Rail electroslag welding shape control method and system coupled with magnetic field and temperature field

Abstract

The present application belongs to the field of welding forming, and particularly discloses a steel rail electric-shield welding shape control method and system coupled with magnetic field and temperature field, which comprises: the steel rails to be welded have a welding gap between them, and the welding gap can accommodate three electric-shield welding nozzles; two water-cooled copper molds are respectively arranged on the two sides of the steel rails to be welded, and a magnetic field stirring device is installed simultaneously; the steel rails to be welded are welded, and for any moment of the steel rails: a molten pool temperature field picture within a first time threshold before the current moment is acquired; welding parameters within a second time threshold before the current moment are acquired; a heat accumulation feature of the current moment is determined based on the acquired welding parameters and molten pool temperature field picture; the heat accumulation feature is input into a trained discrimination model to determine real-time parameter values of the water-cooled copper mold and the magnetic field stirring device, thereby realizing molten pool control at the current moment. The present application adopts a real-time joint control mode of temperature field and magnetic field stirring, and couples heat accumulation information to improve the quality of the welded joint.

Description

Technical Field

[0001] This invention belongs to the field of welding forming, and more specifically, relates to a method and system for controlling the weldability of rail electroslag welds using magnetic field-temperature field coupling. Background Technology

[0002] To reduce the impact of trains on the tracks, welding is used to construct seamless railway lines, making welded joints a crucial component of high-speed seamless tracks. To ensure train safety, the construction of high-speed, heavy-haul, and complex geological seamless tracks places extremely high demands on the performance of welded joints: the tensile strength of the rail welded joints must be no less than 980 MPa, and the microstructure must be free of martensite and bainite. To meet or exceed these performance indicators, the grain size of the rail welded joints must be less than 22 μm, the microstructure must consist of more than 98% pearlite and less than 2% ferrite, and the interlamellar spacing of the pearlite fine structure must be less than 0.1 μm.

[0003] Existing rail welding methods include aluminothermic welding, flash welding, gas pressure welding, and electric arc welding. Aluminothermic welding is mainly used for rail turnout welding; it is simple to operate, but slag inclusions are common in the weld, and the joint grains are coarse, making it prone to martensite formation. Flash welding and gas pressure welding are mainly used for rail production in workshops and on-site welding; their weld grains are fine, but the heat-affected zone has larger grains, sometimes exhibiting martensite and bainite. Electric arc welding is mainly used for rail repair; its heat-affected zone has coarse grains, making it highly susceptible to martensite formation and microcracks. Because existing rail welding methods easily produce martensite and bainite in the microstructure of the weld joints, with coarse grains, and because the weld joint formation is difficult to control, it is difficult to obtain high-quality rail weld joints.

[0004] Therefore, there is an urgent need for a method for welding and shape control of rails to further improve the quality of rail welded joints. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method and system for controlling the shape properties of rail electroslag welds using magnetic field-temperature field coupling. The purpose is to achieve real-time control of the shape properties of rail electroslag welds and improve the quality of rail welded joints.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for controlling the shape properties of electroslag welded rails using a coupled magnetic field and temperature field is proposed, comprising the following steps:

[0007] The steel rails to be welded are placed in the welding position, and there is a welding gap between the steel rails to be welded; a water-cooled copper mold is placed on both sides of the welding position of the steel rails to be welded, and a magnetic stirring device is installed at the same time. The magnetic stirring device is used to magnetically stir the liquid metal in the molten pool.

[0008] Electroslag welding is performed on the rails to be welded. For the rails at any given time:

[0009] Obtain the image of the molten pool temperature field within the threshold time range before the current moment;

[0010] Obtain the welding parameters within the second time threshold prior to the current moment;

[0011] The heat accumulation characteristics at the current moment are determined based on the welding parameters within the second time threshold and the molten pool temperature field image within the first time threshold.

[0012] The accumulated heat features are input into the trained discrimination model to determine the real-time parameter values ​​of the water-cooled copper mold and the magnetic stirring device, thereby achieving molten pool control at the current moment.

[0013] As a further preferred embodiment, the first time threshold is equal to the second time threshold.

[0014] As a further preferred embodiment, the heat accumulation features at the current moment are determined based on the welding parameters within the second time threshold and the molten pool temperature field image within the first time threshold obtained by the heat accumulation feature extraction model; the heat accumulation feature extraction model is a time series-based machine learning model.

