Weight and volume combined belt conveyor mixing metering device

Abstract

The invention belongs to the field of weighing of belt conveyors, and provides a mixed metering device of a belt conveyor, which combines weight and volume. The device is composed of a weight sensing system, a volume sensing system and a data comprehensive processing system, the weight sensing system obtains the instantaneous weight # imgabs0 # of a belt conveyor measuring area, the volume sensing system obtains the instantaneous volume # imgabs1 # of the belt conveyor measuring area, the instantaneous weight # imgabs3 # of materials is obtained by multiplying the instantaneous volume # imgabs1 # by the material density # imgabs2 #, and the instantaneous weight # imgabs0 # and the instantaneous volume # imgabs1 # of the belt conveyor measuring area are obtained. The data comprehensive processing system is responsible for data operation of the weight sensing system and the volume sensing system, data fusion between the weight sensing system and the volume sensing system is completed at the same time, # imgabs4 and # imgabs5 are fused to obtain the final instantaneous material weight # imgabs6, a volume metering module can be used for achieving compensation of a weight metering result, and the volume metering module can be used for measuring the instantaneous material weight # imgabs6. When the weight metering unit is verified and maintained, weight missing data can be supplemented.

Description

Technical Field

[0001] The present invention belongs to the field of belt conveyor weighing, and particularly relates to a belt conveyor hybrid metering device that combines weight and volume. Background Art

[0002] Belt conveyors are widely used in industries such as electronics, machinery, coal, logistics, chemical industry, and food for object assembly, inspection, debugging, packaging, and transportation. Reliable online metering technology can effectively improve the production management efficiency of enterprises and assist in realizing automatic control in each production link. In the process of industrial modernization and intelligentization, with the improvement of enterprise production capacity and belt speed, 90% of the traditional contact belt metering devices use weighing sensors to obtain analog signals, which are not conducive to long-distance transmission, are easily affected by electromagnetic and other factors, and it is difficult to meet the requirements of some enterprises for high metering accuracy when transporting and processing objects. In addition, the contact metering device is affected by various factors such as belt tension, stiffness, self-weight, and deviation, and requires delicate maintenance and frequent calibration during use. During equipment maintenance and calibration, different degrees of metering data loss will occur. In order to overcome problems such as accuracy decline and metering data loss during the operation of the contact metering device, it is necessary to introduce non-contact metering methods for assistance.

[0003] As a typical non-contact metering method, fringe projection profilometry (FPP) has received extensive attention due to its advantages such as low cost, high speed, and simple structure, and has been widely used in fields such as archaeology, medicine, entertainment, and industry. Generally, an FPP system consists of a camera, a projector, and a computer. During the measurement process, the projector actively projects a pre-set fringe pattern onto the surface of the object to be measured, and at the same time, the camera collects the fringe pattern deformed by the height modulation of the object to be measured. These collected deformed fringe patterns contain the depth information of the object to be measured. Using a computer to analyze the collected fringe patterns, the three-dimensional information of the object can be reconstructed, and then the volume and weight of the object to be measured can be calculated according to the actual measurement environment.

[0004] Currently, the improvement of FPP mainly focuses on two aspects: improving the hardware equipment and improving the phase unwrapping algorithm. In addition to replacing high-performance equipment for improving the hardware equipment, it can also be achieved by introducing multiple cameras to build a multi-view vision system.

[0005] Phase unwrapping algorithms are mainly divided into spatial phase unwrapping and temporal phase unwrapping. Spatial phase unwrapping algorithms do not require additional projection maps to calculate the fringe order, but when performing phase unwrapping, they need to analyze the relationship between adjacent points of the calculated wrapped phase, which is affected by surrounding points and is prone to errors. Temporal phase unwrapping algorithms calculate the fringe order point by point, and the calculation of the current point is not affected by surrounding points, effectively solving the problems encountered by spatial phase unwrapping methods. Common temporal phase unwrapping methods include: multi-frequency method, gray coding method, and phase coding method. However, these existing phase unwrapping methods all require additional projection fringe maps to calculate the phase order, which requires adding additional projection equipment and is not conducive to industrial field applications. Summary of the Invention

[0006] In view of the problems existing in the prior art, the present invention proposes a belt conveyor hybrid metering device combining weight and volume. In order to fully combine the advantages of contact metering and non-contact metering, this invention patent designs a hybrid metering device applicable to different scenarios from two perspectives of weight sensing and volume sensing.

[0007] To achieve the above object, the present invention is realized through the following technical solutions: A belt conveyor hybrid metering device combining weight and volume, comprising a weight sensing system, a volume sensing system, and a data integration system;

[0008] The weight sensing system includes belt scale frames fixedly arranged on both sides of the conveyor idler support. Weight sensors are fixedly arranged on the belt scale frames. The two belt scale frames are connected by a support frame, and a speed sensor is fixedly arranged on the support frame. The weight sensors and the speed sensor transmit the acquired data to an integrator, and the integrator transmits the weight calculated by the weight sensing system to the data integration system;

[0009] The volume sensing system includes a projection acquisition module and a calibration module. The projection acquisition module includes a movable support. An industrial projector and an industrial high-speed camera CCD are arranged on the top of the movable support. The industrial projector projects a coded fringe pattern onto the conveyor belt, and the industrial high-speed camera CCD captures the deformed fringe pattern and transmits it to the data integration system;

[0010] The data integration system sends the received deformed fringe pattern into a fringe-depth mapping network F-Dnet combining spatial-frequency domain features. The F-Dnet calculates and obtains the instantaneous volume of the measurement area of the belt conveyor based on the deformed fringe pattern , multiplies it by the material density to obtain the instantaneous weight of the material . The data integration and processing system is responsible for the data operation of the weight sensing system and the volume sensing system, and at the same time completes the data fusion between the weight sensing system and the volume sensing system, and Fuse to obtain the final instantaneous weight of the material , which is achieved through the following formula:

[0011]

[0012] wherein, It is calibrated according to the system accuracy requirements.

