A handheld industrial silicon ingot breaking device and method based on visual recognition and intelligent path optimization

Abstract

A handheld industrial silicon ingot breaking device and method based on visual recognition and intelligent path optimization belong to the technical field of industrial silicon block processing equipment. First, the scene image and depth data of each target silicon block in the work area are collected; second, the target silicon block size analysis and surface local curvature analysis are obtained by scanning; the surface equivalent hardness value analysis and crack determination of the target silicon block are carried out; third, the over-standard silicon block locking frame is displayed, the breaking path and candidate hitting point are generated, and the breaking prompt is carried out; fourth, the amplitude and frequency execution parameters are generated considering the crack surface energy and success probability, and the work is carried out according to the execution parameters; finally, when the target silicon block size in the visual area range meets the requirements, the breaking report is generated, and the power module enters the low power consumption mode. The handheld industrial silicon ingot breaking device is realized. The present application can significantly reduce the artificial judgment error, realize the semi-automatic identification and accurate breaking of over-standard silicon ingot, and improve the efficiency and stability of the breaking operation process.

Description

Technical Field

[0001] This invention belongs to the technical field of industrial silicon ingot processing equipment, and relates to a handheld industrial silicon ingot crushing device and method based on visual recognition and intelligent path optimization. Background Technology

[0002] In the production of industrial silicon, large industrial silicon ingots need to be precisely crushed into blocks with sides of 150-200mm to ensure the efficiency of subsequent processing and storage. This size is a commonly used crushing specification in the industry, which can better match the feeding requirements of subsequent processes such as melting and purification, and can effectively reduce damage from impacts during storage and transportation.

[0003] However, current mainstream crushing methods suffer from low automation and limited crushing efficiency. Most scenarios still rely on the traditional method of visual screening and manual crushing with a hydraulic breaker, depending entirely on the operator's experience to judge the size of the silicon blocks and their subjective control of the striking force. This approach is inefficient for large-scale production and prone to over-crushing or under-crushing, leading to raw material waste and increased costs. Existing automated crushing equipment is mostly large, fixed designs. For example, the dual-power mobile jaw crusher station and its control method described in Chinese invention patent (application number 202511066969.3) are bulky and heavy, making them unsuitable for the flexible production needs of industrial silicon processing companies. Furthermore, for such hard and brittle materials, they are prone to over-crushing and under-crushing, resulting in raw material waste. In addition, a few handheld crushing devices are relatively convenient. For example, a new type of handheld stone crusher for ore mining provided by Chinese utility model patent (application number 202121758319) can solve the problem of flexible movement. However, the size of silicon ingots is still determined by manual visual estimation, which does not improve the crushing accuracy of silicon ingots. The difference in experience and fatigue of operators will also exacerbate the accuracy error of size determination.

[0004] In summary, the crushing process in industrial silicon ingot production faces three major problems: low manual efficiency, inflexible fixed equipment, and unintelligent handheld equipment. Therefore, developing a handheld crushing device and method that can automatically identify silicon ingot size, intelligently plan crushing paths and points, and possess both flexibility and high production efficiency is an urgent need to solve industry pain points and improve the efficiency of the supply chain. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a handheld industrial silicon ingot crushing device and method based on visual recognition and intelligent path optimization. It automatically acquires and identifies the size information of silicon ingots / blocks through a binocular stereo vision algorithm, and intelligently selects the crushing point and adaptively adjusts the crushing parameters by combining a path planning algorithm. This can significantly reduce the error of manual judgment, realize semi-automatic identification and precise crushing of silicon ingots that exceed the standard, and improve the efficiency and stability of the crushing operation process.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A handheld industrial silicon ingot crushing method based on visual recognition and intelligent path optimization, the handheld industrial silicon ingot crushing method comprising the following steps:

[0008] The first step involves the operator wearing AR monitoring and display equipment and carrying a handheld crushing device into the work area to collect scene images and depth data of each target silicon block within the work area. Specifically:

[0009] The work area is scanned using binocular cameras and depth sensors on AR monitoring and display devices to generate scene images and pixel point cloud data. ,in Let be the pixel coordinates of a single pixel in the pixel cloud with respect to the image captured by the binocular camera. This is the distance from the actual location of the pixel to the optical center of the binocular camera, i.e., the depth data.

