b, c, s), b, c, s∈(d, e) and b, c, s∈(d, e) represent the color jitter parameters b, c and s represent the adjustment range of brightness, contrast and saturation respectively, all randomly selected between [d, e]; Positive sample pair generation: Through the above random operations, two different enhancement operations are performed on the same image I to generate two enhanced images I i and I j ; j represents a sample index different from i, i represents a sample index different from j; I i The image after performing one of the above three operations on image I, I j For image I, i After different operations, the enhanced image pairs are generated (I i , I j ); I i =T i (I),I j =T j (I) Among them, T i and T j The random operation is any combination of cropping, rotation or color jittering to ensure that the generated image maintains similar features in the feature space; j represents a sample index different from i, and i represents a sample index different from j; Generate negative sample pairs: During the welding process of railway freight car joints, images at different time points and under different conditions reflect different welding states and changes in welding head position; these images have large differences in feature space, and images I and J collected at different times or under different conditions are selected to form negative sample pairs Select images I and K collected at different times to generate negative sample pairs (I i , I k ), where I i and I k Images from different time points; k represents a sample index different from i, i represents a sample index different from k; the time interval is large enough to ensure that image I i and I k The differences in feature space are significant to avoid content being too similar; During the welding process, the changes in sensor data reflect different welding conditions; select images I and K collected under different welding conditions to generate negative sample pairs (I i , I k ), where I i and I k Images from different welding states; welding parameters vary significantly, ensuring that images I i and I k The difference in feature space is significant; k represents a sample index different from i, and i represents a sample index different from k; Step 3: Model architecture design Neural Architecture Search (NAS) generates and optimizes neural network architectures in an automated way to find the model with the best performance on a specific task. In the integrated forming welding process of railway freight car joint assemblies, NAS helps optimize the control model of the welding head and improve welding quality and efficiency. In the integrated forming welding process of railway freight car joint assembly, the angle and position of the welding head are adjusted through the cooperation of the lifting mechanism and the welding mechanism to ensure the accuracy and stability of welding; the following is a detailed implementation case and parameter description; Define the search space A: including convolutional neural network CNN, recurrent neural network RNN, and Transformer architecture; in the welding control task, convolutional neural network CNN is selected, and its parameters are set as follows: Define the number of convolutional layers in the network, with 3, 5, and 7 layers as candidates; different numbers of convolutional layers extract features of different depths; 3 layers are suitable for shallow feature extraction, while 5 and 7 layers are suitable for deeper feature extraction; the number of convolutional layers L is one of 3, 5, and 7; Define the size of the convolution kernel, with 3×3 and 5×5 as candidates; the 3×3 convolution kernel is suitable for capturing detailed features, and the 5×5 convolution kernel is suitable for capturing features in a larger range; the convolution kernel size K is one of 3×3 and 5×5; Define the type of pooling layer, with maximum pooling and average pooling as candidates; maximum pooling is used to extract the most significant features in the image, and average pooling is used to smooth the feature map; the pooling layer type P is one of MaxPooling and AvgPooling, MaxPooling is maximum pooling, and AvgPooling is average pooling; The formula for maximum pooling is expressed as: Where: Y represents the output feature map after pooling; X represents the input feature map; i, j represent the coordinates of the output feature map; m, n represent the size of the pooling window; maximum pooling extracts the salient features of the image by taking the maximum value within the pooling window; The formula for average pooling is expressed as: Where: k is the size of the pooling window; It means that the input feature map coordinates start from i and j and increase by m and n respectively, and m and n start from zero and end at k-1; Average pooling smoothes the feature map by taking the average value within the pooling window; The candidate architecture α is generated by the controller C. Through the reinforcement learning algorithm, the controller generates and optimizes the candidate neural network architecture, which is expressed as α = C (A). This formula means that the controller continuously generates and evaluates different architectures through exploration and utilization strategies, and selects the architecture with the best performance. Controller C: This is a generative model that continuously generates and evaluates different architectures through exploration and exploitation strategies; Search space A: defines all possible neural network architectures, including various numbers of convolutional layers, convolution kernel sizes, and pooling layer types; Candidate architecture α: a specific neural network architecture generated by the controller in the search space; Reinforcement Learning Process Status S t : Including the number of convolution layers L at time t t , the convolution kernel size K at time t t , pooling layer type P at time t t ; S t =(L t , K t , P t ), S t represents the state at time t; Action A t : Select the architecture parameters of the next layer, add a convolutional layer, change the convolution kernel size or pooling layer type; A t represents the action chosen at time t; Reward R t : Rewards are given based on the performance evaluation results of the architecture, including the integrity and defect-freeness of the weld, the efficiency of the welding process, the stability of the parameters during the welding process, R t =Evaluate(α t ) where Evaluate(α t ) is the evaluation framework α t Function of performance, R t represents the reward calculated at time t; Performance evaluation function Evaluate(α t ) is used to measure the candidate architecture α t Performance on welding tasks; this function comprehensively considers the three indicators of welding quality, welding speed and welding stability; its specific expression is as follows: Evaluate(a t )=w q Q(a t )+w s S(a t )+w r R(a t ) Parameter meaning explanation: Q(α t ): welding quality index function, which measures the integrity and defect-free degree of the weld; S(α t ): welding speed index function, which measures the efficiency of the welding process; R(α r ): welding stability index function, which measures the stability of parameters during welding; w q : Weight coefficient of welding quality index, indicating the importance of welding quality in the comprehensive score; w s : The weight coefficient of the welding speed index, indicating the importance of welding speed in the comprehensive score; w r : The weight coefficient of welding stability index, indicating the importance of welding stability in the comprehensive score; Welding quality index function Q(α t ); Weld quality measures the integrity and defect-freeness of the weld; it is expressed by the following formula: Among them, y i : The true quality value of the i-th weld; The predicted quality value of the i-th weld; N: the total number of welds; Welding speed index function S(α t ): Welding speed measures the efficiency of the welding process; it is expressed by the following formula: L: total length of the weld; T: total time used for welding; Welding stability index function R(α t ): Welding stability measures the stability of parameters during welding; it is expressed by the following formula: Among them, σ j : Standard deviation of the jth welding parameter; μ j : the average value of the jth welding parameter; M: the total number of welding parameters; Reinforcement learning controller optimization formula: The goal of the controller is to maximize the cumulative reward Where: π: strategy, the action selection strategy taken by the controller at each time step t; Indicates the maximum value among all strategies; E: indicates the expected value, which means averaging different possible candidate architectures; It means that the cumulative reward multiplied by the discount factor to the power of t from time t = 0 until t = T; γ: discount factor, used to balance the importance of current rewards and future rewards; γ t : represents the discount factor corresponding to time t; T: total time; Use the images and sensor data from the welding process to train and evaluate the generated candidate architecture α; input the welding image data and sensor data collected at a high frequency to minimize the prediction error and improve the weld quality and the stability of the welding process; Training goal: When training the candidate architecture, the goal is to minimize the prediction error and improve the quality of the weld and the stability of the welding process. This process is achieved by minimizing a loss function L(θ, D). The following is the specific expression of the loss function and the meaning of the related parameters. Formula θ: model parameters, weights and biases of the neural network; D: training data, including image data and sensor data during welding; Indicates selecting the smallest one among all model parameters; L(θ, D): loss function, used to measure the difference between the model's predicted value and the actual value; The specific expression of the loss function is: The loss function L(θ, D) includes multiple parts to comprehensively consider the weld quality, welding speed and stability of the welding process; the following is the specific expression of the loss function: LλLλL L(θ, D)=L q (θ, D)+λ s L s (θ, D)+λ r L r (θ, D) Where: L q (θ, D): weld quality loss; L s (θ, D): welding speed loss; L r (θ, D): welding stability loss; λ s : Weight coefficient of welding speed loss; λ r : Weight coefficient of welding stability loss; Weld quality loss L q (θ, D): The weld quality loss measures the difference between the weld quality predicted by the model and the actual quality; it is expressed using the mean square error (MSE): y i : The true quality value of the i-th weld; The predicted quality value of the ith weld; It means that the number of welds is accumulated from i=1 to i=N; N: the total number of welds; Welding speed loss L s (θ, D): The welding speed loss measures the difference between the welding speed predicted by the model and the expected speed; it is expressed as mean square error (MSE): t i : The actual welding time of the ith weld; Predicted welding time of the ith weld; Welding stability loss L r (θ, D): Welding stability loss measures the fluctuation of parameters during welding; it is expressed as the mean square error (MSE) of the parameter standard deviation: σ j : The true standard deviation of the jth welding parameter; The predicted standard deviation of the jth welding parameter; M: the total number of welding parameters; Performance evaluation indicators: Weld quality: a measure of the integrity and defect-freeness of the weld; Welding speed: the efficiency of the welding process; Welding stability: the stability of parameters during welding; Step 4: Self-supervised contrastive learning Self-supervised