Droplet transfer state detection method

Abstract

The application provides a kind of molten droplet transition state detection method, applied to machine vision and intelligent welding technical field.The method comprises: collecting welding video in welding process, and welding video includes a plurality of continuous video frames, and video frame includes welding molten droplet image;For any video frame, extract the morphological feature information of welding molten droplet image in any video frame in preset feature dimension, to obtain the molten droplet feature vector corresponding to any video frame;According to the time position information of any video frame in the pulse cycle to which it belongs, determine the target sampling time from the plurality of sampling times of the preset normal transition template, and the normal transition template includes the molten droplet standard feature distribution information corresponding to each of the plurality of sampling times in the pulse cycle;Determine the target molten droplet standard feature distribution information corresponding to the target sampling time from the normal transition template, and generate the molten droplet transition state detection result of any video frame according to the molten droplet feature vector and the target molten droplet standard feature distribution information.

Description

Technical Field

[0001] This invention relates to the fields of machine vision and intelligent welding technology, and more specifically, to a method for detecting the transition state of molten droplets. Background Technology

[0002] Pulsed Metal Inert Gas Welding (MIG welding) is widely used in aerospace, shipbuilding, and automotive manufacturing due to its controllable heat input, minimal spatter, and high-quality weld formation. The core process objective of pulsed MIG welding is to achieve a "one-pulse-one-drop" transfer, meaning that exactly one molten droplet is generated and detached into the molten pool during the peak of each pulse. However, in actual welding, droplet transfer often deviates from this ideal state, resulting in abnormal transfer states such as multiple drops per pulse, one drop per multiple pulses, or irregular detachment. This leads to increased spatter, deteriorated weld formation, and even internal defects.

[0003] In related examples, the status is typically determined by acquiring real-time electrical signals during the welding process and analyzing their waveform characteristics. However, electrical signals mainly reflect macroscopic electrical changes in the welding circuit, and status confirmation usually can only be performed after the droplets detach and cause a sudden change in electrical parameters, resulting in a certain detection lag and making it difficult to meet the need for early warning before anomalies occur. Summary of the Invention

[0004] In view of this, the present invention provides a method for detecting the transition state of molten droplets.

[0005] One aspect of the present invention provides a method for detecting the transition state of molten droplets. The method includes: acquiring welding videos during a welding process, wherein the welding videos contain multiple consecutive video frames, and each video frame includes an image of a molten droplet; for any one of the multiple video frames, extracting morphological feature information of the molten droplet image in the video frame under a preset feature dimension to obtain a molten droplet feature vector corresponding to the video frame; determining a target sampling time from multiple sampling times of a preset normal transition template based on the time position information of the video frame in its respective pulse period, wherein the normal transition template includes molten droplet standard feature distribution information corresponding to each of the multiple sampling times within the pulse period; determining the target molten droplet standard feature distribution information corresponding to the target sampling time from the normal transition template; and generating a molten droplet transition state detection result for any video frame based on the molten droplet feature vector and the target molten droplet standard feature distribution information.

[0006] Another aspect of the present invention provides a droplet transition state detection device, the device comprising: an acquisition module for acquiring welding videos during the welding process, wherein the welding videos include multiple consecutive video frames, and each video frame includes a weld droplet image; an extraction module for extracting morphological feature information of the weld droplet image in any one of the multiple video frames under a preset feature dimension, thereby obtaining a droplet feature vector corresponding to any one video frame; a first determination module for determining a target sampling time from multiple sampling times of a preset normal transition template based on the time position information of any one video frame in its respective pulse period, wherein the normal transition template includes droplet standard feature distribution information corresponding to each of the multiple sampling times within the pulse period; and a second determination module for determining the target droplet standard feature distribution information corresponding to the target sampling time from the normal transition template, and generating a droplet transition state detection result for any one video frame based on the droplet feature vector and the target droplet standard feature distribution information.

[0007] Another aspect of the present invention provides an electronic device comprising: one or more processors; and a memory for storing one or more programs, wherein, when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the method described above.

[0008] Another aspect of the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the above-described method.

[0009] Another aspect of the present invention provides a computer program product comprising computer-executable instructions which, when executed, are used to implement the method described above.

[0010] According to embodiments of the present invention, by employing a technical means of acquiring welding videos and extracting droplet morphology feature vectors frame by frame, combining the time position information of video frames in the pulse period, and comparing them frame by frame with a preset normal transition template to generate detection results, the technical problems of state perception lag caused by reliance on macroscopic electrical signals and difficulty in early warning during droplet development in the prior art are at least partially overcome. Thus, the technical effect of real-time and accurate detection of the droplet transition state of each frame is achieved, and timely identification and output of warning signals before anomalies occur is achieved. Attached Figure Description

[0011] The above and other objects, features and advantages of the present invention will become more apparent from the following description of embodiments of the invention with reference to the accompanying drawings, in which:

[0012] Figure 1 An exemplary system architecture for applying a droplet transition state detection method according to an embodiment of the present invention is shown.

[0013] Figure 2 A flowchart of a method for detecting the transition state of molten droplets according to an embodiment of the present invention is shown.

[0014] Figure 3 A flowchart illustrating a method for determining a normal transition template according to an embodiment of the present invention is shown.

[0015] Figure 4 A schematic diagram of single-cycle anomaly warning triggering according to an embodiment of the present invention is shown.

[0016] Figure 5 A schematic diagram of cross-cycle drift warning triggering according to an embodiment of the present invention is shown.

[0017] Figure 6 An overall flowchart of a method for detecting the transition state of molten droplets according to an embodiment of the present invention is shown.

[0018] Figure 7 A block diagram of a droplet transition state detection device according to an embodiment of the present invention is shown.

[0019] Figure 8 A block diagram of an electronic device suitable for implementing the methods described above, according to an embodiment of the present invention, is shown. Detailed Implementation

[0020] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the invention. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the invention for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0021] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0022] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0023] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0024] In the embodiments of this invention, the collection, updating, analysis, processing, use, transmission, provision, disclosure, and storage of data (e.g., including but not limited to user personal information) comply with relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. In particular, necessary measures have been taken to prevent unauthorized access to user personal information data and to maintain the security of user personal information and network security.

[0025] In the embodiments of the present invention, the user's authorization or consent is obtained before acquiring or collecting the user's personal information.

[0026] Pulsed MIG welding is widely used in manufacturing fields such as aerospace, shipbuilding, and automobiles due to its controllable heat input, less spatter, and good weld formation quality. However, when the droplet transfer process deviates from the "one drop per pulse" target, abnormal transfer states such as multiple drops per pulse, multiple drops per pulse, or irregular detachment may occur, leading to increased welding spatter, decreased weld formation quality, and even internal defects.

[0027] Existing technologies typically employ monitoring methods based on electrical signals such as arc voltage and welding current to identify the droplet transfer state. This involves acquiring real-time electrical signals during the welding process and analyzing their waveform characteristics to determine the state. However, this method has the following shortcomings in practical applications:

[0028] Electrical signals mainly reflect macroscopic electrical quantity changes in the welding circuit, making it difficult to directly quantify the geometric morphological evolution of a single molten droplet. Status confirmation is usually only possible after the molten droplet detaches and causes a sudden change in electrical parameters, resulting in detection lag and making it difficult to achieve early warning and proactive intervention during the droplet development process.

[0029] Furthermore, existing signal analysis methods mostly focus on independent state judgments within isolated single-pulse cycles, failing to establish temporal correlations across multiple consecutive pulse cycles. In actual welding processes, influenced by gradually changing factors such as slow fluctuations in wire feed speed and gradual heat accumulation, droplet transition is prone to exhibiting progressive drift across cycles. Existing systems lack the ability to track and predict this gradual trend, making it difficult to achieve effective early warning of progressive anomalies.

[0030] In summary, existing technologies suffer from drawbacks such as lagging morphological perception and insufficient cross-cycle trend tracking capabilities. Therefore, a real-time monitoring and early warning method for droplet transition is needed that can directly quantify the evolution of droplet geometry and integrate continuous multi-cycle temporal characteristics for trend prediction.

[0031] To this end, the present invention provides a visual early warning method for abnormal transition of molten droplets in pulsed MIG welding. By processing the molten droplet image sequence acquired by a high-speed camera, the geometric morphological features of the molten droplets are directly extracted, and an early warning of abnormal transition is achieved within a single pulse cycle. At the same time, the transition state feature quantities of multiple consecutive pulse cycles are analyzed for temporal trends to identify the gradual transition state drift.

[0032] Figure 1 An exemplary system architecture for detecting the molten droplet transition state according to an embodiment of the present invention is shown. It should be noted that... Figure 1 The examples shown are merely examples of system architectures that can be applied to embodiments of the present invention, in order to help those skilled in the art understand the technical content of the present invention, but do not mean that embodiments of the present invention cannot be used in other devices, systems, environments or scenarios.

[0033] like Figure 1 As shown, the system architecture 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing communication links between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0034] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, and / or social media platform software, etc. (for example only).