[0010] The second step involves analyzing the target silicon block size and local surface curvature within the scanned area based on the scene image obtained in the first step. Specifically:

[0011] The scene image obtained in the first step is subjected to denoising and partitioning to identify and separate the target silicon block and its background. Pixel point cloud data of the pixel region where the target silicon block is located is extracted sequentially. First, the minimum circumscribed cuboid model size is analyzed to determine whether the target silicon block size exceeds the standard; then, the local curvature of the target silicon ingot surface is calculated through local three-dimensional surface fitting to determine the selection of the breaking point.

[0012] Step 2.1: Remove noise points from the scene image obtained in the first step by Gaussian filtering, and then segment the scene image by Euclidean clustering algorithm to separate and identify the target silicon ingot from the background.

[0013] Step 2.2: Extract pixel point cloud data of the pixel region where the target silicon block is located. This is transformed into three-dimensional spatial coordinates under a unified main viewpoint coordinate system, as shown in formula (1). The main field of view coordinate system is based on the optical center of the binocular camera as the origin, with the x-axis extending horizontally to the right along the lens of the AR monitoring and interactive device, the y-axis extending vertically upward along the lens, and the z-axis extending outward perpendicular to the lens.

[0014] (1)

[0015] In the formula: The target silicon block's three-dimensional spatial coordinates are in the main field-of-view coordinate system; The coordinates of the center of the scene image; The horizontal focal length of the binocular camera. The focal length is in the vertical direction; Let be a rotation matrix. The translation vector is represented by both of these parameters, which are built into the binocular camera. The superscript T indicates transpose.

[0016] Step 2.3, obtaining 3D spatial coordinates based on pixel point cloud data correction The minimum bounding cuboid model of the target silicon block is constructed. First, the covariance matrix of the pixel point cloud data is calculated as shown in formula (2). Second, the covariance matrix is ​​decomposed as shown in formula (3). Finally, the vertex coordinate difference of the minimum bounding cuboid model in the axial direction is calculated as shown in formula (4). The actual size of the target silicon block is output, and the measurement error is controlled within ±0.5mm.

[0017] (2)

[0018] (3)

[0019] (4)

[0020] In the formula: N represents the total number of pixel point cloud data; Here, p represents the covariance matrix; p represents the three-dimensional spatial coordinates of a pixel. The target silicon block centroid is considered to be a homogeneous material; the superscript T indicates transpose. The principal axis eigenvector; For eigenvalues; This represents the minimum circumscribed cuboid size. These correspond to the length, width, and height of the target silicon block, respectively.

[0021] Determine the size of the target silicon block based on the minimum circumscribed cuboid model size. Does it exceed the size threshold? If the target silicon block size... If the size is greater than 200mm, the target silicon ingot is determined to be out of size, thus providing a basis for marking out-of-size silicon blocks in subsequent AR monitoring and display equipment.

[0022] Step 2.4, obtaining 3D spatial coordinates based on pixel point cloud data correction As shown in formula (5), the three-dimensional surface equation is fitted within a 9×9 scale window centered on each pixel, thus obtaining the second-order term matrix H as shown in formula (6). The eigenvalues ​​of its orthogonal system shown in formulas (7) and (8) are then solved. , That is, they are approximately represented by the maximum and minimum local surface curvature of the pixel.

[0023] (5)

[0024] (6)

[0025] (7)

[0026] (8)

[0027] in, Let each represent a three-dimensional spatial coordinate variable of the three-dimensional surface equation. Substitute the three-dimensional spatial coordinates... The coefficients of each term in the three-dimensional surface equation in formula (5) can be fitted to solve the problem. H is a second-order term matrix; and The maximum and minimum eigenvalues, respectively; and These represent the maximum and minimum local surface curvature of a pixel, respectively.

