A converter installation position deviation detection method using point cloud processing

Abstract

The application discloses a converter installation position deviation detection method using point cloud processing, which uses a HWD module integrated into a network model for extracting features of point clouds after fusing a deep residual network and feature image resolution, and is in parallel with a point cloud extraction layer. In a complex construction environment, the point cloud semantic segmentation is enhanced, the overall accuracy of local and global information extraction of the converter point cloud model is obviously improved, and the required converter model extraction time is saved. Secondly, the optimal model data is intercepted in a circular slice, a fitting circle is solved based on a slice circumferential contour, least square circle center parameters are recorded, the best alignment mode between two axes is determined, the coaxial deviation of the installation position between components is calculated, and the error is effectively reduced. The application still has good monitoring performance in a complex construction environment on site, can well solve the problem of excessive position error of the converter in construction, and has good detection capability.

Description

Technical Field

[0001] This invention relates to a method for detecting positional deviation of a converter body, specifically a method for evaluating converter installation errors using a point cloud processing module. Background Technology

[0002] The converter is a crucial piece of equipment in the steelmaking industry, boasting high efficiency and suitability for smelting various steel grades. The converter body consists of the furnace bottom, furnace body, furnace opening, and supporting ring. The furnace bottom is nearly hemispherical, the furnace body cylindrical, the furnace top a trapezoidal column, and the supporting ring similar to a ring. Converters can weigh hundreds of tons, with diameters and heights exceeding ten meters. Installation is time-consuming, requiring not only minimal alignment deviations but also compliance with thermal expansion requirements, posing significant challenges to adjustments. Currently, the mainstream installation method is the "combination method," where the converter shell is assembled section by section from the furnace bottom to the furnace opening on a temporary support frame. A transport device moves the assembled furnace body to the installation position and then removes the transport device. The entire installation process is difficult to position, slow to adjust, and carries certain construction hazards, severely impacting construction progress. To address these issues, developing intelligent and efficient installation position detection systems has become an inevitable trend.

[0003] With the increasing maturity of engineering site surveying technology, 3D reconstruction has been widely used in engineering. 3D laser scanners can be used to construct 3D point cloud models of construction sites. For large-scale and complex point cloud data, deep learning frameworks for point cloud feature segmentation and extraction have been proposed. However, due to the numerous obstructions on the converter, such as industrial electrical wires, welding boxes, standing support plates, construction workers, converter support frames, and walls on three sides, pooling operations for converter feature extraction can lead to the loss of local spatial information, which is detrimental to point cloud detail prediction and the ability to identify complex sites.

[0004] To facilitate the accurate extraction of converter features from complex point cloud data, the present invention proposes the following technical solution. Summary of the Invention

[0005] The purpose of this invention is to provide a converter installation position deviation detection method using point cloud processing, which solves the problem in the prior art that the pooling operation for converter feature extraction leads to the loss of local spatial information due to numerous obstructions such as industrial wires, welding boxes, standing support plates, construction workers, converter support frames, and walls on three sides, which is detrimental to the ability to predict point cloud details and identify complex sites.

[0006] The objective of this 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 includes the following steps:

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

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

[0010] Step 3: Integrate the HWD module into the feature extraction network model, which is connected 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 from the converter point cloud model.

[0011] Step 4: Determine the accuracy and completeness of the model using the validation set, and select the optimal model as the evaluation model for the installation position; extract the optimal model data into slices, and solve for the fitting circle based on the circumferential contour of the slices, record the coordinate parameters of the center of the fitting circle, evaluate the deviation between the actual data points using these parameters, and calculate the coaxiality of the installation positions between the components.

