Safety pre-detection method and system before electric welding operation of electric power system

Abstract

The invention relates to the field of behavior feature recognition, in particular to a safety pre-detection method and system before electric welding operation of an electric power system, and the method comprises the steps: obtaining the field information data of an operator; on the basis of the field information data, whether an operator prepares for electric welding operation or not is judged; under the condition that it is judged that the operator prepares for the electric welding operation, whether the detection site meets the safety operation requirement or not is judged; and the alarm device is started under the condition that the detection site does not meet the safety operation requirement. According to the embodiment of the invention, by optimizing the OpenPose skeleton point detection algorithm and improving the RT-DETR target detection model, the accuracy and stability of skeleton point detection are improved, the accuracy of small target recognition is improved, and the training cost is reduced, so that the real-time prediction of the electric welding operation process and the effective prevention of on-site potential safety hazards are realized.

Description

Technical Field

[0001] The present invention relates to the field of behavioral feature recognition, and in particular to a safety pre-inspection method and system before electric welding operation in a power system. Background Art

[0002] With the continuous development of power systems, welding operations play a vital role in the installation, maintenance, and overhaul of power equipment. However, welding operations are complex environments and present numerous safety hazards, such as improper operation and equipment failure, which can lead to serious accidents. Therefore, there is an urgent need for a technology that can predict welding operations in real time and prevent safety hazards on site. Recognizing the behavioral characteristics of on-site workers is crucial for this technology. The OpenPose algorithm for human skeleton point detection and the RT-DETR object detection model are commonly used to identify these characteristics.

[0003] OpenPose, a widely used algorithm for human skeleton detection, can detect the 2D poses of multiple people in an image in real time. However, in the complex environment of electric welding operations in power systems, OpenPose suffers from several limitations. Occlusion is a serious issue: welding operations are complex, with frequent camera angles and occlusions. Parts of the worker's body may be hidden, resulting in the inability to detect certain skeleton points. OpenPose simply replaces missing skeleton point coordinates with zero, lacking effective processing methods, which affects subsequent action recognition. Furthermore, skeleton keypoint recognition jitter is a problem: OpenPose is susceptible to changes in lighting, occlusion, and pose when detecting skeleton points. This can cause keypoint jitter, unstable detection, and uneven skeleton trajectories, which in turn affects the stability of subsequent feature extraction. RT-DETR is a real-time object detection model based on the Transformer architecture, offering high accuracy and real-time performance. However, RT-DETR also has several limitations in electric welding operations in power systems. Small objects are difficult to detect: Welding cameras are often mounted at high altitudes, and the captured videos contain numerous small objects (such as welding guns and welding rods). The traditional IoU metric used by RT-DETR is ineffective when evaluating small objects, affecting recognition accuracy. Furthermore, training costs are high: While the RT-DETR model significantly improves accuracy by applying the Transformer, its high number of parameters and computational cost also result in higher training costs than typical models.

[0004] Therefore, the limitations of the human skeleton point detection algorithm OpenPose and the target detection model RT-DETR have greatly affected the accuracy and stability of the behavioral characteristics of on-site workers, restricting the development of intelligent prediction and hidden danger prevention technologies for welding operations. Summary of the Invention

[0005] One of the objectives of the present invention is to provide a method and system for safety pre-inspection before electric welding operations in power systems. This method and system improve the accuracy and stability of skeleton point detection by optimizing the OpenPose skeleton point detection algorithm and improving the RT-DETR target detection model, while also improving the accuracy of small target recognition and reducing training costs, thereby achieving real-time prediction of the welding operation process and effective prevention of on-site safety hazards.

[0006] In order to achieve the above object, the present invention provides a method and system for safety pre-inspection before electric welding operation in a power system, comprising:

[0007] Obtain on-site information data of operators;

[0008] Based on the on-site information data, determining whether the operator is ready for welding operation;

[0009] When judging whether the operator is ready for welding operation, determine whether the inspection site meets the safety operation requirements;

[0010] If it is determined that the inspection site does not meet the safety operation requirements, the alarm device will be activated.

[0011] Optionally, based on the on-site information data, determining whether the operator is preparing for welding operation includes:

[0012] Using a human body posture estimation algorithm to extract skeleton data from the scene information data to obtain information on skeleton key points;

[0013] Using triple interpolation method to complete the data of missing points of the skeleton key points;

[0014] Eliminate jitter on the completed skeleton key points;

[0015] Extract working features based on skeleton key points after eliminating jitter;

[0016] Training an action recognition model based on the extracted work features, and determining whether the operator is holding a welding gun or a welding rod through the trained action recognition model;

[0017] When it is determined that the operator is performing an action of holding a welding gun or a welding rod, the operator performs electric welding.

[0018] Optionally, using a triple interpolation method to complete the data of missing points of the skeleton key points includes:

[0019] Performing symmetrical joint point interpolation on the skeleton key points, including: interpolating corresponding points detectable at the same limb position on the other half of the body using the spinal midline as a vertical symmetry axis;

[0020] Performing asymmetric joint point interpolation on the skeleton key points, including:

[0021] When the frames before and after the missing point can be detected, the coordinate value of the missing point is calculated according to formula (1):

[0022]

[0023] Among them, P m (t+i) is the coordinate value of the missing point, P(t) is the last detectable position coordinate before a key point is missing, i is the position number of the frame counting from the tth frame, t+n is the position number of the first detected frame after the frame where the missing point is located, P(t+n) is the coordinate value of the position number t+n, n is the number of missing points, and T is the total number of frames for each action;

[0024] When only the frame before the missing point can be detected, the coordinate value of the missing point is calculated according to formula (2):

[0025]

[0026] Among them, P m (t+i) is the coordinate of the missing point, and P(t-4) is the coordinate value of position number t-4;

[0027] In the case that only the frame after the missing point can be detected, the coordinate value of the missing point is calculated according to formula (3):

[0028]

[0029] Among them, P m (ti) is the coordinate of the missing point, and P(t+4) is the coordinate value of position number t+4;

[0030] Performing facial missing point interpolation operation on the skeleton key points, wherein the facial missing point interpolation operation method includes: calculating the coordinate value of the missing point according to formula (4) to formula (7):

[0031]

[0032] Among them, E k is the coordinate of the eye, M k are the coordinates of the midpoint between the ear and the nose, For E k Pointing to M k The vector of , j is the modulus, is the horizontal coordinate of the unit vector, is the ordinate of the unit vector, is a vector The horizontal axis, is a vector The vertical coordinate, is the horizontal coordinate of the ear, x N is the horizontal coordinate of the nose, is the vertical coordinate of the ear, y N is the vertical coordinate of the nose, k=1,2.

