Converter installation position deviation detection method applying point cloud processing

Abstract

The invention discloses a converter installation position deviation detection method applying point cloud processing, and the method employs an HWD module integrated with a deep residual network and feature image resolution to be integrated into a network model of point cloud extraction features, and the HWD module is connected with a point cloud extraction layer in parallel. Point cloud semantic segmentation is enhanced in a complex construction environment, it is guaranteed that the overall accuracy of local and global information extraction of a converter point cloud model is obviously improved, and the needed converter model extraction time is saved; and secondly, intercepting a circular slice from the extracted optimal model data, performing fitting circle solution based on the circumferential contour of the slice, recording a least square circle center parameter, determining an optimal alignment mode between two axes, and calculating coaxial deviation of mounting positions between components, so that errors are effectively reduced. The method still has good monitoring performance under the complex environment of site construction, can well solve the problem that the error of the assembly position of the converter exceeds the standard in construction, and has good detection capability.

Description

Technical Field

[0001] The invention relates to a method for detecting a position deviation of a converter body, and specifically relates to a method for evaluating a converter installation error by applying a point cloud processing module. Background Art

[0002] The converter is an extremely important equipment in the steelmaking industry. It has high working efficiency and is suitable for smelting multiple types of steel. The converter body is composed of a furnace bottom, a furnace body, a furnace mouth and a support ring. The furnace bottom is nearly hemispherical, the furnace body is cylindrical, the furnace top is a trapezoidal column, and the support ring is similar to a circular ring. The weight of the converter can reach several hundred tons, and the diameter and height are both more than ten meters. It takes a long time to install it. Not only must it ensure a very small matching deviation, but it must also meet the requirements of thermal expansion, which brings great difficulties to the adjustment. The current mainstream installation method is the "combination method", that is, when the converter shell is assembled on the temporary support frame, it is carried out from the furnace bottom section to the furnace mouth section, from bottom to top. The transport device moves the assembled furnace body as a whole to the installation position, and then removes the transport device. The entire installation process is difficult to locate, slow to adjust, and accompanied by certain construction risks, which seriously affects the construction progress. In order to solve the above problems, it has become an inevitable trend to develop intelligent and efficient installation position detection.

[0003] With the increasing maturity of on-site surveying and mapping technology, 3D reconstruction has been widely used in engineering. A 3D point cloud model of the construction site can be constructed using a 3D laser scanner. For large-scale and complex point cloud data, a deep learning framework for point cloud feature segmentation and extraction has been proposed. Since the converter is covered with industrial wires, welding boxes, standing support plates, construction workers, converter support frames, and three walls, the pooling operation of converter feature extraction will lead to local spatial information loss, which is not conducive to the prediction of point cloud details and the recognition ability of complex sites.

[0004] In order to facilitate accurate extraction of converter features in complex point cloud data, the present invention proposes the following technical solution. Summary of the invention

[0005] The purpose of the present invention is to provide a method for detecting the installation position deviation of a converter using point cloud processing, so as to solve the problem in the prior art that the pooling operation of converter feature extraction will lead to local spatial information loss, which is not conducive to the prediction of point cloud details and the recognition ability of complex sites, because the converter is covered with numerous obstructions such as industrial wires, welding boxes, standing support plates, construction workers, converter support frames, and three sides are against walls.

[0006] The purpose of the present invention can be achieved through the following technical solutions:

[0007] A method for detecting converter installation position deviation using point cloud processing, the method specifically comprises the following steps:

[0008] Step 1: Collect the point cloud data model information of the converter at the construction site, and use the deep residual network and feature image resolution to solve the problem of unbalanced data sets for aggregation;

[0009] Step 2: Integrate the deep residual network and feature image processing conversion into the Haar wavelet module, and use the integrated HWD module to process the local input feature signals that are easy to lose, perform feature conversion and reduce spatial resolution;

[0010] Step 3: Integrate the HWD module into the feature extraction network model, which is in parallel with the overall feature extraction layer. Use the integrated network model to train and identify the point cloud data of the converter; extract local and global information features of the converter point cloud model;

[0011] Step 4: Use the validation set to determine whether the model is accurate and complete, and select the optimal model as the installation position evaluation model; slice the extracted optimal model data, perform a fitted circle solution based on the sliced ​​circular contour, record the coordinate parameters of the center of the fitted circle, and use these parameters to evaluate the deviation from the actual data points and calculate the coaxiality of the installation positions between components.