[0015] As a further preferred embodiment, the heat accumulation features at the current moment are determined based on the welding parameters within the second time threshold obtained by processing the trained heat accumulation feature extraction model and the molten pool temperature field image within the first time threshold, including:

[0016] Obtain the temperature image features from each molten pool temperature field image;

[0017] Obtain the welding parameter characteristics corresponding to each welding parameter;

[0018] Based on the trained heat accumulation feature extraction model, the temperature image features, welding parameter features, and the interrelationship between temperature image features and welding parameter features are processed to determine the heat accumulation features.

[0019] As a further preferred embodiment, the heat accumulation feature extraction model and the discrimination model are trained in a joint adversarial manner, including:

[0020] Obtain a first training set, which includes multiple training groups. Each training group includes a first training sample and a second training sample. Both the first and second training samples include welding parameters, temperature field images for the corresponding time period, and label values. The label values ​​reflect the parameters of the water-cooled copper mold and the magnetic stirring device.

[0021] Based on the first training set, the first neural network is trained through multiple rounds of iteration to generate a trained heat accumulation feature extraction model and a discrimination model;

[0022] The first neural network includes an initialized first cumulative heat feature extraction model, an initialized second cumulative heat feature extraction model, and an initialized discrimination model; after training, the trained first cumulative heat feature extraction model or the trained second cumulative heat feature extraction model is used as the trained cumulative heat feature extraction model.

[0023] As a further preferred option, the first neural network is trained through multiple rounds of iteration to generate a trained heat accumulation feature extraction model and a discrimination model, wherein each round of iteration includes:

[0024] Obtain the updated first neural network generated in the previous iteration;

[0025] For each training group,

[0026] The first training sample in the training group is processed using the updated first heat accumulation feature extraction model to obtain the first heat accumulation feature;

[0027] The updated second heat accumulation feature extraction model is used to process the second training samples in the same training group to obtain the second heat accumulation feature.

[0028] The updated discrimination model is used to process the first heat accumulation feature and the second heat accumulation feature to generate a discrimination result, which reflects the degree of difference between the first heat accumulation feature and the second heat accumulation feature.

[0029] Based on the discrimination result and the label value, determine whether to proceed to the next iteration or determine the trained heat accumulation feature extraction model and discrimination model.

[0030] As a further preferred embodiment, in each iteration, after obtaining the discrimination result through forward propagation, a loss function is constructed based on the discrimination result and the label value, and the model parameters are updated by backpropagation based on the loss function.

[0031] As a further preferred option, the characteristics of the accumulated heat at the current moment are determined, including:

[0032] Obtain the current welding gap of the rail.

[0033] The heat accumulation characteristics at the current moment are determined based on the welding gap of the rail at the current moment, the welding parameters within the second time threshold obtained, and the molten pool temperature field image within the first time threshold.

[0034] As a further preferred embodiment, the parameters of the water-cooled copper mold include water flow rate, water flow direction, and water flow pulse speed.

[0035] According to another aspect of the present invention, a magnetic field-temperature field coupled rail electroslag weldability control system is provided, comprising:

[0036] The welding preparation module is used to place the rails to be welded in the welding position, with a welding gap between the rails; two water-cooled copper molds are placed on both sides of the two rails to be welded, and a magnetic stirring device is installed at the same time.

[0037] The welding decision module is used to acquire the molten pool temperature field image within a first time threshold before the current moment when electroslag welding the rail to be welded; acquire the welding parameters within a second time threshold before the current moment; determine the heat accumulation characteristics at the current moment based on the acquired welding parameters within the second time threshold and the molten pool temperature field image within the first time threshold; input the heat accumulation characteristics into the trained discrimination model to determine the real-time parameter values ​​of the water-cooled copper mold and the magnetic field stirring device, thereby realizing the molten pool control at the current moment.

[0038] In summary, compared with the prior art, the above-described technical solutions conceived by this invention mainly possess the following technical advantages:

[0039] 1. Welding using electroslag heating of molten slag has a large heat input and a slow cooling rate, which can avoid the formation of martensite and bainite in the weld joint when welding high alloy steel. At the same time, the stirring process and cooling behavior of the metal in the electroslag weld pool are controlled in real time by combining temperature field and electromagnetic stirring, so as to realize the real-time adjustment of pearlite lamellar structure and properties, and improve the quality of weld joint.

[0040] 2. Forced cooling is achieved by using a water-cooled copper mold, and the corresponding parameters of the water-cooled copper mold are precisely controlled by combining temperature images and neural networks, which improves the forming control effect of the weld.

[0041] 3. A time-series neural network was used to realize the cumulative calculation and control of heat, avoiding the deviation of fitting results caused by the temperature measurement deviation at a certain point. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of an exemplary magnetic field-temperature field coupled rail electroslag weldability control device, as shown in some embodiments of the present invention.

