Depth learning-based detection method and system for drill rod of converter air hole punching machine

Abstract

The invention discloses a deep learning-based detection method for a poking drill rod of a converter air hole poking machine, and the method comprises the following steps: scanning poking drill rods in different states, and obtaining a point cloud data set; preprocessing the point cloud data set to obtain a preprocessed point cloud data set; marking the preprocessed point cloud data set to obtain a training data set; based on a Pointnet network, an initial poking drill rod detection model is constructed; training the initial drill rod detection model by using the training data set to obtain a drill rod detection model; and carrying out actual detection on the poking drill rod of the converter air hole poking machine by using the poking drill rod detection model. According to the method, multiple scanning data weighted average fusion, local feature enhancement and adaptive bilateral filtering technologies are adopted, the influence of a complex environment on point cloud data is effectively reduced, and the detection precision and reliability in a severe environment are improved. According to the method, the deep learning algorithm is adopted, the bending and deviation conditions of the poking drill rod can be rapidly predicted, and the method is suitable for real-time industrial detection.

Description

Technical Field

[0001] The present invention belongs to the field of industrial detection and monitoring in nonferrous metal smelting production, and specifically relates to a converter tuyere poking machine poking drill rod detection method and system based on deep learning. Background Art

[0002] The tuyere poke machine is a very important auxiliary equipment in the smelting process. It is mainly used to clean the adhesive in the tuyere of the converter to ensure the stability of heat exchange and reaction in the furnace. The mechanical tuyere poke machine has been successfully used in most smelters in China that adopt the converter blowing process, achieving the purpose of reducing the labor intensity of employees, shortening the blowing time, and improving the product qualification rate.

[0003] In actual operation, the poking rod of the air eye poking machine may deviate and bend. These problems will affect the normal operation and efficiency of the air eye poking machine. The deviation of the poking rod refers to the angle between the poking rod and the straight air duct, which may cause the steel chisel to pass through the ball chamber unsmoothly, thereby generating vertical or horizontal forces. This force may cause the screws between the ball chamber and the air eye seat to loosen, affecting the normal operation of the air eye poking machine. The bending of the poking rod refers to the process of poking the air eye. If the steel chisel is not poked out at the air eye, or when it pokes the ball chamber gland or other parts, due to heat transfer, the steel chisel itself has a higher temperature during the working process, and the rigidity is weakened, which can easily cause the steel chisel to bend or break.

[0004] At present, the detection technologies for the bending or offset of the drill rod of the wind hole drilling machine mainly include visual detection technology, mechanical sensor detection technology and ultrasonic detection technology. Visual detection technology obtains the image of the drill rod by installing a camera or an industrial camera, and calculates the bending and offset of the drill rod by combining image processing algorithms (such as edge detection, shape matching, etc.). Visual detection is greatly disturbed by changes in ambient light, dust, reflection, etc., resulting in reduced detection accuracy and stability. Due to the two-dimensional characteristics of the image, it is difficult to accurately reflect the three-dimensional bending and offset of the drill rod. Mechanical sensor detection technology directly measures the displacement, deformation or stress change of the drill rod by installing displacement sensors, strain gauges or other mechanical sensors on the drill rod or bracket. Mechanical sensors can usually only detect the displacement or stress of specific points of the drill rod, and cannot provide information on the bending and offset of the drill rod as a whole, especially limited perception of deformation in a larger range. In addition, mechanical sensors are easily damaged under high temperature, strong vibration and harsh environment, resulting in high maintenance costs. Ultrasonic detection technology uses ultrasonic sensors to measure the surface wave reflection of the drill rod to detect the deformation and offset of the drill rod. Ultrasonic testing is effective for local detection of drill rods, but the overall deformation detection accuracy is not high. Ultrasonic transmission is unstable in high temperature or dusty environments, and the detection effect is limited. Summary of the invention

[0005] One of the purposes of the present invention is to provide a method for detecting the drill rod of a converter tuyere machine based on deep learning, so as to realize real-time and efficient detection of the bending and displacement of the drill rod in harsh environments, thereby significantly improving production efficiency and detection accuracy.