[0028] The third step involves analyzing the surface equivalent hardness of the target silicon block and determining any cracks. Specifically:

[0029] Step 3.1 further processes the scene image obtained from denoising and partitioning in Step 2.1 into grayscale. Utilizing the characteristic that cracks typically have low grayscale and high contrast, a local adaptive threshold is used to determine the cracks on the surface of the target silicon ingot. Specifically: for each pixel corresponding to the identified target silicon block, all pixels of the entire target silicon block are taken. The subscript n represents the total number of pixels, as shown in formulas (9) and (10) for calculating the local mean. with standard deviation The local adaptive threshold of the pixel is calculated as shown in formula (11). The system then analyzes whether a pixel represents a crack. The calculation formula is shown below:

[0030] (9)

[0031] (10)

[0032] (11)

[0033] in, This represents the grayscale dynamic range constant, used to normalize the grayscale variation range; express The grayscale value of the pixel at the location; This represents the sensitivity coefficient, used to adjust the degree of response of the threshold to changes in contrast, and is typically between 0.2 and 0.5.

[0034] Within the scene image, the local adaptive threshold for all pixels within the target silicon block should be the same; therefore, it is calculated only once and stored in the system data to avoid repeated calculations. When the grayscale value of a pixel... When the grayscale value is high, the pixel is determined to be a crack region; when the grayscale value is low, the pixel is determined to be a crack region. When this happens, the pixel is determined to be a non-cracked region.

[0035] Step 3.2: Based on the grayscale processing of the scene images described above, extract the contrast within each scene image partition as shown in formulas (12) and (13). Entropy Two texture roughness features, as shown in Formula (14), are input into a preset convolutional neural network model CNN, and the surface equivalent hardness value of each partition of the silicon ingot is output as shown in Formula (15).

[0036] (12)

[0037] (13)

[0038] (14)

[0039] (15)

[0040] In the formula: For the first Contrast of the partitions; represents the probability of the gray-level co-occurrence matrix of this partition; i and j are gray-level indices, ranging from 0 to 255; For the first The entropy value of the partition; To prevent zero constant, take ; This is the surface equivalent hardness value; This refers to the Convolutional Neural Network (CNN) model. The input feature vector; This is the normal roughness of the partition.

[0041] Furthermore, the Convolutional Neural Network (CNN) model includes two convolutional layers, two pooling layers, two fully connected layers, and an output layer. The convolutional layers use a 3×3 method to extract surface texture features from the target silicon block; the pooling layers remove local noise such as random bright spots, light points, and slight reflections; the fully connected layers map the surface texture features after removing local noise to equivalent surface hardness values; and the output layer uses linear regression to predict the equivalent surface hardness value for each region. The prediction method of the output layer is shown in formula (16):

[0042] (16)

[0043] In the formula, To output the predicted surface equivalent hardness value, a real number output, ranging from 0 to 10; These are the feature vectors of the fully connected layer. These are the output layer weights, obtained through network training; superscript. Indicates transpose; The bias term is also obtained from the training of the convolutional neural network model.

[0044] Furthermore, based on the predicted output of the equivalent surface hardness values ​​for each partition... Then, classify the hardness by region. If If so, it is classified as a high-hardness zone; if If so, it is classified as a medium hardness zone; if Then it is classified as a low hardness zone.

[0045] The fourth step involves displaying a locking frame for the excessive silicon block within the field of view of the AR monitoring and display device. A breakage path and candidate impact points are generated through comprehensive optimization of the measurement strategy, providing a breakage warning. Specifically:

[0046] Step 4.1: Based on the out-of-size silicon blocks screened out in Step 2.3, a red rectangular locking frame matching the outline of the out-of-size silicon block (i.e., the minimum circumscribed cuboid model) is generated in the AR monitoring and display device. If part of the target silicon block is outside the field of view and it is impossible to determine whether its size exceeds the limit, an arrow is displayed on the edge of the locking frame in the AR monitoring and display device to guide the operator to adjust the field of view to see the complete target silicon block.

[0047] Step 4.2: Based on the calculation, analysis and identification of the target silicon ingot size, local surface curvature, surface equivalent hardness value and cracks completed in steps 2 and 3, a comprehensive optimization strategy considering multiple factors is adopted to generate the fracture path, as shown in formula (17):

[0048] (17)

[0049] In the formula: Prioritize based on overall criteria; For the benefit of sizing; Surface equivalent hardness value Weighted terms; For crack weighting; These are safety penalties, and are mandatory conditions that must be met. The weights are for each item.