[0033] Optionally, performing de-jitter processing on the completed skeleton point data includes:

[0034] Calculate the smoothing factor according to formula (8) to formula (10),

[0035]

[0036] Among them, α t is the smoothing factor, r is the smoothing factor adjustment parameter, f C is the dynamic cutoff frequency, Δt is the time interval between adjacent frames, f min is the minimum cutoff frequency, β is the adjustment coefficient, is the rate of change of key point coordinates after filtering, dx is the rate of change of input key point coordinates, is the rate of change of key point coordinates after filtering in the previous frame, α d is the smoothing factor calculated with a fixed cutoff frequency;

[0037] Calculate the filtered signal according to formula (11):

[0038]

[0039] in, is the filtered keypoint coordinate of the current frame, x t is the original keypoint coordinate of the current frame, are the filtered keypoint coordinates of the previous frame.

[0040] Optionally, the minimum cutoff frequency and the adjustment coefficient can be dynamically adjusted, wherein the dynamic adjustment method of the minimum cutoff frequency includes: dynamically adjusting the minimum cutoff frequency according to formula (12) to formula (15),

[0041]

[0042] Among them, f m ′ m in is the adjusted minimum cutoff frequency, l1 is the first learning rate, E j is the jitter error of the jth window, E j+1 is the jitter error of the j+1th window, E is the jitter error of the window, e iis the jitter at the i-th moment, u is the average error in the sub-window, is the filtered key point coordinate value at the i-th moment, x i is the coordinate value of the key point before filtering at the i-th moment;

[0043] The dynamic adjustment method of the adjustment coefficient includes: dynamically adjusting the adjustment coefficient according to formula (16),

[0044]

[0045] Among them, β ′ is the adjusted regulation coefficient, l2 is the second learning rate, is the variance of the delay degree of the i-th sub-window, is the variance of the delay degree of the i+1th window.

[0046] Optionally, the working features extracted based on the de-jittered skeleton key point data include:

[0047] Extract the coordinate features of the key points of the skeleton;

[0048] According to formula (17), the directional features of the skeleton key points are extracted.

[0049]

[0050] Among them, f ori is the directional feature of the skeleton key points, is the y coordinate of the i-th bone key point, is the y coordinate of the first bone key point, is the x-coordinate of the i-th bone key point, is the x-coordinate of the first skeleton key point, and G is the selected key point group;

[0051] According to formula (18), the distance features of the skeleton key points are extracted.

[0052]

[0053] Among them, f dis is the distance of the key points of the skeleton, is the x-coordinate of the j-th bone key point, is the x coordinate of the kth skeleton key point, P j y is the y coordinate of the jth bone key point, is the y coordinate of the kth skeleton key point, and V is the selected key point group;

[0054] According to formula (19), the trajectory features of the skeleton key points are extracted.

[0055]

[0056] in, is the displacement vector of the lth skeleton key point in frame t+1, is the x coordinate of the lth skeleton key point in frame t+1, is the x coordinate of the lth skeleton key point in frame t, is the y coordinate of the lth skeleton key point in the t+1 frame, is the y coordinate of the lth bone key point in frame t.

[0057] Optionally, training an action recognition model based on the extracted work features, and determining whether the operator is holding a welding gun or a welding rod by using the trained action recognition model includes:

[0058] The working features are used to train the LSTM network to classify the actions;

[0059] Inputting the on-site information data of the operator into the trained LSTM network to obtain the probability distribution of the action category;

[0060] Whether to perform a corresponding action is determined based on the probability distribution and the probability threshold.

[0061] Optionally, when determining whether the operator is preparing for welding operation, determining whether the inspection site meets the safety operation requirements includes:

[0062] Using a data annotation tool to annotate the on-site information data;

[0063] Build an object detection model and optimize the loss function;

[0064] Determine whether the safety equipment at the inspection site is complete based on the optimized target detection model;

[0065] If it is judged that the safety equipment at the testing site is not complete, the testing site does not meet the safety operation requirements;

[0066] If it is determined that the safety equipment at the inspection site is complete, determine whether the operator has taken the welding gun and welding rod;

[0067] When it is determined that the operator is holding a welding gun and a welding rod, determining whether the operator is wearing fireproof gloves correctly;

[0068] If it is determined that the operator did not wear fire-resistant gloves correctly, the inspection site does not meet the safety operation requirements;

[0069] If it is determined that the operator is wearing fireproof gloves correctly, the operator is continuously checked after an interval of 5 seconds to see whether the operator is wearing a protective mask correctly;

[0070] If the operator is not wearing the protective mask correctly after continuous detection at an interval of 5 seconds, the detection site does not meet the safety operation requirements.

[0071] Optionally, building an object detection model and optimizing the loss function includes:

[0072] The Gaussian distribution of the bounding box is modeled according to formulas (20) to (22).

[0073]

[0074] Among them, μ is the mean vector, Σ is the covariance matrix, c x is the horizontal coordinate of the center point of the bounding box, c y is the ordinate of the center point of the bounding box, σ w is the initial standard deviation of width, σ h is the initial standard deviation of height, w is the width of the bounding box, and h is the height of the bounding box;

[0075] According to formula (23) to formula (25), a new dynamic covariance matrix is ​​obtained.