[0043] Figure 2 This is a flowchart illustrating an exemplary method for controlling the formability of three-wire electroslag welds using magnetic field-temperature field coupling, as shown in some embodiments of the present invention.

[0044] Figure 3 The flowchart illustrates an exemplary method for controlling the formability of three-wire electroslag welds using magnetic field-temperature field coupling, as shown in other embodiments of the present invention.

[0045] Figure 4 These are schematic diagrams of neural network models shown in some embodiments of the present invention;

[0046] Figure 5 This is a schematic diagram illustrating the neural network model training process in some embodiments of the present invention.

[0047] In all the accompanying drawings, the same reference numerals are used to denote the same elements or structures, wherein: 1-rail head, 2-welding wire, 3-water-cooled copper mold, 4-outlet, 5-thermocouple, 6-electromagnet, 7-water pipe, 8-rail waist, 9-rail bottom, 10-rail bottom copper mold, 11-No. 1 melting nozzle, 12-No. 2 melting nozzle, 13-No. 3 melting nozzle, 14-inlet. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0049] Flowcharts are used in this embodiment of the invention to illustrate the operations performed by the system according to the embodiments of the invention. It should be understood that the preceding or following operations are not necessarily executed precisely in sequence. Instead, the steps can be processed in reverse order or simultaneously. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0050] Figure 1 This is a schematic diagram of an exemplary three-wire electroslag weld shape control device for magnetic field-temperature field coupling of steel rails.

[0051] The apparatus may include: a 3-wire welding nozzle, a water-cooled copper mold, an electromagnetic stirring device, and a temperature detection device; the rail to be welded includes a rail base 9, a rail web 8, and a rail head 1. Preferably, the rail base, rail web, and rail head are all welded using 20# steel pipes with an inner diameter of 5.0 mm and an outer diameter of 10.0 mm, coated with flux to control the slag properties. The welding wire 2 is a specially developed Φ4.0 mm flux-cored welding wire. Welding nozzles 11 and 3 are used for welding the two sides of the rail base, and welding nozzle 12 is used for welding the middle of the rail base, the rail web, and the rail head, until the rail welding is completed, forming a welded joint.

[0052] The water-cooled copper mold 3 has an inner cavity that mimics the shape of a steel rail and includes a built-in water circulation network. The copper molds are connected together by water pipes 7 to form an integrated circuit, and are equipped with an inlet 14 and an outlet 4. By adjusting the water flow rate in the copper mold, the temperature field and cooling rate of the weld and heat-affected zone are controlled. An electromagnet 6 is built into the copper mold to form a non-uniform magnetic field, which magnetically stirs the liquid metal in the molten pool and promotes heat conduction in the heat-affected zone. A thermocouple 5 is built into the copper mold to measure the real-time temperature and welding thermal cycle.

[0053] The near-net-shape forming of rail welds, the integration of temperature and magnetic fields using copper molds, and the real-time monitoring of welding current, voltage, water flow rate, magnetic field strength, slag pool depth, and the temperature of welds and heat-affected zones via PCs form an integrated process technology.

[0054] In some embodiments, the apparatus may further include an industrial control system. The industrial control system may include computer devices capable of processing programs, such as PCs and PLCs. The industrial control system may include a welding preparation module and a welding decision module. The welding preparation module places two workpieces to be welded in a welding position, with a welding gap between them. This gap accommodates at least three electroslag welding wires, including one main wire and two auxiliary wires distributed on either side of the main wire. Two water-cooled copper molds are placed on the back sides of the two workpieces, and a magnetic stirring device is installed on their back sides. The welding decision module, when welding the workpiece at the current moment, acquires a temperature field image within a first time threshold prior to the current moment; acquires welding parameters within a second time threshold prior to the current moment; determines the heat accumulation characteristics at the current moment based on the acquired welding parameters within the second time threshold and the temperature field image within the first time threshold; inputs the heat accumulation characteristics into a trained first neural network to determine the real-time parameter values ​​of the water-cooled copper mold parameters and the magnetic stirring data; and performs molten pool control at the current moment based on the real-time parameter values ​​of the water-cooled copper mold parameters and the magnetic stirring data.

[0055] It should be understood that the systems and modules of one or more embodiments of the present invention can be implemented in various ways. For example, in some embodiments, the systems and modules can be implemented by hardware, software, or a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the methods and systems described above can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The systems and modules of the present invention can be implemented not only by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., but also by software, for example, executed by various types of processors, or by a combination of the aforementioned hardware circuits and software (e.g., firmware).