[0012] Furthermore, step one specifically includes:

[0013] Step 101: The input of the nth layer is x n , when the input and output dimensions are the same, the input is directly added to the output, then the output formula of the residual block of this layer is as follows:

[0014] H(x n )=F(x n )+z n

[0015] When the dimensions are different, a 1×1 convolution is added to adjust the number of channels. The residual block is expressed as follows:

[0016] y n =H(x n )=F(x n )+g(x n )

[0017] Step 102: Input x of the n+1th layer n+1 =f(y n ), g(x n ) is a direct mapping, g(x n )=W′ n x,W′ n is a 1×1 convolution operation, f(y n ) is the activation function, expressed using Relu, and the formula is as follows:

[0018] x n+1 =xn +F(x n )

[0019] Step 103: For any deeper layer N, its output is the sum of the input of the shallower layer n plus the residual of each subsequent layer, and the formula is as follows:

[0020]

[0021] Step 104: For the backward pass, derive the loss function with respect to the output x of any layer n n The gradient of is described as follows;

[0022]

[0023] Furthermore, step 2 specifically includes:

[0024] Step 201: Haar wavelet uses a scaling function for signal processing and data analysis. The scaling function is used to describe the change of a signal over time or space. When the scale changes, it can change its shape. The description formula is as follows:

[0025]

[0026] Where: i and j represent the order and dimension of the Haar basis function respectively. This function set is defined on the half-open interval [0,1). The value of each piecewise constant function is 1 in a small range and 0 in other areas, as shown below:

[0027]

[0028] Step 202: The scaling function and the wavelet basis function are shown in the following formula:

[0029]

[0030] Furthermore, step three specifically includes:

[0031] Step 301: uniformly sample s points from the input n point set, take a point in the point set s as the center point, take out m points within the radius R, normalize them to the unit sphere, and convert them into a point cloud represented by spherical coordinates as input. The output shape is B×N×M×3, where 3 is the channel coordinates of x, y, and z.

[0032] Step 302: Learn local point set features, with input N′ and output N′×(d+C′); extract features through MLP, and use maximum pooling to aggregate the local features of each point with the features of other points in its neighborhood;

[0033] Step 303: The local neighborhood information lost by the maximum pooling is recycled and stored in the HWD network module and added to the global features; finally, the global features are mapped to the output space through the fully connected layer, and segmentation and classification are performed to obtain the features of the entire point set.

[0034] Furthermore, step four specifically includes:

[0035] Step 401: construct a converter cylindrical model H, where the reference axis is a, the measured axis b is a space curve, the tolerance range is T, the ideal axis is constructed according to the measured axis b, and the contact point e is the maximum error range of the measured axis b with the ideal axis as the center;

[0036] Step 402: Obtain two groups of equal-height slice models of the converter body, determine E1 and E2 at appropriate positions of the furnace shell and the furnace body along the height direction of the converter, and then intercept point cloud slices at equal intervals; perform coaxiality detection on both ends of the furnace top and the furnace body. The ideal axis and the reference axis are extended over a long distance, and there are errors in the x-axis and y-axis;

[0037] Step 403: Analyze the two slices obtained; solve the fitting circle based on the circular contour, and calculate the coordinate parameters of the center of multiple fitting circles;

[0038] Step 404: According to the calculated fitting circle, perform cylinder slice fitting, divide the cylinder into n sampling section cylinders in the OZ direction for the reference element; take m points on each section cylinder for measurement; record multiple sets of point cloud slice fitting circle center coordinates, and obtain a set of data about x and y (x i ,y i )(i=1,2,…,n), assuming that they have a linear relationship; using The method to obtain (x i ,y i ) to fit it so that it satisfies Finding the least squares solution is essentially finding the optimal parameters a and b so that x i ,y i and maximum proximity;

[0039] Step 405: The description target is the effect of the fitting parameters a and b of the sum of squares. An optimization function needs to be defined while avoiding the offset of positive and negative values. The residual sum of squares needs to be used to define it. The residual is the difference between the actual observed value and the fitted value.