[0006] The second object of the present invention is to provide a system for realizing the converter tuyere machine drill rod detection method based on deep learning.

[0007] The present invention provides a method for detecting a converter tuyere machine drill rod based on deep learning, comprising the following steps:

[0008] S1. Scan the poking rod in different states to obtain a point cloud data set;

[0009] S2. Preprocessing the point cloud data set obtained in step S1 to obtain a preprocessed point cloud data set;

[0010] S3. Annotate the preprocessed point cloud data set obtained in step S2 to obtain a training data set;

[0011] S4. Based on the Pointnet network, an initial poking rod detection model is constructed;

[0012] S5. Use the training data set obtained in step S3 to train the initial poking rod detection model obtained in step S4 to obtain a poking rod detection model;

[0013] S6. Using the drill rod detection model obtained in step S5, actual converter tuyere drill rod detection is performed.

[0014] Step S1 specifically includes: scanning the poking rod in different states; scanning the poking rod in a single state several times to obtain several frames of point cloud data in the state; and combining several frames of point cloud data in all states to obtain a point cloud data set.

[0015] Step S2 is specifically as follows: firstly weighted average fusion is performed on the point cloud data set obtained in step S1, then local feature enhancement is performed to increase local information, and finally adaptive bilateral filtering is performed to remove noise to obtain a preprocessed point cloud data set.

[0016] The weighted average fusion is to process several frames of point cloud data obtained by scanning the poking rod several times in the same state in the point cloud data set, and the following formula is used to calculate each point P i The coordinates of are weighted averaged:

[0017]

[0018] in, Point P obtained from the kth scan iThe coordinate on the x-axis; Point P obtained from the kth scan i The coordinate on the y-axis; Point P obtained from the kth scan i The coordinate on the z-axis; is the weight coefficient; (x i ,y i ,z i ) is the point P after weighted average fusion processing i The coordinates of

[0019] The local feature enhancement is to add the normal vector information norm(i) of the point cloud coordinates after weighted average fusion processing to obtain a new point cloud data coordinate (x i ,y i ,z i ,norm(i));

[0020] The adaptive bilateral filtering process is used to remove noise in the point cloud data, and the filtering is performed using the following formula:

[0021]

[0022] Where P(i) is the coordinate of the i-th point in the point cloud data; P(j) is the coordinate of the j-th point in the point cloud data; ||P(i)-P(j)|| is the Euclidean distance between the i-th point and the j-th point in the point cloud data; ||norm(i)-norm(j)|| is the difference in normal vectors between the i-th point and the j-th point in the point cloud data; σ s is the weight controlling the spatial proximity; σ r is the weight for controlling the normal vector difference; P filtered (i) is the coordinate of the i-th point in the point cloud data after adaptive bilateral filtering;

[0023] Step S3 is specifically as follows: for the point cloud data corresponding to the drilling rod in the same state in the preprocessed point cloud data set, the curvature and offset of the drilling rod in the corresponding state are marked to obtain a training data set.

[0024] The initial poking rod detection model in step S4 includes a feature extraction module, a feature aggregation module and a poking rod prediction module;

[0025] Construct a feature extraction module and introduce a dynamic weight generation mechanism into the multi-layer perceptron so that the weights can be adaptively adjusted according to the local features of the point cloud during the feature extraction process. Then, the output of each layer of the multi-layer perceptron is batch normalized to speed up the convergence and finally obtain the feature vector.

[0026] The feature aggregation module aggregates the input feature vector through the maximum pooling layer to generate global features;

[0027] The poking rod prediction module includes a curvature prediction submodule and an offset prediction submodule; the curvature prediction submodule inputs the input global features into a multi-layer perceptron and outputs the predicted curvature; the offset prediction submodule inputs the input global features into a multi-layer perceptron and outputs the predicted offset;

[0028] The predicted curvature and predicted offset are outputs of the initial poking rod detection model.