[0050] Furthermore, the scaling benefit, as shown in formula (18), can be approximated as:

[0051] (18)

[0052] In the formula: This represents the minimum circumscribed cuboid side length of the current target silicon block. In order to be in Updated edge length predicted by the local fracture model after impact.

[0053] Furthermore, the calculation of the hardness weighting term is the aforementioned surface equivalent hardness value.

[0054] Furthermore, the approximate calculation of the crack weighting term takes into account both the crack and the local curvature, as shown in formula (19):

[0055] (19)

[0056] In the formula: the clip function is a numerically limited function; Confidence level; These are curvature weighting coefficients; This represents the maximum principal curvature.

[0057] Furthermore, the safety penalty items mainly consider three points: (1) the distance between the hammer of the crushing device and personnel or other non-target silicon block objects is less than 100mm; (2) the elevation angle between the hammer and the ground is less than 30°; (3) if the amplitude and frequency parameters exceed the upper limit of the equipment, the device will be triggered to stop suddenly, and the penalty item will be infinitely large.

[0058] Step 4.3, based on the comprehensive priority of candidate hitting points Sorting conditions, plan path nodes, and generate Node sequence. Arrows appear on the AR monitoring and display equipment to guide operators to perform crushing operations in sequence.

[0059] The fifth step involves generating amplitude and frequency execution parameters based on the crack surface energy and success probability, transmitting these parameters to the handheld crushing device, and then the operator performs the operation according to these parameters. Specifically:

[0060] Step 5.1: Based on the surface equivalent hardness value obtained in step 3.2, estimate the crack energy required for the candidate impact point, as shown in formula (20):

[0061] (20)

[0062] In the formula: Candidate hitting points Required crack energy; is the critical energy release rate, and is the unit surface energy required for silicon material to fracture. The local equivalent thickness is measured in real time by the minimum circumscribed cuboid model of the target silicon ingot; The equivalent crack length is related to the contact surface of the hammer head and is taken as 0.3 to 0.6 times the diameter of the hammer head. Energy utilization efficiency; This is the hardness scaling factor; The surface equivalent hardness value of the partition r where the candidate impact point is located, obtained from step 3.2. .

[0063] Based on the expected success rate of breaking The target energy required for a single strike is calculated as shown in formula (21):

[0064] (twenty one)

[0065] In the formula: The target energy for a single strike; The crack surface energy requirement is calculated using formula (20); Candidate hitting points; The value represents the expected success probability of crushing, and should be chosen to balance crushing efficiency and the risk of over-crushing, ranging from 0.6 to 0.9. The shape parameter of the silicon ingot is 1 to 3.

[0066] Step 5.2, considering the equivalent impact model, the crack energy required at the candidate impact point is mapped to the frequency and amplitude required by the handheld crusher, as shown in formula (22):

[0067] (twenty two)

[0068] In the formula: The expected target energy for a single strike is obtained by formula (21); For kinetic energy transfer efficiency, we take 0.6; For the hammerhead mass; It represents the displacement amplitude; For frequency.

[0069] Furthermore, the selection of the specific execution parameters for the displacement amplitude A and frequency f should take into account reducing noise and vibration during equipment operation, and preferably using a smaller recommended frequency. Therefore, the corresponding Execution amplitude of the striking point With execution frequency Select using the following three steps: (1) Use the recommended frequency The amplitude of the action is calculated according to formula (23). (2) If the solution is obtained Upper limit of amplitude range Then take the execution frequency. (3) If the obtained execution amplitude Upper limit of amplitude range Then Substituting into formula (24), we obtain the execution frequency. .

[0070] (twenty three)

[0071] (twenty four)

[0072] In the formula: The specific amplitude of a single strike; The specific execution frequency for a single strike; This is the upper limit of the amplitude, which is adjusted according to the driving and damping systems used. This is the upper limit of the frequency, which is also adjusted according to the drive and damping systems used. , , The meanings of the parameters are the same as in formula (22).

[0073] Step 5.3: After the operator starts the handheld crushing device, the handheld crushing device drives the breaker hammer to impact according to the predetermined control parameters. During the impact process, the AR monitoring and display equipment continuously monitors the crushing progress and dynamically adjusts the path and impact plan. Specifically, the AR monitoring and display equipment acquires 10 frames of images of the target silicon block impact area per second, and updates the pixel point cloud data of the target area once according to steps 2.2 and 2.3, identifies and calculates the real-time size of the target silicon block, and determines whether it is qualified. Based on the actual crushing situation of the target silicon block on site, the crushing path and crushing point are dynamically adjusted according to step 4.2.