[0035] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0036] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0037] It should be noted that the droplet transition state detection method provided in this embodiment of the invention can generally be executed by server 105. Correspondingly, the droplet transition state detection device provided in this embodiment of the invention can generally be located in server 105. The droplet transition state detection method provided in this embodiment of the invention can also be executed by a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Correspondingly, the droplet transition state detection device provided in this embodiment of the invention can also be located in a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Alternatively, the droplet transition state detection method provided in this embodiment of the invention can also be executed by the first terminal device 101, the second terminal device 102, or the third terminal device 103, or by other terminal devices different from the first terminal device 101, the second terminal device 102, or the third terminal device 103. Accordingly, the droplet transition state detection device provided in the embodiments of the present invention can also be set in the first terminal device 101, the second terminal device 102 or the third terminal device 103, or in other terminal devices different from the first terminal device 101, the second terminal device 102 or the third terminal device 103.

[0038] For example, the welding video during the welding process can be originally stored in any one of the first terminal device 101, the second terminal device 102, or the third terminal device 103 (e.g., the first terminal device 101, but not limited thereto), or it can be stored on an external storage device and imported into the first terminal device 101. Then, the first terminal device 101 can locally execute the droplet transition state detection method provided in the embodiments of the present invention, or send the welding video during the welding process to other terminal devices, servers, or server clusters, and have the other terminal devices, servers, or server clusters that receive the welding video execute the droplet transition state detection method provided in the embodiments of the present invention.

[0039] It should be understood that Figure 1The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0040] Figure 2 A flowchart of a method for detecting the transition state of molten droplets according to an embodiment of the present invention is shown.

[0041] like Figure 2 As shown, the droplet transition state detection method of this embodiment includes operations S201 to S204.

[0042] In operation S201, welding video is acquired during the welding process. The welding video contains multiple consecutive video frames, and each video frame includes an image of the weld droplets.

[0043] The welding video during the welding process can be the raw video obtained by continuously shooting the pulsed MIG welding process using a high-speed camera. For example, the shooting frame rate can be 5000 frames / second, and the image resolution is preferably 1024×1024 pixels to clearly capture the transient evolution of the molten droplets.

[0044] In some embodiments, the acquisition of welding video during the welding process can involve continuous filming of the droplet transition region during pulsed MIG welding. The droplet transition region is the spatial region between the welding wire tip and the surface of the molten pool, which contains the complete trajectory of the droplet as it detaches from the welding wire tip and falls into the molten pool.

[0045] A welding droplet image refers to the area within a video frame containing a molten droplet. The droplet typically appears as a dark area against the background of arc light, and its outline reflects its geometric shape. For example, in a video frame, a near-spherical molten metal is suspended from the tip of the welding wire; this spherical area is the welding droplet image in that frame.

[0046] In some embodiments, the welding video during the welding process can also be a video obtained by preprocessing the original video obtained by continuously shooting the pulse MIG welding process using a high-speed camera.

[0047] For example, after acquiring the original welding video, in order to improve the accuracy and stability of subsequent droplet region extraction, each original video frame in the original video can be preprocessed, and the droplet region in the original video frame can be extracted and the corresponding droplet binary image can be generated. The sequence composed of the droplet binary images of each frame in chronological order is used as the droplet video for subsequent steps.

[0048] In some embodiments, generating the binary image of the molten droplet corresponding to each original video frame can be performed as follows:

[0049] First, each original video frame is subjected to grayscale linear normalization using the following formula (1) to obtain the normalized video frame:

[0050] (1);

[0051] In equation (1), This represents the original grayscale value of any pixel coordinate in the current original video frame. , These are the minimum and maximum grayscale values ​​in the current original video frame, respectively. This represents the normalized grayscale value of any pixel coordinate in the current original video frame. Normalization ensures a uniform grayscale dynamic range across all original video frames, eliminating overall brightness fluctuations caused by changes in arc intensity with pulse phase, thereby improving the stability of subsequent droplet region segmentation.

[0052] Then, the droplet region is extracted for each normalized video frame. During pulsed MIG welding, the droplet always appears as a dark area in the image against the background of the arc light, so a fixed segmentation threshold can be used to extract the droplet region.

[0053] Specifically, the normalized video frames are binarized using a fixed segmentation threshold T. Pixels with grayscale values ​​below T are identified as droplet regions, while pixels with grayscale values ​​above or equal to T are identified as background regions. Connectivity analysis is then performed on the pixels identified as droplet regions, and each connected region is marked as a candidate droplet region. Extensive experimental verification shows that the grayscale values ​​of the normalized droplet regions are stably distributed in the low grayscale range, exhibiting a significant grayscale difference from the background regions. The preferred segmentation threshold T is 40. In practical applications, it can be adjusted within the range of 30 to 50 based on specific welding process parameters (such as welding current and voltage) and operating conditions (such as arc intensity and shielding gas flow rate).

[0054] Finally, the extracted droplet candidate regions are subjected to erosion and dilation morphological processing in sequence. Erosion removes isolated noise points caused by image noise or arc reflection, while dilation repairs the shrinking droplet region boundaries caused by erosion, restoring the droplet region to its true shape. This morphological processing (opening operation) effectively removes isolated noise points and contour adhesion between regions within the droplet candidate regions, resulting in independent droplet regions with complete shapes and clear boundaries. Based on this, the closed contours of each independent droplet region are extracted to generate a binary image of the droplet; the pixel grayscale value within the droplet region is set to 1, while the pixel grayscale value in the background region is set to 0.

[0055] The binary image of the molten droplet obtained after the above preprocessing can be used as the basis for subsequent molten droplet morphology feature extraction and molten droplet transition state analysis, and can be called upon in subsequent operation steps.

[0056] In operation S202, for any video frame among multiple video frames, the morphological feature information of the welding droplet image in any video frame under the preset feature dimension is extracted to obtain the droplet feature vector corresponding to any video frame.

[0057] Preset feature dimensions refer to a set of physical quantities selected in advance to quantify the geometry and motion of the molten droplet. These feature dimensions together constitute the feature space describing the state of the molten droplet.

[0058] Morphological feature information refers to the numerical information extracted from welding droplet images that is used to quantitatively describe the geometry, size, motion state, and morphological changes of the droplets.

[0059] It should be noted that in operation S202, morphological feature information can be extracted either directly from the original video frame or from the binary image of the molten droplet. Since the binary image of the molten droplet has separated the droplet region from the background region, feature extraction from it can avoid interference from background noise such as arc light and splashing, which is beneficial to improving the accuracy and stability of feature extraction. Therefore, it is the preferred implementation method. The following explanation uses the binary image of the molten droplet as an example.

[0060] In some embodiments, the preset feature dimensions can be at least one of the following: equivalent diameter, aspect ratio, profile non-circularity, longitudinal displacement velocity of the center of gravity, diameter change rate, and necking ratio. It should be noted that the above six feature dimensions are preferred embodiments. In practical applications, some of these dimensions can be selected according to specific welding process conditions and inspection requirements, or other feature dimensions that can characterize the evolution of the droplet morphology can be added. This invention does not limit this.

[0061] The following section uses six preset feature dimensions, including equivalent diameter, aspect ratio, non-circularity of the profile, longitudinal displacement velocity of the center of gravity, diameter change rate, and necking ratio, as an example to illustrate the specific implementation method of extracting the morphological feature information of the welding droplet image in any video frame under the preset feature dimensions, and obtaining the droplet feature vector corresponding to any video frame.

[0062] First, define the image moments of the binary image of the molten droplet. Let... For the binary image of the molten droplet in pixel coordinates The normalized gray value at the point (1 inside the contour, 0 outside the contour) defines the zeroth moment. and longitudinal first moment The formulas are (2) and (3) as follows:

[0063] (2);

[0064] (3).

[0065] The morphological feature information of each preset feature dimension is calculated based on the above image moments.

[0066] equivalent diameter The diameter of a circle with the same area as the droplet profile is used to characterize the size of the droplet. It is determined by the total number of pixels within the droplet profile. The result is obtained by reverse deduction, as shown in formula (4):

[0067] (4);

[0068] in .

[0069] Aspect Ratio The height of the minimum bounding rectangle of the droplet profile With width The ratio is used to characterize the degree of stretching of the molten droplet, as shown in formula (5):

[0070] (5);

[0071] Non-circularity of the outline The total number of pixels within the droplet contour With the perimeter of the outline The calculation is used to characterize the degree to which the shape of the molten droplet deviates from a circle, as shown in formula (6):

[0072] (6);

[0073] when When it represents a perfect circle, The larger the value, the more irregular the outline.

[0074] in, The Euclidean distance can be calculated segment by segment from the droplet profile point sequence and then accumulated. Specifically, the method for calculating the profile perimeter is as follows: After obtaining the closed profile point sequence of the droplet region, calculate the Euclidean distance (in pixels) between each pair of adjacent profile points in the order of the profile points, and sum the distances between all adjacent points. The sum is the profile perimeter.

[0075] longitudinal displacement velocity of center of gravity The rate at which the droplet descends in the arc column is used is calculated by the difference in the ordinate of the centroid of adjacent frames, as shown in formula (7):

[0076] (7);

[0077] In equation (7), This represents the time interval between adjacent frames. For example, when using a shooting frame rate of 5000 frames per second, It is 0.2ms. The first frame of each pulse cycle segment The initial value is 0. Let y be the ordinate of the centroid of the droplet in the nth video frame. c,n-1 Let be the ordinate of the centroid of the molten droplet in the (n-1)th video frame. Where, is the ordinate of the centroid of the molten droplet in the nth video frame. The zeroth moment corresponding to the nth video frame pointer and longitudinal first moment The result is shown in formula (8):

[0078] (8).