[0012] Furthermore, step one specifically includes:

[0013] Step 101: The input of the nth layer is When the input and output dimensions are the same, the input is directly added to the output. The output formula for this residual block is as follows: ;

[0014] When the dimensions are different, a 1×1 convolution is added to adjust the number of channels. The residual block representation formula is as follows: ;

[0015] Step 102: Input of the (n+1)th layer For direct mapping, This is a 1×1 convolution operation. It is the activation function, represented by ReLU, and the formula is as follows: ;

[0016] Step 103: For any deeper layer N, its output is the sum of the input of the shallower layer n plus the residuals of each subsequent layer, as shown in the following formula: ;

[0017] Step 104: For backpropagation, derive the output of the loss function for any layer n. The gradient of is described by the following formula: .

[0018] Furthermore, step two specifically includes:

[0019] Step 201: Haar wavelets use scaling functions for signal processing and data analysis. Scaling functions describe how a signal changes over time or space; they can change shape as the scale changes. The formula is as follows: In the formula: i and j represent the order and dimension of the Haar basis functions, respectively. This set of functions 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 regions, as shown below:

[0020] ;

[0021] Step 202: The scaling function and wavelet basis functions are shown in the following equations:

[0022] .

[0023] Furthermore, step three specifically includes:

[0024] Step 301: Sample s points uniformly from the input set of n points. Using the points in the set of points s as the center point, take out m points within the radius R, normalize them to the unit sphere, and convert them into point clouds represented by spherical coordinates as input. Then the output shape is B×N×M×3, where 3 is the channel coordinate of x, y, z.

[0025] Step 302: Learn local point set features, with the input being... The output is Feature extraction is performed using MLP, and max pooling is used to aggregate the local features of each point with the features of other points in its neighborhood.

[0026] Step 303: The local neighborhood information of the maximum pooling loss 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 a fully connected layer to perform segmentation and classification, so as to obtain the features of the entire point set.

[0027] Furthermore, step four specifically includes:

[0028] Step 401: Construct a converter cylindrical model H, with reference axis a, measured axis b being a space curve with tolerance zone T, ideal axis being constructed based on measured axis b, and contact point e being the maximum error range of measured axis b centered on ideal axis.

[0029] Step 402: Obtain two sets of equal-height slice models of the converter body, and determine the appropriate positions of the converter along the height direction in the furnace shell and furnace body respectively. and Then, point cloud slices are cut at equal intervals; coaxiality is tested at both ends of the furnace top and furnace body. The ideal axis and the reference axis have errors in the x-axis and y-axis after a long extension.

[0030] Step 403: Analyze the two obtained slices; solve for the fitted circle based on the circular contour, and calculate the center coordinate parameters of multiple fitted circles;

[0031] Step 404: Following the calculated fitted circle, perform cylindrical slice fitting, dividing the cylinder into n sampling cross-section cylinders along the OZ direction for the reference features; take m points on each cross-section cylinder for measurement; record the center coordinates of multiple sets of point cloud slice fitted circles to obtain a set of data regarding x and y. Set them to have a linear relationship; use The method of obtaining Perform a fitting to satisfy the following conditions: Finding the least squares solution essentially means finding the optimal parameters a and b, such that... Approaching as closely as possible;

[0032] Step 405: Describe the effect of fitting parameters a and b on the sum of squares. An optimization function needs to be defined. To avoid the cancellation of positive and negative values, the residual sum of squares needs to be used. The residual is the difference between the actual observed value and the fitted value.

[0033] The beneficial effects of this invention are:

[0034] Compared to most methods that employ classic point cloud feature extraction and segmentation, this invention integrates a deep residual network and a feature image-resolved HWD module into the point cloud feature extraction network model, operating in parallel with the point cloud extraction layer. This enhances point cloud semantic segmentation in complex construction environments, significantly improving the overall accuracy of local and global information extraction from the converter point cloud model and saving time. Secondly, the extracted optimal model data is sliced ​​into circular pieces, and a fitted circle is solved based on the circumferential contour of the slices. The least-squares center parameters are recorded to determine the optimal alignment between two axes, and the coaxial deviation of the installation positions between components is calculated, effectively reducing errors. Finally, experiments show that this invention maintains good monitoring performance even in complex on-site construction environments, effectively solving the problem of excessive errors in converter assembly positions during construction and demonstrating excellent detection capabilities. Attached Figure Description

[0035] The invention will now be further described with reference to the accompanying drawings.