[0040] Beneficial effects of the present invention:

[0041] Compared with most methods that use classic point cloud feature extraction and segmentation, the present invention uses the HWD module integrated into the network model of point cloud feature extraction after incorporating deep residual network and feature image resolution, which is in parallel with the point cloud extraction layer. In the complex construction environment, point cloud semantic segmentation is enhanced to ensure that the overall accuracy of local and global information extraction of the converter point cloud model is significantly improved, saving the required converter model extraction time; secondly, the extracted optimal model data is cut into circular slices, and the fitting circle is solved based on the slice circumference contour, the least squares center parameter is recorded, the optimal alignment between the two axes is determined, and the coaxial deviation of the installation position between the components is calculated, so that the error is effectively reduced. Finally, the experiment shows that the present invention still has good monitoring performance under the complex environment of on-site construction, can well solve the problem of excessive converter assembly position error during construction and has good detection capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The present invention will be further described below in conjunction with the accompanying drawings.

[0043] Figure 1 It is a schematic diagram of the research method architecture of the present invention;

[0044] Figure 2 This is a schematic diagram of the converter on-site construction environment;

[0045] Figure 3 is a schematic diagram of a deep residual network unit according to the present invention;

[0046] Figure 4 It is a schematic diagram of the wavelet decomposition process of the present invention;

[0047] Figure 5 is a schematic diagram of the network structure of the HWD module of the present invention;

[0048] Figure 6 is a schematic diagram of the improved PointNet++ network structure of the present invention;

[0049] Figure 7 is a schematic diagram of the coaxiality model error curve of the present invention;

[0050] Figure 8 It is a schematic diagram of point cloud slices at the upper and lower ends of the converter of the present invention;

[0051] Fig. 9 It is a schematic diagram of the slice curves at both ends of the converter body of the present invention;

[0052] Fig.10 It is a schematic diagram of cylindrical slice fitting after the fitting circle calculation described in the present invention;

[0053] Fig.11is a schematic diagram summarizing the least squares method of the present invention;

[0054] Fig.12 is a schematic diagram of the cylindrical geometric model contour detection according to the present invention;

[0055] Fig.13 It is a schematic diagram of the converter three-dimensional reconstruction measurement work of the present invention;

[0056] Fig.14 is a schematic diagram of the converter training set annotation according to the present invention;

[0057] Fig.15 It is a schematic diagram of the change trend of the loss function described in the present invention;

[0058] Fig.16 : is a schematic diagram of the comparison of F1 values ​​before and after the improvement of the PointNet++ model described in the present invention;

[0059] Fig.17 is a schematic diagram comparing converter point clouds extracted by the network model before and after the improvement described in the present invention;

[0060] Fig.18 is a schematic diagram of a fitting circle test experiment according to the present invention;

[0061] Fig.19 is a schematic diagram of the test results of the center point of the fitting circle according to the present invention;

[0062] Fig. 20 It is a schematic diagram of the calculation result of the coaxiality of the circular section according to the present invention;

[0063] Fig.21 It is a schematic diagram for comparing the coaxiality errors described in the present invention. DETAILED DESCRIPTION

[0064] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] A method for detecting converter installation position deviation using point cloud processing, such as Figure 1 As shown, the following steps are included:

[0066] Step 1: First, place the laser scanner around the converter, and then start shooting around it. Figure 2The purpose is to obtain its point cloud data, use deep residual network and feature image resolution to solve the problem of imbalanced data sets for aggregation, and improve the ability of sampling operations and learning feature transformations;

[0067] Step 1 specifically includes: resolving the problem of imbalanced data sets for aggregation, improving the ability of sampling operations and learning feature transformations.