[0056] It should be noted that the above description of the processing device and its modules is for ease of description only and should not be construed as limiting the invention to the scope of the embodiments described. It is understood that those skilled in the art, after understanding the principle of the system, may arbitrarily combine the various modules or construct subsystems connected to other modules without departing from this principle.

[0057] Figure 2 This is an exemplary schematic flowchart of a three-wire electroslag weld form control method for magnetic field-temperature field coupling. In some embodiments, the method 200 may be further executed by an industrial control system.

[0058] Step 210: Place the two workpieces to be welded in the welding position.

[0059] In some embodiments, step 210 may be performed by the welding preparation module.

[0060] In some implementation scenarios, one or more embodiments of the present invention can be used for welding rails in scenarios such as welding railways, bullet trains, and high-speed railways. In some embodiments, one or more embodiments of the present invention can also be used for welding large and thick parts (such as large boilers, large trusses, and steel structures).

[0061] The workpieces to be welded can be two long sections of welding rails that have already been placed and ground smooth before being placed at the welding position. In some embodiments, the workpieces to be welded need to undergo certain pre-welding preparation work to ensure welding quality. In some embodiments, pre-welding preparation work includes, but is not limited to: removing oxide scale and laying welding lining.

[0062] After the pre-welding preparations are completed, the two workpieces to be welded are placed in the welding position. The welding position refers to the relative distance between the two workpieces and the welding location. In some embodiments, the two workpieces are placed horizontally. In some embodiments, the two workpieces may also be at any angle.

[0063] Furthermore, the gap to be welded contains at least three electroslag welding wires, the three electroslag welding wires include one main wire and two auxiliary wires, the two auxiliary wires are respectively distributed on both sides of the main wire, and the two water-cooled copper molds are respectively placed on the back side of the two rails to be welded.

[0064] In some embodiments, the welding preparation module can be clamped and placed using fixed tooling, such as welding assembly fixtures or welding fixing fixtures. In some embodiments, the welding preparation module can also be placed using a freely movable tool, such as a 4-axis, 5-axis, or 6-axis robot. In some embodiments, the workpiece to be welded can be placed manually, guided and positioned by the welding preparation module. Such methods are still within the scope of protection of this invention and will not be elaborated further here.

[0065] In some embodiments, the workpiece to be welded is a steel rail, which is Ω-shaped and includes a rail head, a rail web, and a rail base. The two auxiliary wires are used to weld the rail base, and the main welding wire is used to weld the rail head and rail web. In this embodiment, the water-cooled copper mold also includes a rail base copper mold 10, which is used to quickly establish a stable weld pool.

[0066] Step 220: Start welding and acquire real-time images of the temperature field within the molten pool of the workpiece to be welded, and adjust the welding parameters, temperature field, and magnetic field in real time accordingly.

[0067] In some embodiments, step 220 may be performed by the welding decision module.

[0068] In some embodiments, the welding decision module can directly perform the three-wire electroslag welding based on a fixed process specification. In some embodiments, the welding decision module can perform the three-wire electroslag welding based on preset parameters. The welding process is divided into the following three welding stages: Stage I, arc ignition and slag formation stage, where the two auxiliary wires and the main wire simultaneously ignite and maintain a stable arc; Stage II, normal welding stage, where the two auxiliary wires extinguish their arcs, and only the main wire continues to maintain an arc; Stage III, lead-out stage, where the main wire continues welding with a small current and small heat input to fill the crater as welding is nearing completion.

[0069] In some embodiments, the welding decision module can slice the weld based on the digital model corresponding to the weld spacing, obtain the shape profile of the current weld layer corresponding to each slice from bottom to top, and then obtain the corresponding welding process parameters based on the shape profile of the current weld layer.

[0070] The welding decision module can acquire the temperature field of the workpiece to be welded based on one or more thermistors or thermocouples coupled to the workpiece. It is understood that when the required data is the temperature field of the entire surface of the workpiece, using multiple thermistors or thermocouples is preferred. In some embodiments, the welding decision module can also use an infrared temperature sensor (such as an infrared temperature camera) to acquire a real-time image of the temperature field within the molten pool of the workpiece.

[0071] Furthermore, at each welding position during the welding process, the welding decision module can adjust the welding parameters at each point within that position, such as... Figure 3 As shown, the specific steps include the following sub-steps 221 to 225.

[0072] Step 221: Obtain the temperature field image within the first time threshold before the current time.

[0073] In some embodiments, the first time threshold can be a specific time set by the process engineer, such as 5s, 10s, 20s, etc. In some embodiments, the first time threshold can be calculated based on the thermal conductivity of the welding material. For example, for materials with higher thermal conductivity, the heat accumulated in the preceding welding has a greater impact on subsequent welding, so the corresponding first time threshold range is larger, and vice versa.

