Elevator traction sheave abrasion laser measurement method based on artificial intelligence

Abstract

The invention discloses an artificial intelligence-based elevator traction sheave wear laser measurement method, which realizes zero-target, high-speed and high-precision sheave groove wear detection through coordination of a three-stage neural network of'point cloud layer-parameter layer-diagnosis layer 'and laser triangular scanning: a laser emits a radial fan-shaped laser surface, an encoder synchronously outputs angle information, and the angle information of the laser surface is detected; collecting a whole circle of pixel point cloud; after the point cloud is subjected to coordinate transformation, solving the pitch circle radius of the traction wheel through least square iteration to serve as a reference; de-noising the point cloud layer by using a K proximity graph and a graph edge convolutional network, and outputting a pixel-level mask; the parameter layer predicts the pulse interval through two layers of LSTM, and the angle is corrected through Kalman filtering; and five-layer convolution sharing characteristics of the diagnosis layer are realized, the three detection heads output the maximum wear depth and the maximum eccentric wear index respectively, the results of the first two ends are spliced with the pitch radius and the elevator rated parameters, and the remaining life is regressed by the third detection head.

Description

TECHNICAL FIELD

[0001] The application relates to an elevator traction sheave wear laser measurement method based on artificial intelligence and belongs to the field of elevator traction sheave wear laser measurement. BACKGROUND

[0002] The traction sheave is a key component for normal operation of an elevator, directly affects safe and stable operation of the elevator, and is important to perform wear detection. Traction sheave groove wear is a primary quantitative signal for degradation of elevator traction capacity, and the depth directly determines the steel wire rope contact stress and the traction coefficient. Excessive wear will induce steel wire rope slip, car squat or top collision, is the first quantitative threshold for elevator safety, and is the only measurable entrance for predictive maintenance. However, current research and field application still generally remain at the stage of plug gauges, angle gauges or handheld laser profilometers, need manual target marking, have long elevator stopping time, are significantly disturbed by oil stains and burrs, and the measurement uncertainty and accuracy are difficult to meet actual needs. Meanwhile, the current laser measurement method has a loophole that encoder jitter is not quantified, so that the arc length error cannot be inhibited and the measurement result is affected. Moreover, wear and remaining life mapping still relies on experience, and predictive maintenance lacks uncertainty support, resulting in coexistence of excessive maintenance and sudden stop. In view of the above problems in traction sheave wear detection, such as low accuracy, complex process, significant oil stain and burr disturbance, unquantified encoder jitter and experience-based wear-life mapping, the application provides a traction sheave wear laser detection method based on artificial intelligence, designs a laser point cloud imaging system and a traction sheave wear artificial intelligence detection method, and embeds three-level artificial intelligence modules of neural network denoising, encoder angle correction and multi-task prediction into a laser triangulation full link. High-speed, accurate and zero-target traction sheave groove wear detection is achieved, and has wide application market space and economic value. SUMMARY

[0003] The application is a traction sheave wear laser measurement method based on artificial intelligence.

[0004] The technical scheme adopted by the application is as follows:

[0005] A traction sheave wear laser measurement method based on artificial intelligence comprises the following steps:

[0006] A laser fixed to the side wall of the traction machine emits a radial fan-shaped laser surface to the bottom of the traction sheave groove, an encoder installed at the end of the traction machine main shaft synchronously acquires rotation angle information of the traction sheave, and a laser receiving device collects point cloud data under a pixel coordinate system of one rotation of the traction sheave;

[0007] Based on the laser triangulation constraint relationship, the point cloud data is converted from the pixel coordinate system to the internal coordinate system of the laser, the rotation angle information is used as the polar angle to convert the point cloud data into the polar coordinate system, and then the three-dimensional world coordinate system coordinates are obtained;

[0008] Based on the three-dimensional world coordinate system coordinates, the traction wheel pitch circle radius is solved;