[0079] Diameter change rate The rate of growth or contraction of the droplet volume is used to characterize the growth or contraction rate of the droplet, calculated from the difference in the equivalent diameter between adjacent frames, as shown in formula (9):

[0080] (9);

[0081] In equation (9), d n d n-1 These represent the equivalent diameters of the nth and (n-1)th video frames, respectively. The first frame of each pulse cycle... The initial value is 0.

[0082] neck reduction ratio Used to characterize the degree of droplet necking. The droplet profile is measured within its longitudinal range. Scan line by line, record the outline width of each line, and take the minimum line width. With maximum line width The ratio is as follows, specifically as shown in formula (10):

[0083] (10);

[0084] In equation (10), and These are the pixel coordinates of the highest and lowest rows of the outline, respectively.

[0085] The physical meaning of η is as follows: When the droplet is in a stable growth stage, its overall profile is relatively uniform (i.e., the droplet shape is regular), necking has not yet formed, the difference in profile width between rows is small, and the value of η is relatively large, close to 1. As necking gradually develops, the root of the droplet (where it connects with the welding wire) is continuously thinned, and the row width at this point decreases sharply, while the main body of the droplet still maintains a relatively large width. At this time, the ratio of the minimum row width to the maximum row width, η, decreases accordingly and approaches 0, indicating that necking is obvious and the droplet is about to detach.

[0086] By concatenating the morphological feature information corresponding to the above six preset feature dimensions, the droplet feature vector of the current video frame is obtained. .

[0087] Arrange the morphological feature information corresponding to the above six preset feature dimensions according to the frame number to obtain the droplet morphological evolution curve within the pulse cycle, thereby completely describing the morphological change process of the droplet from formation, growth, necking to shedding within a pulse cycle.

[0088] After obtaining the droplet feature vector corresponding to any video frame, the droplet feature vector is compared with the normal transition template to determine the detection result. Specifically, this includes the following operations S203 and S204.

[0089] In operation S203, based on the time position information of any video frame in its respective pulse period, the target sampling time is determined from multiple sampling times of the preset normal transition template. The normal transition template includes the droplet standard feature distribution information corresponding to each of the multiple sampling times within the pulse period.

[0090] Temporal position information refers to the relative temporal position of the current video frame within its pulse period, represented by the normalized time t, with a value range of [0,1]. Here, t=0 corresponds to the start of the pulse period, and t=1 corresponds to the end of the pulse period. Since the actual number of frames in different pulse periods may differ (e.g., one period contains 50 frames, another contains 48 frames), directly using the frame number for cross-period comparison will lead to misalignment. Therefore, the position of each frame within the period needs to be normalized to a unified interval, allowing comparisons between periods of different durations under the same time reference. For example, if the current frame is the 15th frame captured within its pulse period, and a total of 50 frames are captured in that period, then the temporal position information of this frame is 15 / 50=0.3, indicating that the frame is located at the 30% time position of the pulse period, in the rising peak phase.

[0091] A normal transition template is a dataset pre-constructed during the offline phase, describing the standard distribution of droplet morphological characteristics at various moments within the pulse period under a normal "one pulse, one droplet" transition state. The normal transition template is equivalent to a "standard model of normal state," used during online detection to compare the droplet characteristics of the current frame with it to determine whether the current droplet state deviates from the normal range. For example, a normal transition template may contain standard feature distribution information corresponding to 100 sampling moments, corresponding to normalized moments t=0.01, 0.02, ..., 1.00, covering the complete process from the start to the end of the pulse.

[0092] The sampling time refers to a pre-selected discrete time point on the normalized time axis of the pulse period. This is used to discretize the continuous time axis, allowing pulse periods of different durations and frame numbers to be compared and analyzed on the same standard time scale. The position of each sampling time on the normalized time axis is represented by the normalized time value t.j This is represented as t, where j is the normalized time number. For example, if the normalized time axis [0,1] of the pulse period is divided into 100 discrete points at equal intervals, then the normalized time value for each sampling time is t. j =j / 100 (j=1,2,…,100), that is, t1=0.01, t2=0.02,…,t 100 =1.00. Each sampling moment corresponds to a set of standard feature distribution information in the normal transition template, thus forming a standard trajectory of the features in the template as the pulse phase changes.

[0093] The standard feature distribution information of a molten droplet refers to the set of information in the normal transition template that corresponds to a specific sampling time and describes the statistical regularities of the morphological characteristics of a normal molten droplet at that time. The standard feature distribution information includes a mean vector and a covariance matrix. The mean vector describes the average value of each feature dimension of the normal sample at that time; for example, at the normalized time t=0.3, the mean of the equivalent diameter is 150 pixels. The covariance matrix describes the fluctuation range of each feature dimension of the normal sample at that time and the correlation between the dimensions. For example, there may be a negative correlation between the equivalent diameter and the necking ratio at that time; that is, when the equivalent diameter is large, the necking ratio is usually small. This correlation can be reflected by the off-diagonal elements in the covariance matrix. Using the mean vector and the covariance matrix, a normal distribution ellipsoid centered on the mean can be constructed, which can then be used to measure the degree to which the feature vector of the current frame deviates from the normal distribution using Mahalanobis distance.

[0094] The target sampling time refers to the sampling time that most sequentially corresponds to the current video frame after matching its temporal position information with multiple sampling times in the normal transition template. In actual detection, the temporal position information of the current frame often cannot precisely coincide with the sampling times in the template (e.g., the current frame's t=0.273, while the template's sampling times are 0.01, 0.02, ..., 1.00). Therefore, it is necessary to select a time from the template's discrete sampling times that is closest to the current frame's temporal position as the target sampling time for subsequent comparison. For example, the normalized time of the current frame is t=0.273, and the adjacent sampling time in the template is t... 27 =0.27 and t 28 =0.28, and after comparison, the sampling time t with the smallest difference from 0.273 was selected. 27 =0.27 is used as the target sampling time.

[0095] In some implementations, determining the target sampling time from multiple sampling times of a preset normal transition template may include: calculating the absolute difference between the normalized time of the current video frame and each sampling time in the normal transition template, and selecting the sampling time that minimizes the difference as the target sampling time.

[0096] As an alternative to the nearest neighbor matching method described above, linear interpolation matching can also be used. This involves determining which two adjacent sampling times the normalized time of the current video frame falls between, and then performing linear interpolation on the mean vector and covariance matrix corresponding to these two times to obtain the continuous standard feature distribution information corresponding to the current frame. This method can eliminate discretization errors, but it increases the computational load.

[0097] After determining the target sampling time, the target droplet standard feature distribution information corresponding to that time can be selected from the normal transition template for subsequent operation S204 to determine anomalies.

[0098] The normal transition template is prepared offline before performing operation S203. The specific preparation process of the normal transition template will be described in detail below.

[0099] Figure 3 A flowchart illustrating a method for determining a normal transition template according to an embodiment of the present invention is shown.

[0100] like Figure 3 As shown, the method for determining the normal transition template in this embodiment includes operations S301 to S305.

[0101] During operation S301, multiple sample video clips are acquired. Each sample video clip is a video clip showing the normal transition of the molten droplet within a complete pulse cycle.

[0102] The sample video segments are pulse cycle segments in a normal droplet transition state. Each segment contains all video frames from the start frame to the end frame of that pulse cycle. In this embodiment, each sample video segment corresponds to one complete "one pulse, one droplet" transition process.

[0103] In some embodiments, several pulse cycle segments of a normal "one pulse, one drop" transition are collected as sample video segments. "Normal" refers to a droplet transition state that has been manually labeled or verified through other methods to conform to the "one pulse, one drop" target. To ensure the statistical representativeness of the template, the number of sample video segments is preferably no less than 100. The larger the number, the higher the statistical stability of the template in describing the normal state, and the better it can cover various minor fluctuations under normal operating conditions (such as slight changes in wire feed speed, morphological differences caused by heat accumulation).

[0104] In operation S302, the sample droplet feature vector of each sample video frame in each sample video segment is determined, and the sample droplet feature sequence of each sample video segment is obtained.

[0105] The sample droplet feature vector refers to the droplet morphology feature vector extracted for a certain video frame in the sample video segment according to the method in step S202 mentioned above. It can include six dimensions: equivalent diameter, aspect ratio, non-circularity of the outline, longitudinal displacement velocity of the center of gravity, diameter change rate, and necking ratio.

[0106] The sample droplet feature vectors of each frame in this video clip are arranged in frame order to form the sample droplet feature sequence of this clip, which is used to describe the complete trajectory of the droplet morphology evolution over time within this pulse period.

[0107] In some embodiments, determining the sample droplet feature vector of each sample video frame in each sample video segment and obtaining the sample droplet feature sequence of each sample video segment includes: for each sample video segment acquired, extracting the sample droplet feature vector of each video frame in the segment according to the method of the aforementioned step S202, and obtaining the sample droplet feature sequence corresponding to each sample video segment.

[0108] In operation S303, the time axis of each sample droplet feature sequence is aligned to obtain the aligned sample droplet feature sequence, which includes sample droplet feature vectors at multiple sampling times.