[0029] In the feature extraction module, the dynamic weight generation mechanism calculates the dynamic weight W(i) according to the local feature relationship of the point cloud, and adaptively adjusts the point importance in the feature extraction process; wherein, the dynamic weight W(i) is generated by a two-layer multi-layer perceptron (MLP) based on the neighborhood relationship, and the weight W(i) is calculated using the following formula:

[0030]

[0031] Among them, R ij is the neighborhood relationship between point i and point j in the point cloud data; P filtered (j) is the coordinate of the jth point in the point cloud data after adaptive bilateral filtering; N(i) is the total number of points in the point cloud data; MLP ω is the weight processing of multi-layer perceptron; domain relationship R ij Use the following formula to calculate:

[0032]

[0033] ||P filtered (i)-P filtered (j)|| is the Euclidean distance between the i-th point and the j-th point in the point cloud data after adaptive bilateral filtering;

[0034] At the same time, the input features of each point in the point cloud data after adaptive bilateral filtering are processed by a multi-layer perceptron to extract local features, which can be expressed by the following formula:

[0035] f i =MLP(P filtered (i))

[0036] Among them, f i is the initial local feature of the i-th point in the point cloud data after adaptive bilateral filtering;

[0037] Combined with the corresponding dynamic weight W(i), the adjustment is performed using the following formula:

[0038] f i '=W(i)f i

[0039] Among them, fi ' is the weighted local feature of the i-th point in the point cloud data after adaptive bilateral filtering;

[0040] Finally, the weighted local features of each layer of MLP output are batch normalized, using the following formula:

[0041]

[0042] in, is the local feature of the i-th point in the point cloud data after adaptive bilateral filtering; μ is the mean of the feature in the current batch; σ 2 is the variance of the feature in the current batch; ε is a preset constant with a value range of 10 -5 to 10 -8 ;

[0043] The feature extraction module outputs the local features of each point As the input of the feature aggregation module;

[0044] The feature aggregation module aggregates the input feature vector through the maximum pooling layer to generate global features, and uses the following formula for calculation:

[0045]

[0046] max(·) is the maximum value of all points in each dimension; f global is the global eigenvector;

[0047] The curvature prediction submodule is expressed using the following formula:

[0048] C=MLP curve (f global )

[0049] Among them, MLP curve is a multi-layer perceptron for predicting curvature; f global is the global feature; C is the predicted curvature;

[0050] The offset prediction submodule is expressed using the following formula:

[0051] O=MLP offset (f global )

[0052] Among them, MLP offset is a multi-layer perceptron for predicting the offset; O is the predicted offset;

[0053] Step S6 specifically includes the following steps:

[0054] Scan the drill rod several times to obtain multiple frames of point cloud data;

[0055] Preprocessing the obtained multi-frame point cloud data as described in step S2 to obtain preprocessed point cloud data;

[0056] The pre-processed point cloud data is input into the poking rod detection model obtained in step S5 to obtain the predicted curvature and offset.

[0057] The present invention also provides a system for realizing the deep learning-based converter tuyere poking drill rod detection method, comprising a data acquisition module, a data preprocessing module, a data annotation module, a model building module, a model training module, and a poking drill rod detection module;

[0058] The data acquisition module scans the poking rods in different states to obtain a point cloud data set, and uploads the data to the data preprocessing module;

[0059] The data preprocessing module preprocesses the point cloud data set according to the received data, obtains the preprocessed point cloud data set, and uploads the data to the data annotation module;

[0060] The data annotation module annotates the preprocessed point cloud data set according to the received data, obtains the training data set, and uploads the data to the model construction module;

[0061] The model building module builds an initial poking rod detection model based on the Pointnet network according to the received data, and uploads the data to the model training module;

[0062] The model training module trains the initial poking rod detection model using the training data set according to the received data, obtains the poking rod detection model, and uploads the data to the poking rod detection module;

[0063] The drill rod detection module uses the drill rod detection model to perform actual converter tuyere drill rod detection based on the received data.