contrastive learning is an unsupervised learning method that enhances the feature extraction capability of the model by constructing positive and negative sample pairs. In the integrated forming welding process of railway freight car joint assembly, self-supervised contrastive learning is used to enhance the feature extraction capability of the welding head control model, improve the weld quality and the stability of the welding process. The following is the application process: Construct positive sample pairs and negative sample pairs: According to step 2, two different enhancement operations T are performed on the same image I. i and T j , generate positive sample pairs (I i , I j ); Select different image pairs from images at different times or under different conditions (I i , I k ) as a negative sample pair; Feature extraction: Use a k-layer convolutional neural network to extract image features. The convolution kernel size is k×k, and the number of channels is 64, 128, and 256 respectively. Generate the feature vector h of image I at time i and time j. i and h j ; Calculate similarity: Use the cosine similarity function to calculate the similarity between feature vectors: j represents a sample index different from i; i represents a sample index different from j; h i and h j : Feature representation of positive sample pairs, representing the feature vectors obtained from the same image through different enhancement methods; sim(h i ,h j ): Similarity function, usually cosine similarity; h i ·h j Indicates h i and h j The dot product of ||h i || and ||h j || means h i and h j The L2 norm of ; Calculate contrast loss: Using the contrastive loss function L defined contrastive Perform training, optimize the feature extraction model, and define the contrast loss function L contrastive The contrast loss function is used to enhance the feature extraction ability of the model by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs for training; The formula means: Parameter Description: τ: temperature parameter, used to adjust the smoothness of the similarity distribution; L contrastive : Contrastive loss function, used to measure the performance of the feature extraction model, trained by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs; h i and h j : Feature representation of positive sample pairs, feature vectors obtained from the same image through different data enhancement methods; h i and h j It represents the feature vector of the same image after being rotated, cropped, and color-dithered in different enhancement methods; sim(h i ,h j ): eigenvector h i and h j The similarity between The denominator of the negative sample pair, performs exponential operation and sums the similarities of all negative sample pairs; 1 [k≠i] : indicator function, which takes the value of 1 when k≠i and 0 otherwise, and is used to exclude the self-comparison of positive sample pairs; j represents a sample index different from i; i represents a sample index different from j; Step 5: Meta-learning optimization Meta-learning optimization is a technique that allows models to adapt quickly on multiple tasks by training them. In welding control tasks, meta-learning helps models adapt quickly under different welding conditions, improving welding quality and efficiency. The following is a detailed implementation case and parameter description. Define the meta-learning task T The meta-learning tasks include a variety of different welding tasks, welding of different materials, different welding speeds, and different welding environments. By training on multiple tasks, the model learns how to quickly adapt to new welding tasks. Meta-learning algorithm optimizes model parameters The meta-learning algorithm optimizes multiple tasks to enable the model parameters to quickly adapt to new tasks. In welding control tasks, the meta-learning algorithm is used to allow the model to quickly adjust under different welding conditions to improve the welding effect. The objective formula for optimizing model parameters is: in: θ * : Optimal model parameters; these are the parameters obtained through the optimization process, which make the model perform best on multiple tasks; Find the parameter θ that minimizes the objective function; this is an optimization operation, which means finding the parameter θ that minimizes the following sum; θ: model parameter; represents the parameters that need to be trained, such as the weights and biases of the neural network; For all tasks T sampled from the task distribution p(T) i Sum; This symbol indicates the accumulation of losses for multiple tasks; T i : the i-th task; the specific task sampled from the task distribution; p(T): task distribution; represents the probability distribution of task T, used to sample tasks from it; In Task T i On, model f θ The loss function of the model; represents the error or performance index of the model on a specific task; L: loss function, which measures the difference between the model output and the true value; f θ : model function with parameter θ; Perform model training in the welding control task and use the data collected in step 1 to perform meta-learning model training: Initialize model parameters θ; Task sampling: sampling a set of tasks from the task distribution p(T); Task training: For each task T i , use its data to train the model and calculate the loss Parameter update: Update the model parameters based on the sum of the losses of all tasks: in, θ: model parameters; represents the weights and biases of the neural network; ←: Assignment operator, which means updating the value on the right to the variable on the left; α: learning rate; this is a hyperparameter that controls the step size of each update; The gradient operator represents the gradient of the model parameter θ; the gradient points in the direction of increasing loss function, and updating parameters in the opposite direction reduces loss.

Images