[0009] The K-neighbor graph is constructed for each laser point through the point cloud layer to search for the nearest neighbor points to form an undirected edge and generate an edge feature, the edge feature is processed through a graph edge convolution network to identify abnormalities, remove noise and extract local geometric features, and a pixel-level denoising mask is outputted;

[0010] The parameter layer adopts a long short-term memory network to process the adjacent pulse time interval sequence of the encoder, predict the next pulse interval, and output a correction angle and a confidence degree, and a Kalman filter is used to optimally weight and fuse the correction angle and the original angle coding pulse according to the confidence degree, to obtain accurate angle data;

[0011] The pixel-level denoising mask and the accurate angle data are inputted into a five-layer convolution module of the diagnosis layer to extract basic features; the output features of the five-layer convolution module are respectively supplied to three detection heads, the first detection head outputs the maximum wear depth of the traction wheel, the second detection head outputs the eccentric wear index, and the third detection head outputs the remaining life estimation value after splicing the first detection head output, the second detection head output, the traction wheel pitch circle radius and the rated parameters of the elevator with the basic features.

[0012] Further, the radial fan-shaped laser surface emitted by the laser passes through the center of the traction wheel main shaft and is perpendicular to the wheel groove axis, and the laser 、 The pixel equivalent in the direction is obtained by factory calibration.

[0013] Further, the encoder outputs 1024 angle coding pulses when the traction wheel rotates one revolution, and the rotation angle of the traction wheel corresponding to adjacent laser pulses is 360° / 1024.

[0014] Further, in the point cloud layer, the K-neighbor graph searches for 20 nearest neighbor points for each laser point to form an undirected edge, and the edge feature is the distance between two laser point feature vectors, and the feature vector includes the transverse coordinate and the depth coordinate of the laser point in the internal coordinate system of the laser.

[0015] Further, the long short-term memory network of the parameter layer is a two-layer structure, and the hidden layer dimension of each long short-term memory network is 64 dimensions, and a Dropout layer is arranged after the two long short-term memory networks of the parameter layer.

[0016] Further, the optimal weighted fusion of the Kalman filter, specifically, the weight proportion of the corrected angle and the original angle encoding pulse is dynamically adjusted according to the confidence, so that the fusion of the LSTM prediction result and the original pulse data is realized.

[0017] Further, in the five-layer convolution module of the diagnosis layer, each layer of convolution adopts a configuration of a 3*3 convolution kernel, a step of 1 and padding of 1, and after each layer of convolution, batch normalization and ReLU activation operations are sequentially performed, and a 2*2 maximum pooling layer is arranged between adjacent two convolution modules.

[0018] The three detection heads share the convolution features output by the five-layer convolution module, and the three detection heads adopt a differentiated loss weighting manner.

[0019] Further, in the solving process of the pitch circle radius of the traction sheave, a least square objective function is constructed by calculating the residual of the point set on the wheel groove to the wheel shaft center, and the objective function is solved by using a Gauss-Newton iteration method, and the Gauss-Newton iteration method involves the element index of the Jacobian vector and the mean value calculation function.

[0020] Further, the radius of the polar coordinate system is the depth coordinate of the laser point in the internal coordinate system of the laser, and the polar angle is the rotation angle output by the encoder; when the polar coordinates are converted to the three-dimensional world coordinate system, the rotation center of the main shaft of the traction sheave is taken as the origin of the world coordinate system.

[0021] Further, the elevator rated parameters specifically include the hardness of the elevator traction sheave, the rated running speed of the elevator and the rated load of the elevator.

[0022] The present application has the following beneficial effects:

[0023] (1) By cooperating laser triangulation scanning with a three-level neural network, the traction sheave wear detection can be completed without attaching a target and stopping the elevator, the on-site operation time is shortened, and the artificial intervention intensity is reduced.

[0024] (2) By adopting point cloud layer denoising and parameter layer angle correction, the influence of oil stains, burrs and encoder jitter on the measurement result can be reduced, and the data consistency is improved.