[0109] Specifically, time axis alignment algorithms such as Dynamic Time Warping (DTW) are used to align the time axis, and the time axis of all sample droplet feature sequences is uniformly normalized to K equally spaced sampling times within the preset interval [0,1] (K is preferably 100 in this embodiment).

[0110] Since the actual number of frames in different pulse periods may differ, nonlinear mapping is performed on each sequence so that frames with the same morphological evolution stage in each sample video segment correspond to the same normalized time.

[0111] The specific operation of time axis alignment is as follows: taking the average sequence length of all sample video segments as a reference, the DTW algorithm is used to determine the corresponding position of each frame in each sample video segment on the normalized time axis, and then the sample droplet feature sequence of each sample video segment is mapped to a unified K sampling time by interpolation.

[0112] After time-axis alignment, each sample video segment has a corresponding droplet feature vector at K sampling times on the normalized time axis. Let the aligned feature sequence be denoted as... ,in For sample number, This is the normalized time number.

[0113] At the same time, the average peak frame rate of all sample video segments was recorded. The average peak frame count This is used to assist in estimating the normalized time of the current frame when the total number of frames in the current pulse period is unknown during real-time detection. The average peak period frame count is calculated by statistically averaging the number of frames from the pulse start frame to the peak period end frame in each sample video segment.

[0114] In operation S304, for any sampling time, the mean vector and covariance matrix corresponding to any sampling time are determined based on the sample droplet feature vector of the aligned sample droplet feature sequence at any sampling time.

[0115] Specifically, for each sampling time t on the normalized time axis j ( Based on the droplet feature vectors of all M sample video segments at that moment, the mean vector at that moment is calculated according to the following formulas (11) and (12). Covariance Matrix .

[0116] (11);

[0117] (12);

[0118] Among them, the mean vector This represents the normal molten droplet at normalized time t. j The standard morphological characteristics of the location; Describe the normal molten droplet at normalized time t j The statistical distribution shape and fluctuation range of the eigenvectors.

[0119] In operation S305, a normal transition template is determined based on the mean vector and covariance matrix at each sampling time. Specifically, the mean vector and covariance matrix obtained from operation S304 above are combined to form the normal transition template. This template describes the characteristic distribution range of a normal "pulse-droplet" transition at various moments on the normalized time axis. It covers the entire morphological evolution of a normal droplet from pulse initiation, peak growth, necking formation to detachment (or resetting), and can be used as a reference standard for normal state during subsequent online detection.

[0120] In operation S204, the target droplet standard feature distribution information corresponding to the target sampling time is determined from the normal transition template, and the droplet transition state detection result of any video frame is generated based on the droplet feature vector and the target droplet standard feature distribution information.

[0121] The standard feature distribution information of the target droplet refers to the standard feature distribution information corresponding to the target sampling time in the normal transition template.

[0122] The detection result refers to the judgment conclusion generated for the current video frame, reflecting whether the droplet transition state of that frame is normal. This result is usually a binary judgment, including two states: normal and abnormal. When it is judged to be abnormal, a warning signal can be further triggered.

[0123] In some embodiments, generating the droplet transition state detection result for any video frame based on the droplet feature vector and the target droplet standard feature distribution information may include at least one of the following two detection methods: intra-cycle anomaly detection and cross-cycle drift detection. Accordingly, the detection result may include intra-cycle detection sub-results and / or cross-cycle drift detection sub-results. Intra-cycle anomaly detection is used to determine whether there is an immediate anomaly in the current frame relative to the normal transition template, while cross-cycle drift detection is used to determine whether there is a gradual drift trend across multiple consecutive pulse cycles. By combining the two detection methods, both sudden anomalies and slowly changing anomalies can be covered simultaneously, improving the comprehensiveness and reliability of the detection.

[0124] According to embodiments of the present invention, by employing a technical means of acquiring welding videos and extracting droplet morphology feature vectors frame by frame, combining the time position information of video frames in the pulse period, and comparing them frame by frame with a preset normal transition template to generate detection results, the technical problems of state perception lag caused by reliance on macroscopic electrical signals and difficulty in early warning during droplet development in the prior art are at least partially overcome. Thus, the technical effect of real-time and accurate detection of the droplet transition state of each frame is achieved, and timely identification and output of warning signals before anomalies occur is achieved.

[0125] In step S202, the temporal position information of any video frame within its respective pulse period can only be obtained after determining the pulse period to which the video frame belongs, as well as its start and end frames. Therefore, before executing step S202, the acquired welding video needs to be automatically segmented into pulse periods to determine the start and end frames of each pulse period, thereby dividing the welding video into multiple independent pulse periods.

[0126] Based on this, according to an embodiment of the present invention, the above method further includes an automatic pulse period segmentation step, specifically including: determining the grayscale mean of each video frame in the welding video to obtain a brightness sequence that varies with the frame number; performing a sliding mean filtering process on the brightness sequence to obtain a smooth brightness sequence; determining the start frame and end frame of each pulse period according to the smooth brightness sequence and a preset brightness discrimination threshold; and dividing the welding video into multiple pulse periods according to the start frame and end frame.

[0127] A brightness sequence refers to the sequence of grayscale averages obtained by calculating the global grayscale average for each frame in a welding video and arranging them according to their frame numbers. In pulsed MIG welding, the arc brightness during the peak period is significantly higher than that during the baseline period. Therefore, the brightness sequence exhibits a periodic alternation of high and low values ​​corresponding to the pulse rhythm.

[0128] A smoothed luminance sequence is a smoothed sequence obtained by applying a moving average filter to a luminance sequence. It is used to eliminate random noise interference between frames and make the periodic characteristics of luminance changes clearer.

[0129] In some embodiments, the sliding filter window length is preferably 5 frames. That is, for each frame, the average brightness of itself and the previous 4 frames is taken as the smoothed brightness value of that frame.

[0130] The preset brightness discrimination threshold is a brightness judgment benchmark value used to distinguish between the peak period of the pulse and the base period. It is automatically calculated by the average grayscale value within a preset number of frames in the initial stage of the welding process.

[0131] The starting frame of a pulse cycle refers to the frame corresponding to the continuous rise of the smooth brightness sequence and the exceeding of the preset brightness discrimination threshold, marking the beginning of the pulse cycle from the base period to the peak period.

[0132] The pulse cycle termination frame refers to the frame corresponding to the continuous decrease of the smooth brightness sequence and the result below the preset brightness discrimination threshold, marking the end of the pulse cycle from the peak period back to the base period.

[0133] In some embodiments, determining the start and end frames of each pulse period based on the smoothed brightness sequence and a preset brightness discrimination threshold may include: using frames where the smoothed brightness sequence continuously increases and exceeds the preset brightness discrimination threshold as the start frame of the pulse period, and using frames where the smoothed brightness sequence continuously decreases and falls below the preset brightness discrimination threshold as the end frame of the pulse period. This "continuous" determination is used to avoid false triggering caused by fluctuations in a single frame, thus improving the reliability of boundary detection.

[0134] In other embodiments, determining the start and end frames of each pulse period based on the smoothed brightness sequence and a preset brightness discrimination threshold may include: using the preset brightness discrimination threshold as a boundary between the peak period and the base period; determining the peak period when the smoothed brightness value is higher than the preset brightness discrimination threshold, and the base period when it is lower than the preset brightness discrimination threshold. Based on this, the start and end frames of the pulse period are determined by detecting the process of the smoothed brightness sequence crossing the preset brightness discrimination threshold. Specifically, determining the current video frame as the start frame of a pulse period requires simultaneously meeting the following three conditions: the previous frame was in the base period, i.e., the smoothed brightness value of the previous frame was lower than or equal to the preset brightness discrimination threshold; the current video frame entered the peak period, i.e., the smoothed brightness value of the current video frame was higher than the preset brightness discrimination threshold; and the brightness showed a continuous upward trend, i.e., the smoothed brightness value of the current video frame was greater than the smoothed brightness value of the previous frame, ensuring that the increase in brightness was a genuine trend change rather than random fluctuation. Similarly, the following three conditions must be met simultaneously to determine the current video frame as the pulse period termination frame: the previous frame is in the peak period, that is, the smoothed brightness value of the previous frame is higher than or equal to the preset brightness discrimination threshold; the current video frame enters the base period, that is, the smoothed brightness value of the current video frame is lower than the preset brightness discrimination threshold; and the brightness shows a continuous decreasing trend, that is, the smoothed brightness value of the current video frame is less than the smoothed brightness value of the previous frame.

[0135] Using the above method, the start and end frames of each pulse cycle can be detected and determined frame by frame based on the smooth brightness sequence and the preset brightness discrimination threshold. Through the above boundary detection, the welding video is divided into several pulse cycle segments, each pulse cycle segment corresponding to a complete droplet formation and shedding process, for subsequent processing steps.

[0136] The following explains how the preset brightness discrimination threshold is determined.

[0137] According to an embodiment of the present invention, the preset brightness discrimination threshold is determined by the following operation:

[0138] First, acquire a preset number of video frames within a preset time period to obtain an initial video frame sequence.

[0139] The preset time period can be the initial stage of welding, that is, a transition period after the welding process begins. During this period, the welding state gradually stabilizes and the arc intensity has not yet been significantly affected by subsequent heat accumulation and other factors.