[0064] The present invention discloses a method and system for detecting the drill rod of a converter tuyere machine based on deep learning, which adopts weighted average fusion of multiple scan data, local feature enhancement and adaptive bilateral filtering technology, effectively reduces the influence of complex environment on point cloud data, and improves the detection accuracy and reliability in harsh environment. The method of the present invention adopts deep learning algorithm, can quickly predict the bending and deviation of the drill rod, and is suitable for real-time industrial detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a schematic diagram of the process of the present invention;

[0066] Figure 2 It is a schematic diagram of the structure of the system of the present invention. DETAILED DESCRIPTION

[0067] The present invention provides a method for detecting a converter tuyere drilling rod based on deep learning, and a schematic diagram of the process is shown in FIG. Figure 1 As shown, the following steps are included:

[0068] S1. Scan the poking rods in different states to obtain point cloud data sets, specifically:

[0069] Scan the poking rod in different states; scan the poking rod in a single state several times to obtain several frames of point cloud data in this state; and collect several frames of point cloud data in all states to obtain a point cloud data set.

[0070] S2. Preprocess the point cloud data set obtained in step S1 to obtain a preprocessed point cloud data set, specifically:

[0071] The point cloud data set obtained in step S1 is first subjected to weighted average fusion, then local feature enhancement is performed to increase local information, and finally adaptive bilateral filtering is performed to remove noise to obtain a preprocessed point cloud data set.

[0072] The weighted average fusion is to process several frames of point cloud data obtained by scanning the poking rod several times in the same state in the point cloud data set, and the following formula is used to calculate each point P i The coordinates of are weighted averaged:

[0073]

[0074] in, Point P obtained from the kth scan i The coordinate on the x-axis; Point P obtained from the kth scan i The coordinate on the y-axis; Point P obtained from the kth scan i The coordinate on the z-axis; is the weight coefficient; (x i ,y i ,z i ) is the point P after weighted average fusion processing i The coordinates of

[0075] The local feature enhancement is to add the normal vector information norm(i) of the point cloud coordinates after weighted average fusion processing to obtain a new point cloud data coordinate (x i ,y i ,z i ,norm(i));

[0076] The adaptive bilateral filtering process is used to remove noise in the point cloud data, and the filtering is performed using the following formula:

[0077]

[0078] Where P(i) is the coordinate of the i-th point in the point cloud data; P(j) is the coordinate of the j-th point in the point cloud data; ||P(i)-P(j)|| is the Euclidean distance between the i-th point and the j-th point in the point cloud data; ||norm(i)-norm(j)|| is the difference in normal vectors between the i-th point and the j-th point in the point cloud data; σ s is the weight controlling the spatial proximity; σ r is the weight for controlling the normal vector difference; P filtered (i) is the coordinate of the i-th point in the point cloud data after adaptive bilateral filtering;

[0079] S3. Annotate the preprocessed point cloud data set obtained in step S2 to obtain a training data set, specifically:

[0080] For the point cloud data corresponding to the poking rod in the same state in the preprocessed point cloud data set, the curvature and offset of the poking rod in the corresponding state are marked to obtain the training data set.

[0081] S4. Based on the Pointnet network, an initial poking rod detection model is constructed;

[0082] The initial poking rod detection model includes a feature extraction module, a feature aggregation module and a poking rod prediction module;

[0083] Construct a feature extraction module and introduce a dynamic weight generation mechanism into the multi-layer perceptron so that the weights can be adaptively adjusted according to the local features of the point cloud during the feature extraction process. Then, the output of each layer of the multi-layer perceptron is batch normalized to speed up the convergence and finally obtain the feature vector.

[0084] The feature aggregation module aggregates the input feature vector through the maximum pooling layer to generate global features;

[0085] The poking rod prediction module includes a curvature prediction submodule and an offset prediction submodule; the curvature prediction submodule inputs the input global features into a multi-layer perceptron and outputs the predicted curvature; the offset prediction submodule inputs the input global features into a multi-layer perceptron and outputs the predicted offset;

[0086] The predicted curvature and predicted offset are outputs of the initial poking rod detection model.