[0068] Step 10 1: The input of the nth layer is x n , when the input and output dimensions are the same, directly adding the input to the output is equivalent to simply performing the same mapping, which does not generate additional parameters and does not increase the computational complexity. The output formula of the residual block of this layer is as follows:

[0069] H(x n )=F(x n )+z n

[0070] When the dimensions are different, 1×1 convolution is added to adjust the number of channels. For forward transmission, this residual learning structure can be implemented by adding direct edges to the forward neural network, and the entire network can still be trained through end-to-end backpropagation, such as Figure 3 As shown. At this time, the residual block is expressed as follows:

[0071] y n =H(x n )=F(x n )+g(x n )

[0072] Step 102: Input x of the n+1th layer n+1 =f(y n ), g(x n ) is a direct mapping, g(x n )=W′ n x,W′ n is a 1×1 convolution operation, f(y n ) is the activation function, expressed using Relu, and the formula is as follows:

[0073] x n+1 =x n +F(x n )

[0074] Step 103: For any deeper layer N, its output is the sum of the input of the shallower layer n plus the residual of each subsequent layer, and the formula is as follows:

[0075]

[0076] Step 104: For reverse transmission, the above idea can be used to derive the loss function with respect to the output x of any layer nn The gradient of is described as follows;

[0077]

[0078] It can be seen that the gradient of the loss function with respect to the deeper high-level layer N can be directly transmitted to any shallower layer. The residual network can directly propagate the signal to any layer in both forward and reverse transmission;

[0079] Step 2: Integrate the deep residual network and feature image into the Haar wavelet module, and use the integrated Haar wavelet transform module (HWD) to process the input local feature signals that are easy to lose, perform feature conversion and reduce spatial resolution.

[0080] Step 201: Haar wavelet uses a scaling function for signal processing and data analysis. The scaling function is used to describe the change of a signal over time or space. When the scale changes, it can change shape, such as Figure 4 As shown, the description formula is as follows:

[0081]

[0082] Where: i and j represent the order and dimension of the Haar basis function respectively. This function set is defined on the half-open interval [0,1), and the value of each piecewise constant function is 1 in a small range and 0 in other areas, as shown below:

[0083]

[0084] Step 202: The basis function is the core component of the wavelet transform, which can capture the local characteristics of the signal at different time and frequency scales. The scale function and the wavelet basis function are shown in the following formula:

[0085]

[0086] Haar wavelet transform has good localization properties in edge information detection, feature texture analysis, and both in time domain and frequency domain. It has a simple structure and high computational efficiency, and can focus on any detail of an object, such as Figure 5 shown.

[0087] Step 3: Integrate the HWD module into the feature extraction network model, which is connected in parallel with the extraction layer, such as Figure 6 The improved network model is used for training to identify the converter point cloud data, enhance the semantic segmentation of point cloud, improve the ability to predict point cloud details and identify complex sites, and effectively extract the local and global information features of the converter point cloud model;

[0088] Step 301: uniformly sample c points from the input n point set, and the grouping layer uses the point in the point set c as the center point, and takes out k points within the radius R. Normalize to the unit sphere, convert to a point cloud represented by spherical coordinates as input, and the output shape is B×N×K×3, where 3 is the channel coordinates of x, y, and z;

[0089] Step 302: Learn local point set features, with input N′ and output N′×(d+C′). Perform feature extraction through MLP, and use maximum pooling to aggregate the local features of each point with the features of other points in its neighborhood to obtain a more global feature representation;

[0090] Step 303: The local neighborhood information lost by pooling is recycled and stored in the HWD network module, and as much local feature data as possible is retained to supplement the global features. Finally, the global features are mapped to the output space through the fully connected layer, and segmentation and classification are performed until the features of the entire point set are obtained;

[0091] Step 304: Divide the data set in a ratio of 7:2:1, with the training set accounting for 70%, the validation set accounting for 20%, and the test set accounting for 10%.