[0013] Furthermore, both the industrial projector and the industrial high-speed camera are fixed on the top of the movable bracket through an adjustment mechanism. The adjustment mechanism includes an adjustment frame fixed on the top of the movable bracket. A screw rod driven by a stepping motor is installed on the adjustment frame. A slider is sleeved on the screw rod. A camera fixing frame is fixedly arranged on the slider. An angle adjuster is arranged on the top of the camera fixing frame. One end of the angle adjuster is hinged to the top of the camera fixing frame, and the other end is hinged to the industrial projector or the industrial high-speed camera.

[0014] Furthermore, the calibration module includes a checkerboard and four legs. Each of the four legs is provided with a screw rod slider mechanism driven by a stepping motor. The rectangular plate is placed on the slider and is driven by the slider to move up and down.

[0015] Furthermore, the stripe-depth mapping network combining spatial and frequency domain features includes three modules: feature extraction, feature mapping, and feature reshaping. The feature extraction module adopts a three-branch structure. The first branch directly introduces stripe texture information. The second branch increases the feature richness by adding details on the basis of the stripe spatial domain information. The third branch adopts a learnable filter convolution kernel to perform cyclic convolution with the original image to remove redundancy and highlight the phase feature on the basis of the stripe frequency domain information. The feature mapping module realizes the conversion from stripe features to depth features and adopts an encoder-decoder structure. The encoding end gradually downsamples the features to realize the degradation of stripe information. The decoding end gradually increases the feature scale and enriches the depth information layer by layer in a feature supplement way. Feature reshaping mainly realizes the reconstruction of depth features into depth data.

[0016] Furthermore, the specific structure of the feature extraction module is as follows:

[0017] The first branch directly introduces stripe texture information;

[0018] The second branch adopts a convolutional residual block CRB. The CRB learns the detailed features of the stripe pattern through two layers of convolution, then superimposes with the original stripe pattern to realize detail enhancement, and finally activates to introduce non-linear changes. Its data processing process can be represented by Equation (1):

[0019] (1)

[0020] wherein, represents The convolution operation, represents the batch normalization operation, is the Relu activation function, represents the residual operation, is the input image The output after passing through the CRB;

[0021] The third branch adopts a learnable frequency-domain auxiliary block LFAB. First, the input fringe pattern is Fourier-transformed to obtain the spectral information of the fringe pattern. Guided by the position of the fundamental frequency component in the spectral diagram and the distance from the zero frequency, a spatially learnable filtering convolution kernel is designed to perform circular convolution with the original image. Among them, after the Fourier transform of the learnable filtering convolution kernel, the frequency-domain characteristics satisfy the Gaussian distribution and can be represented by Equation (2):

[0022] (2)

[0023] Among them, represents the frequency-domain coordinate system, is the center position of the filter, is a learnable parameter used to control the filter window size.

[0024] Furthermore, the encoding end of the encoder-decoder structure of the feature mapping module uses two local attention modules LAM and two downsampling convolution blocks DCB for downsampling, and four upsampling convolution blocks UCB are used for upsampling at the decoding end of the encoder-decoder structure;

[0025] The data processing process of DCB is expressed as Equation (3):

[0026] (3)

[0027] Among them, represents the convolution operation with a convolution kernel size of ³, and the superscript represents the number of the convolution block, represents the max pooling operation with a pooling kernel of ², represents instance normalization, represents element-wise addition, represents the output of the th DCB, output of the

[0028] UCB extracts features through two convolutional layers combined with residual connections, and then restores the feature resolution through transposed convolution operations. The data processing process of UCB is expressed as Equation (4):

[0029] (4)

[0030] Among them, represents the output of the -th UCB, represents the output of the -th UCB, represents a transposed convolution operation for restoring the feature resolution.

[0031] Furthermore, the feature reshaping block FRB first uses convolution to extract high-level depth data, and then reduces the number of channels through convolution, and outputs a depth data matrix equal to the stripe size. The data processing process of FRB is expressed as Equation (5):

[0032] (5)

[0033] Among them, is the output of the feature mapping stage, is convolution of, is the depth of the upper surface of the bulk material relative to the zero reference plane.

[0034] Compared with the prior art, the beneficial effects of the present invention are:

[0035] 1. A flexible mechanical structure more suitable for stripe projection 3D reconstruction is designed. This structure allows the projection angle of the projector and the acquisition angle of the camera to be freely adjusted. At the same time, their heights can also be adjusted according to requirements, and the projection acquisition module can move as a whole. The calibration module is located outside the measurement area, and it only needs to ensure that the calibration module, the projection acquisition module, and the belt conveyor are on the same horizontal plane. During calibration, the projection acquisition module is moved above the calibration module. The rectangular plate of the calibration module can move up and down flexibly to ensure that the calibration space covers both the area above the belt and the area below the belt, improving the accuracy of subsequent measurements.