[0076]

[0077]

[0078] Among them, λ w is the width dynamic scaling factor, λ h is the height dynamic scaling factor, a1 is the width scaling amplitude control parameter, and a2 is the height scaling amplitude control parameter;

[0079] Calculate the loss function according to formula (26) to formula (28),

[0080]

[0081] Among them, L NWD is the loss function, is the similarity measure, For two Gaussian distributions and The Gaussian distance between them, C is the normalization constant, ‖·‖ F is the Frobenius norm, μ1 is the Gaussian distribution The mean vector of μ2 is Gaussian distribution. The mean vector, Σ ′ 1 is Gaussian distribution The dynamic covariance matrix, Σ′ 2 is Gaussian distribution The dynamic covariance matrix of .

[0082] On the other hand, the present invention further provides a safety pre-inspection system before electric welding operation in a power system, wherein the system includes a processor configured to execute any of the above-mentioned methods.

[0083] Beneficial effects of the present invention:

[0084] The embodiments of the present invention improve the accuracy and stability of skeleton point detection by optimizing the OpenPose skeleton point detection algorithm and improving the RT-DETR target detection model, while also enhancing the precision of small target recognition and reducing training costs, thereby achieving real-time prediction of the welding process and effective prevention of on-site safety hazards.

[0085] Other features and advantages of the embodiments of the present invention will be described in detail in the subsequent detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] The accompanying drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the following detailed description, they are used to explain the embodiments of the present invention, but do not constitute a limitation of the embodiments of the present invention. In the accompanying drawings:

[0087] Figure 1 is a flow chart of a method for safety pre-inspection before electric welding operation in a power system according to one embodiment of the present invention;

[0088] Figure 2 is a flowchart of a method for determining whether an operator is ready for welding work according to one embodiment of the present invention;

[0089] Figure 3 is a schematic diagram of facial interpolation according to one embodiment of the present invention;

[0090] Figure 4 2 is a schematic diagram of the filtering effect of skeleton key points according to one embodiment of the present invention;

[0091] Figure 5 is a flow chart of a working feature extraction method according to one embodiment of the present invention;

[0092] Figure 6 1 is a schematic diagram of skeletal key point directional feature extraction according to one embodiment of the present invention;

[0093] Figure 7 1 is a schematic diagram of skeleton key point distance feature extraction according to one embodiment of the present invention;

[0094] Figure 8is a flowchart of a method for determining whether an operator is holding a welding gun or a welding rod according to one embodiment of the present invention;

[0095] Figure 9 is a flow chart of a method for determining whether a detection site meets safety operation requirements according to one embodiment of the present invention;

[0096] Figure 10 2 is a schematic diagram of detecting the correct wearing of protective equipment according to an embodiment of the present invention. DETAILED DESCRIPTION

[0097] The following describes the embodiments of the present invention with reference to the accompanying drawings and preferred embodiments. Those skilled in the art will readily appreciate the other advantages and benefits of the present invention from the disclosure herein. The present invention may also be implemented or applied through various other specific embodiments, and the various details in this specification may be modified or altered based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are intended only to illustrate the present invention and are not intended to limit the scope of protection of the present invention.

[0098] It should be noted that the illustrations provided in the following embodiments are merely schematic illustrations of the basic concept of the present invention. Therefore, the illustrations only show components related to the present invention and are not drawn according to the number, shape, and size of components in actual implementation. In actual implementation, the type, quantity, and proportion of each component may be changed arbitrarily, and the component layout may also be more complex.

[0099] like Figure 1 Shown is a flow chart of a method for safety pre-inspection before welding operation in a power system according to one embodiment of the present invention. Figure 1 In the method, the method may include the following steps:

[0100] In step S10, the on-site information data of the operator is obtained;

[0101] In step S11, based on the on-site information data, it is determined whether the operator is ready for welding operation;

[0102] In step S12, when it is determined that the operator is preparing for welding work, it is determined whether the inspection site meets the safety operation requirements;

[0103] In step S13, when it is determined that the inspection site does not meet the safety operation requirements, the alarm device is activated.

[0104] In this Figure 1In the illustrated method for pre-welding safety checks on power systems, step S10 is used to obtain on-site operator information data. Since the present invention aims to perform pre-welding safety checks on operators, it is necessary to obtain operator behavior information and on-site safety equipment information. In this embodiment, multiple video clips of the worksite can be captured using a dome camera, including operator behavior data and safety equipment information data.

[0105] Step S11 is used to determine whether the operator is preparing for welding. When the operator is identified as holding a welding gun or a welding rod, it is considered that the operator is preparing for welding. The specific method for determining whether the operator is preparing for welding can be various forms known to those skilled in the art. In one embodiment of the present invention, the method may include: Figure 2 The steps shown. Figure 2 In the step S11, the following steps may be included:

[0106] In step S20, a human body posture estimation algorithm is used to extract skeleton data from the scene information data to obtain information on skeleton key points;

[0107] In step S21, a triple interpolation method is used to complete the data of the missing points of the skeleton key points;

[0108] In step S22, the completed skeleton key points are subjected to de-jittering processing;

[0109] In step S23, working features are extracted based on the skeleton key points after the jitter is eliminated;

[0110] In step S24, an action recognition model is trained based on the extracted work features, and the trained action recognition model is used to determine whether the operator is holding a welding gun or a welding rod.

[0111] In step S25 , when it is determined that the worker is holding a welding gun or a welding rod, the worker performs electric welding.

[0112] In this Figure 2 In the method shown, step S20 is used to extract human skeleton key points using a pre-trained human pose estimation algorithm model, and record the extracted human skeleton key point coordinate data for each frame of image. In this example, the human pose estimation algorithm model can be OpenPose.