[0092] Step 4: Slice the extracted optimal model data, solve the fitting circle based on the sliced ​​circular contour, record the coordinate parameters of the center of the fitting circle, evaluate the deviation from the actual data points through these parameters, and calculate the coaxiality of the installation positions between the components. The process is as follows:

[0093] Step 401: Construct a converter cylindrical model H, whose reference axis is a, the measured axis b is a space curve, the tolerance range is T, the ideal axis is constructed according to the measured axis b, and the contact point e is the maximum error range of the measured axis b with the ideal axis as the center, such as Figure 7 shown.

[0094] Step 402: Determine E1 and E2 at appropriate positions of the furnace shell and furnace body along the height direction of the converter, and then intercept the converter body point cloud slices at equal intervals, such as Figure 8 As shown in the figure, for both ends of the furnace top and furnace bottom, the ideal axis and the reference axis are extended over a long distance, and there are errors in the x-axis and y-axis.

[0095] Step 403: Analyze the two slices obtained. The slice curves at both ends are as follows: Fig. 9 As shown. Based on the circular contour, the fitting circle is solved, as shown Fig.10 As shown. The coordinate parameters of the center of multiple fitting circles are calculated, and there is a deviation between these parameter evaluations and the actual data points. Therefore, the accuracy of the circle fitting center has a great influence on the accuracy of the axis fitting and the coaxiality calculation. The description formula is as follows:

[0096] A(x2 +y 2 )+Bx+Cy+D=0

[0097] Where: Center coordinates radius (A≠0, otherwise it is fitted into a circle with a curvature of 0, B 2 +C 2 -4AD>0). A, B, C, and D only need to be determined to a scalar multiple, and constraint B can be added. 2 +C 2 -4AD=1. Let A=1 and minimize the objective function under the added constraints. The descriptive formula is:

[0098]

[0099]

[0100] The constraint B 2 +C 2 -4AD=1 is transformed into matrix form, and the description formula is:

[0101]

[0102] In order to obtain the best fitting circle, the optimization problem is solved according to the Lagrange multiplier method, and the description formula is as follows:

[0103] L(P,λ)=P T NP-λ(P T QP-1)

[0104] Taking the derivative with respect to P, we get From |Q|≠0, we can conclude that Q -1 NP=λP,λ is Q -1 N is the eigenvalue, and P is the eigenvector. P T NP-AP T QP=0, under the constraint of formula (2) T NP=λ, that is, λ is P T NP extreme value. Since the result of formula (1) is non-negative, in order for the fitting circle to be meaningful, λ must be greater than 0. Therefore, the eigenvector corresponding to the minimum non-negative eigenvalue λ is P, and the fitting circle can be accurately extracted.

[0105] Step 404: According to the calculated fitting circle, fit the cylinder slice model, and divide the cylinder into n sampling section cylinders in the OZ direction for the reference element. Take m points on each section cylinder for measurement. Record multiple sets of point cloud slice fitting circle center coordinates to obtain a set of data about x and y (x i ,y i)(i=1,2,…,n), assuming they have a linear relationship. The method to obtain (x i ,y i ) to fit it so that it satisfies

[0106] beg The method for this set of data The least squares method, is the least squares solution. Finding the least squares solution is essentially finding the optimal parameters a and b so that x i ,y i and As close as possible, Fig.11 shown.

[0107] Step 405: To describe the effect of least squares (the target is the sum of squares) fitting parameters a and b, an optimization function needs to be defined while avoiding the offset of positive and negative values. This function needs to be defined using the sum of squares of the residuals (the difference between the actual observed value and the estimated value (fitted value)).