[0036] 2. Instead of directly calculating the height of the upper surface of the material relative to the working surface of the belt, the system introduces a zero reference plane and calculates the distance of the upper surface of the material relative to the zero reference plane and the distance of the working surface of the belt relative to the zero reference plane respectively. This method can ensure the accuracy of the zero reference plane and improve the measurement accuracy.

[0037] 3. When calculating the height of the upper surface of the material relative to the zero reference plane, an end-to-end deep learning network, namely F-Dnet, is adopted. Different from traditional networks, this network does not have a simple encoding-decoding structure. Instead, it consists of three steps: feature extraction, feature mapping, and feature reshaping. First, stripe features are extracted, then the stripe features are mapped into depth features, and finally, the depth data is reshaped based on the depth features. Feature extraction fully considers information richness and comprehensively extracts features by combining the stripe itself, the stripe after detail enhancement, and the stripe with redundant information filtered out. Feature mapping adopts an encoding-decoding structure. The encoding structure gradually degrades the stripe features through downsampling, and the decoding structure gradually supplements the depth features through upsampling. Feature reshaping uses 2-layer convolution to extract features, and finally, 1*1 convolution is used to complete channel dimensionality reduction. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 FIG. is a schematic diagram of the hardware architecture of the present invention.

[0039] Figure 2 FIG. is a schematic diagram of the structure of the weight sensing system.

[0040] Figure 3 FIG. is a schematic diagram of the structure of the projection acquisition module.

[0041] Figure 4 FIG. is a schematic diagram of the structure of the calibration module.

[0042] Figure 5 FIG. is the principle of obtaining the material depth.

[0043] Figure 6 FIG. is a framework diagram of the stripe-depth mapping network combining spatial-frequency domain features.

[0044] Figure 7 FIG. is a schematic diagram of the movement of the rectangular plate.

[0045] Figure 8 FIG. is a schematic diagram of obtaining the weight at different stages.

[0046] Figure 9 FIG. is a comparison diagram of different methods.

[0047] In the figures, 1 is the weight sensing system, 11 is the belt scale frame, 12 is the weight sensor, 13 is the support frame, 14 is the speed measurement sensor, 2 is the projection acquisition module, 21 is the movable support, 22 is the adjustment frame, 23 is the screw, 24 is the slider, 25 is the camera fixing frame, 26 is the angle adjuster, 3 is the calibration module, 31 is the leg, 32 is the screw-slider mechanism, and 33 is the rectangular block. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0048] The embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings and embodiments, so as to fully understand how the present invention uses technical means to solve technical problems and achieve the implementation process of technical effects and implement accordingly.

[0049] As Figure 1 shown, a belt conveyor hybrid metering device combining weight and volume includes a weight sensing system, a volume sensing system, and a data integration system.

[0050] As Figure 2 shown, the weight sensing system includes a belt scale frame 11 fixedly arranged on both sides of the conveyor idler frame. The belt scale frame 11 adopts the ICS20 type. Four weight sensors 12 are fixedly arranged on the belt scale frame 11. The weight sensors 12 adopt the SQB model. The two belt scale frames 11 are connected by a support frame 13. A speed measurement sensor 14 is fixedly arranged on the support frame 13. The speed measurement sensor 14 adopts the T12C model. The weight sensors 12 and the speed measurement sensor 14 transmit the acquired data to an integrator, and the integrator transmits the weight calculated by the weight sensing system to the data integration system, and the measurement result is the average value of the measurement values of the 4 sensors.

[0051] The volume sensing system includes a calibration module 3 and a projection acquisition module 2. The calibration module 3 is used to assist the volume sensing system to realize the calibration of system parameters. The required calibration system parameters include the phase-height mapping matrix , , , the internal parameter matrix of the camera, including the focal length , the principal point and the distortion coefficient, and the external parameter matrix, including the rotation matrix and the translation vector . The phase-height mapping matrix is used to realize the mapping of the phase of the deformed fringe pattern to the depth information of the bulk material. The internal and external parameters of the camera are used to realize the mapping from the image coordinate system of the camera image plane to the world coordinate system, and further realize the calculation of the actual area of each pixel unit in the image.

[0052] As Figure 3 shown, the projection acquisition module 2 includes a movable support 21. An industrial projector and an industrial high-speed camera CCD are arranged on the top of the movable support 21. The industrial projector projects a coded fringe pattern onto the conveyor belt, and the industrial high-speed camera CCD captures the deformed fringe pattern and transmits it to the data integration system.

[0053] Both the industrial projector and the industrial high-speed camera are fixed on the top of the movable support 21 through an adjustment mechanism. The adjustment mechanism includes an adjustment frame 22 fixed on the top of the movable support 21. A screw rod 23 driven by a stepping motor is installed on the adjustment frame 22. A slider 24 is sleeved on the screw rod 23. A camera fixing frame 25 is fixedly arranged on the slider 24. An angle adjuster 26 is arranged on the top of the camera fixing frame 25. One end of the angle adjuster 26 is hinged to the top of the camera fixing frame 25, and the other end is hinged to the industrial projector or the industrial high-speed camera. The stepping motor is used to control the up and down movement of the industrial camera and the projector, and the two angle adjusters 26 above the camera and the projector are used to control the shooting angle of the camera and the projection angle of the projector.