[0140] In this embodiment, the preset frame number can be the first 100 frames of the welding video. The welding start stage is chosen as the basis for threshold calculation because the welding state at this stage has not been significantly affected by subsequent heat accumulation and other factors, the arc brightness is relatively stable and representative, and can provide a reliable benchmark reference for the brightness changes of the entire welding process.

[0141] Then, the grayscale mean of each initial video frame in the initial video frame sequence is calculated to obtain the grayscale mean of each frame; subsequently, the maximum grayscale mean is determined from the grayscale mean of each initial video frame in the initial video frame sequence. and minimum gray mean .

[0142] Then, the difference between the maximum and minimum grayscale mean values ​​is multiplied by a preset weighting coefficient and added to the minimum grayscale mean value to obtain the preset brightness discrimination threshold. Specifically, the preset brightness discrimination threshold can be calculated according to formula (13). .

[0143] (13);

[0144] In equation (13), The value is a weighting coefficient, and in this embodiment, it is preferably set to 0.6.

[0145] After pulse period segmentation is completed, the pulse period to which any video frame belongs is determined from multiple pulse periods to obtain the target pulse period. Each pulse period segment has a definite start frame and end frame, and any video frame uniquely belongs to one of the pulse period segments. Finding this pulse period segment is to determine the target pulse period.

[0146] Next, the starting frame number of the target pulse period is obtained, and the cumulative frame count from the starting frame to any video frame is calculated. The cumulative frame count is calculated by subtracting the starting frame number from the frame number of any video frame and then adding 1.

[0147] Then, the time position information is determined based on the ratio of the cumulative frame count to the preset frame count.

[0148] The preset frame rate is the average peak frame rate determined during the normal transition template phase. Based on this, the temporal location information t of any video frame j The cumulative frame rate can be divided by the average peak frame rate. Determined. When t j When the value is greater than 1, take t. j =1.

[0149] In some embodiments, generating the droplet transition state detection result for any video frame based on the droplet feature vector and the target droplet standard feature distribution information can be achieved using a single-cycle anomaly detection method.

[0150] Anomaly detection within a single cycle refers to independently calculating the Mahalanobis distance between the droplet feature vector and the normal transition template for each video frame, and determining whether the current frame is an abnormal frame based on whether the Mahalanobis distance exceeds a certain value.

[0151] In some embodiments, anomaly detection within a single cycle specifically includes the following steps:

[0152] First, obtain the droplet feature vector of any video frame and the standard feature distribution information of the target droplet corresponding to any video frame; and calculate the Mahalanobis distance between the droplet feature vector and the standard feature distribution information of the target droplet based on the droplet feature vector, mean vector and covariance matrix. Specifically, the Mahalanobis distance can be determined by the following formula (14).

[0153] (14);

[0154] In equation (14), t j This represents the sampling time of any video frame on the normalized time axis; This represents the Mahalanobis distance of any video frame. Represents the droplet feature vector of any video frame; This represents the mean vector in the standard feature distribution information of the target droplet corresponding to any video frame, i.e., the sampling time t on the normalized time axis in the normal transition template. j The mean vector at point; This represents the covariance matrix in the standard feature distribution information of the target droplet corresponding to any video frame, i.e., the sampling time t on the normalized time axis in the normal transition template. j The covariance matrix at that location.

[0155] The smaller the Mahalanobis distance, the closer the droplet feature vector of any video frame is to the standard distribution of the normal transition template at that sampling time, that is, the more normal the current state is; the larger the Mahalanobis distance, the more significantly any video frame deviates from the normal state.

[0156] Then, after obtaining the Mahalanobis distance, the Mahalanobis distance is compared with a preset anomaly threshold. In some embodiments, the preset anomaly threshold... The method for determining this is as follows: when the six-dimensional eigenvectors follow a multivariate normal distribution, the square of the Mahalanobis distance is... Approximately follows the degree of freedom of Chi-square distribution The chi-square distribution quantile corresponding to a 99% confidence level is used as the preset anomaly threshold. As shown in formula (15):

[0157] (15).

[0158] The preset anomaly threshold indicates that, in normal transition samples, 99% of the frames have a Mahalanobis distance lower than [a certain threshold]. ,Exceed The frame was deemed abnormal. In actual use, the value can be adjusted within the range of 3.5 to 5.0 based on the validation results of the training data.

[0159] After obtaining the Mahalanobis distance and comparing it with a preset anomaly threshold, it is necessary to further combine the temporal position information of any video frame within the pulse period for comprehensive judgment to avoid triggering false alarms due to a brief increase in Mahalanobis distance during the normal droplet detachment phase. Specifically, a single-cycle anomaly warning is triggered when the following two conditions are met simultaneously:

[0160] Condition 1: The Mahalanobis distance of any video frame exceeds a preset anomaly threshold, i.e. ;

[0161] Condition 2: Any video frame is earlier than the expected droplet detachment time of this pulse period, i.e. ,in, The normalized parameter for the expected detachment time is the sampling time corresponding to the droplet detachment in the normal transition template. In this embodiment, the process parameter range for pulsed MIG welding of steel is obtained from experimental statistics, and the preferred value is 0.85.

[0162] Both conditions must be met simultaneously to ensure that the warning occurs before the molten droplets actually detach, providing an advance warning period so that operators can take timely intervention measures.

[0163] Using the above method, a corresponding droplet transition state detection result is generated for each video frame. When both of the above conditions are met simultaneously, the current frame is determined to be abnormal and an early warning is triggered; otherwise, it is determined to be normal. By repeating the above process frame by frame, real-time anomaly detection and early warning of the droplet transition state in pulsed MIG welding can be achieved within a single cycle.

[0164] The following combination Figure 4 The triggering timing of single-cycle abnormal warnings is explained.

[0165] Figure 4 A schematic diagram of single-cycle anomaly warning triggering according to an embodiment of the present invention is shown.

[0166] like Figure 4 As shown, the horizontal axis represents the normalized time t, which ranges from [0,1] and corresponds to the complete process of the pulse period from start to end; the vertical axis represents the Mahalanobis distance. Figure 4 The gray area 401 in the diagram represents the normal transition range of the molten droplet. Within this range, the droplet feature vector of the current frame falls within the normal distribution range, indicating a normal state. When the Mahalanobis distance curve exceeds the normal transition range, it indicates that the current frame deviates from the normal state, satisfying condition one.

[0167] However, not all frames exceeding the normal transition range should trigger an alarm. In the later stages of the pulse cycle, during the normal droplet detachment phase, the Mahalanobis distance may briefly increase due to drastic changes in droplet morphology. However, this is a normal transition process and should not trigger an anomaly alarm. Therefore, this embodiment uses condition two to differentiate: only when the deviation from the normal transition range occurs before the expected droplet detachment time is it considered an anomaly. Figure 4 As shown, if the Mahalanobis distance exceeds the normal transition range before the expected detachment time, it is judged as abnormal and an early warning is triggered; however, if the Mahalanobis distance exceeds the normal transition range after the expected detachment time, no early warning is triggered.

[0168] Figure 4 The area indicated by the double-headed arrows represents the warning lead time, which is the time interval between the triggering of the warning and the expected time of droplet detachment. This warning lead time provides sufficient response time for operators or control systems to take timely intervention measures (such as adjusting welding parameters) before the droplet actually detaches, thereby effectively avoiding welding defects caused by abnormal transitions.

[0169] In this embodiment of the invention, the condition that the target sampling time is earlier than the expected droplet detachment time is added, which can avoid triggering false alarms due to a brief increase in Mahalanobis distance during the normal droplet detachment stage, thereby effectively reducing the false alarm rate while ensuring the anomaly detection rate.

[0170] In some embodiments, after an anomaly detection within a single cycle determines that the droplet transition state is abnormal and triggers an early warning, the specific transition type to which the anomaly belongs can be further determined so that operators can take targeted intervention measures according to different anomaly types.

[0171] In embodiments of the present invention, the abnormal transition type includes one of the following three: single-pulse multi-droplet abnormality, droplet accumulation abnormality, and short-circuit precursor abnormality.

[0172] Among them, the "multiple drops in one pulse" anomaly indicates that the molten droplets detach prematurely before the expected detachment time within the current pulse cycle. At the feature level, this manifests as a sudden decrease in the equivalent diameter of the current video frame relative to the standard value at the corresponding moment in the normal template, indicating that the first molten droplet has detached prematurely.

[0173] Abnormal droplet accumulation indicates that the droplet size continuously exceeds the upper limit of the normal range and is still increasing, failing to detach at the expected time. Characteristically, this manifests as the droplet's equivalent diameter being greater than the upper limit of the equivalent diameter, and the rate of change of the equivalent diameter being positive (i.e., the size is still continuously increasing).

[0174] The short-circuit precursor anomaly indicates that the molten droplet has not yet formed a significant necking, but the longitudinal displacement velocity of the center of gravity has increased abnormally, indicating that the molten droplet is accelerating towards the molten pool and a short-circuit transition may be imminent. At the characteristic level, this manifests as the droplet's non-circularity being less than a preset non-circularity threshold, and the longitudinal displacement velocity of the center of gravity exceeding the upper velocity limit.

[0175] The following explains how to determine the type of exception.