[0108] The description formula is as follows:

[0109]

[0110] In order to achieve the best fitting effect, the residual sum of squares is obviously minimized, which is converted to the value of a and b when S is the extreme value. To solve the extreme point problem, the partial derivatives of a and b are calculated respectively and set them equal to 0. The description formula is:

[0111]

[0112]

[0113]

[0114]

[0115]

[0116]

[0117] The least squares center position of the Tth slice section of the actual axis being measured is O t (a t , b t , z t ),,like Fig.12 The intersection of the reference axis and the XOY plane of the Tth slice is A0 (x0, b0, z0). The direction vector of the reference axis is Point O t (a t, b t , z t ) to the reference axis and the coaxiality error description formula of the least square method is:

[0118]

[0119] φ f =2max{R t}

[0120] Example

[0121] S1: Prepare the dataset

[0122] In order to verify the effectiveness and accuracy of the method of the present invention, the CISDI converter data set is used for experimental verification. Fig.13 As shown, the total weight of this type exceeds 200 tons, and a laser scanner is used to obtain the converter point cloud data.

[0123] The data set was obtained under four working positions. Since the size and model of the converter are consistent, a sample of the converter was selected for research. Each point cloud occupies more than three million points. 5000 samples were selected, of which 3500 samples were used as training sets, 1000 samples were used as validation sets, and 500 samples were used as test sets. The specific sample division for research is shown in Table 1.

[0124] Table 1 Converter size parameters and data set division

[0125]

[0126] S2: Data preprocessing

[0127] Through Haar wavelet processing, irrelevant signals are filtered out, the overall information is retained, and local detail information is recorded for recycling and storage, retaining as much local feature data as possible to supplement the global features. More attention is paid to important features, the recognition accuracy is improved, the image features are clearly expressed, and good localization properties are achieved in both the time domain and the frequency domain. Under the scale change formula:

[0128]

[0129] Where: i and j represent the order and dimension of the Haar basis function respectively. This function set is defined on the half-open interval [0,1), and the value of each piecewise constant function is 1 in a small range and 0 in other areas.

[0130] S3: Experimental Analysis

[0131] The scanned converter point cloud image data is annotated, such as Fig.14As shown. Training and testing are carried out on the annotated point cloud data to build a converter point cloud extraction and segmentation model. The loss rate is a key indicator for evaluating the network model, such as Fig.15 As shown in Figure 2, the loss rate of point cloud extraction after splicing tends to be stable after 140 iterations. The experimental environment of this paper is shown in Table 2.

[0132] Table 2 Network training environment configuration

[0133]

[0134]

[0135] The F1 value takes into account the proportion of correctly predicted sample results to the total samples, and the proportion of predicted true positive examples (TP) to all actual positive examples. It is a comprehensive coordinated mean evaluation indicator. The final comparison results are as follows: Fig.16 As shown in the figure, the comparison results show that the improved feature extraction model proposed in this paper is better than the original feature extraction model in all converter classifications, and the F1 value is improved. The formula is as follows:

[0136]

[0137] Where: R is the precision and P is the recall.

[0138] In order to verify the recognition performance of converter extraction, this paper trains and tests the model under the condition of many on-site obstructions and analyzes it. The segmentation result data of converter extraction recognition under complex conditions is shown in Table 2.

[0139] Table 2 Network model test results

[0140]

[0141] Based on the multi-scale feature operation processing of point cloud feature extraction and segmentation, the converter point cloud data collected from multiple locations is pre-processed and applied to the improved point cloud feature extraction and segmentation network model, and multi-scale features are intelligently aggregated according to the sparsity of the point cloud quantity. This network model has good segmentation and extraction capabilities, effectively separates background point clouds, and has stronger robustness to the input of point cloud density. The integrated method extracts the comparison before and after Fig.17 shown.

[0142] The slice profile taken from the converter generates a series of data points {x i ,y i}, these data points approximately fall on a circle, and irrelevant noise points are removed to estimate the shape of the actual circumference of the converter as much as possible. Fig.18As shown in the figure, the noise data according to this contour point distribution can accurately calculate the results, and the effect is significant for the data points with errors. The calculation results of the deviation of the center coordinates of 10 sets of converter point cloud fitting circles are collected for inspection. The stability of the center coordinates is tested in the experiment, such as Fig.19 shown.