[0054] As Figure 4 shown, the calibration module 3 includes a checkerboard and four legs 31. Screw-slider mechanisms 32 driven by stepping motors are arranged on all the four legs 31. The rectangular plate is placed on the slider 24 and is driven by the slider 24 to move up and down.

[0055] The data integration system is a graphics workstation including an Intel Core i7-9700 CPU and an NVIDIA GeForce RTX 2080 Ti GPU graphics card, and is used for data operation of the belt conveyor bulk material mixing metering device combining weight and volume. The data integration system sends the received deformed fringe pattern into the fringe-depth mapping network F-Dnet combining spatial-frequency domain features. The F-Dnet calculates and obtains the instantaneous volume of the measurement area of the belt conveyor according to the deformed fringe pattern , and multiplies it by the material density to obtain the instantaneous weight of the material . The data comprehensive processing system is responsible for data operation of the weight sensing system and the volume sensing system, and simultaneously completes data fusion between the weight sensing system and the volume sensing system. And fuses them to obtain the final instantaneous weight of the material , which is realized through the following formula:

[0056]

[0057] where is calibrated according to the system accuracy requirements.

[0058] The overall measurement structure and principle of the mixing metering device of the present invention are briefly introduced above. The principle of the weight sensing system of the present invention is basically the same as that of the existing belt scale. The principle of the volume sensing system of the present invention will be introduced in detail below.

[0059] During volume acquisition, based on the principle of fringe projection profilometry, the projection acquisition module 2 projects a sinusoidal fringe pattern and captures the deformed fringe pattern modulated by the depth of the bulk material. The instantaneous stacking depth of the material in the measurement area relative to the belt is obtained through phase analysis. , and Integrate the actual area of each pixel unit to obtain the volume of the material. . The measurement principle diagram of Figure 5 is shown as follows.

[0060] According to the zero reference plane defined in the phase-height mapping process, is divided into and . is the distance from the working surface of the belt to the zero reference plane, and this distance is negative; is the distance from the upper surface of the bulk material to the zero reference plane, and this distance is positive. Therefore, .

[0061] Since the distance of the belt relative to the zero reference plane is a fixed value, is only measured once. The specific measurement method is as follows: Use the fringe pattern of the zero reference plane taken during the calibration of the phase-height mapping system parameters to calculate the phase distribution of the zero reference plane fringes . Without adding materials on the belt, take the deformed fringe pattern of the belt surface during idling and calculate the phase of the belt surface. After subtracting it from the phase of the zero reference plane, the phase difference between the belt and the zero reference plane is obtained, and after phase-height mapping, is obtained.

[0062] Since the bulk material on the belt changes in real time with time, it is necessary to give the instantaneous measurement value in real time. When calculating and using the traditional phase-shift algorithm, it is necessary to project phase-shifted fringe patterns at high speed and use the full-cycle equal-phase-shift algorithm for calculation.

[0063] In order to reduce the number of projected fringe frames and improve the measurement speed, the present invention designs a fringe-depth mapping network combining spatial-frequency domain features (Depth prediction Network considering frequency feature, F-Dnet), which can directly learn the instantaneous depth information of the bulk material relative to the zero reference plane from a deformed fringe pattern modulated by the surface contour of the material. . The structure of F-Dnet is shown in Figure 6As shown in (a), it mainly includes three parts: feature extraction, feature mapping, and feature reshaping. Feature extraction adopts a three-branch structure: the first branch directly introduces stripe texture information, the second branch increases the feature richness by adding details based on the stripe spatial domain information, and the third branch adopts a filter to remove redundancy and highlight the phase feature based on the stripe frequency domain information. Feature mapping mainly realizes the conversion from stripe features to depth features, adopting an encoder-decoder structure. The encoding end gradually downsamples the features to realize the degradation of stripe information, and the decoding end gradually increases the feature scale and enriches the depth information layer by layer in a feature supplement way. Feature reshaping mainly realizes the reconstruction from depth features to depth data. The specific structures of the three parts of the network are as follows:

[0064] During the feature extraction process, considering that there is redundant information such as background light and noise in the deformed stripe pattern, these information will significantly increase the complexity of network learning. In order to effectively remove these redundant components and enhance the effective phase features, a frequency domain filtering technology is introduced in the third branch of feature extraction, and a learnable frequency domain assistant block (LFAB) is designed. By performing the inverse Fourier transform on the window learnable filter to obtain the spatial domain system function and circularly convolving it with the original stripe pattern, the background light is filtered out to achieve the purpose of enhancing the phase information.

[0065] Specifically, feature extraction adopts a three-branch structure: the first branch directly introduces stripe texture information. The second branch starts from spatial domain feature extraction and designs a convolution residual block (CRB) in a way of detail supplement, as Figure 6 shown in (b). The CRB learns the detail features of the stripe pattern through two layers of convolutions, then superimposes with the original stripe pattern to realize detail enhancement, and finally activates to introduce non-linear changes to increase the feature expression ability. Its data processing process can be represented by Equation (1):

[0066] (1)

[0067] where represents the convolution operation of , represents the batch normalization operation, is the Relu activation function, represents the residual operation. is the output of the input image passing through the CRB, represents the input stripe pattern.