[0025] (3) Taking the pitch circle radius as the reference, the least square iteration is used for real-time solving, the reference deviation caused by different elevator manufacturing tolerances is avoided, and the wear amount calculation is more universal.

[0026] (4) The diagnosis layer outputs the maximum wear depth, the eccentric wear index and the remaining life in a multi-task parallel manner, and provides the quantitative information required for maintenance decision-making at one time, and reduces the repeated work of step-by-step detection.

[0027] (5) Hardware deployment only involves side wall fixed laser and shaft end encoder, without modification of traction machine body, simple installation and removal process, suitable for batch installation on in-use elevators. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 is a schematic diagram of a laser measurement system structure.

[0029] Figure 2 is a schematic diagram of a laser triangulation measurement principle.

[0030] Figure 3 is a schematic diagram of a traction sheave point cloud processing artificial intelligence algorithm structure.

[0031] Figure 4 is a traction sheave detection flowchart. DETAILED DESCRIPTION

[0032] The application will be further described below in conjunction with the drawings.

[0033] An elevator traction sheave wear laser measurement method based on artificial intelligence, comprising the following steps:

[0034] Step 1: Construct a traction sheave groove wear laser measurement system.

[0035] Traction sheave groove wear directly affects the contact stress of elevator steel wire rope and the traction coefficient, thereby bringing safety hazards to the operation of the elevator. Since the wear of the groove 1 is usually a process from quantitative change to qualitative change, the change of the wear amount is very subtle, and common detection methods are difficult to accurately, timely and reliably measure. Laser measurement has the advantages of non-contact and high precision, and is one of the optimal measurement schemes under the sub-millimeter wear standard, so the application adopts a laser measurement method to detect the wear of the traction sheave groove, and the designed laser measurement system structure is as shown in Figure 1 The system includes a laser 2 and an encoder 3. The laser is fixed on the side wall of the traction machine, and the light fan emitted by the laser is directed to the bottom of the groove along the radial direction (normal direction) of the traction sheave, perpendicular to the groove axis, so that laser measurement can be performed on the bottom of the traction sheave groove. Figure 1 The dashed rectangle A is the laser scanning area, and B is the receiving device (such as a CMOS / CCD image sensor, etc.). The receiving device rotates through a full circle along the main shaft D to achieve one-frame coverage and zero-splicing high-precision sampling. The encoder is installed at the shaft end (non-load end) of the main shaft of the traction machine and is coaxially connected to the main shaft through an elastic coupling, and rotates with the wheel. Each rotation provides 1024 angle encoding pulses.

[0036] Step 11: Measurement point coordinate conversion

[0037] Let the center of the traction sheave main shaft be the origin of the world coordinate system , and the internal coordinate system of the laser be The pixel coordinate system of the laser receiving device is Let a point on the traction sheave be , and its coordinates in the laser internal coordinate system be , respectively representing the lateral coordinate and the depth coordinate; and the pixel coordinates of the optical axis center point be . According to the laser measurement triangular constraint relationship, there are:

[0038] ,

[0039] wherein, , are the pixel equivalents of the laser in the and directions, representing the conversion relationship between the pixel coordinates and the physical coordinates, which are calibrated by the laser manufacturer at the time of factory delivery. The measurement principle diagram of the laser triangulation is shown in Figure 2 .

[0040] Because the laser fan plane passes through the main shaft center and is perpendicular to the sheave axis, a polar coordinate system can be established, which can be regarded as the polar coordinate system radius , is the height , and the angle is given by the encoder, which is set as , so the formula for converting the laser plane coordinates into polar coordinates is:

[0041] ,

[0042] wherein, is the absolute angle of the kth laser line, is the angle of the traction sheave rotation between adjacent laser pulses, and in the present application, the traction sheave rotates one revolution, and the laser emits 1024 pulses, so .