[0113] Step S21 is used to use the triple interpolation method to complete the data of the missing points of the skeleton key points. Due to the complex on-site environment of the electric welding operation, and due to the influence of the viewing angle of the ball camera and occlusion, part of the operator's body may be hidden and cannot be detected. The current mainstream skeleton point detection algorithm OpenPose only replaces the missing coordinates with 0 for the undetectable skeleton points, and there is no good way to deal with it, which will affect the subsequent action recognition. Therefore, in order to solve the problem of missed detection of skeleton key points due to occlusion problems, the triple interpolation method is used to complete the skeleton key points. In this embodiment, the triple interpolation method can be an interpolation for symmetrical joint missing points, asymmetrical joint missing points, and facial missing points when the operator is facing away from the camera. Specifically, the method of using the triple interpolation method to complete the missing point data of skeleton key points can include the following steps:

[0114] In step S30, symmetrical joint point interpolation is performed on the skeleton key points, including: interpolation is performed by taking the spinal midline as the vertical symmetry axis and combining corresponding points that can be detected at the same limb position on the other half of the body;

[0115] In step S31, asymmetric joint point interpolation is performed on the skeleton key points, including:

[0116] When the frames before and after the missing point can be detected, the coordinate value of the missing point is calculated according to formula (1):

[0117]

[0118] Among them, P m (t+i) is the coordinate value of the missing point, P(t) is the last detectable position coordinate before a key point is missing, i is the position number of the frame counting from the tth frame, t+n is the position number of the first detected frame after the frame where the missing point is located, P(t+n) is the coordinate value of the position number t+n, n is the number of missing points, and T is the total number of frames for each action;

[0119] When only the frame before the missing point can be detected, the coordinate value of the missing point is calculated according to formula (2):

[0120]

[0121] Among them, P m (t+i) is the coordinate of the missing point, and P(t-4) is the coordinate value of position number t-4;

[0122] In the case that only the frame after the missing point can be detected, the coordinate value of the missing point is calculated according to formula (3):

[0123]

[0124] Among them, P m (ti) is the coordinate of the missing point, and P(t+4) is the coordinate value of position number t+4;

[0125] In step S32, a facial missing point interpolation operation is performed on the skeleton key points, wherein the facial missing point interpolation operation method includes: calculating the coordinate values ​​of the missing points according to formulas (4) to (7):

[0126]

[0127]

[0128] Among them, E k is the coordinate of the eye, M k are the coordinates of the midpoint between the ear and the nose, For E k Pointing to M k The vector of , j is the modulus, is the horizontal coordinate of the unit vector, is the ordinate of the unit vector, is a vector The horizontal axis, is a vector The vertical coordinate, is the horizontal coordinate of the ear, x N is the horizontal coordinate of the nose, is the vertical coordinate of the ear, y N is the vertical coordinate of the nose, k=1,2.

[0129] Step S30 is used to perform corresponding point interpolation using body symmetry in the case where only one side of a symmetrical joint point is covered.

[0130] Step S31 is used to interpolate for the case where all symmetrical joint points are covered and the left and right hand movements are inconsistent. Specifically, in the case where the frames before and after the missing point can be detected: there are a finite number of n frames with missing points between two frames with detectable points P, assuming that the coordinate values ​​between the first and second detectable frames increase at a constant speed, linear interpolation is used to calculate the missing point P. m The coordinate values ​​of the missing point are calculated using the formula (1). In the case where only the frame before the missing point can be detected: add the values ​​obtained from the five detectable frames before the missing point and use linear interpolation to calculate the value of the missing point. The calculation formula is (2). In the case where only the frame after the missing point can be detected: add the values ​​obtained from the five detectable frames after the missing point and use linear interpolation to calculate the value of the missing point. The calculation formula is (3).

[0131] Step S32 is used to interpolate the situation when the operator is facing away from the camera. When the operator is facing away from the camera, OpenPose cannot identify the key points of the face because the operator's facial information cannot be detected. According to actual observations of the human face structure, the positions of the nose, ears, and eyes form an isosceles triangle on the left and right sides of the face. Therefore, when the missing points are both eyes and nose, this structural relationship is used for interpolation. In this example, based on common sense in daily operations, it is assumed that the position of the eyes in two-dimensional space is always above the ears and nose. Figure 3 In the facial interpolation diagram shown, R1 and R2 are ears, N is the nose, and E1 and E2 are eyes. Assuming R1M1 = M1N = NM2 = R2M2, line segments E1M1 and E2M2 always maintain a perpendicular relationship with R1N and R2N respectively, and each eye and nose point forms a triangle with the ear point. Point M k The coordinates are When k=1, M1 represents the left side, and when k=2, it represents the right side. Based on the vertical properties of the two-dimensional vector and the eye E k Assuming that it is always above R1N and R2N, the vector By swapping the unit vectors , and set a negative sign on the y-axis, and then multiply it by the corresponding modulus length to calculate, In the Cartesian coordinate system, the direction is And through The unit vector of the non-zero vector normalized by the length of . In this example, the length of the modulus may be 1.2 m, where m is one quarter of the distance between the ears R1 and R2.

[0132] Step S22 is used to eliminate jitter processing of the completed skeleton key points. Since the OpenPose algorithm may have jitter problems in skeleton key point recognition under complex operating environments, it is necessary to eliminate jitter processing of the skeleton key points to improve recognition accuracy. In this embodiment, the steps of the jitter elimination processing method can be various forms known to those skilled in the art, including but not limited to methods such as Gaussian filtering and Kalman filtering. In one example of the present invention, the real-time smoothing capability of the OneEuro filter is used to eliminate jitter. Specifically, step S22 may include the following steps:

[0133] In step S40, the smoothing factor is calculated according to formula (8) to formula (10),

[0134]

[0135]

[0136] Among them, α tis the smoothing factor, r is the smoothing factor adjustment parameter, f C is the dynamic cutoff frequency, Δt is the time interval between adjacent frames, f min is the minimum cutoff frequency, β is the adjustment coefficient, is the rate of change of key point coordinates after filtering, dx is the rate of change of input key point coordinates, is the rate of change of key point coordinates after filtering in the previous frame, α d is the smoothing factor calculated with a fixed cutoff frequency;

[0137] In step S41, the filtered signal is calculated according to formula (11),

[0138]

[0139] in, is the filtered keypoint coordinate of the current frame, x t is the original keypoint coordinate of the current frame, are the filtered keypoint coordinates of the previous frame.