[0068] The LFAB structure designed for the third branch is as Figure 6As shown in (c), first, the input fringe pattern is Fourier-transformed to obtain the spectral information of the fringe pattern. Guided by the position of the fundamental frequency component in the spectrum, the redundant information such as the background light of the fringe pattern is removed by circularly convolving the original image with a window-learnable filter, thereby strengthening the phase information of the image. Among them, the window-learnable filter can be represented by Equation (2):

[0069] (2)

[0070] Among them, represents the pixel coordinates at the position with the largest amplitude of the zero-frequency component in the frequency-domain coordinate system, is the coordinate of the pixel with the largest amplitude of the fundamental frequency component, is a learnable parameter used to control the size of the filter window.

[0071] In the design of the feature map, the core idea is to map the extracted fringe features to deep features through an encoder-decoder structure. When downsampling at the encoding end, on the one hand, considering that the shallow features have a large scale and rich information, a local attention module (Local Attention Block, LAM) is adopted, as shown in Figure 6 (d). LAM extracts fringe features more comprehensively and highlights important regions by integrating global and local information, providing rich details for feature mapping; on the other hand, aiming at the problem that traditional convolution is prone to gradient disappearance in deep networks, a Figure 6 (e)-shown downsampling convolution block (DownsamplingConvolution Block, DCB) is designed to ensure the stable transmission of information in the deep network by using residual connections during convolution. When upsampling at the decoding end, an Figure 6 (f)-shown upsampling convolution block (Unsampling ConvolutionBlock, UCB) is designed to gradually restore the multi-scale feature sizes obtained at the encoding end and combine residual connections to alleviate the problem of gradient disappearance caused by the increase in network depth.

[0072] Specifically, the structure of each part of the feature map is as follows: The LAM module adopts a double-branch structure to jointly extract the detailed features of the key regions by combining local attention and global attention mechanisms, improving the model's ability to analyze complex texture features. In order to effectively extract detailed features in the early stage and balance the model performance and the number of parameters, two LAM modules are selected to be deployed in the shallow layer of the encoding end. The DCB adopts a double-layer convolutional residual structure, by Residual connections are used between the convolutional blocks, effectively alleviating the problem of information degradation during feature transmission. Considering that instance normalization (IN) processes each individual sample independently, it can retain more details and avoid interference between different samples. Therefore, DCB uses IN to replace the traditional batch normalization (BN). The data processing process of DCB is expressed as Equation (3):

[0073] (3)

[0074] where represents a convolutional operation with a kernel size of 3, and the superscript indicates the number of the convolutional block. represents a max pooling operation with a pooling kernel of 2, represents instance normalization,[[]] represents element-wise addition. represents the output of the th DCB, represents the output of the th DCB.

[0075] UCB extracts features through two convolutional layers combined with residual connections, and then restores the feature resolution through a transposed convolution operation. The data processing process of UCB is expressed as Equation (4):

[0076] (4)

[0077] where represents the output of the th UCB, represents the output of the th UCB. represents a transposed convolution operation used to restore the feature resolution.

[0078] The core idea of feature reshaping is to reshape deep features into deep data. This network constructs a Feature Reshaping Block (FRB). This module first uses convolution to extract high-level deep data, and then uses convolution to reduce the number of channels, outputting a depth matrix equal to the stripe size. The data processing process of FRB is expressed as Equation (5):

[0079] (5)

[0080] where is the output of the feature mapping stage, is convolution, is the depth of the bulk material relative to the zero reference plane.

[0081] The above has introduced in detail the belt conveyor hybrid metering device that combines weight and volume. On the basis of the above structure, the belt conveyor hybrid metering method that combines weight and volume is introduced in detail, which is carried out according to the following steps:

[0082] Step S1, system calibration: During the system calibration process, first place the calibration module 3 and the projection acquisition module 2 in the actual measurement environment, and make the working surface of the belt parallel to the rectangular plate of the calibration module 3. During the system calibration process, the calibration module 3 is placed directly below the projection acquisition module 2, and the height of the rectangular plate is adjusted by the stepping motor so that its height is about 2 cm higher than the working surface of the belt conveyor. After adjusting to the horizontal, this surface is defined as the zero reference surface for the entire measurement process. The projection acquisition module 2 projects standard phase-shifted sine fringes, acquires the phase-shifted fringes, and obtains the fringe phase distribution corresponding to the zero reference surface through the operation of formula (6):

[0083] (6)

[0084] where, is the wrapped phase of the zero reference surface, is the th frame fringe pattern of the zero reference surface, is the th frame phase shift amount of the zero reference surface fringe pattern;

[0085] After measuring the phase of the zero reference surface, the stepping motor drives the rectangular plate to move up and down a specified distance. As shown in Figure 7 , it shows the process of the rectangular plate moving. Each time it moves, the phase change caused by the rectangular plate relative to the zero reference surface at the specified height is . Substitute it into the phase-height mapping equation, and obtain the phase-height mapping matrix parameters , , by the least squares method. The phase-height mapping equation is as shown in formula (7):