[0143] On the scanned converter 3D point cloud model elements, measure and calculate multiple cross-sectional circles. Then, use the least square method to construct a 3D straight line as the common axis based on the center of each cross-sectional circle, and then calculate the coaxiality error, such as Fig. 20 The obtained converter point cloud data is used for coaxiality detection. The method used in this paper is compared with the traditional method to obtain multiple sets of data, such as Fig.21 It is shown that the intelligent detection and calculation method of coaxiality has higher accuracy and can be applied to the evaluation of converter installation position accuracy.

[0144] In summary, in order to verify the effectiveness of the method of the present invention, the following conclusions were drawn by comparing the results with those of existing research literature: the error of the deviation detection method proposed in the present invention is lower than that of polar coordinate measurement and other methods, and the accuracy is improved by 6.2% compared with the accuracy before the introduction of wavelet transform processing. The time required is also greatly reduced, which proves that the method can well capture the characteristics and models of point clouds and has high detection capabilities.

[0145] The above contents are merely examples and explanations of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the invention or exceed the scope defined by the claims, they shall all fall within the protection scope of the present invention.

Claims

1. A method for detecting converter installation position deviation using point cloud processing, characterized in that: The method specifically comprises the following steps: Step 1: Collect the point cloud data model information of the converter at the construction site, and use the deep residual network and feature image resolution to solve the problem of unbalanced data sets for aggregation; Step 2: Integrate the deep residual network and feature image processing conversion into the Haar wavelet module, and use the integrated HWD module to process the local input feature signals that are easy to lose, perform feature conversion and reduce spatial resolution; Step 3: Integrate the HWD module into the feature extraction network model, which is in parallel with the overall feature extraction layer. Use the integrated network model to train and identify the point cloud data of the converter; extract local and global information features of the converter point cloud model; Step 4: Use the validation set to determine whether the model is accurate and complete, and select the optimal model as the installation position evaluation model; slice the extracted optimal model data, perform a fitted circle solution based on the sliced ​​circular contour, record the coordinate parameters of the center of the fitted circle, and use these parameters to evaluate the deviation from the actual data points and calculate the coaxiality of the installation positions between components.

2. The method for detecting converter installation position deviation using point cloud processing according to claim 1, characterized in that: Step 1 specifically includes: Step 101: The input of the nth layer is x n , when the input and output dimensions are the same, the input is directly added to the output, then the output formula of the residual block of this layer is as follows: H(x n )=F(x n )+z n When the dimensions are different, a 1×1 convolution is added to adjust the number of channels. The residual block is expressed as follows: y n =H(x n )=F(x n )+g(x n ) Step 102: Input x of the n+1th layer n+1 =f(y n ), g(x n ) is a direct mapping, g(x n )=W n ′x,W n ′ is a 1×1 convolution operation, f(y n ) is the activation function, expressed using Relu, and the formula is as follows: x n+1 =x n +F(x n ) Step 103: For any deeper layer N, its output is the sum of the input of the shallower layer n plus the residual of each subsequent layer, and the formula is as follows: Step 104: For the backward pass, derive the loss function with respect to the output x of any layer n n The gradient of is described as follows; 3. The method for detecting converter installation position deviation using point cloud processing according to claim 1, characterized in that: Step 2 specifically includes: Step 201: Haar wavelet uses a scaling function for signal processing and data analysis. The scaling function is used to describe the change of a signal over time or space. When the scale changes, it can change its shape. The description formula is as follows: Where: i and j represent the order and dimension of the Haar basis function respectively. This function set is defined on the half-open interval [0,1). The value of each piecewise constant function is 1 in a small range and 0 in other areas, as shown below: Step 202: The scaling function and the wavelet basis function are shown in the following formula:

4. The method for detecting converter installation position deviation using point cloud processing according to claim 1, characterized in that: Step three specifically includes: Step 301: uniformly sample s points from the input n point set, take a point in the point set s as the center point, take out m points within the radius R, normalize them to the unit sphere, and convert them into a point cloud represented by spherical coordinates as input. The output shape is B×N×M×3, where 3 is the channel coordinates of x, y, and z. Step 302: Learn local point set features, with input N′ and output N′×(d+C′); extract features through MLP, and use maximum pooling to aggregate the local features of each point with the features of other points in its neighborhood; Step 303: The local neighborhood information lost by the maximum pooling is recycled and stored in the HWD network module and added to the global features; finally, the global features are mapped to the output space through the fully connected layer, and segmentation and classification are performed to obtain the features of the entire point set.