[0140] Step S40 is used to calculate the smoothing factor α t , the smoothing factor determines the weighted ratio of the current signal to the past signal. The calculation formula of the dynamic cutoff frequency is Among them, f min is the minimum cutoff frequency, β is the adjustment coefficient, in this example, f min The values ​​of f and β can be set based on empirical values. min =1.0, β=0.007. To calculate the dynamic cutoff frequency, so that the dynamic cutoff frequency changes with the rate of change of the key point coordinates Dynamic adjustment improves the response speed when the signal changes rapidly to avoid excessive delay, and reduces the frequency when the signal changes slowly to remove high-frequency noise. The smoothing calculation formula for the key point coordinate change rate is Among them, α d By fixing the cutoff frequency The smoothing factor is calculated. The signal change rate is smoothed by this formula to reduce the influence of noise in the signal change rate on the cutoff frequency adjustment. Step S41 is used to calculate the filtered signal. In the filtering process, by reducing the minimum cutoff frequency f min It can significantly reduce jitter, but there will be a delay, while increasing the speed coefficient β can reduce the delay, but the de-jitter effect will be greatly reduced. Therefore, in order to balance the signal smoothness and delay, it is necessary to adjust f min In order to combine the actual scene requirements and enhance the adaptability of the filter, it is necessary to introduce an adaptive parameter adjustment mechanism so that f minand β change dynamically, so that better filtering effect can be obtained in different situations. min The steps of β and β can be various forms known to those skilled in the art. In one example of the present invention, the steps can include:

[0141] In step S50, the minimum cutoff frequency is dynamically adjusted according to formulas (12) to (15).

[0142]

[0143]

[0144] Among them, f′ min is the minimum cutoff frequency after adjustment, l1 is the first learning rate, E j is the jitter error of the jth window, E j+1 is the jitter error of the j+1th window, E is the jitter error of the window, e i is the jitter at the i-th moment, u is the average error in the sub-window, is the filtered key point coordinate value at the i-th moment, x i is the coordinate value of the key point before filtering at the i-th moment;

[0145] In step S51, the adjustment coefficient is dynamically adjusted according to formula (16).

[0146]

[0147] Among them, β ′ is the adjusted regulation coefficient, l2 is the second learning rate, is the variance of the delay degree of the i-th sub-window, is the variance of the delay degree of the i+1th window.

[0148] Step S50 is used to determine the minimum frequency f min Specifically, in this example, a sliding window of size 10 is taken and divided into two sub-windows. The jitter variance E of each sub-window is calculated. The jitter variance reflects the degree of fluctuation of the jitter value. Then the change of the jitter variance between adjacent sub-windows is calculated. When the change exceeds the jitter change threshold, the adjustment mechanism is triggered. According to the formula To dynamically adjust f min In this example, the jitter change threshold may be 10%, and the first learning rate l1 may be set to 0.05.

[0149] Step S51 is used to dynamically adjust the adjustment coefficient. Specifically, in this example, a sliding window of size 10 is taken and divided into 2 sub-windows, and the variance of the delay degree change of each sub-window is calculated. delay , which is the time offset between the original signal and the filtered signal. This offset reflects the tracking delay of the filter. Similarly, the variance of the delay between each sub-window is compared. When the delay change threshold is exceeded, the adjustment mechanism is triggered. According to the formula To dynamically adjust β. In this example, the delay change threshold can be 10%, and the second learning rate l2 can be set to 0.01. Figure 4 The filtering effect diagram shows the filtering effect of the optimized OneEuro.

[0150] Step S23 is used to extract work features based on the skeleton key point data after the jitter is eliminated. Work features are parameters that can reflect the behavior of the operator. In this embodiment, the steps for extracting work features can be various forms known to those skilled in the art. In this embodiment, step S23 can include the following: Figure 5 The steps shown in Figure 5 In the step S23, the following steps may be performed:

[0151] In step S60, the coordinate features of the skeleton key points are extracted;

[0152] In step S61, the directional features of the skeleton key points are extracted according to formula (17):

[0153]

[0154] Among them, f ori is the directional feature of the skeleton key points, is the y coordinate of the i-th bone key point, is the y coordinate of the first bone key point, is the x-coordinate of the i-th bone key point, is the x-coordinate of the first skeleton key point, and G is the selected key point group;

[0155] In step S62, the distance features of the skeleton key points are extracted according to formula (18),

[0156]

[0157] Among them, f dis is the distance of the key points of the skeleton, is the x-coordinate of the j-th bone key point, is the x-coordinate of the k-th skeleton key point, is the y coordinate of the jth bone key point, is the y coordinate of the kth skeleton key point, and V is the selected key point group;

[0158] In step S63, the trajectory features of the skeleton key points are extracted according to formula (19),

[0159]

[0160] in, is the displacement vector of the lth skeleton key point in frame t+1, is the x coordinate of the lth skeleton key point in frame t+1, is the x coordinate of the lth skeleton key point in frame t, is the y coordinate of the lth skeleton key point in the t+1 frame, is the y coordinate of the lth bone key point in frame t.

[0161] In this Figure 5 In the method shown, step S60 is used to extract coordinate features. The coordinate values ​​of the 25 processed skeleton key points (P x ,P y ) to represent the position of the limbs and torso, thereby obtaining basic spatial information from the skeleton and extracting direction, distance and trajectory features based on these coordinate values.

[0162] Step S61 is used to extract directional features. In this example, the radian value is extracted as the directional feature based on the coordinates of the skeleton key points. Figure 6 As shown, the neck position (point 1) is set as the axis point, according to the formula To calculate the angles between the left arm key points 2, 3, 4 and the right arm key points 5, 6, 7 and the axis point, where G is the selected key point group, including key points 2, 3, 4, 5, 6 and 7.

[0163] Step S62 is used to extract distance features. In this example, the distance feature includes two parts, such as Figure 7 As shown in the figure, the first is the change in distance between skeleton key points, which reflects the characteristics of the action and involves key points 1-9 and 12. The second is the change in distance between hand key points 4 and 7 and the center position of the welding gun and welding rod, which is used to assist in verification. The center position of the welding gun and welding rod is detected by the subsequent object detection model. Figure 7 The medium yellow lines represent the distance edges between the selected keypoints.