[0086] (7)

[0087] where, , , are the parameters of the phase-height mapping, is the phase difference between the zero reference surface and the rectangular plate, is the known moving distance of the rectangular plate, is the number of times the rectangular plate moves;

[0088] After obtaining the system parameters of the phase height mapping, continue to calibrate the camera to obtain the internal and external parameters of the camera for subsequent calculation of the volume of the material. The calibration of the internal and external parameters of the camera is assisted by a checkerboard, and the Zhang Zhengyou camera calibration toolbox is used for the calibration algorithm. The specific process is as follows: First, place the checkerboard closely against the reference surface, and then the camera of the projection acquisition module 2 collects the image of the checkerboard. After that, remove the calibration module 3, and continuously change the placement position and angle of the checkerboard manually, and take 15 - 20 images of the checkerboard and send them to the data comprehensive processing system for calculation to obtain the internal parameter matrix of the camera, including the focal length , the principal point, and the distortion coefficient. Use the external parameters captured by placing the checkerboard image at the zero reference surface position as the final external parameters, including the rotation matrix and the translation vector .

[0089] Step S2. Measurement of the distance from the belt working surface to the zero reference surface

[0090] When measuring the distance from the belt working surface to the zero reference surface during the idle rotation of the belt, only the projection acquisition module 2 is required. First, the projector projects 5 phase-shifted sine fringes onto the belt working surface, and the high-speed industrial camera collects 5 frames of fringe images of the belt working surface. After calculation using formula (8), the phase distribution of the belt working surface is obtained:

[0091] (8)

[0092] where is the wrapped phase of the belt working surface, is the -th frame of the fringe image of the belt working surface, is the phase shift amount of the -th frame of the fringe image of the belt working surface.

[0093] The data comprehensive processing system unfolds the wrapped phase change of the belt working surface relative to the zero reference surface ( ) to obtain the continuous phase. Then substitute it into the phase height mapping equation of formula (9) to obtain the distance from the belt working surface to the zero reference surface.

[0094] (9)

[0095] Step S3. Measurement of the distance from the upper surface of the bulk material to the zero reference surface

[0096] When measuring the distance from the upper surface of the bulk material to the zero reference surfaceWhen a deformed fringe pattern of bulk materials captured by a high-speed industrial camera CCD is used as the input of the F-Dnet, it can directly predict ; However, before the F-Dnet realizes the material depth prediction, it needs to train the network and fix the network parameters;

[0097] The acquisition process of the training dataset is as follows: when the materials on the working surface of the belt enter the measurement area, the projector projects 5 phase-shifted sine fringes on the surface of the materials, and the high-speed industrial camera captures 5 frames of deformed fringe patterns. The corresponding phase distribution on the surface of the materials is calculated by formula (10):

[0098] (10)

[0099] where is the wrapped phase of the deformed fringe pattern of the materials, is the th frame of the deformed fringe pattern;

[0100] The data comprehensive processing system unfolds the wrapped phase change of the upper surface of the materials relative to the zero reference plane to obtain the continuous phase, and then substitutes it into the phase-height mapping equation to obtain the distance from the upper surface of the bulk materials to the zero reference plane ; ;

[0101] The F-Dnet is fully trained with the dataset obtained by the above traditional algorithm. After the training is completed, the F-Dnet only needs a deformed fringe pattern of the materials as the input to output high-precision material depth data;

[0102] Step S4, instantaneous volume acquisition

[0103] After using the F-Dnet to predict the material depth data in the belt measurement area , from the depth of the materials relative to the working surface of the belt is obtained , and then it is converted from the pixel coordinate system to the world coordinate system through the camera calibration parameters. The instantaneous volume of the bulk materials in the belt measurement area is obtained by the integral formula ;

[0104] Step S5, instantaneous weight acquisition

[0105] The weight sensing system obtains the instantaneous weight of the belt conveyor measurement area , and the volume sensing system obtains the instantaneous volume of the belt conveyor measurement area , which is multiplied by the material density to obtain the instantaneous weight of the materials , and then the data comprehensive processing system combines with The final instantaneous weight of the material is obtained through fusion and is achieved by formula (11):

[0106] (11)

[0107] where is calibrated according to the system accuracy requirements;

[0108] Step S6, Obtaining the stage weight

[0109] The stage weight of the bulk material is obtained from the sum of the instantaneous weights of the material within the measurement area, as Figure 8 shown. The width of the belt is 0.8 meters, and the length of the belt that the volume measurement system can capture is 1 meter. Therefore, at time, the instantaneous weight of the material at that time can be obtained from the deformed fringe pattern of the bulk material captured . Combining with the speed of the belt it is possible to obtain . Therefore, after passing time, the instantaneous weight of the material at time can be calculated based on the captured deformed fringe pattern . Finally, the instantaneous weights at all times are added together to obtain the stage weight of the material

[0110] (12)

[0111] where is the instantaneous weight of the material at time t, is the instantaneous weight of the material at time is the instantaneous weight of the material at time

[0112] To evaluate the prediction performance of F-Dnet for the distance between the upper surface of the bulk material and the zero reference plane, a comparative experiment was conducted with four methods: AEN, U-Net, hNet, and GAN. The material model was used as the analysis object to compare the prediction results of the five methods. As Figure 9 shown, the results indicate that F-Dnet performs best in the task of predicting the distance between the upper surface of the material and the reference plane, significantly outperforming other methods