[0043] The polar coordinates can be converted into three-dimensional world coordinate system coordinates using the following formula:

[0044] ,

[0045] Therefore, the coordinates of the point on the kth frame point column output by the laser scanning are as follows, wherein is the index of the point.

[0046] ,

[0047] ​Wherein, u is the transverse pixel column number of the laser point in the pixel coordinate system (original image column index). K is the encoder pulse number (0-1023), corresponding to the kth laser line. V is the vertical pixel row number of the laser point in the pixel coordinate system (original image row index). X, y, z are the coordinates of the laser point in the three-dimensional world coordinate system after conversion (the origin is set at the rotation center of the main shaft). I is the serial number (index) of the current laser point in the frame. K is the frame number (i.e. the encoder pulse number, 0-1023). X i k , Y i k , Z i k are the three-dimensional world coordinates of the i th laser point in the k th frame.

[0048] Step 12: Calculation of the groove wear amount

[0049] Let the radius of the traction sheave pitch circle be , and the residual error of the point set on the groove to the wheel shaft center is calculated to construct a least squares objective function:

[0050] ,

[0051] Solve the above equation using the Gauss-Newton iterative method:

[0052] ,

[0053] Wherein, is the element index of the Gauss-Newton iterative method Jacobian vector, and mean is the mean value calculation function. The value of can be quickly obtained by the Gauss-Newton iterative method, providing a basis for further processing.

[0054] Step 2: Traction sheave point cloud processing based on artificial intelligence.

[0055] The common types of traction sheave groove wear include wear, burr, and dirt, and the vibration generated during the operation of the traction sheave may cause unstable laser point cloud imaging.

[0056] To solve these problems, the present application proposes a three-stage neural network processing algorithm of "point cloud layer-parameter layer-diagnosis layer", and the overall structure of the system is as shown in Figure 3 , wherein the point cloud layer is used to clarify the traction sheave point cloud data obtained by the laser, the parameter layer corrects the encoder pulse angle, and the diagnosis layer estimates the maximum wear depth point of the traction sheave, the eccentric wear index, and the remaining life by jointly predicting the data of the point cloud layer and the parameter layer.

[0057] Step 21: Structure and working principle of the point cloud layer.

[0058] Firstly, the point cloud data collected by the laser receiving device is processed through the point cloud layer, which mainly consists of K-Nearest Neighbor Graph (KNN-Graph) and Edge Convolution Network (EdgeConv). K-Nearest Neighbor Graph searches for the 20 nearest neighbors of each laser point to form an undirected edge and generate edge features As shown in the following formula:

[0059]

[0060] wherein, is the index of the laser point, is the feature vector of the point, represents the distance between the feature vectors of two feature points.

[0061] Edge Convolution Network further processes the generated edge features to complete the functions of abnormal point recognition, noise removal, and local geometric feature extraction, converting the originally unordered point cloud data into a graph that can perceive local geometric structures. Finally, it outputs a pixel-level denoising mask for the laser point cloud image frame, effectively shielding the influence of interference factors such as oil stains and burrs on the measurement results.

[0062] Step 22: Structure and working principle of parameter layer.

[0063] The parameter layer is used to correct the angle error of the encoder pulse. Due to the vibration caused by the operation of the traction wheel, the encoder may vibrate, which may cause errors in the angle encoding pulse. Therefore, the parameter layer uses two layers of Long Short-Term Memory Network (LSTM) to process the time interval sequence of adjacent pulses of the encoder. The hidden layer dimension of each layer of Long Short-Term Memory Network is 64, and a Dropout layer is set after the two layers of Long Short-Term Memory Network, with a dropout rate of 0.2 to avoid network overfitting. Through the learning and analysis of the two layers of Long Short-Term Memory Network on the adjacent pulse time interval sequence, the next pulse interval is predicted, and the corresponding correction angle and confidence are output. If there is a difference between the predicted correction angle and the original angle encoding pulse, it is processed through the subsequent Kalman filter. The Kalman filter dynamically adjusts the weight proportion of the correction angle and the original angle encoding pulse according to the confidence, realizes the optimal weighted fusion of the Long Short-Term Memory Network prediction result and the original pulse data, and finally obtains accurate angle data, effectively suppressing the arc length error caused by the encoder vibration.