5. The method for detecting converter installation position deviation using point cloud processing according to claim 1, characterized in that: Step 4 specifically includes: Step 401: construct a converter cylindrical model H, where the reference axis is a, the measured axis b is a space curve, the tolerance range is T, the ideal axis is constructed according to the measured axis b, and the contact point e is the maximum error range of the measured axis b with the ideal axis as the center; Step 402: Obtain two groups of equal-height slice models of the converter body, determine E1 and E2 at appropriate positions of the furnace shell and the furnace body along the height direction of the converter, and then intercept point cloud slices at equal intervals; perform coaxiality detection on both ends of the furnace top and the furnace body. The ideal axis and the reference axis are extended over a long distance, and there are errors in the x-axis and y-axis; Step 403: Analyze the two slices obtained; solve the fitting circle based on the circular contour, and calculate the coordinate parameters of the center of multiple fitting circles; Step 404: According to the calculated fitting circle, perform cylinder slice fitting, divide the cylinder into n sampling section cylinders in the OZ direction for the reference element; take m points on each section cylinder for measurement; record multiple sets of point cloud slice fitting circle center coordinates, and obtain a set of data about x and y (x i ,y i )(i=1,2,…,n), assuming that they have a linear relationship; using The method to obtain (x i ,y i ) to fit it so that it satisfies Finding the least squares solution is essentially finding the optimal parameters a and b so that x i ,y i and maximum proximity; Step 405: The description target is the effect of the fitting parameters a and b of the sum of squares. An optimization function needs to be defined while avoiding the offset of positive and negative values. The residual sum of squares needs to be used to define it. The residual is the difference between the actual observed value and the fitted value.

6. The method for detecting converter installation position deviation using point cloud processing according to claim 5, characterized in that: The formula in step 403 is as follows: A(x 2 +y 2 )+Bx+Cy+D=0 Where: Center coordinates radius A≠0,B 2 +C 2 -4AD>0, A, B, C, D only need to be determined to a scalar multiple, and constraint B can be added 2 +C 2 -4AD=1, let A=1, and minimize the objective function under the added constraints; the description formula is: The constraint B 2 +C 2 -4AD=1 is transformed into matrix form, and the description formula is: In order to obtain the best fitting circle, the optimization problem is solved according to the Lagrange multiplier method, and the description formula is as follows: L(P,λ)=P T NT-λ(P T QP-1) Taking the derivative with respect to P, we get From |Q|≠0, we can conclude that Q -1 NP=λP,λ is Q -1 N is the eigenvalue, P is the eigenvector; since P T NP-λP T QP=0, under the constraint of formula (2) T NP=λ, that is, λ is P T NP extreme value; the eigenvector corresponding to the minimum non-negative eigenvalue λ is P, and the fitting circle can be accurately extracted.

7. The method for detecting converter installation position deviation using point cloud processing according to claim 5, characterized in that: The residual sum of squares is described as follows: In order to achieve the best fitting effect, the residual sum of squares is obviously minimized. At this time, it is converted to finding the values ​​of a and b when S is the extreme value. To solve the extreme point problem, the partial derivatives of a and b are calculated respectively, and they are set equal to 0. The description formula is: The least squares center position of the Tth slice section of the actual axis being measured is O t (a t ,b t ,z t ), the intersection of the reference axis and the XOY plane of the Tth slice is A0(x0,b0,z0), and the direction vector of the reference axis is Point O t (a t ,b t ,z t ) to the reference axis and the coaxiality error description formula of the least square method is: f f =2max{R t }。。

Images