[0074] Furthermore, during the operation, the system automatically records information such as scene images, impact points, impact time, number of impacts, execution amplitude, and execution frequency into the process log.

[0075] Step 6: Once the target silicon block sizes within the viewing area all meet the requirements, the AR monitoring and display equipment generates a breakage report, and the power module enters low-power mode. Specifically:

[0076] When the AR monitoring and display device detects three consecutive times, at 500ms intervals, that the size of all target silicon blocks within the field of view is within the acceptable range of less than 200mm, the operation is automatically determined to be complete, and the handheld crushing device stops operating. Subsequently, key data is extracted from the process log, and a crushing report is output in a preset format. Its core information includes basic operation information and quality judgment information, such as operation time, number of strikes, pass rate, and abnormal alarms. After the operator has reviewed the report, they can manually close it. The report is then stored in the system history, and the AR monitoring and display device enters low-power mode.

[0077] A handheld industrial silicon ingot crushing device based on visual recognition and intelligent path optimization is disclosed. This device implements the aforementioned handheld industrial silicon ingot crushing method. The handheld industrial silicon ingot crushing device comprises two core components: an AR monitoring and display device and the handheld crushing device itself. The two components can achieve bidirectional data transmission via a high-interference-resistant data cable or Bluetooth pairing. Specifically:

[0078] The AR monitoring and display device includes a control module, an image perception module, a power supply module, a display module, and a communication module. The image perception module includes a binocular camera and a depth sensor to scan the work area and acquire scene images and depth information. The control module includes an image processor, a path planning module, and a power supply module, receiving data from the image perception module for target silicon ingot identification, size determination, path planning, and generation of crushing parameter commands. The display module is primarily a transflective silicon-based organic light-emitting diode (OLED) display screen, used to present the target silicon ingot's locking frame, impact point guidance prompts, and crushing progress information within the wearer's field of view. The power supply module supports entering a low-power mode after operation. The communication module supports two transmission methods: Bluetooth pairing and data cable transmission.

[0079] The handheld crushing device includes a communication module, a crushing execution module, a drive module, a grip and shock absorption module, and a sensing module. The communication module receives crushing parameter commands from the AR monitoring and display device. The crushing execution module is primarily a detachable tungsten carbide alloy impact breaker; the type and size of the hammerhead can be replaced and adjusted according to actual construction needs. The drive module includes a servo driver, which drives the breaker to perform crushing operations at a predetermined frequency and amplitude according to the crushing parameter commands received from the AR monitoring and display device. The grip and shock absorption module buffers high-frequency vibrations during the crushing process by wrapping the handle with an anti-slip rubber sleeve and incorporating shock-absorbing springs inside the device, reducing operational discomfort. Feedback is provided through the sensing module.

[0080] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0081] (1) By using visual recognition and image processing technology, the size of target silicon blocks in a certain area can be automatically calculated and screened. Combined with the projection marking guidance of AR monitoring and display equipment, the judgment error caused by human error can be significantly reduced.

[0082] (2) By taking the target silicon block size, crack condition, surface curvature and surface equivalent hardness as a comprehensive optimization strategy, a reference crushing path, candidate impact point and crushing execution parameters are generated. At the same time, the crushing condition of the target silicon block is dynamically corrected and the crushing strategy is adjusted in time, which can effectively reduce the over-crushing / under-crushing rate of industrial silicon ingots and improve the product qualification rate.

[0083] (3) AR monitoring and display equipment provides intuitive guidance within the field of view, which can effectively reduce the reliance of the production line on the operator's work experience and proficiency. At the same time, it automatically records operation data and generates breakage reports, which is beneficial for summarizing industrial production.