[0064] Step 23: Structure and working principle of diagnosis layer.

[0065] ​​The diagnostic layer receives the pixel-level denoising mask output by the point cloud layer and the accurate angle data output by the parameter layer. First, the two types of data are input into a five-layer convolution module for basic feature extraction. In the five-layer convolution module, each layer of convolution adopts a configuration of a 3x3 convolution kernel, a step of 1, and padding of 1. Batch normalization (BatchNorm) and ReLU activation operations are sequentially performed after each convolution operation. A 2x2 max pooling layer is set between adjacent two convolution modules. Through this structure, the features related to wear such as texture, steps, and grooves can be fully extracted. The output features of the five-layer convolution module are supplied to three detection heads. The three detection heads share the above convolution features, but use different loss weighting methods. The first detection head outputs the maximum wear depth of the traction sheave based on the extracted basic features. The second detection head outputs the eccentric wear index based on the same basic features. The third detection head outputs the remaining life estimate value by concatenating the maximum wear depth output by the first detection head, the eccentric wear index output by the second detection head, the previously solved traction sheave pitch radius R, and the elevator rated parameters (specifically including the hardness of the elevator traction sheave, the rated running speed of the elevator, and the rated load of the elevator) with the basic features output by the five-layer convolution module, and then processing them through an internal network.

[0066] The entire detection process is shown in Figure 4 From the start of elevator operation and encoder pulse emission, the data is collected through laser triggering, and then processed through coordinate conversion, pitch radius solving, and AI three-stage point cloud processing to finally output the maximum wear depth of the traction sheave, the eccentric wear index, and the remaining life estimate value. This realizes comprehensive and accurate detection of the wear state of the traction sheave, provides reliable data support for elevator predictive maintenance, and effectively avoids excessive maintenance and sudden downtime. To further verify the effectiveness and accuracy of the method and device of the present application, 150 in-use elevators were selected for trial, and the detection results were compared with those of traditional manual detection methods and pure laser detection methods. The comparison results are shown in Table 1.

[0067] Table 1 Comparison of detection effects of different methods on traction sheave surface

[0068]

[0069] The above description is only the preferred embodiments of the present application. It should be noted that those skilled in the art can make several improvements without departing from the principles of the present application, and these improvements should also be considered within the scope of protection of the present application.

Claims

1. A laser measurement method for elevator traction sheave wear based on artificial intelligence, characterized in that: Includes the following steps: A radial fan-shaped laser surface is emitted towards the bottom of the traction wheel groove by a laser fixed to the side wall of the traction machine. The rotation angle information of the traction wheel is synchronously acquired by an encoder installed on the end of the main shaft of the traction machine. The point cloud data in the pixel coordinate system of the traction wheel is collected by the laser receiving device after one rotation of the traction wheel. Based on the laser triangulation constraint relationship, the point cloud data is converted from the pixel coordinate system to the laser internal coordinate system. The rotation angle information is used as the polar angle to convert the point cloud data into polar coordinate system representation, and then the three-dimensional world coordinate system coordinates are obtained. Solve for the pitch circle radius of the traction sheave based on the three-dimensional world coordinate system. A K-nearest neighbor graph is constructed for each laser point through a point cloud layer to search for the nearest neighbor point, forming undirected edges and generating edge features. The edge features are then processed by a graph edge convolutional network to identify anomalies, remove noise, and extract local geometric features, outputting a pixel-level denoising mask. The parameter layer uses a long short-term memory network to process the sequence of adjacent pulse time intervals of the encoder, predicts the next pulse interval, and outputs the corrected angle and confidence level. The Kalman filter performs optimal weighted fusion of the corrected angle and the original angle encoded pulse based on the confidence level to obtain accurate angle data. The pixel-level denoising mask and precise angle data are input into the five-layer convolutional module of the diagnostic layer to extract basic features. The output features of the five-layer convolutional module are respectively supplied to three detection heads. The first detection head outputs the maximum wear depth of the traction sheave, the second detection head outputs the wear index, and the third detection head splices the outputs of the first and second detection heads, the pitch circle radius of the traction sheave, and the elevator rated parameters with the basic features to output the estimated remaining life value.