[0074] Step 222: Obtain the welding parameters within the second time threshold before the current time.

[0075] Similarly, the second time threshold can also be set similarly to the first time threshold. In some embodiments, the second time threshold can be the same as the first time threshold. In this scenario embodiment, the matrix formed by the welding parameters has the same dimension as the matrix resulting from temperature accumulation, which facilitates subsequent matrix processing.

[0076] In other embodiments, the second time threshold may differ from the first time threshold. In this particular scenario, the second time threshold can more intuitively reflect the overall extent of heat input from welding parameters, without being limited by temperature image data. It is understood that since temperature image data reflects a higher data dimension, the number of temperature images that can be used in a single calculation is limited, while the amount of data required for storing welding parameters is small. This approach allows for the intermittent acquisition of temperature image data, acquiring more welding parameter data within the intervals, thereby reducing the amount of data required. Correspondingly, due to the different matrix dimensions, a feature normalization step needs to be added during subsequent feature representation.

[0077] Step 223: Determine the heat accumulation characteristics at the current moment based on the welding parameters within the second time threshold and the temperature field image within the first time threshold.

[0078] In an embodiment of the present invention, the welding decision module can process the welding parameters within the second time threshold and the temperature field images within the first time threshold based on a trained heat accumulation feature extraction model to determine the heat accumulation feature at the current moment. Specifically, the welding decision module can acquire temperature image features in each temperature field image; acquire welding parameter features corresponding to each welding parameter; and process the temperature image features, welding parameter features, and the interrelationships between image features and welding parameter features based on the trained heat accumulation feature extraction model to determine the heat accumulation feature.

[0079] For example, the subsequent description is based on the premise that the second time threshold is equal to the first time threshold. Specifically, it may include the following steps:

[0080] Step 1): Feature the temperature field image corresponding to the first time threshold to obtain the corresponding temperature field image matrix. Feature extraction processing refers to processing the original information and extracting feature data. Feature extraction processing can improve the expression of the original information to facilitate subsequent tasks. In some embodiments, feature extraction processing can adopt statistical methods (e.g., principal component analysis), dimensionality reduction techniques (e.g., linear discriminant analysis), feature normalization, data binning, etc. For example, taking the temperature in the temperature field image as an example, the welding decision module can proportionally correspond peak temperature values ​​within 0 to 800℃ to [1,0,0], peak temperature values ​​within 800 to 1200℃ to [0,1,0], and peak temperature values ​​above 1200℃ to [0,0,1].

[0081] Step 2): Based on the trained heat accumulation feature extraction model, the temperature field image matrix and the welding parameter matrix are coupled and sorted according to the time series. Further, the time series-based matrix clusters are represented using a chained implicit representation to obtain a fixed-length matrix representation output.

[0082] It's understandable that a well-trained heat accumulation feature extraction model can be a time-series-based machine learning model that can transform a variable-length input into a fixed-length vector representation for output. Processing this data through the trained heat accumulation feature extraction model transforms it into a fixed-length vector representation, which facilitates subsequent feedforward propagation.

[0083] For example, a time-series-based machine learning model can be a Gate Repeating Unit (GRU) model. Specifically, the features obtained in step 2) (such as features 1, 2, 3, ..., n) and their relationships (such as sequential sequence and / or temporal order) are input into the heat accumulation feature extraction model, which can then output the sequence of encoded hidden states at each time step (such as k1 to kn). n Furthermore, kn It contains all the information corresponding to the temperature field image at this moment, for example, such as k n It can include information on the lowest and highest temperature values ​​in a temperature field image and their corresponding locations, such as average temperature information, or temperature contour information.

[0084] In this way, the heat accumulation feature extraction model can transform multiple temperature image features and welding parameter features over a period of time into a fixed-length vector representation k. n (i.e., heat accumulation characteristics).

[0085] Step 224: Input the accumulated heat features into the trained discrimination model to determine the real-time parameter values ​​of the water-cooled copper mold parameters and the magnetic field stirring data.

[0086] In one or more embodiments of the present invention, the discriminant model is described as a pre-trained classification cluster model. It can classify the input heat accumulation features and then retrieve the corresponding output results, wherein the output results include real-time parameter values ​​of the water-cooled copper mold parameters and the magnetic field stirring data.

[0087] For example, if R has been pre-generated in the discriminative model c,1 R c,2 , ...R c,i …R c,m There are m cluster centers, and the neighborhood parameter (R) of the clusters. c,i ,∈ c,i ).