[0084] (4) The device connection method with wired and wireless dual links complementing each other and the design of replaceable impact hammer head make the present invention adaptable to industrial silicon ingot crushing under different production lines and material conditions. Attached Figure Description

[0085] Figure 1 This is a flowchart illustrating the usage of the present invention;

[0086] Figure 2 This is a schematic diagram of the circuit module composition;

[0087] Figure 3 This is a schematic diagram of the structure of the present invention;

[0088] Figure 4 This is a schematic diagram illustrating the identification and path planning of excessive silicon blocks according to the present invention;

[0089] In the picture: 1. AR monitoring and display equipment; 2. Handheld crushing equipment; 3. High anti-interference data cable. Detailed Implementation

[0090] The present invention will be further described below with reference to specific implementation examples.

[0091] A handheld industrial silicon ingot crushing method based on visual recognition and path optimization, comprising the following steps:

[0092] Step 1: The operator, wearing AR monitoring and display device 1 and carrying handheld crushing equipment 2, enters the work area to collect scene images and depth data of each target silicon block within the work area. Specifically:

[0093] The AR monitoring and display device 1 uses a binocular camera and a depth sensor to scan the work area, generating scene images and pixel point cloud data. ,in Let be the pixel coordinates of a single pixel in the pixel cloud with respect to the image captured by the binocular camera. This is the distance from the actual location of the pixel to the optical center of the binocular camera, i.e., the depth data.

[0094] The second step involves analyzing the target silicon block size and local surface curvature within the scanned area based on the scene image obtained in the first step. Specifically:

[0095] The scene image obtained in the first step is subjected to denoising and partitioning to identify and separate the target silicon block and its background. Pixel point cloud data of the pixel region where the target silicon block is located is extracted sequentially. First, the minimum circumscribed cuboid model size is analyzed to determine if the target silicon ingot size exceeds the limit. Then, the local curvature of the target silicon ingot surface is calculated through local three-dimensional surface fitting, which is used to select the impact point for breakage. Specifically:

[0096] Step 2.1: Remove noise points from the scene image obtained in the first step by Gaussian filtering, and then segment the scene image by Euclidean clustering algorithm to separate and identify the target silicon ingot from the background.

[0097] Step 2.2: Extract pixel point cloud data of the pixel region where the target silicon block is located. This is transformed into three-dimensional spatial coordinates under a unified main viewpoint coordinate system, as shown in formula (1). The main field of view coordinate system is based on the optical center of the binocular camera as the origin, with the x-axis extending horizontally to the right along the lens of the AR monitoring and interactive device 1, the y-axis extending vertically upward along the lens, and the z-axis extending outward perpendicular to the lens.

[0098] Step 2.3, obtaining 3D spatial coordinates based on pixel point cloud data correction The minimum bounding cuboid model of the target silicon block is constructed. First, the covariance matrix of the pixel point cloud data is calculated as shown in formula (2). Second, the covariance matrix is ​​decomposed as shown in formula (3). Finally, the vertex coordinate difference of the minimum bounding cuboid model in the axial direction is calculated as shown in formula (4). The actual size of the target silicon block is output, and the measurement error is controlled within ±0.5mm.

[0099] Determine the size of the target silicon block based on the minimum circumscribed cuboid model size. Does it exceed the size threshold? If so... This determines that the target silicon ingot size exceeds the standard, thus providing a basis for marking the excessive silicon blocks in the subsequent AR monitoring and display device 1.

[0100] Step 2.4, obtaining 3D spatial coordinates based on pixel point cloud data correction As shown in formula (5), the three-dimensional surface equation is fitted within a 9×9 scale window centered on each pixel, thus obtaining the second-order term matrix H as shown in formula (6). The eigenvalues ​​of its orthogonal system shown in formulas (7) and (8) are then solved. , That is, they are approximately represented by the maximum and minimum local surface curvature of the pixel.

[0101] The third step involves analyzing the surface equivalent hardness of the target silicon block and determining any cracks. Specifically:

[0102] Step 3.1 further processes the scene image obtained from denoising and partitioning in Step 2.1 into grayscale. Utilizing the characteristic that cracks typically have low grayscale and high contrast, a local adaptive threshold is used to determine the cracks on the surface of the target silicon ingot. Specifically: for each pixel corresponding to the identified target silicon block, all pixels of the entire target silicon block are taken. The subscript n represents the total number of pixels, as shown in formulas (9) and (10) for calculating the local mean. with standard deviation The local adaptive threshold of the pixel is calculated as shown in formula (11). And perform analysis to determine whether a pixel is a crack.