[0087] In the feature extraction module, the dynamic weight generation mechanism calculates the dynamic weight W(i) according to the local feature relationship of the point cloud, and adaptively adjusts the point importance in the feature extraction process; wherein, the dynamic weight W(i) is generated by a two-layer multi-layer perceptron (MLP) based on the neighborhood relationship, and the weight W(i) is calculated using the following formula:

[0088]

[0089] Among them, R ij is the neighborhood relationship between point i and point j in the point cloud data; P filtered (j) is the coordinate of the jth point in the point cloud data after adaptive bilateral filtering; N(i) is the total number of points in the point cloud data; MLP ω is the weight processing of multi-layer perceptron; domain relationship R ij Use the following formula to calculate:

[0090]

[0091] ||P filtered (i)-P filtered (j)|| is the Euclidean distance between the i-th point and the j-th point in the point cloud data after adaptive bilateral filtering;

[0092] At the same time, the input features of each point in the point cloud data after adaptive bilateral filtering are processed by a multi-layer perceptron to extract local features, which can be expressed by the following formula:

[0093] f i =MLP(P filtered (i))

[0094] Among them, f i is the initial local feature of the i-th point in the point cloud data after adaptive bilateral filtering;

[0095] Combined with the corresponding dynamic weight W(i), the adjustment is performed using the following formula:

[0096] f i '=W(i)f i

[0097] Among them, f i ' is the weighted local feature of the i-th point in the point cloud data after adaptive bilateral filtering;

[0098] Finally, the weighted local features of each layer of MLP output are batch normalized, using the following formula:

[0099]

[0100] in, is the local feature of the i-th point in the point cloud data after adaptive bilateral filtering; μ is the mean of the feature in the current batch; σ 2 is the variance of the feature in the current batch; ε is a preset constant with a value range of 10 -5 to 10 -8 ;

[0101] The feature extraction module outputs the local features of each point As the input of the feature aggregation module;

[0102] The feature aggregation module aggregates the input feature vector through the maximum pooling layer to generate global features, and uses the following formula for calculation:

[0103]

[0104] max(·) is the maximum value of all points in each dimension; f global is the global eigenvector;

[0105] The curvature prediction submodule is expressed using the following formula:

[0106] C=MLP curve (f global )

[0107] Among them, MLP curve is a multi-layer perceptron for predicting curvature; f global is the global feature; C is the predicted curvature;

[0108] The offset prediction submodule is expressed using the following formula:

[0109] O=MLP offset (f global )

[0110] Among them, MLP offset is a multilayer perceptron for predicting the offset; O is the predicted offset.

[0111] S5. Use the training data set obtained in step S3 to train the initial poking rod detection model obtained in step S4 to obtain a poking rod detection model;

[0112] S6. Using the poking rod detection model obtained in step S5, the actual converter poking tuyere machine poking rod detection is performed, which specifically includes the following steps:

[0113] Scan the drill rod several times to obtain multiple frames of point cloud data;

[0114] Preprocessing the obtained multi-frame point cloud data as described in step S2 to obtain preprocessed point cloud data;

[0115] The pre-processed point cloud data is input into the poking rod detection model obtained in step S5 to obtain the predicted curvature and offset.

[0116] The present invention also provides a system for realizing the converter tuyere poking machine poking drill rod detection method based on deep learning, and its structural schematic diagram is as follows: Figure 2 As shown, it includes a data acquisition module, a data preprocessing module, a data annotation module, a model building module, a model training module, and a poking rod detection module;

[0117] The data acquisition module scans the poking rods in different states to obtain a point cloud data set, and uploads the data to the data preprocessing module;

[0118] The data preprocessing module preprocesses the point cloud data set according to the received data, obtains the preprocessed point cloud data set, and uploads the data to the data annotation module;

[0119] The data annotation module annotates the preprocessed point cloud data set according to the received data, obtains the training data set, and uploads the data to the model construction module;

[0120] The model building module builds an initial poking rod detection model based on the Pointnet network according to the received data, and uploads the data to the model training module;

[0121] The model training module trains the initial poking rod detection model using the training data set according to the received data, obtains the poking rod detection model, and uploads the data to the poking rod detection module;

[0122] The drill rod detection module uses the drill rod detection model to perform actual converter tuyere drill rod detection based on the received data.