[0036] Figure 1 This is a schematic diagram of the research method architecture described in this invention;

[0037] Figure 2 This is a schematic diagram of the construction environment at the converter site;

[0038] Figure 3This is a schematic diagram of the deep residual network unit described in this invention;

[0039] Figure 4 This is a schematic diagram of the wavelet decomposition process described in this invention;

[0040] Figure 5 This is a schematic diagram of the HWD module network structure described in this invention;

[0041] Figure 6 This is a schematic diagram of the improved PointNet++ network structure described in this invention;

[0042] Figure 7 This is a schematic diagram of the coaxiality model error curve described in this invention;

[0043] Figure 8 This is a schematic diagram of point cloud slices at the upper and lower ends of the converter described in this invention;

[0044] Figure 9 This is a schematic diagram of the slice curves at both ends of the converter body described in this invention;

[0045] Figure 10 This is a schematic diagram of fitting a cylindrical slice after the fitting circle calculation as described in this invention;

[0046] Figure 11 This is a schematic diagram summarizing the least squares method described in this invention;

[0047] Figure 12 This is a schematic diagram of the contour detection of the cylindrical geometric model described in this invention;

[0048] Figure 13 This is a schematic diagram of the converter three-dimensional reconstruction measurement work described in this invention;

[0049] Figure 14 This is a schematic diagram of the converter training set annotation described in this invention;

[0050] Figure 15 This is a schematic diagram illustrating the changing trend of the loss function described in this invention;

[0051] Figure 16 This refers to the PointNet++ model before and after improvements described in this invention. Value comparison diagram;

[0052] Figure 17 This is a comparative diagram of the extraction of converter point clouds by the network model before and after the improvement described in this invention;

[0053] Figure 18 This is a schematic diagram of the fitting circle test experiment described in this invention;

[0054] Figure 19This is a schematic diagram of the test results of the center point of the fitted circle as described in this invention;

[0055] Figure 20 This is a schematic diagram of the coaxiality calculation results of the circular cross section described in this invention;

[0056] Figure 21 This is a schematic diagram of coaxiality error comparison according to the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] A method for detecting converter installation position deviation using point cloud processing, such as Figure 1 As shown, it includes the following steps:

[0059] Step 1: First, place the laser scanner around the converter, and then begin taking pictures around it, such as... Figure 2 As shown. The goal is to acquire its point cloud data, use deep residual networks and feature image resolution to address the problem of imbalanced datasets by aggregation, and improve the ability of sampling operations and learning feature transformations;

[0060] Step one specifically includes: identifying and resolving the problem of imbalanced datasets through aggregation, and improving the ability of sampling operations and learning feature transformations.

[0061] Step 101: The input of the nth layer is When the input and output dimensions are the same, the input is directly added to the output. This is equivalent to performing a simple equivalent mapping, without generating additional parameters or increasing computational complexity. Therefore, the output formula for this residual block is as follows:

[0062] ;

[0063] When the dimensions are different, a 1×1 convolution is added to adjust the number of channels. For forward propagation, this residual learning structure can be implemented by adding direct edges to a feedforward neural network, and the entire network can still be trained through end-to-end backpropagation, such as... Figure 3 As shown. The residual block representation formula is as follows: ;

[0064] Step 102: Input of the (n+1)th layer For direct mapping, This is a 1×1 convolution operation. It is the activation function, represented by ReLU, and the formula is as follows: ;

[0065] Step 103: For any deeper layer N, its output is the sum of the input of the shallower layer n plus the residuals of each subsequent layer, as shown in the following formula: ;

[0066] Step 104: For reverse propagation, the loss function with respect to any layer n can be derived from the above approach. The gradient of is described by the following formula: .