[0164] Step S63 is used to extract trajectory features. In this example, the trajectory features describe the movement of key points in the time series. For each key point, its displacement vector between adjacent frames is calculated. Assume that the coordinates of key point i in frame t and frame t+1 are respectively and Then the displacement vector

[0165] Step S24 is used to train the action recognition model based on the extracted work features, and to determine whether the operator is holding the welding gun or the welding rod through the trained action recognition model. In this embodiment, the method of using the action recognition model to determine the operator's behavior can be various forms known to those skilled in the art. In one example of the present invention, step S24 can include the following: Figure 8 The steps shown in Figure 8 In the step S24, the following steps may be performed:

[0166] In step S70, the LSTM network is trained using the working features to classify the actions;

[0167] In step S71, the on-site information data of the operator is input into the trained LSTM network to obtain the probability distribution of the action category;

[0168] In step S72 , whether to perform a corresponding action is determined based on the probability distribution and the probability threshold.

[0169] In this Figure 8 In the illustrated method, step S70 is used to train the LSTM network and classify actions. The input to the LSTM is a three-dimensional tensor with the dimensions [number of samples, time steps, number of features]. Here, number of samples is the total number of training samples, time steps is the number of frames in each action sequence, and number of features is the dimensionality of the extracted features.

[0170] Step S71 is used to obtain the probability distribution of the operator's action categories. After extracting the four work characteristics of the operator, they are input into the trained LSTM network model to obtain a prediction result, which is a probability distribution. For example, prediction = [0.9, 0.1] indicates that the probability of the first category is the highest, that is, the recognized action is category 1. This recognition process is continuous. The video captured by the dome camera is continuously passed through the model. Using a sliding window, the model detects 60 frames of data at a time, and then moves 30 frames at a time for further detection.

[0171] Step S72 is used to determine whether to perform a corresponding action based on the probability distribution and the probability threshold.

[0172] After the action determination in step S24, if it is determined in step S25 that the operator is taking the welding gun or the welding rod, it means that the operator is performing electric welding.

[0173] Step S12 is used to determine whether the inspection site meets the safety operation requirements. Safe operation requirements include the safety equipment being complete, the workers wearing fireproof gloves correctly, and the workers wearing protective masks correctly. In this embodiment, the method for determining whether the inspection site meets the safety operation requirements can be various forms known to those skilled in the art. In one example of the present invention, step S12 can include: Figure 9 The steps shown. Figure 9 In the step S12, the following steps may be included:

[0174] In step S80, a data annotation tool is used to annotate the on-site information data;

[0175] In step S81, a target detection model is constructed and a loss function is optimized;

[0176] In step S82, it is determined whether the safety equipment at the detection site is complete based on the optimized target detection model;

[0177] In step S83, if it is determined that the safety equipment at the testing site is not complete, the testing site does not meet the safety operation requirements;

[0178] If the safety equipment at the inspection site is complete, determine whether the operator is holding the welding gun and welding rod;

[0179] In step S84, when it is determined that the operator is holding the welding gun and the welding rod, it is determined whether the operator is wearing fireproof gloves correctly;

[0180] If it is determined that the workers are not wearing fire-resistant gloves correctly, the inspection site does not meet the safety operation requirements;

[0181] In step S85, if it is determined that the worker is wearing fireproof gloves correctly, the operator is continuously checked after an interval of 5 seconds to see whether the worker is wearing a protective mask correctly;

[0182] In step S86, if the worker is not wearing the protective mask correctly after continuous detection for 5 seconds, the detection site does not meet the safety operation requirements.

[0183] In this Figure 9 In the illustrated method, step S80 is used to label the on-site information data. In this embodiment, the objects to be detected include fire masks, fire gloves, welding torches, welding rods, fire extinguishers, and fire blankets. Specifically, in this example, the data collected by the on-site dome camera can be labeled using LabelImg software, and the data format can be YOLO format.

[0184] Step S81 is used to build a target detection model and optimize the loss function. In order to improve the accuracy of small target positioning and recognition, in this embodiment, the traditional IOU loss function can be replaced by a regression loss function NWD based on normalized Gaussian distance. At the same time, a dynamic scaling factor is introduced to dynamically adjust the covariance matrix. Therefore, in this embodiment, the method of building a target detection model and optimizing the loss function can include the following steps:

[0185] In step S90, Gaussian distribution modeling of the bounding box is performed according to formulas (20) to (22).

[0186]

[0187] Among them, μ is the mean vector, Σ is the covariance matrix, c x is the horizontal coordinate of the center point of the bounding box, c y is the ordinate of the center point of the bounding box, σ w is the initial standard deviation of width, σ h is the initial standard deviation of height, w is the width of the bounding box, and h is the height of the bounding box;

[0188] In step S91, a new dynamic covariance matrix is ​​obtained according to formula (23) to formula (25),

[0189]

[0190] Among them, λ w is the width dynamic scaling factor, λ h is the height dynamic scaling factor, a1 is the width scaling amplitude control parameter, and a2 is the height scaling amplitude control parameter;

[0191] In step S92, the loss function is calculated according to formula (26) to formula (28),

[0192]

[0193] Among them, L NWD is the loss function, is the similarity measure, For two Gaussian distributions and The Gaussian distance between them, C is the normalization constant, ‖·‖ F is the Frobenius norm, μ1 is the Gaussian distribution The mean vector of μ2 is Gaussian distribution. The mean vector of Σ′1 is Gaussian distribution The dynamic covariance matrix, Σ ′ 2 is Gaussian distribution The dynamic covariance matrix of .