[0103] Furthermore, as described in formula (11) This represents the grayscale dynamic range constant; in this specific embodiment, it is taken as... ; express The grayscale value of the pixel at the location; This represents the sensitivity coefficient, which is typically taken as 0.2 to 0.5. In this embodiment, it is taken as... .

[0104] Within the scene image, the local adaptive threshold for all pixels within the target silicon block should be the same; therefore, it is calculated only once and stored in the system data to avoid repeated calculations. When the grayscale value of a pixel... When the grayscale value is high, the pixel is determined to be a crack region; when the grayscale value is low, the pixel is determined to be a crack region. When this happens, the pixel is determined to be a non-cracked region.

[0105] Step 3.2: Based on the grayscale processing of the scene images described above, extract the contrast within each scene image partition as shown in formulas (12) and (13). Entropy Two texture roughness features, as shown in Formula (14), are input into a preset convolutional neural network model CNN, and the surface equivalent hardness value of each partition of the silicon ingot is output as shown in Formula (15).

[0106] Furthermore, in the formula (13) To prevent zero constant, take This specific embodiment takes .

[0107] Furthermore, the Convolutional Neural Network (CNN) model comprises two convolutional layers, two pooling layers, two fully connected layers, and an output layer. The convolutional layers use a 3×3 method to extract surface texture features from the target silicon block; the pooling layers remove local acquisition noise such as random bright spots, light spots, and slight reflections; the fully connected layers map the surface texture features after removing local acquisition noise to equivalent surface hardness values; and the output layer uses linear regression to predict the equivalent surface hardness value of each partition. The prediction method of the output layer is shown in formula (16).

[0108] Furthermore, based on the predicted output of the equivalent surface hardness values ​​for each partition... Then, classify the hardness by region. If If so, it is classified as a high-hardness zone; if If so, it is classified as a medium hardness zone; if Then it is classified as a low hardness zone.

[0109] The fourth step involves displaying the locked frame of the excessive silicon block in the field of view of AR monitoring and display device 1. A breakage path and candidate impact points are generated through comprehensive optimization of the measurement strategy, providing a breakage warning. Specifically:

[0110] Step 4.1: Based on the out-of-size silicon blocks screened out in Step 2.3, a red rectangular locking frame matching the outline of the out-of-size silicon block, i.e., the minimum circumscribed cuboid model, is generated in AR monitoring and display device 1. If part of the target silicon block is outside the field of view and it is impossible to determine whether its size exceeds the standard, an arrow is displayed on the edge of the locking frame in AR monitoring and display device 1 to guide the operator to adjust the field of view to see the complete target silicon block.

[0111] Step 4.2: Based on the calculation, analysis and identification of the target silicon ingot size, local surface curvature, surface equivalent hardness value and cracks completed in the second and third steps, a comprehensive optimization strategy considering multiple factors is adopted to generate the breakage path, as shown in formula (17).

[0112] Furthermore, in the formula (17) The weights for each item are 0.4, 0.3, and 0.3 in this specific embodiment.

[0113] Furthermore, the size gain is shown in formula (18).

[0114] Furthermore, the calculation of the hardness weighting term is the aforementioned surface equivalent hardness value.

[0115] Furthermore, the approximate calculation of the crack weighting term takes into account the crack and local curvature, as shown in formula (19).

[0116] Furthermore, in the formula (19), when a crack is identified according to step 3.1, the confidence level is... Set the confidence level to 0.8; if no detection is found, then... Take 0.2; Take 0.2; Take the maximum local principal curvature of the surface obtained in step 2.4, that is If there is no curvature, then take 0.

[0117] Furthermore, the safety penalty items mainly consider three points: (1) the distance between the hammer of the crushing device and personnel or other non-target silicon block objects is less than 100mm; (2) the elevation angle between the hammer and the ground is less than 30°; (3) if the amplitude and frequency parameters exceed the upper limit of the equipment, the device will be triggered to stop suddenly, and the penalty item will be infinitely large.

[0118] Step 4.3, based on the comprehensive priority of candidate hitting points Sorting conditions, plan path nodes, and generate Node sequence. Arrows appear on AR monitoring and display device 1 to guide operators to perform crushing operations in sequence.