[0088] For any heat accumulation characteristic k n , will k n The similarity distance is obtained by comparing the vectors with the centers of the m clusters mentioned above, thereby determining the cluster to which the vector belongs. In some embodiments, the welding decision module can also use vector similarity coefficients to determine the degree of similarity between two vectors. The similarity coefficient refers to the calculation of the similarity between samples using a formula; the smaller the similarity coefficient value, the smaller the similarity between individuals and the greater the difference. When the similarity coefficient between two vectors is large, it can be determined that the two vectors are highly similar. In some embodiments, the similarity coefficients used include, but are not limited to, simple matching similarity coefficient, Jaccard similarity coefficient, cosine similarity, adjusted cosine similarity, Pearson correlation coefficient, etc.

[0089] Furthermore, based on the parameter values ​​of the water-cooled copper mold and magnetic field stirring data corresponding to the center of the cluster, the real-time parameter value corresponding to the current parameter is determined.

[0090] Step 225: Control the molten pool at the current moment based on the real-time parameter values ​​of the water-cooled copper mold parameters and the magnetic field stirring data.

[0091] In one or more embodiments of the present invention, the heat accumulation feature, in addition to being based on the welding parameters within a second time threshold and the temperature field image within a first time threshold, also needs to consider the workpiece gap during these time periods. It is understood that in some implementation scenarios, the welding gap between rails is non-uniform, therefore the heat input per unit area / volume must also be considered. Therefore, when characterizing the temperature field feature or welding parameters, temperature data or welding parameters divided by the welding gap can be used as the temperature input feature, and then input into the heat accumulation feature extraction model. Alternatively, the welding gap can be directly input into the heat accumulation feature extraction model as a parameter for characterization.

[0092] Figure 4 This is a schematic diagram of an exemplary model structure. Figure 4 The system includes a first cumulative calorie feature extraction model, a second cumulative calorie feature extraction model, and a discriminant model. The first and second cumulative calorie feature extraction models have the same model structure. Joint training is further based on, for example, Figure 5 The method shown in process 500 is used for training.

[0093] Step 510: Obtain the first training set.

[0094] The first training set refers to the training sample set used to train the first neural network. The first training set includes multiple training groups, each containing a first training sample and a second training sample. Each training sample includes welding parameters and a temperature field image for the corresponding time period. The label values ​​reflect the output results of the molten pool control under the two parameters. Welding parameters include water-cooled copper mold parameters, magnetic field stirring parameters, and process parameters for the three welding wires, etc. Image features corresponding to the temperature field images can be obtained using feature extraction processing. Further explanation of image features and feature extraction processing can be found in the detailed description of step 223, and will not be repeated here.

[0095] The first neural network can be understood as an untrained neural network model or a neural network model that has not been fully trained. It further includes an initialized first heat accumulation feature extraction model, an initialized second heat accumulation feature extraction model, and an initialized discriminant model. Each layer of the first neural network can be set with initial parameters, which can be continuously adjusted during training until training is complete.

[0096] Step 520: Based on the first training set, train the first neural network through multiple iterations to generate a trained first neural network. Each iteration further includes:

[0097] Step 521: Use the updated first heat accumulation feature extraction model to process the first training sample in the training group to obtain the corresponding first heat accumulation feature.

[0098] The first cumulative heat feature extraction model is a sequence-based machine learning model, such as the previously described GRU gate unit model, which can transform a variable-length input into a fixed-length vector representation for output. Specifically, in one or more embodiments of the present invention, the first cumulative heat feature extraction model performs forward propagation based on data in the first training sample to obtain the corresponding first cumulative heat features.

[0099] Step 522: Process the second training samples in the same training group using the updated second heat accumulation feature extraction model to obtain the second heat accumulation features. Similar to step 521, the second heat accumulation feature extraction model performs forward propagation based on the second training samples in the same training group to obtain the corresponding second heat accumulation features.

[0100] Step 523: Process the first heat accumulation feature and the second heat accumulation feature using the updated discrimination model to generate a discrimination result, which is used to reflect the degree of similarity between the two.

[0101] The discriminant model can determine the similarity between the first and second cumulative heat features. For example, it can obtain the distance between the first and second cumulative heat features in the vector space. The smaller the distance between the vector spaces, the more similar the two features are. Furthermore, the discriminant model can also determine the similarity probability value between the first and second cumulative heat features.

[0102] Step 524: Based on the discrimination result and the label value, determine whether to proceed to the next iteration or to determine the first trained neural network.