[0194] Step S90 is used to model Gaussian distribution. x ,c y ,w,h) is modeled as a two-dimensional Gaussian distribution. Step S91 is used to introduce a dynamic scaling factor λ w and λ h , forming a new dynamic covariance matrix, where the scaling factor is dynamically generated by the target size: Among them, a1 is the width scaling amplitude control parameter, and a2 is the height scaling amplitude control parameter. In this embodiment, a1 and a2 can be set to 0.5. After the dynamic scaling factor is introduced, NWD is upgraded from a static distribution metric to a dynamic learnable metric tool, which gives a larger height variance to slender objects such as welding guns and welding rods, and the recognition effect is better. The loss function is then calculated in step S92. In this example, the detection box and the true box are A=(c xa ,c ya ,w a ,h a ) and B=(c xb ,c yb ,w b ,h b ), the Gaussian distribution after modeling is and Calculate the Gaussian distance between these two bounding boxes Since the calculated distance metric cannot be used directly for similarity, the normalized index is used to obtain the similarity metric. Where C is the normalization constant. In this example, C can be set to 12.8. Finally, calculate the loss function L NWD In order to reduce the number of model parameters, the model is lightweighted. The specific steps of the lightweight improvement can be various forms known to those skilled in the art. In one example of the present invention, the standard convolution in the Resnet18 backbone network can be replaced by a depthwise separable convolution. Specifically, the standard convolution parameter calculation formula is: N SC =3·3·C1·C2. The calculation formula of the lightweight improved depth-separable convolution parameter is: N DSC =C1×3×3+C1×1×1×C2, where C1 and C2 are the number of input and output channels respectively. After calculation, the number of parameters of Resnet18 is about 1.1×10 7 After replacement, the number of parameters is reduced to 1.2×10 6 , reducing training costs.

[0195] Step S82 is used to determine whether the safety equipment at the inspection site is complete based on the optimized target detection model. The improved RT-DETR model is used to detect whether there are fire masks, fire gloves, fire extinguishers and fire blankets at the work site. Furthermore, fire masks and fire gloves are protective equipment. If they are not detected, a protective equipment missing alarm is triggered; fire extinguishers and fire blankets are fire-fighting equipment. If they are not detected, a fire-fighting equipment missing alarm is triggered. Step S83 is used to determine whether the operator performs the two actions of holding the welding gun and the welding rod at the same time. Step S84 is used to determine whether the operator wears fire-fighting gloves correctly. When both the actions of holding the welding gun and the welding rod are detected, it is determined that the electric welding operation is about to begin. Considering that there may be a situation where the protective equipment is available but not worn or worn incorrectly, the correct wearing of the fire-fighting gloves is first detected, such as Figure 10 As shown in the diagram for detecting the correct wearing of protective equipment, the operator is determined to be wearing the fire-resistant gloves correctly by checking whether key points 4 and 7 on the hand are within the fire-resistant glove detection frame. If the detection frame does not completely cover key points 4 and 7, an alarm is issued. Then, behavioral detection continues through steps S85 and S86. After a 5-second interval, the operator begins to determine whether the fire mask is being worn correctly by checking whether key points 0, 15, and 16 on the face are within the fire mask detection frame. If not, an alarm is issued, reminding the operator to wear a fire mask.

[0196] Step S13 is used to activate the alarm device when it is determined that the inspection site does not meet the safety operation requirements. Based on the judgment of steps S10 to S12, an alarm reminder of possible safety hazards is implemented before the welding operation occurs.

[0197] On the other hand, an embodiment of the present invention further provides a safety pre-inspection system before electric welding operation in a power system, the system comprising a processor configured to execute any of the above methods.

[0198] Beneficial effects of the present invention:

[0199] The embodiments of the present invention improve the accuracy and stability of skeleton point detection by optimizing the OpenPose skeleton point detection algorithm and improving the RT-DETR target detection model, while also enhancing the precision of small target recognition and reducing training costs, thereby achieving real-time prediction of the welding process and effective prevention of on-site safety hazards.

[0200] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0201] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0202] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0203] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0204] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0205] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0206] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0207] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0208] The above are merely embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.

Claims

1. A method for safety pre-inspection before electric welding operation in a power system, characterized in that: The method comprises: Obtain on-site information data of operators; Based on the on-site information data, determining whether the operator is ready for welding operation; When judging whether the operator is ready for welding operation, determine whether the inspection site meets the safety operation requirements; If it is determined that the inspection site does not meet the safety operation requirements, the alarm device will be activated.

2. The method according to claim 1, characterized in that Based on the on-site information data, determining whether the operator is ready for welding includes: Using a human body posture estimation algorithm to extract skeleton data from the scene information data to obtain information on skeleton key points; Using triple interpolation method to complete the data of missing points of the skeleton key points; Eliminate jitter on the completed skeleton key points; Extract working features based on skeleton key points after eliminating jitter; Training an action recognition model based on the extracted work features, and determining whether the operator is holding a welding gun or a welding rod through the trained action recognition model; When it is determined that the operator is performing an action of holding a welding gun or a welding rod, the operator performs electric welding.

3. The method according to claim 2, characterized in that The data for completing the missing points of the skeleton key points using the triple interpolation method include: Performing symmetrical joint point interpolation on the skeleton key points, including: interpolating corresponding points detectable at the same limb position on the other half of the body using the spinal midline as a vertical symmetry axis; Performing asymmetric joint point interpolation on the skeleton key points, including: When the frames before and after the missing point can be detected, the coordinate value of the missing point is calculated according to formula (1): Among them, P m (t+i) is the coordinate value of the missing point, P(t) is the last detectable position coordinate before a key point is missing, i is the position number of the frame counting from the tth frame, t+n is the position number of the first detected frame after the frame where the missing point is located, P(t+n) is the coordinate value of the position number t+n, n is the number of missing points, and T is the total number of frames for each action; When only the frame before the missing point can be detected, the coordinate value of the missing point is calculated according to formula (2): Among them, P m (t+i) is the coordinate of the missing point, and P(t-4) is the coordinate value of position number t-4; In the case that only the frame after the missing point can be detected, the coordinate value of the missing point is calculated according to formula (3): Among them, P m (ti) is the coordinate of the missing point, and P(t+4) is the coordinate value of position number t+4; Performing facial missing point interpolation operation on the skeleton key points, wherein the facial missing point interpolation operation method includes: calculating the coordinate value of the missing point according to formula (4) to formula (7): Among them, E k is the coordinate of the eye, M k are the coordinates of the midpoint between the ear and the nose, For E k Pointing to M k The vector of , j is the modulus, is a unit vector The horizontal axis, is a unit vector The vertical coordinate, is a vector The horizontal axis, is a vector The vertical coordinate, is the horizontal coordinate of the ear, x N is the horizontal coordinate of the nose, is the vertical coordinate of the ear, y N is the vertical coordinate of the nose, k=1,2.