[0113] To objectively evaluate the performance differences between F-Dnet and other methods, a quantitative analysis was conducted on 80 test data, comparing the reconstruction accuracy, network training time, and model complexity. The results are summarized in Table 1. The reconstruction accuracy was measured by the average root mean square error (RMSE) and the average absolute error (MAE), calculated based on the mean of 80 test data sets. The model complexity was evaluated by the number of parameters and the memory size. The optimal results in the table are marked in bold, and the sub-optimal results are marked in italics. The analysis shows that the proposed method is superior to the methods of AEN, U-Net, and hNet in terms of the RMSE and MAE metrics. Moreover, it improved by 6.7% compared to GAN in terms of RMSE, with higher prediction accuracy. In terms of training time, although F-Dnet is longer than AEN, U-Net, and hNet, it is 58% shorter than GAN. In summary, F-Dnet has significant advantages in terms of prediction accuracy, model complexity, and training efficiency.

[0114] Table 1. Averages of different methods on 80 test data sets Comparison of reconstruction accuracy and model complexity

[0115]

[0116] Verification of the effectiveness of each module of F-Dnet

[0117] The innovations of the present invention include designing a window-learnable frequency domain filter to enhance the phase information of the image and introducing LAM. To verify the effectiveness of these innovations, ablation experiments were conducted, and the results are shown in Table 2. In the experiment, Scheme 1 is the traditional U-Net model, serving as the basic method. Scheme 2 adds frequency domain processing to the basic method, and the results show that its RMSE and MAE are reduced by 0.2981 mm and 0.158 mm respectively compared to Scheme 1, proving the effectiveness of frequency domain processing. Scheme 3 further sets the filter window as a learnable parameter, and the RMSE and MAE are further reduced. Scheme 4 introduces LAM and the network performance is further improved. The results show that these innovations effectively improve the accuracy of F-Dnet in predicting depth information, and the number of model parameters does not increase significantly.

[0118] Table 2. Comparison of average reconstruction accuracy and number of parameters of different network schemes on 80 test data sets Comparison of reconstruction accuracy and number of parameters

[0119]

[0120] Comparison of instantaneous volume measurement results of different algorithms

[0121] After verifying the advantages of the F-Dnet, a comparative experiment on the instantaneous volume of bulk materials was further carried out. Table 3 compares the results of predicting the instantaneous volume and the standard volume of materials by five different methods. The results show that the material volume predicted by the F-Dnet is the closest to the standard volume.

[0122] Table 3. Prediction and error of the instantaneous volume of bulk materials at the same moment by different methods ( )

[0123]

[0124] Instantaneous weight based on the volume sensing system Measurement results

[0125] To verify the accuracy of the volume sensing system in predicting the weight of bulk materials at different times, a comparative experiment on volume-based weight measurement was carried out. The depth data of the materials from to (set to 0.5 s) were obtained in real time through the F-Dnet, and the instantaneous volume of the materials at each moment was calculated. By multiplying it by the material density, the instantaneous weight of the materials at different times was obtained. Table 4 shows that the proposed volume sensing system-based weight measurement method can accurately predict the material weight, approaching the standard value.

[0126] Table 4. Error between the predicted weight and the standard weight of the instantaneous weight of materials at different times (kg)

[0127]

[0128] Stage weight Measurement

[0129] To verify the effectiveness of the proposed belt conveyor bulk material hybrid metering device combining weight and volume, stage weight measurements were carried out. Specifically, the experiment was conducted by taking measurements every two days at different time periods, as shown in Table 5. are the measurement results of the weight sensor 12, are the measurement results of the volume sensing system, and the hybrid metering is the result calculated using the proposed formula (11). It can be seen from the data that as the system operates, the belt will deviate, resulting in a decrease in the measurement accuracy of the weight sensing system, while the proposed volume-based weight measurement method always maintains an ideal effect. Combining it with the weight sensing system for hybrid metering can effectively improve the measurement accuracy.

[0130] Table 5. Errors between weight measurement, volume-based weight measurement, and hybrid metering weight measurement and the standard weight at different time periods (t)

[0131]

[0132] The experiment shows that the measurement accuracy of the belt conveyor bulk material mixing metering device combining weight and volume of the present invention is higher. In the case of various factors such as electromagnetic interference, belt tension, stiffness, self-weight, and deviation, compensation for the weight measurement result can be achieved. When the weight metering unit is calibrated and maintained, supplementary filling for the missing weight data can also be realized.

[0133] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claimed rights.