[0123] The method of the present invention can effectively improve the detection accuracy and real-time performance of the brazing rod of the copper smelting converter. The optimized PointNet network can adaptively extract meaningful local features from the point cloud data, dynamically adjust the noise in the point cloud data, and accurately predict the curvature and offset of the brazing rod through global feature aggregation. The method can significantly improve the reliability of brazing rod bending and offset detection, and provide strong support for the precise control of the brazing rod in the production process.

Claims

1. A method for detecting a converter tuyere drilling rod based on deep learning, characterized in that: The following steps are involved: S1. Scan the poking rod in different states to obtain a point cloud data set; S2. Preprocessing the point cloud data set obtained in step S1 to obtain a preprocessed point cloud data set; S3. Annotate the preprocessed point cloud data set obtained in step S2 to obtain a training data set; S4. Based on the Pointnet network, an initial poking rod detection model is constructed; S5. Use the training data set obtained in step S3 to train the initial poking rod detection model obtained in step S4 to obtain a poking rod detection model; S6. Using the drill rod detection model obtained in step S5, actual converter tuyere drill rod detection is performed.

2. The converter tuyere poking machine poking drill rod detection method based on deep learning according to claim 1 is characterized in that: Step S1 specifically includes: scanning the poking rod in different states; scanning the poking rod in a single state several times to obtain several frames of point cloud data in the state; and combining several frames of point cloud data in all states to obtain a point cloud data set.

3. The converter tuyere poking machine poking drill rod detection method based on deep learning according to claim 2 is characterized in that: The point cloud data set obtained in step S1 is first subjected to weighted average fusion, then local feature enhancement is performed to increase local information, and finally adaptive bilateral filtering is performed to remove noise to obtain a preprocessed point cloud data set.

4. The converter tuyere poking machine poking drill rod detection method based on deep learning according to claim 3 is characterized in that: The weighted average fusion is to process several frames of point cloud data obtained by scanning the poking rod several times in the same state in the point cloud data set, and the following formula is used to calculate each point P i The coordinates of are weighted averaged: in, Point P obtained by the kth scan i The coordinate on the x-axis; Point P obtained by the kth scan i The coordinate on the y-axis; Point P obtained by the kth scan i The coordinate on the z-axis; is the weight coefficient; (x i ,y i ,z i ) is the point P after weighted average fusion processing i The coordinates of The local feature enhancement is to add the normal vector information norm(i) of the point cloud coordinates after weighted average fusion processing to obtain a new point cloud data coordinate (x i ,y i ,z i ,norm(i)); The adaptive bilateral filtering process is used to remove noise in the point cloud data, and the filtering is performed using the following formula: Where P(i) is the coordinate of the i-th point in the point cloud data; P(j) is the coordinate of the j-th point in the point cloud data; ||P(i)-P(j)|| is the Euclidean distance between the i-th point and the j-th point in the point cloud data; ||norm(i)-norm(j)|| is the difference in normal vectors between the i-th point and the j-th point in the point cloud data; σ s is the weight for controlling spatial proximity; σ r is the weight for controlling the normal vector difference; P filtered (i) is the coordinate of the i-th point in the point cloud data after adaptive bilateral filtering.

5. The converter tuyere drilling rod detection method based on deep learning according to claim 1 is characterized in that: Step S3 is specifically as follows: for the point cloud data corresponding to the drilling rod in the same state in the preprocessed point cloud data set, the curvature and offset of the drilling rod in the corresponding state are marked to obtain a training data set.

6. The method for detecting converter tuyere drilling rod based on deep learning according to claim 1, characterized in that: The initial poking rod detection model in step S4 includes a feature extraction module, a feature aggregation module and a poking rod prediction module; Construct a feature extraction module and introduce a dynamic weight generation mechanism into the multi-layer perceptron so that the weights can be adaptively adjusted according to the local features of the point cloud during the feature extraction process. Then, the output of each layer of the multi-layer perceptron is batch normalized to speed up the convergence and finally obtain the feature vector. The feature aggregation module aggregates the input feature vector through the maximum pooling layer to generate global features; The poking rod prediction module includes a curvature prediction submodule and an offset prediction submodule; the curvature prediction submodule inputs the input global features into a multi-layer perceptron and outputs the predicted curvature; the offset prediction submodule inputs the input global features into a multi-layer perceptron and outputs the predicted offset; The predicted curvature and predicted offset are outputs of the initial poking rod detection model.