2. The laser measurement method for elevator traction sheave wear based on artificial intelligence as described in claim 1, characterized in that: The radially fan-shaped laser surface emitted by the laser passes through the center of the traction wheel's main shaft and is perpendicular to the wheel groove axis. , The pixel equivalent for the orientation is obtained from the factory calibration.

3. The laser measurement method for elevator traction sheave wear based on artificial intelligence as described in claim 1, characterized in that: The encoder outputs 1024 angular encoded pulses as the traction wheel rotates one revolution, and the rotation angle of the traction wheel corresponding to an adjacent laser pulse is 360° / 1024.

4. The laser measurement method for elevator traction sheave wear based on artificial intelligence as described in claim 1, characterized in that: In the point cloud layer, the K-nearest neighbor graph searches for 20 nearest neighbor points for each laser point to form an undirected edge. The edge feature is the distance between the feature vectors of two laser points, and the feature vector contains the horizontal coordinate and depth coordinate of the laser point in the laser's internal coordinate system.

5. The laser measurement method for elevator traction sheave wear based on artificial intelligence as described in claim 1, characterized in that: The parameter layer's long short-term memory network has a two-layer structure, with each long short-term memory network having a hidden layer dimension of 64. Dropout layers are set after both long short-term memory networks in the parameter layer.

6. The laser measurement method for elevator traction sheave wear based on artificial intelligence as described in claim 1, characterized in that: The optimal weighted fusion of the Kalman filter specifically involves dynamically adjusting the weight ratio of the correction angle and the original angle encoded pulse based on the confidence level, thereby achieving the fusion of the LSTM prediction result and the original pulse data.

7. The laser measurement method for elevator traction sheave wear based on artificial intelligence as described in claim 1, characterized in that: In the five-layer convolutional module of the diagnostic layer, each convolutional layer adopts a configuration of 3×3 convolutional kernel, stride 1, and padding 1. After each convolutional layer, batch normalization and ReLU activation operations are performed in sequence. A 2×2 max pooling layer is set between two adjacent convolutional modules. The three detectors share the convolutional features output by the five-layer convolutional module, and the three detectors use differentiated loss weighting methods.

8. The laser measurement method for elevator traction sheave wear based on artificial intelligence as described in claim 1, characterized in that: In the process of solving the pitch circle radius of the traction wheel, the least squares objective function is constructed by calculating the residual from the point set on the wheel groove to the center of the wheel axle, and the Gauss-Newton iteration method is used to solve the objective function. The Gauss-Newton iteration method involves the element index and mean calculation function of the Jacobian vector.

9. The laser measurement method for elevator traction sheave wear based on artificial intelligence as described in claim 1, characterized in that: The radius of the polar coordinate system is the depth coordinate of the laser point in the internal coordinate system of the laser, and the polar angle is the rotation angle output by the encoder; when converting from polar coordinates to the three-dimensional world coordinate system, the origin of the world coordinate system is the rotation center of the traction wheel spindle.

10. The laser measurement method for elevator traction sheave wear based on artificial intelligence as described in claim 1, characterized in that: The elevator's rated parameters specifically include the hardness of the elevator traction sheave, the elevator's rated operating speed, and the elevator's rated load.

Images