[0176] If the droplet transition state detection result of any video frame is determined to be abnormal, the specific abnormal transition type is determined based on the deviation direction of each feature component in the droplet feature vector of any video frame relative to the corresponding feature component in the standard feature distribution information of the target droplet. The preset feature dimensions include equivalent diameter, aspect ratio, contour non-circularity, longitudinal displacement velocity of the center of gravity, diameter change rate, and necking ratio. The determination method includes the following steps:

[0177] First, the mean of the standard equivalent diameter at the target sampling time is obtained from the standard characteristic distribution information of the target droplet. Standard deviation of the standard equivalent diameter and the average longitudinal displacement velocity of the standard center of gravity The standard deviation of the longitudinal displacement velocity of the standard center of gravity .in, and The covariance matrix at the corresponding time point in the normal transition template It is obtained by taking the square root of the diagonal element. Specifically, This is the square root of the diagonal element corresponding to the equivalent diameter feature in the covariance matrix. It is the square root of the diagonal element corresponding to the longitudinal displacement velocity characteristic of the centroid in the covariance matrix.

[0178] Then, based on the mean of the standard equivalent diameter Standard deviation of the standard equivalent diameter Determine the upper and lower limits of the equivalent diameter. Specifically, the upper limit of the equivalent diameter is... +2 The lower limit of the equivalent diameter is -2 Based on the average longitudinal displacement velocity of the standard center of gravity. The standard deviation of the longitudinal displacement velocity of the standard center of gravity Determine the speed limit; specifically, the speed limit is... +2 .

[0179] Based on this, the anomaly type is determined according to the deviation of each feature component of the current frame from the above thresholds in the following manner.

[0180] If the equivalent diameter of the droplets in any video frame is less than the lower limit of the equivalent diameter, the abnormal transition type is determined to be a single-pulse multi-droplet anomaly. In some embodiments, when and When this occurs, it indicates that the droplet diameter suddenly shrinks and the first droplet has detached, thus identifying the abnormal transition type as a multi-droplet anomaly.

[0181] In any video frame, the equivalent diameter of the molten droplet is greater than the upper limit of the equivalent diameter, and the rate of change of the equivalent diameter is greater than the preset rate of change threshold (i.e., In the case of ), that is and The abnormal transition type was determined to be droplet accumulation anomaly;

[0182] When the longitudinal displacement velocity of the droplet's center of gravity in any video frame exceeds the upper velocity limit, and the non-circularity of the droplet's outline in any video frame is less than a preset non-circularity threshold, i.e. and This indicates that the droplet's downward movement speed is abnormally accelerated and its shape is close to a regular sphere (without obvious necking), thus identifying the abnormal transition type as a short-circuit precursor anomaly.

[0183] The determination of the above three anomaly types is based on the physical laws of droplet morphological evolution. In the single-droplet anomaly, the equivalent diameter suddenly shrinks to below the lower limit of the normal range, indicating that the droplet has detached prematurely and the welding wire tip has begun to melt again. This corresponds to the physical phenomenon of a sudden reduction in diameter and the first droplet having detached. In the droplet accumulation anomaly, the equivalent diameter continuously exceeds the upper limit of the normal range and the rate of diameter change is positive, indicating that the droplet continues to grow without detaching. This corresponds to the physical phenomenon of the size continuously exceeding the upper limit of the normal range and still increasing. In the short-circuit precursor anomaly, the longitudinal displacement velocity of the center of gravity exceeds the upper limit of the velocity while the non-circularity of the profile is below the threshold, indicating that the droplet has accelerated its downward movement before forming obvious necking. This corresponds to the physical phenomenon of abnormally accelerated downward movement and a shape close to a regular sphere, indicating that it has accelerated its approach to the molten pool before necking.

[0184] By using the above method, the specific abnormal transition type can be accurately identified based on the deviation direction of each feature component in the current frame, providing a targeted basis for subsequent process adjustments.

[0185] In some embodiments, to identify the gradual drift of the droplet transition state across cycles caused by slowly fluctuating wire feed speeds, gradual heat accumulation, and other slowly changing interference factors, in addition to the above-mentioned anomaly detection within a single cycle, periodic statistical feature extraction and cross-cycle trend analysis can also be performed on the completed pulse cycles. Cross-cycle drift detection is performed after each pulse cycle ends to capture slowly changing trends that cannot be identified by single-cycle detection, thus complementing the anomaly detection within a single cycle.

[0186] According to an embodiment of the present invention, the above method further includes: after the end of any pulse period, statistically summarizing the droplet feature vectors and Mahalanobis distances of each video frame within any pulse period to obtain the periodic statistical feature vector of any pulse period.

[0187] A periodic statistical feature vector is used to describe the feature set of the overall state of the pulse period. In this embodiment, the periodic statistical feature vector may include the following statistical components: the mean equivalent diameter of each video frame in any pulse period; the mean aspect ratio of each video frame in any pulse period; the mean non-circularity of the contour of each video frame in any pulse period; the sampling time corresponding to the droplet detachment in any pulse period; and the maximum value of the Mahalanobis distance of each video frame in any pulse period.

[0188] For example, a periodic statistical eigenvector can be represented as ,in The pulse period is numbered, and the statistical components in the periodic statistical feature vector are defined as follows.

[0189] For the first Equivalent diameter of all frames within a pulse period The mean value reflects the average size level of the molten droplet within the pulse cycle, as shown in formula (16):

[0190] (16).

[0191] For the first Aspect Ratio of All Frames within a Pulse Period The average value reflects the average stretching degree of the molten droplet within the pulse cycle, as shown in formula (17):

[0192] (17);

[0193] For the first Non-circularity of the outline of all frames within the pulse period The mean value reflects the average irregularity of the droplet shape within the pulse cycle, as shown in formula (18):

[0194] (18);

[0195] Among them, in equations (16), (17), and (18) For the first The total number of frames within a pulse period.

[0196] For the first The normalized detachment time of the droplet within a pulse cycle, i.e., the position of the actual detachment frame on the normalized time axis of that pulse cycle, reflects the timing of droplet detachment. The detachment time is determined by the necking ratio. Decreased below the preset neck constriction threshold In this embodiment, a necking threshold is preset for frame determination. The preferred value is 0.2; if within this period Not dropped The following will be taken .

[0197] For the first The peak value of the Mahalanobis distance within the pulse period is shown in formula (19):

[0198] (19);

[0199] In equation (19), For the first The first pulse within the pulse period Mahalanobis distance of frames This reflects the maximum deviation of the feature vector from the normal template within the pulse period, even if it does not exceed a preset anomaly threshold. The periodic variation trend of this value can still serve as an early signal of system state drift.

[0200] After obtaining the periodic statistical eigenvectors of each pulse cycle, cross-cycle trend analysis is performed on the periodic statistical eigenvectors of each pulse cycle to obtain the cross-cycle trend analysis results. Specifically, the steps are as follows.

[0201] First, maintain a length of A sliding window for each pulse cycle stores the most recent... A sequence of statistical eigenvectors for each period Based on the range of pulsed MIG welding process parameters and experimental verification, The preferred value is 50. The cumulative number of cycles after welding begins has not reached [a certain value]. Previously, all currently accumulated periodic data were used instead of the complete sliding window for calculation.

[0202] Then, trend tests are performed on each statistical component in the periodic statistical feature vector of each pulse period within the sliding window to obtain the test statistics of each statistical component. The test statistics are then standardized to obtain the standardized statistics of each statistical component.

[0203] In some embodiments, the Mann-Kendall trend test is used for nonparametric trend testing. The calculation method for the trend test is explained using a statistical component in a periodic statistical eigenvector as an example.

[0204] Let the time series data of a certain statistical component be denoted as First, iterate through all data pairs in the time series data. Where i is less than The test statistic S is calculated using the following formula (20):

[0205] (20).

[0206] Among them, the sign function Defined as formula (21):

[0207] (twenty one);

[0208] in, for .

[0209] When S is a large positive number, it indicates that the time series data is on an upward trend; when S is a large negative number, it indicates that the time series data is on a downward trend.

[0210] Considering the possible tandem values ​​in the actual data, such as the variance of the test statistic S as shown in formula (22) below. Make corrections:

[0211] (twenty two);

[0212] Among them, t p This represents the number of identical data points in the p-th group with the same size.

[0213] Next, the test statistic S is standardized to the standardized statistic Z, as shown in formula (23):

[0214] (twenty three).

[0215] After standardization, Z-likely follows a standard normal distribution N(0,1), which can be directly compared with the critical value of the standard normal distribution.

[0216] The standardized statistics corresponding to each statistical component can be determined using the method described above.

[0217] After obtaining the standardized statistics corresponding to each statistical component, the standardized statistics are compared with a preset trend threshold. If the standardized statistics corresponding to at least one statistical component are greater than the preset trend threshold, a drift warning is generated, indicating that there is a drift trend in the welding process.

[0218] Specifically, the preset trend threshold is set as follows: In this embodiment, the corresponding significance level is... The critical value for a two-sided test, The preferred value is 1.96. Considering that the drift risk scenarios of concern in this invention all exhibit a continuous increase in eigenvalues ​​(such as a gradual increase in droplet size and a gradual increase in shape irregularity), it is preferable to only test the upward trend. When the standardized statistic of a certain statistical component within the sliding window satisfies... When a statistical component is determined to have a statistically significant upward trend, a cross-period drift warning is triggered. At this time, the statistical characteristic values ​​of each period do not exceed the preset anomaly threshold. However, the monotonous upward trend within the sliding window has been captured by trend testing, demonstrating that the cross-cycle drift detection of this invention effectively supplements the blind zone of single-cycle detection.