[0067] This demonstrates that the gradient of the loss function with respect to deeper layers N can be directly propagated to any shallower layer. Residual networks can propagate signals directly to any layer, whether forward or backward.

[0068] 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 easily lost local input feature signals, perform feature transformation and reduce spatial resolution.

[0069] Step 201: Haar wavelets use scaling functions for signal processing and data analysis. Scaling functions describe how a signal changes over time or space; they can change shape when the scale changes, such as... Figure 4 As shown, the formula is described as follows:

[0070] In the formula, i and j represent the order and dimension of the Haar basis functions, respectively. This set of functions 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 regions, as shown below:

[0071] ;

[0072] Step 202: Basis functions are the core components of wavelet transform, capable of capturing local features of a signal at different time and frequency scales. The scaling function and wavelet basis functions are shown in the following equations:

[0073] .

[0074] The Haar wavelet transform exhibits excellent localization properties in edge detection, feature texture analysis, and both the time and frequency domains. It boasts a simple structure, high computational efficiency, and the ability to focus on arbitrary details of an object, such as... Figure 5 As shown.

[0075] Step 3: Integrate the HWD module into the feature extraction network model, where it is connected in parallel with the extraction layer, such as... Figure 6As shown, the improved network model is used to train and identify point cloud data of the converter, enhancing point cloud semantic segmentation, improving the ability to predict point cloud details and identify complex sites, and realizing the effective extraction of local and global information features of the converter point cloud model;

[0076] Step 301: Uniformly sample c points from the input set of n points. For the grouped layer, take a point from the set of c as the center point and extract k points within a radius R. Normalize to a unit sphere and convert to a point cloud represented by spherical coordinates as input. The output shape is B×N×K×3, where 3 represents the channel coordinates of x, y, and z.

[0077] Step 302: Learn local point set features, with the input being... The output is Feature extraction is performed using MLP, and max pooling is used to aggregate the local features of each point with the features of other points in its neighborhood to obtain a more global feature representation;

[0078] Step 303: The local neighborhood information from the pooling loss is recycled and stored in the HWD network module, retaining as much local feature data as possible to supplement the global features. Finally, the global features are mapped to the output space through a fully connected layer for segmentation and classification until the features of the entire point set are obtained;

[0079] Step 304: Divide the dataset into three parts according to 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%.

[0080] Step 4: Extract the optimal model data into slices, perform a fitting circle calculation based on the circular contour of the slices, record the coordinate parameters of the fitted circle's center, and use these parameters to evaluate the deviation from the actual data points, calculating the coaxiality of the installation positions between components. The process is as follows:

[0081] Step 401: Construct a converter cylindrical model H, with its reference axis as a, the measured axis b as a space curve, and a tolerance zone of T. The ideal axis is constructed based on the measured axis b, and the contact point e represents the maximum error range of the measured axis b centered on the ideal axis. Figure 7 As shown.

[0082] Step 402: Determine the appropriate positions for the converter along the height direction on the furnace shell and furnace body respectively. and Then, point cloud slices of the converter body are cut at equal intervals, such as... Figure 8 As shown, for both ends of the furnace top and bottom, the ideal axis and the reference axis extend over a long distance, resulting in errors in the x-axis and y-axis.

[0083] Step 403: Analyze the two obtained slices, and the slice curves at both ends are as follows. Figure 9 As shown. The method involves fitting a circle based on a circular profile, as follows: Figure 10 As shown. The coordinate parameters of the centers of multiple fitted circles are calculated, and these parameter estimates deviate from the actual data points. Therefore, the accuracy of the circle fitting centers has a significant impact on the accuracy of the axis fitting and the coaxiality calculation. The formula is as follows:

[0084] Where: coordinates of the center of the circle ,radius ( Otherwise, it will be fitted to a circle with a curvature of 0. For A, B, C, and D, only a scalar multiple needs to be determined; constraints can be added. .make And minimize the objective function under the added constraints. The formula is as follows:

[0085] ;