[0103] After obtaining the discrimination result through forward propagation of the discriminative model, a loss function can be constructed based on the discrimination result and sample labels. Backpropagation is then performed based on the loss function to update the model parameters. In some embodiments, the training sample label data can be represented as y1, and the discrimination result as... The calculated loss function value is represented as Loss1. In some embodiments, different loss functions can be selected depending on the model type, such as the mean squared error loss function or the cross-entropy loss function, etc., and no limitation is imposed in this invention. For example,

[0104] In some embodiments, the gradient backpropagation algorithm can be used to update model parameters. The backpropagation algorithm compares the prediction results of a specific training sample with the label data to determine the update magnitude of each weight in the model. In other words, the backpropagation algorithm is used to determine how the loss function changes relative to each weight (also known as the gradient or error derivative), denoted as . Furthermore, the gradient backpropagation algorithm can backpropagate the value of the loss function layer by layer through the output layer to the hidden layer and then to the input layer, thereby determining the correction value (or gradient) of the model parameters in each layer. The correction value (or gradient) of the model parameters in each layer includes multiple matrix elements (such as gradient elements), each corresponding one-to-one with a model parameter. Each gradient element reflects the direction (increase or decrease) and amount of parameter correction. In one or more embodiments of the present invention, after the discriminant model completes the gradient backpropagation, it further backpropagates the model parameters layer by layer to the first heat accumulation feature extraction model and the second heat accumulation feature extraction model to complete one round of iterative update. Compared to training each model individually, using a unified loss function for training by jointly training the first heat accumulation feature extraction model, the second heat accumulation feature extraction model, and the discriminant model results in higher training efficiency.

[0105] Taking a single training iteration as an example. For a training set, the label values ​​corresponding to the two welding parameters X1 and X2 in the training samples are Y1 and Y2, respectively. The neural network processes X1 and X2 through forward propagation to obtain the predicted labels. and The corresponding loss values ​​are Loss1 and Loss2. The discriminant model performs a difference assessment on the output predicted labels, resulting in a loss value Loss3. Loss3 reflects the difference between the theoretical and actual calculated values ​​of the difference between two training samples in the same training group, i.e., (Y1-Y2) and... The difference between them is obviously small if all sub-models are trained accurately.

[0106] During backpropagation, Loss3 is used to update the discriminant model, while Loss1 and Loss2 are used to update the two models respectively, thus updating the first neural network as a whole. In simpler terms, the first and second cumulative heat feature extraction models generate the most accurate features, while the discriminant model identifies the subtlest differences between the two feature extraction models; their training results are mutually exclusive. Training terminates only when the training results of both the feature extraction model and the discriminant model meet preset conditions. This setup aims to obtain the most accurate training results.

[0107] After training, either the first heat accumulation feature extraction model or the second heat accumulation feature extraction model can be used directly. In some embodiments, the training samples in the training set can be specifically configured to allow the two feature extraction models to most accurately adapt to different parameter ranges. For example, the first training sample in the training set may mainly use a small current, while the second training sample may use a large current, or a small magnetic field and a large magnetic field. Such configurations enable the two feature extraction models to adapt to more specific application scenarios and achieve higher accuracy for those scenarios. Combinations like these are all within the scope of this specification.

[0108] Furthermore, if it is necessary to make the entire model converge as quickly as possible, the loss value corresponding to the first heat accumulation feature extraction model can be the weighted result between Loss1 and Loss3; the loss value corresponding to the second heat accumulation feature extraction model can be the weighted result between Loss2 and Loss3. This setting can reduce the degree of conflict between the models, but it will sacrifice some training accuracy, but the training process will be greatly reduced and the number of training samples required will also be greatly reduced.

[0109] Furthermore, the first and second cumulative calorie feature extraction models can be pre-trained. For example, a cumulative calorie feature extraction model can be pre-trained using 1,000 samples, and then replicated as the initial model for the first and second cumulative calorie feature extraction models. Then, different training samples from the same training pair can be used for training until the training results converge.

[0110] In some embodiments, the decision to proceed to the next iteration or to determine the trained discriminant model can be based on the discrimination result and label values. The criteria for this decision may include whether the preset number of iterations has been reached, whether a termination instruction has been received, or whether the updated model meets a preset performance threshold; for example, all three loss values ​​may be below a preset threshold, or the weighted sum of the three loss values ​​may be below a threshold. If it is determined that the next iteration is needed, the next iteration can be performed based on the first part of the model updated in the current iteration. In other words, the updated model obtained in the current iteration will be used as the initial model for the next iteration. If it is determined that the next iteration is not needed, the updated model obtained in the current iteration can be used as the final trained model.