7. The method for detecting converter tuyere drilling rod based on deep learning according to claim 6 is characterized in that: In the feature extraction module, the dynamic weight generation mechanism calculates the dynamic weight W(i) according to the local feature relationship of the point cloud, and adaptively adjusts the point importance in the feature extraction process; wherein, the dynamic weight W(i) is generated by a two-layer multi-layer perceptron (MLP) based on the neighborhood relationship, and the weight W(i) is calculated using the following formula: Among them, R ij is the neighborhood relationship between point i and point j in the point cloud data; P filtered (j) is the coordinate of the jth point in the point cloud data after adaptive bilateral filtering; N(i) is the total number of points in the point cloud data; MLP ω is the weight processing of multi-layer perceptron; domain relationship R ij Use the following formula to calculate: ||P filtered (i)-P filtered (j)|| is the Euclidean distance between the i-th point and the j-th point in the point cloud data after adaptive bilateral filtering; At the same time, the input features of each point in the point cloud data after adaptive bilateral filtering are processed by a multi-layer perceptron to extract local features, which can be expressed by the following formula: f i =MLP(P filtered (i)) Among them, f i is the initial local feature of the i-th point in the point cloud data after adaptive bilateral filtering; Combined with the corresponding dynamic weight W(i), the adjustment is performed using the following formula: f i '=W(i)f i Among them, f i ' is the weighted local feature of the i-th point in the point cloud data after adaptive bilateral filtering; Finally, the weighted local features of each layer of MLP output are batch normalized, using the following formula: in, is the local feature of the i-th point in the point cloud data after adaptive bilateral filtering; μ is the mean of the feature in the current batch; σ 2 is the variance of the feature in the current batch; ε is a preset constant with a value range of 10 -5 to 10 -8 ; The feature extraction module outputs the local features of each point As the input of the feature aggregation module; The feature aggregation module aggregates the input feature vector through the maximum pooling layer to generate global features, and uses the following formula for calculation: max(·) is the maximum value of all points in each dimension; f global is the global eigenvector; The curvature prediction submodule is expressed using the following formula: C=MLP curve (f global ) Among them, MLP curve is a multi-layer perceptron for predicting curvature; f global is the global feature; C is the predicted curvature; The offset prediction submodule is expressed using the following formula: O=MLP offset (f global ) Among them, MLP offset is a multilayer perceptron for predicting the offset; O is the predicted offset.

8. The method for detecting converter tuyere drilling rod based on deep learning according to claim 1, characterized in that: Step S6 specifically includes the following steps: Scan the drill rod several times to obtain multiple frames of point cloud data; Preprocessing the obtained multi-frame point cloud data as described in step S2 to obtain preprocessed point cloud data; The pre-processed point cloud data is input into the poking rod detection model obtained in step S5 to obtain the predicted curvature and offset.

9. A system for implementing the deep learning-based converter tuyere poking drill rod detection method according to any one of claims 1 to 8, characterized in that: It includes data acquisition module, data preprocessing module, data annotation module, model building module, model training module and drilling rod detection module; The data acquisition module scans the poking rods in different states to obtain a point cloud data set, and uploads the data to the data preprocessing module; The data preprocessing module preprocesses the point cloud data set according to the received data, obtains the preprocessed point cloud data set, and uploads the data to the data annotation module; The data annotation module annotates the preprocessed point cloud data set according to the received data, obtains the training data set, and uploads the data to the model construction module; The model building module builds an initial poking rod detection model based on the Pointnet network according to the received data, and uploads the data to the model training module; The model training module trains the initial poking rod detection model using the training data set according to the received data, obtains the poking rod detection model, and uploads the data to the poking rod detection module; The drill rod detection module uses the drill rod detection model to perform actual converter tuyere drill rod detection based on the received data.

Images