[0086] , (1)

[0087] Constraints Transforming into matrix form, the formula is as follows:

[0088] (2)

[0089] To obtain the best-fit circle, the optimization problem is solved using the Lagrange multiplier method, described by the following formula:

[0090] ;

[0091] Differentiating with respect to P, we get ,Depend on The conclusion is , yes The eigenvalues ​​of are , and P is the eigenvector; since ,Place Under the constraints of equation (2) That is, λ is Extreme values. Since the result of equation (1) is non-negative, λ must be greater than 0 for the fitted circle to be meaningful. Therefore, the eigenvector corresponding to the smallest non-negative eigenvalue λ is P, which can accurately extract the fitted circle.

[0092] Step 404: Following the calculated fitted circle, fit the cylindrical slice model, dividing the cylinder into n sampling cross-section cylinders along the OZ direction based on the reference features. Take m points on each cross-section cylinder for measurement. Record the center coordinates of multiple sets of point cloud slice fitted circles to obtain a set of data regarding x and y. (i=1, 2, ..., n), assuming they have a linear relationship. Using The method of obtaining Perform a fitting to satisfy the following conditions: .

[0093] beg The method is for this set of data Least squares method, It is a least squares solution. Finding the least squares solution essentially means finding the optimal parameters a and b such that... As close as possible, such as Figure 11 As shown.

[0094] Step 405: To describe the effect of least squares (target: sum of squares) fitting of parameters a and b, an optimization function needs to be defined. To avoid offsetting positive and negative values, it needs to be defined using the sum of squares of the residuals (the difference between the actual observed values ​​and the estimated values ​​(fitted values)). The formula is as follows:

[0095] ;

[0096] To achieve the best fit, the sum of squared residuals must be minimized. This transforms the problem into finding the values ​​of a and b when S is an extremum. To solve the extremum problem, we take the partial derivatives of a and b respectively and set them equal to 0. The formula is:

[0097] ;

[0098] The least-squares center position of the Tth slice section of the actual measured axis is ,like Figure 12 The intersection point of the reference axis shown and the XOY plane of the T-th slice is: The direction vector of the reference axis is ,point The distance to the reference axis and the coaxiality error described by the least squares method are as follows:

[0099] ;

[0100] Example

[0101] S1: Preparing the dataset

[0102] To verify the effectiveness and accuracy of the method of this invention, experimental verification was conducted using the CISDI converter dataset. The converter experimental equipment is as follows: Figure 13 As shown, this type has a total weight of over 200 tons and uses a laser scanner to acquire converter point cloud data.

[0103] This dataset was obtained from four work locations. Since the converters were of the same size and model, samples from one converter were selected for the study. Each point cloud contains more than three million points, and 5,000 samples were selected, of which 3,500 samples were used as the training set, 1,000 samples as the validation set, and 500 samples as the test set. The specific sample division for the study is shown in Table 1.

[0104] Table 1. Division of converter size parameters and datasets

[0105] ;

[0106] S2: Data Preprocessing

[0107] Haar wavelet processing filters out irrelevant signals, preserves overall information, and records local details for recycling and storage, retaining as much local feature data as possible to supplement the global features. It focuses more on important features, improves recognition accuracy, clearly expresses image features, and exhibits good localization properties in both the time and frequency domains. The scaling formula is as follows:

[0108] In the formula, i and j represent the order and dimension of the Haar basis functions, respectively. This set of functions 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 regions.

[0109] S3: Experimental Analysis

[0110] The scanned converter point cloud image data is labeled, such as... Figure 14 As shown. Training and testing were conducted on the labeled point cloud data to construct a converter point cloud extraction and segmentation model. The loss rate is a key metric for evaluating the network model, such as... Figure 15 The loss rate of point cloud extraction after stitching stabilizes after 140 iterations. The experimental environment of this paper is shown in Table 2.