[0111] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for controlling the magnetic-temperature field coupling of a rail electroslag welding profile, characterized in that, The method comprises the following steps: Placing the steel rails to be welded in a welding position, the steel rails to be welded having a gap to be welded between them; placing a water-cooled copper mold on both sides of the steel rails to be welded at the welding position, and simultaneously installing a magnetic field stirring device for magnetically stirring the liquid metal in the molten pool; Performing electroslag welding on the steel rails to be welded, and for any instant of time: Obtaining a molten pool temperature field picture within a first time threshold before the instant of time; Obtaining welding parameters within a second time threshold before the instant of time; Determining a heat accumulation feature of the instant of time based on the obtained welding parameters within the second time threshold and the molten pool temperature field picture within the first time threshold; Inputting the heat accumulation feature into a trained discrimination model to determine real-time parameter values of the water-cooled copper mold and the magnetic field stirring device, thereby realizing molten pool control at the instant of time; The first time threshold is equal to the second time threshold; Determining the heat accumulation feature of the instant of time based on the trained heat accumulation feature extraction model processing the obtained welding parameters within the second time threshold and the molten pool temperature field picture within the first time threshold, comprising: Obtaining temperature image features in each molten pool temperature field picture; Obtaining welding parameter features corresponding to each welding parameter; Processing the temperature image features, the welding parameter features, and the mutual relationship between the temperature image features and the welding parameter features based on the trained heat accumulation feature extraction model to determine the heat accumulation feature; the heat accumulation feature extraction model is a machine learning model based on time series; The heat accumulation feature extraction model and the discrimination model are trained in a joint adversarial manner, comprising: Obtaining a first training set, the first training set comprising a plurality of training groups, each training group comprising a first training sample and a second training sample, the first training sample and the second training sample each comprising welding parameters, temperature field images of a corresponding time period, and a label value, the label value reflecting water-cooled copper mold and magnetic field stirring device parameters; Based on the first training set, a first neural network is trained through multiple rounds of iteration to generate a trained heat accumulation feature extraction model and a discrimination model; The first neural network comprises an initialized first heat accumulation feature extraction model, an initialized second heat accumulation feature extraction model, and an initialized discrimination model; after training, the trained first heat accumulation feature extraction model or the second heat accumulation feature extraction model is used as the trained heat accumulation feature extraction model.

2. The magnetic field-temperature field coupled rail electroslag welding shape control method of claim 1 wherein, The first neural network is trained through multiple rounds of iteration to generate a trained heat accumulation feature extraction model and a discrimination model, wherein each round of iteration comprises: Obtaining an updated first neural network generated in the previous round of iteration; For each training group, The first training sample in the training group is processed using the updated first heat accumulation feature extraction model to obtain a first heat accumulation feature; The second training sample in the same training group is processed using the updated second heat accumulation feature extraction model to obtain a second heat accumulation feature; The updated discriminant model is used to process the first heat accumulation feature and the second heat accumulation feature to generate a discriminant result, which reflects the difference between the first heat accumulation feature and the second heat accumulation feature. Based on the discriminant result and the label value, it is determined whether to perform the next round of iteration or to determine the trained heat accumulation feature extraction model and discriminant model.

3. The magnetic field-temperature field coupled rail electroslag welding shape control method of claim 2, wherein, In each round of iteration, after obtaining the discriminant result through forward propagation, a loss function is constructed based on the discriminant result and the label value, and the model parameters are updated based on the loss function through back propagation.

4. The magnetic field-temperature field coupled rail electroslag welding shape control method according to any one of claims 1 to 3, characterized in that, The water-cooled copper mold parameters include water flow size, water flow direction, and water flow pulse speed.

5. A magnetic field-temperature field coupled rail electroslag welding shape controllability system for implementing the magnetic field-temperature field coupled rail electroslag welding shape controllability method according to any one of claims 1 to 4, characterized by, The method comprises the following steps: A welding preparation module is used to place the steel rails to be welded in a welding position, and the water-cooled copper molds are placed on both sides of the two steel rails to be welded, and a magnetic field stirring device is installed at the same time. A welding decision module is used to obtain the molten pool temperature field picture within a first time threshold before the current time when the steel rails to be welded are welded by the electroslag welding, to obtain the welding parameters within a second time threshold before the current time, and to determine the heat accumulation feature of the current time based on the obtained welding parameters within the second time threshold and the molten pool temperature field picture within the first time threshold. The heat accumulation feature is input into the trained discriminant model to determine the real-time parameter value of the water-cooled copper mold and the magnetic field stirring device, thereby realizing the molten pool control at the current time.

Images