[0219] The following combination Figure 5 The triggering mechanism of cross-cycle drift warning is explained.

[0220] Figure 5 A schematic diagram of cross-cycle drift warning triggering according to an embodiment of the present invention is shown.

[0221] like Figure 5 The horizontal axis represents the pulse period number k, and the vertical axis represents the mean value of the period characteristics (such as the mean value of the equivalent diameter). Figure 5 The horizontal dashed line in the diagram represents the single-cycle threshold, i.e., the preset anomaly threshold τ. Figure 5 As shown, the feature values ​​of each period are all lower than the single-period threshold. Therefore, if only anomaly detection within a single period is relied upon, none of the periods will be judged as anomalies.

[0222] However, within the current sliding window (such as Figure 5 (The window range is marked in the image). Although the mean values ​​of each periodic feature did not exceed the single-period threshold, they generally showed a continuous upward trend with the increase of the period number, such as... Figure 5 The trend line is shown in the diagram.

[0223] At this point, a trend test is performed on the time-series data within the current sliding window, and standardized statistics are calculated. When a large monotonically upward trend in the data within the window reaches a preset trend threshold, a cross-period drift warning is triggered, such as... Figure 5 As shown, the drift warning is triggered between cycle 40 and cycle 50.

[0224] Based on this, the periodic statistical eigenvectors The five statistical components are subjected to the above trend test respectively, and the drift type is determined based on the statistical component that triggers the drift warning and its trend direction.

[0225] Showing a significant upward trend ( The average size of the molten droplets continues to increase, which may be caused by excessive wire feed speed or heat accumulation in the workpiece.

[0226] Showing a significant upward trend ( The average stretching of the molten droplet continues to increase, and the molten droplet is abnormally stretched by electromagnetic force during the peak period, which may be caused by the high pulse peak current.

[0227] Showing a significant upward trend ( The droplet shape continues to become increasingly irregular, leading to decreased transition stability.

[0228] Showing a significant upward trend ( The timing of droplet detachment continues to be delayed, and the risk of large droplet accumulation gradually increases.

[0229] Showing a significant upward trend ( The peak value of the Mahalanobis distance in each cycle continues to rise, and the overall state of the system is approaching the abnormal boundary, which is the highest priority drift warning signal.

[0230] Furthermore, after generating single-cycle anomaly warning signals and cross-cycle drift warning signals, graded warning signals are output according to the triggering status of the two signals and in accordance with preset priority rules, as follows.

[0231] Level 1 warning, triggered only for single-cycle anomalies, not for cross-cycle drift. This indicates a sudden abnormal transition in the current pulse cycle. It is recommended to immediately adjust the peak current or pulse width to restore the normal "one pulse, one drop" transition state.

[0232] Level 2 warning, triggered only by cross-cycle drift warning, not by single-cycle anomaly warning. This indicates a gradual transition drift in the welding process. Although the characteristic values ​​of each cycle have not yet exceeded the single-cycle warning threshold, they show a statistically significant monotonic trend. It is recommended to fine-tune process parameters such as wire feed speed and shielding gas flow rate according to the type of drift component output by the cross-cycle drift warning. In cross-cycle drift detection, the drift component type is determined by trend test to be a statistically significant upward trend.

[0233] A Level 3 warning is triggered simultaneously by both a single-cycle anomaly warning and a cross-cycle drift warning. This indicates that the welding process exhibits both a sudden anomaly in the current cycle and a systematic drift trend, constituting the highest level warning. It is recommended to suspend welding and check the welding equipment and process parameter settings.

[0234] The aforementioned level 1, 2, or 3 warning signals are output to the welding control system in real time through the control interface. The control system then performs corresponding adjustment actions based on the warning level to achieve closed-loop control of abnormal droplet transition in pulsed MIG welding.

[0235] This invention combines single-cycle sudden anomaly detection with cross-cycle gradual drift detection and outputs graded early warning signals, realizing full-process monitoring and early warning coverage of the transition state of pulsed MIG welding droplets from transient to gradual, providing a basis for advance intervention in intelligent welding quality control.

[0236] The following combination Figure 6 The method for detecting the molten droplet transition state of the present invention will be further described.

[0237] Figure 6 An overall flowchart of a method for detecting the transition state of molten droplets according to an embodiment of the present invention is shown.

[0238] like Figure 6 As shown, the droplet transition state detection method of this embodiment includes operations S601 to S608.

[0239] In operation of S601, image acquisition and preprocessing are performed. Welding images are acquired using a high-speed camera, and the independent closed contours of the molten droplets are extracted through grayscale normalization and threshold segmentation.

[0240] In operation S602, pulse period automatic segmentation is performed. Based on the periodic fluctuations of the global grayscale mean of the image, it is automatically segmented into image sequence segments corresponding to a single pulse period.

[0241] In operation S603, droplet morphology features are extracted. The equivalent diameter, necking ratio, and other six-dimensional morphological features of the profile are calculated, and the droplet morphology evolution curves within a single cycle are generated in frame order.

[0242] In operation S604, a normal transition template is established. Offline, normal samples are aligned along the time axis, and the mean vector and covariance matrix at each normalized time step are calculated to construct the normal transition template.

[0243] In S605 operation, single-cycle anomaly detection is performed. The Mahalanobis distance between the current frame features and the normal transition template is calculated in real time, and a single-cycle sudden anomaly warning is triggered and classified before the droplet falls off.

[0244] In operation S606, periodic statistical features are extracted. After the pulse cycle ends, the mean and peak values ​​of the morphological and distance features within that cycle are calculated, and a five-dimensional periodic statistical feature vector is extracted.

[0245] In operating S607, cross-cycle drift detection is performed. A sliding window of continuous multi-cycle features is maintained, and a non-parametric trend test is used to identify asymptotic state drift and output the drift type.

[0246] When operating S608, graded early warning signals are output. Based on the combined detection results of single-cycle sudden anomalies and cross-cycle gradual drift, graded early warning signals are output according to preset priorities to guide closed-loop control.

[0247] Through the above operations, this invention achieves real-time monitoring and graded early warning of the entire process of pulsed MIG welding droplet transition state from sudden abnormality to gradual drift, providing reliable technical support for online control of welding quality.

[0248] Figure 7 A block diagram of a droplet transition state detection device according to an embodiment of the present invention is shown.

[0249] like Figure 7 As shown, the droplet transition state detection device of this embodiment includes a data acquisition module 710, an extraction module 720, a first determination module 730, and a second determination module 740.

[0250] The acquisition module 710 is used to acquire welding videos during the welding process. The welding videos consist of multiple consecutive video frames, and each video frame includes an image of the weld droplets.

[0251] The extraction module 720 is used to extract the morphological feature information of the welding droplet image in any one of multiple video frames under a preset feature dimension, so as to obtain the droplet feature vector corresponding to any one video frame.

[0252] The first determining module 730 is used to determine the target sampling time from multiple sampling times of a preset normal transition template based on the time position information of any video frame in its respective pulse period. The normal transition template includes the standard feature distribution information of the molten droplets corresponding to each of the multiple sampling times within the pulse period.

[0253] The second determining module 740 is used to determine the target droplet standard feature distribution information corresponding to the target sampling time from the normal transition template, and generate the droplet transition state detection result of any video frame based on the droplet feature vector and the target droplet standard feature distribution information.

[0254] According to embodiments of the present invention, any plurality of the acquisition module 710, extraction module 720, first determination module 730, and second determination module 740 can be combined into one module / unit / subunit, or any one of these modules / units / subunits can be split into multiple modules / units / subunits. Alternatively, at least part of the functionality of one or more of these modules / units / subunits can be combined with at least part of the functionality of other modules / units / subunits and implemented in one module / unit / subunit. According to embodiments of the present invention, at least one of the acquisition module 710, extraction module 720, first determination module 730, and second determination module 740 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or any other reasonable means of integrating or packaging circuitry, or implemented in hardware or firmware, or in any one of software, hardware, and firmware implementations, or in a suitable combination of any of these. Alternatively, at least one of the acquisition module 710, extraction module 720, first determination module 730 and second determination module 740 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0255] It should be noted that the droplet transition state detection device part in the embodiments of the present invention corresponds to the droplet transition state detection method part in the embodiments of the present invention. For a detailed description of the droplet transition state detection device part, please refer to the droplet transition state detection method part, which will not be repeated here.

[0256] Figure 8 A block diagram of an electronic device suitable for implementing the methods described above, according to an embodiment of the present invention, is shown. Figure 8 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0257] like Figure 8 As shown, an electronic device 800 according to an embodiment of the present invention includes a processor 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 802 or a program loaded from a storage portion 808 into a random access memory (RAM) 803. The processor 801 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 801 may also include onboard memory for caching purposes. The processor 801 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present invention.

[0258] RAM 803 stores various programs and data required for the operation of electronic device 800. Processor 801, ROM 802, and RAM 803 are interconnected via bus 804. Processor 801 executes various operations of the method flow according to embodiments of the present invention by executing programs in ROM 802 and / or RAM 803. It should be noted that the programs may also be stored in one or more memories other than ROM 802 and RAM 803. Processor 801 may also execute various operations of the method flow according to embodiments of the present invention by executing programs stored in said one or more memories.