[0111] Table 2 Network Training Environment Configuration

[0112] ;

[0113] The value considers the proportion of correctly predicted samples to the total sample and the proportion of correctly predicted positive examples (TP) to all actual positive samples, making it a comprehensive and balanced mean evaluation index. The final comparison results are as follows: Figure 16 As shown in the figure, the comparison results demonstrate that the improved feature extraction model proposed in this paper outperforms the original feature extraction model in all converter classifications. The value has increased. The formula is as follows: In the formula: R is precision and P is recall.

[0114] To test the converter extraction and recognition performance, this paper trains and tests the model under conditions with numerous obstructions on site and analyzes the results. Table 3 shows the converter extraction, recognition, and segmentation results under complex conditions.

[0115] Table 3 Network Model Test Results

[0116] ;

[0117] Based on point cloud feature extraction and segmentation with multi-scale feature processing, converter point cloud data collected from multiple locations is preprocessed and applied to an improved point cloud feature extraction and segmentation network model. Multi-scale features are intelligently aggregated according to the sparsity of the point cloud. This network model has excellent segmentation and extraction capabilities, effectively separating background point clouds and exhibiting stronger robustness to point cloud density inputs. The comparison before and after extraction using the ensemble method is shown in the figure. Figure 17 As shown.

[0118] A series of data points are generated from the slice outline cut by the converter. These data points approximately fall on a circle, and irrelevant noise points are removed to estimate the true circumference shape of the converter as accurately as possible. Actual tests are as follows... Figure 18 As shown, the results can be accurately calculated based on the noise data of this contour point distribution, and the effect is significant for data points with errors. The results of the calculated deviation of the center coordinates of 10 sets of converter point agglomerations were collected for verification. The stability of the center coordinates was tested in the experiment. Figure 19 As shown.

[0119] On the 3D point cloud model elements of the converter obtained by scanning, circles of multiple cross-sections are measured and calculated. Then, using the centers of each cross-sectional circle, a 3D straight line is constructed using the least squares method as a common axis. The coaxiality error is then calculated, such as... Figure 20 As shown. The acquired converter point cloud data is subjected to coaxiality detection. Multiple sets of data are compared between the method used in this paper and traditional methods, such as... Figure 21 As shown, the intelligent detection and calculation method for coaxiality has higher accuracy and can be applied to the assessment of converter installation position accuracy.

[0120] In summary, to verify the effectiveness of the method of this invention, the results were compared with those of existing research literature, and the following conclusions were drawn: the deviation detection method proposed in this invention has a lower error rate than methods such as polar coordinate measurement, and the accuracy is improved by 6.2% compared with the method before introducing wavelet transform processing. The time required is also greatly reduced, which proves that the method can capture the features and model of point clouds very well and has a high detection capability.

[0121] The above description is merely an example and illustration of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

Claims

1. A method for detecting a deviation of a mounting position of a converter by using a point cloud process, characterized by, The method specifically comprises the following steps: Step one: collect the point cloud data model information of the converter at the construction site, use the deep residual network and feature image resolution to solve the problem of unbalanced data set and aggregate; Step two: convert the deep residual network and feature image processing into the Haar wavelet module, use the processed HWD module to process the easily lost local input feature signal, perform feature conversion and reduce spatial resolution; Step three: 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 the local and global information features of the converter point cloud model; Step four: judge whether the model is accurate and complete through the validation set, select the optimal model as the installation position evaluation model; cut the extracted optimal model data into slices, solve the fitting circle based on the circular contour, record the fitting circle center coordinate parameters, evaluate the deviation between these parameters and the actual data points, and calculate the coaxiality of the installation position between the components.