Claims

1. Belt conveyor mixing metering device combining weight and volume, characterized in that: It includes a weight sensing system, a volume sensing system and a data integration system; A weight sensing system, comprising a belt scale frame fixedly arranged on both sides of a conveying idler bracket, a weight sensor fixedly arranged on the belt scale frame, a support frame connecting between the two belt scale frames, a speed measuring sensor fixedly arranged on the support frame, the weight sensor and the speed measuring sensor transmitting the acquired data to an integrator, and the integrator transmitting the weight calculated by the weight sensing system to a data integration system The volume perception system includes a projection acquisition module and a calibration module. The projection acquisition module includes a movable support. At the top of the movable support, there is an industrial projector and an industrial high-speed camera CCD. The industrial projector projects a coded fringe pattern onto the conveyor belt, and the industrial high-speed camera CCD captures the deformed fringe pattern and transmits it to the data integration system; The data integration system sends the received deformed fringe pattern into the fringe-depth mapping network F-Dnet that combines the spatial-frequency domain features. F-Dnet calculates and obtains the instantaneous volume of the measurement area of the belt conveyor based on the deformed fringe pattern. , which is multiplied by the material density to obtain the instantaneous weight of the material. The data integration and processing system is responsible for the data operation of the weight sensing system and the volume sensing system, and simultaneously completes the data fusion between the weight sensing system and the volume sensing system. It is fused with to obtain the final instantaneous weight of the material. This is achieved through the following formula: ； Among them, Obtained by calibration according to the system accuracy requirements.

2. The belt conveyor hybrid metering device combining weight and volume according to claim 1, characterized in that: Both the industrial projector and the industrial high-speed camera are fixed to the top of the movable support through an adjustment mechanism. The adjustment mechanism includes an adjustment frame fixed to the top of the movable support. A screw rod driven by a stepping motor is installed on the adjustment frame. A slider is sleeved on the screw rod. A camera fixing frame is fixedly arranged on the slider. An angle adjuster is arranged at the top of the camera fixing frame. One end of the angle adjuster is hinged to the top of the camera fixing frame, and the other end is hinged to the industrial projector or the industrial high-speed camera.

3. The belt conveyor hybrid metering device combining weight and volume according to claim 2, characterized in that: The calibration module includes a checkerboard and four legs. Each of the four legs is provided with a screw rod slider mechanism driven by a stepping motor. A rectangular plate is placed on the slider and is driven by the slider to move up and down.

4. The belt conveyor hybrid metering device combining weight and volume according to claim 3, characterized in that: The stripe-depth mapping network combining spatial-frequency domain features includes three modules: feature extraction, feature mapping, and feature reshaping. The feature extraction module adopts a three-branch structure. The first branch directly introduces stripe texture information. The second branch, based on the stripe spatial domain information, improves the feature richness by adding details. The third branch, based on the stripe frequency domain information, uses a learnable filtering convolution kernel to perform circular convolution with the original image to remove redundancy and highlight the phase feature; The feature mapping module realizes the conversion from stripe features to depth features and adopts an encoder-decoder structure. The encoding end gradually downsamples the features to realize the degradation of stripe information. The decoding end gradually increases the feature scale and enriches the depth information layer by layer in a feature supplement way. Feature reshaping mainly realizes the reconstruction from depth features to depth data.

5. The belt conveyor hybrid metering device combining weight and volume according to claim 4, characterized in that: The specific structure of the feature extraction module is as follows: The first branch directly introduces stripe texture information; The second branch adopts a convolutional residual block (CRB). The CRB learns the detailed features of the fringe pattern through two layers of convolution, then superimposes with the original fringe pattern to achieve detail enhancement, and finally activates to introduce non-linear changes. Its data processing process can be expressed by Equation (1): （1） Among them, represents the convolution operation of represents the batch normalization operation, is the Relu activation function, represents the residual operation, is the input image the output after passing through the CRB; The third branch adopts a learnable frequency domain auxiliary block LFAB. First, the input stripe pattern is subjected to Fourier transform to obtain the spectral information of the stripe pattern. Guided by the position of the fundamental frequency component in the spectral diagram and the distance from the zero frequency, a spatially learnable filtering convolution kernel is designed to perform circular convolution with the original image. Among them, after the learnable filtering convolution kernel is subjected to Fourier transform, its frequency domain characteristics satisfy the Gaussian distribution and can be represented by Equation (2): （2） Among them, represents the frequency domain coordinate system, is the center position of the filter, is a learnable parameter for controlling the filter window size.

6. The belt conveyor hybrid metering device combining weight and volume according to claim 5, characterized in that: In the encoding end of the encoder-decoder structure of the feature mapping module, downsampling adopts two local attention modules LAM and two downsampling convolution blocks DCB. When upsampling in the decoding end of the encoder-decoder structure, four upsampling convolution blocks UCB are adopted; The data processing process of DCB is expressed as Equation (3): （3） Among them, represents a convolution operation with a convolution kernel size of 3, and the superscript represents the number of the convolution block, represents a max pooling operation with a pooling kernel of 2, represents instance normalization, represents element-wise addition, represents the output of the th DCB, represents the output of the th DCB; UCB extracts features through two convolutional layers combined with residual connections, and then restores the feature resolution through transposed convolution operations. The data processing process of UCB is expressed as Equation (4): （4） Among them, represents the output of the th UCB, represents the output of the th UCB, represents a transposed convolution operation for restoring the feature resolution.

7. The belt conveyor hybrid metering device combining weight and volume according to claim 6, characterized in that: The feature reshaping block FRB first uses convolution to extract high-level depth data, and then passes through convolution to reduce the number of channels and output a depth data matrix with the same size as the stripe. The data processing process of FRB is expressed as Equation (5): （5） Among them, is the output of the feature mapping stage, is convolution of is the depth of the upper surface of the bulk material relative to the zero reference plane.

Images