[0259] According to an embodiment of the present invention, the electronic device 800 may further include an input / output (I / O) interface 805, which is also connected to a bus 804. The electronic device 800 may also include one or more of the following components connected to the input / output (I / O) interface 805: an input section 806 including a keyboard, mouse, etc.; an output section 807 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN card, modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the input / output (I / O) interface 805 as needed. A removable medium 811, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 810 as needed so that computer programs read from it can be installed into the storage section 808 as needed.

[0260] According to embodiments of the present invention, the method flow according to embodiments of the present invention can be implemented as a computer software program. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the method shown in the flowchart. In such embodiments, the computer program can be downloaded and installed from a network via communication section 809, and / or installed from removable medium 811. When the computer program is executed by processor 801, it performs the functions defined in the system of the embodiments of the present invention. According to embodiments of the present invention, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0261] The present invention also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of the present invention.

[0262] According to embodiments of the present invention, the computer-readable storage medium may be a non-volatile computer-readable storage medium. Examples include, but are not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In the present invention, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0263] For example, according to embodiments of the present invention, a computer-readable storage medium may include the ROM 802 and / or RAM 803 described above and / or one or more memories other than ROM 802 and RAM 803.

[0264] Embodiments of the present invention also include a computer program product comprising a computer program containing program code for performing the methods provided in the embodiments of the present invention. When the computer program product is run on an electronic device, the program code is used to enable the electronic device to implement the methods provided in the embodiments of the present invention.

[0265] When the computer program is executed by the processor 801, it performs the functions defined in the system / apparatus of this embodiment of the invention. According to embodiments of the invention, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0266] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 809, and / or installed from a removable medium 811. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0267] According to embodiments of the present invention, program code for executing the computer programs provided in the embodiments of the present invention can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0268] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions. Those skilled in the art will understand that the features described in the various embodiments of the present invention can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in the present invention. In particular, the features described in the various embodiments of the present invention can be combined and / or combined in various ways without departing from the spirit and teachings of the present invention. All such combinations and / or pairings fall within the scope of this invention.

[0269] The embodiments of the present invention have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of the invention. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. The present invention does not depart from its scope, and those skilled in the art can make various substitutions and modifications, all of which should fall within the scope of the present invention.

Claims

1. A method for detecting the transition state of a molten droplet, characterized in that, The method includes: The welding process is captured via video, wherein the welding video comprises multiple consecutive video frames, and each video frame includes an image of the weld droplet. For any one of the multiple video frames, extract the morphological feature information of the welding droplet image in the video frame under a preset feature dimension to obtain the droplet feature vector corresponding to the video frame. Based on the time position information of any video frame in its respective pulse period, the target sampling time is determined from multiple sampling times of a preset normal transition template, wherein the normal transition template includes the droplet standard feature distribution information corresponding to each of the multiple sampling times within the pulse period. The target droplet standard feature distribution information corresponding to the target sampling time is determined from the normal transition template, and the droplet transition state detection result of any video frame is generated based on the droplet feature vector and the target droplet standard feature distribution information.

2. The method according to claim 1, characterized in that, The target droplet standard feature distribution information includes the mean vector and covariance matrix; The step of generating the droplet transition state detection result for any video frame based on the droplet feature vector and the target droplet standard feature distribution information includes: Based on the droplet feature vector, the mean vector, and the covariance matrix, calculate the Mahalanobis distance between the droplet feature vector and the standard feature distribution information of the target droplet; If the Mahalanobis distance is greater than a preset anomaly threshold and the target sampling time is earlier than the expected droplet detachment time, the droplet transition state detection result of any video frame is determined to be abnormal, and an anomaly warning is triggered. The expected droplet detachment time is the sampling time corresponding to the droplet detachment in the normal transition template.

3. The method according to claim 2, characterized in that, The method further includes: If the droplet transition state detection result of any video frame is determined to be abnormal, the abnormal transition type is determined based on the deviation direction of each feature component in the droplet feature vector of any video frame relative to the corresponding standard feature information in the target droplet standard feature distribution information. The abnormal transition type includes one of the following: single-pulse multi-droplet anomaly, droplet accumulation anomaly, or short-circuit precursor anomaly. The single-pulse multi-droplet anomaly indicates that the molten droplet falls off before the expected detachment time; the molten droplet accumulation anomaly indicates that the equivalent diameter of the molten droplet is greater than the upper limit of the equivalent diameter and does not fall off at the expected detachment time; the short-circuit precursor anomaly indicates that the non-circularity of the molten droplet profile is less than the preset non-circularity threshold and the longitudinal displacement velocity of the center of gravity is greater than the upper limit of the velocity.

4. The method according to claim 3, characterized in that, The preset feature dimensions include at least one of the following: equivalent diameter, aspect ratio, contour non-circularity, longitudinal displacement velocity of the center of gravity, diameter change rate, and necking ratio; The step of determining the abnormal transition type based on the deviation direction of each feature component in the droplet feature vector of any video frame relative to the corresponding standard feature information in the target droplet standard feature distribution information includes: The mean and standard deviation of the standard equivalent diameter at the target sampling time, as well as the mean and standard deviation of the longitudinal displacement velocity of the standard centroid, are obtained from the standard characteristic distribution information of the target droplet. The upper and lower limits of the equivalent diameter are determined based on the mean and standard deviation of the standard equivalent diameter, and the upper limit of the velocity is determined based on the mean and standard deviation of the longitudinal displacement velocity of the standard centroid. If the equivalent diameter of the droplet in any video frame is less than the lower limit of the equivalent diameter, the abnormal transition type is determined to be a single-pulse multi-droplet anomaly. If the equivalent diameter of the droplet in any video frame is greater than the upper limit of the equivalent diameter and the rate of change of the equivalent diameter is greater than the preset rate of change threshold, the abnormal transition type is determined to be an abnormal droplet accumulation. When the longitudinal displacement velocity of the center of gravity of the droplet in any video frame is greater than the upper velocity limit, and the non-circularity of the droplet profile in any video frame is less than a preset non-circularity threshold, the abnormal transition type is determined to be a short-circuit precursor anomaly.

5. The method according to claim 1, characterized in that, The method further includes: The average grayscale value of each video frame in the welding video is determined to obtain a brightness sequence that varies with the frame number. The brightness sequence is subjected to a moving average filter to obtain a smooth brightness sequence; Based on the smoothed brightness sequence and the preset brightness discrimination threshold, the start frame and end frame of each pulse period are determined; The welding video is divided into multiple pulse cycles based on the start frame and the end frame.

6. The method according to claim 5, characterized in that, The method further includes: The target pulse period is obtained by determining the pulse period to which any video frame belongs from the plurality of pulse periods; The time position information is determined by the ratio of the cumulative number of frames from the start frame of the target pulse period to any video frame to a preset number of frames.

7. The method according to claim 5, characterized in that, The preset brightness discrimination threshold is determined using the following operation: Obtain a preset number of video frames from the welding video within a preset time period to obtain an initial video frame sequence; The maximum and minimum grayscale average values ​​are determined from the average grayscale values ​​of each initial video frame in the initial video frame sequence. The difference between the maximum and minimum grayscale average values ​​is multiplied by a preset weighting coefficient and then added to the minimum grayscale average value to obtain the preset brightness discrimination threshold.

8. The method according to claim 1, characterized in that, The normal transition template is determined using the following operation: Multiple sample video clips are acquired, wherein each sample video clip is a video clip showing the normal transition of the molten droplet within a complete pulse cycle; Determine the sample droplet feature vector of each sample video frame in each sample video segment to obtain the sample droplet feature sequence of each sample video segment; Align the time axis of each sample droplet feature sequence to obtain an aligned sample droplet feature sequence, wherein the aligned sample droplet feature sequence includes sample droplet feature vectors at multiple sampling times; For any sampling time, the mean vector and covariance matrix corresponding to any sampling time are determined based on the sample droplet feature vector of the aligned sample droplet feature sequence at that sampling time. The normal transition template is determined based on the mean vector and covariance matrix at each sampling time.

9. The method according to claim 1, characterized in that, The method further includes: After any pulse period ends, the droplet feature vectors and Mahalanobis distances of each video frame within that pulse period are statistically summarized to obtain the periodic statistical feature vector of that pulse period, wherein the periodic statistical feature vector includes at least one of the following statistical components: The average equivalent diameter of each video frame within any pulse period; The average aspect ratio of each video frame within any given pulse period; The mean of the non-circularity of the contour of each video frame within any pulse period; The sampling time corresponding to the droplet detachment within any pulse period; The maximum value of the Mahalanobis distance of each video frame within any pulse period; Cross-cycle trend analysis is performed on the periodic statistical feature vectors of each pulse cycle to obtain the cross-cycle trend analysis results.

10. The method according to claim 9, characterized in that, The cross-cycle trend analysis of the periodic statistical feature vectors of each pulse period yields the following results: Trend tests are performed on each statistical component in the periodic statistical feature vector of each pulse period within the sliding window to obtain the test statistics of each statistical component. The test statistics are then standardized to obtain the standardized statistics of each statistical component. If at least one of the statistical components has a standardized statistic greater than a preset trend threshold, a drift warning is generated, indicating that there is a drift trend in the welding process.

Images