2. The method of claim 1, wherein the method is a method of detecting a deviation of a mounting position of a converter using point cloud processing, characterized by, Step one specifically comprises: Step 101: the input of the nth layer is When the input and output dimensions are consistent, the input is directly added to the output, and then the output formula of the residual block of this layer is as follows: ; When the dimensions are different, a 1x1 convolution is added to adjust the number of channels, and the residual block expression formula is as follows: ; Step 102: input of the nth+1 layer For direct mapping, For 1x1 convolution operation, is an activation function, represented by Relu, with the formula as follows: ; Step 103: For any deeper layer N, its output is the input of the shallower layer n plus the sum of the residuals of each layer after it, as follows: ; Step 104: For the backward pass, the gradient of the loss function with respect to the output of any layer n is derived, described by the equation: .​ 3. The method of claim 1, wherein the method is a method of detecting a deviation of a mounting position of a converter using point cloud processing, characterized by, Step two specifically comprises: Step 201: Haar wavelet uses scale function for signal processing and data analysis, scale function is used to describe signal change with time or space, it can change shape when scale changes, description formula is as follows: In the formula: i and j respectively represent order and dimension of Haar base function, the 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 is 0 in other areas, as follows: ； Step 202: the scale function and wavelet base function are as follows: 。 4. The method of claim 1, wherein the method is a method of detecting a deviation of a mounting position of a converter using point cloud processing, characterized by, Step three specifically comprises: Step 301: uniformly sample s points in the input n point set, take the points in the point set s as the center point, take m points within the radius R, normalize to the unit sphere, and convert the point cloud represented by the spherical coordinates as input, then the output shape is BxNXMx3, 3 is the channel coordinates of x, y and z; Step 302: learn local point set features, input is , output is ; feature extraction by MLP, aggregate each point's local features with other points' features in its neighborhood using max pooling; Step 303: recycle and store the local neighborhood information of the maximum pooling loss in the HWD network module, and supplement it to the global feature; finally, map the global feature to the output space through the fully connected layer, implement segmentation and classification, and obtain the features of the entire point set.

5. The method of claim 1, wherein the method is a method of detecting a deviation of a mounting position of a converter using point cloud processing, characterized by, Step four specifically comprises: Step 401: construct a converter cylindrical model H, the reference axis is a, the measured axis b is a space curve, the tolerance band 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, and determine the converter along the height direction at appropriate positions of the furnace shell and the furnace body and Then take the point cloud slices at equal intervals; The coaxiality of the furnace top and the ends of the furnace body is detected, and the ideal axis and the reference axis extend through a long distance, and there is an error in the x-axis and the y-axis; Step 403: analyze the two parts of the slice obtained; based on the circular contour, solve the fitting circle, and calculate the center coordinate parameters of multiple fitting circles; Step 404: In accordance with the calculated fitting circle, the cylindrical slice fitting is performed to divide the cylinder into n sampling cross-sectional cylinders in the OZ direction of the reference element; m points are measured on each cross-sectional cylinder; a plurality of sets of point cloud slice fitting circle center coordinates are recorded to obtain a set of data about x and y , it is set that they have a linear relationship; fitting is performed on the obtained in a manner of , so that it satisfies ; the essence of finding the least square solution is to find the optimal parameters a and b, so that is closest. Step 405: describe the effect of fitting parameters a and b which is the square sum of the target, need to define an optimization function, at the same time avoid positive and negative offset, need to use residual sum of squares to define, residual is the difference between actual observation value and fitting value.

6. The method of claim 5, wherein the method is a method of detecting a deviation of a mounting position of a converter using point cloud processing, characterized by, The formula in step 403 is as follows: Wherein: the center coordinate of the circle is , the radius is , , , A, B, C, D only need to be determined to a scalar multiple, a constraint can be added , let , and minimize the objective function under the added constraint condition; the formula is described as: , , ; Constraints Transforming into matrix form, the formula is as follows: (2) To obtain the best fitting circle, the optimization problem is solved according to the Lagrange multiplier method, and the formula is described as follows: ; Taking derivative of P gives , and , , is eigenvalue, P is eigenvector; since , so , under the constraint of formula (2) , that is, λ is extreme value; the eigenvector corresponding to the minimum non-negative eigenvalue λ is P, which can accurately extract the fitting circle.

Images