4. The method according to claim 2, characterized in that The de-jitter processing of the completed skeleton point data includes: Calculate the smoothing factor according to formula (8) to formula (10), Among them, α t is the smoothing factor, r is the smoothing factor adjustment parameter, f C is the dynamic cutoff frequency, Δt is the time interval between adjacent frames, f min is the minimum cutoff frequency, β is the adjustment coefficient, is the rate of change of key point coordinates after filtering, dx is the rate of change of input key point coordinates, is the rate of change of key point coordinates after filtering in the previous frame, α d is the smoothing factor calculated with a fixed cutoff frequency; Calculate the filtered signal according to formula (11): in, is the filtered keypoint coordinate of the current frame, x t is the original keypoint coordinate of the current frame, are the filtered keypoint coordinates of the previous frame.

5. The method according to claim 4, characterized in that The minimum cutoff frequency and the adjustment coefficient can be dynamically adjusted, wherein the dynamic adjustment method of the minimum cutoff frequency includes: The minimum cutoff frequency is dynamically adjusted according to formula (12) to formula (15), Among them, f′ min is the minimum cutoff frequency after adjustment, l1 is the first learning rate, E j is the jitter error of the jth window, E j+1 is the jitter error of the j+1th window, E is the jitter error of the window, e i is the jitter at the i-th moment, u is the average error in the sub-window, is the filtered key point coordinate value at the i-th moment, x i is the coordinate value of the key point before filtering at the i-th moment; The dynamic adjustment method of the adjustment coefficient includes: According to formula (16), the adjustment coefficient is dynamically adjusted. Among them, β′ is the adjusted regulation coefficient, l2 is the second learning rate, is the variance of the delay degree of the i-th sub-window, is the variance of the delay degree of the i+1th window.

6. The method according to claim 2, characterized in that The working features of extracting skeleton key point data after eliminating jitter include: Extract the coordinate features of the key points of the skeleton; According to formula (17), the directional features of the skeleton key points are extracted. Among them, f ori is the directional feature of the skeleton key points, P i y is the y coordinate of the i-th bone key point, is the y coordinate of the first bone key point, is the x-coordinate of the i-th bone key point, is the x-coordinate of the first skeleton key point, and G is the selected key point group; According to formula (18), the distance features of the skeleton key points are extracted. Among them, f dis is the distance of the key points of the skeleton, is the x-coordinate of the j-th bone key point, is the x coordinate of the kth skeleton key point, P j y is the y coordinate of the jth bone key point, is the y coordinate of the kth skeleton key point, and V is the selected key point group; According to formula (19), the trajectory features of the skeleton key points are extracted. in, is the displacement vector of the lth skeleton key point in frame t+1, is the x coordinate of the lth skeleton key point in frame t+1, is the x coordinate of the lth skeleton key point in frame t, is the y coordinate of the lth skeleton key point in the t+1 frame, is the y coordinate of the lth bone key point in frame t.

7. The method according to claim 2, characterized in that Training an action recognition model based on the extracted work features, and determining whether the operator is holding a welding gun or a welding rod by the trained action recognition model includes: The working features are used to train the LSTM network to classify the actions; Inputting the on-site information data of the operator into the trained LSTM network to obtain the probability distribution of the action category; Whether to perform a corresponding action is determined based on the probability distribution and the probability threshold.

8. The method according to claim 1, characterized in that When judging whether the operator is ready for welding work, the following are the procedures to determine whether the inspection site meets the safety operation requirements: Using a data annotation tool to annotate the on-site information data; Build an object detection model and optimize the loss function; Determine whether the safety equipment at the inspection site is complete based on the optimized target detection model; If it is judged that the safety equipment at the testing site is not complete, the testing site does not meet the safety operation requirements; If it is determined that the safety equipment at the inspection site is complete, determine whether the operator has taken the welding gun and welding rod; When it is determined that the operator is holding a welding gun and a welding rod, determining whether the operator is wearing fireproof gloves correctly; If it is determined that the operator did not wear fire-resistant gloves correctly, the inspection site does not meet the safety operation requirements; If it is determined that the operator is wearing fireproof gloves correctly, the operator is continuously checked after an interval of 5 seconds to see whether the operator is wearing a protective mask correctly; If the operator is not wearing the protective mask correctly after continuous detection at an interval of 5 seconds, the detection site does not meet the safety operation requirements.

9. The method according to claim 8, characterized in that Building a target detection model and optimizing the loss function include: The Gaussian distribution of the bounding box is modeled according to formulas (20) to (22). Among them, μ is the mean vector, Σ is the covariance matrix, c x is the horizontal coordinate of the center point of the bounding box, c y is the ordinate of the center point of the bounding box, σ w is the initial standard deviation of width, σ h is the initial standard deviation of height, w is the width of the bounding box, and h is the height of the bounding box; According to formula (23) to formula (25), a new dynamic covariance matrix is ​​obtained. Among them, λ w is the width dynamic scaling factor, λ h is the height dynamic scaling factor, a1 is the width scaling amplitude control parameter, and a2 is the height scaling amplitude control parameter; Calculate the loss function according to formula (26) to formula (28), Among them, L NWD is the loss function, is the similarity measure, For two Gaussian distributions and The Gaussian distance between them, C is the normalization constant, ‖·‖ F is the Frobenius norm, μ1 is the Gaussian distribution The mean vector of μ2 is Gaussian distribution. The mean vector of Σ′1 is Gaussian distribution The dynamic covariance matrix of Σ′2 is Gaussian distribution The dynamic covariance matrix of .

10. A safety pre-inspection system before electric welding operation in a power system, characterized in that: The system comprises a processor configured to execute the method according to any one of claims 1 to 9.

Images