[0119] The fifth step involves generating amplitude and frequency execution parameters based on the crack surface energy and success probability, and transmitting these parameters to the handheld crushing device 2. The operator then performs the operation according to these parameters. Specifically:

[0120] Step 5.1: Based on the surface equivalent hardness value obtained in step 3.2, estimate the crack energy required for the candidate impact point, as shown in formula (20).

[0121] Furthermore, in the formula (20) The critical energy release rate is taken as 10 J / m³ in this specific embodiment. 2 ; The equivalent crack length is taken as 0.3 to 0.6 times the hammer diameter, and in this specific embodiment, it is taken as 0.5 times the hammer diameter; For energy utilization efficiency, this specific embodiment uses 0.25; The coefficient for hardness scaling is 0.2 in this specific embodiment.

[0122] Based on the expected success rate of breaking The target energy required for a single strike is calculated as shown in formula (21).

[0123] Furthermore, in the formula (21) The desired success rate of crushing should be determined by balancing crushing efficiency and the risk of over-crushing, and should be between 0.6 and 0.9. In this embodiment, 0.8 is chosen. The shape parameter of the silicon ingot is 1 to 3, and in this embodiment it is 1.6.

[0124] Step 5.2, considering the equivalent impact model, the crack energy required for the candidate impact point is mapped to the frequency and amplitude required by the handheld crusher 2, as shown in formula (22).

[0125] Furthermore, in the formula (22) For the kinetic energy transfer efficiency, a value of 0.6 is used in this specific embodiment; For displacement amplitude, in this specific embodiment, the amplitude range is 10-20mm; The frequency is specified in this specific embodiment, which uses a frequency range of 5-10Hz.

[0126] Furthermore, the selection of the specific execution parameters for the displacement amplitude A and frequency f should take into account reducing noise and vibration during equipment operation, and preferably using a smaller recommended frequency. This specific embodiment takes =5Hz. Therefore, the corresponding frequency is... Execution amplitude of the striking point With execution frequency Select using the following three steps: (1) Use the recommended frequency =5Hz, calculate the amplitude according to formula (23) (2) If the solution is obtained The upper limit of the amplitude range is taken in this specific embodiment. Then take the execution frequency. (3) If the obtained execution amplitude Upper limit of amplitude range Then Substituting into formula (24), we obtain the execution frequency. .

[0127] Furthermore, in the formula (23) As the upper limit of the amplitude, this specific embodiment takes 20mm; in the formula (24) As the upper limit of frequency, this specific embodiment uses 10Hz.

[0128] Step 5.3: After the operator starts the handheld crushing device 2, the handheld crushing device 2 drives the breaker hammer to impact according to the predetermined control parameters. During the impact process, the AR monitoring and display device 1 continuously monitors the crushing progress and dynamically adjusts the path and impact plan. Specifically, the AR monitoring and display device 1 acquires 10 frames of images of the target silicon block impact area per second, and updates the pixel point cloud data of the target area once according to steps 2.2 and 2.3, identifies and calculates the real-time size of the target silicon block, and determines whether it is qualified. Based on the actual crushing situation of the target silicon block on site, the crushing path and crushing point are dynamically adjusted according to step 4.2.

[0129] Furthermore, during the operation, the system automatically records information such as scene images, impact points, impact time, number of impacts, execution amplitude, and execution frequency into the process log.

[0130] Step 6: When the dimensions of all target silicon blocks within the viewing area meet the requirements, AR monitoring and display device 1 generates a breakage report, and the power module enters low-power mode. Specifically:

[0131] When the AR monitoring and display device 1 measures all target silicon blocks within its field of view to be within the acceptable range of less than 200mm for three consecutive times at 500ms intervals, it automatically determines that the operation is complete and stops the operation of the handheld crushing device 2. Subsequently, it extracts key data from the process log and outputs a crushing report in a preset format. The core information includes basic operation information and quality judgment information, such as operation time, number of blows, pass rate, and abnormal alarms. After the operator has reviewed the report, they can manually close it. The report is then stored in the system history, and the AR monitoring and display device 1 enters low-power mode.

[0132] A handheld industrial silicon ingot crushing device based on visual recognition and intelligent path optimization is disclosed. This device implements the aforementioned handheld industrial silicon ingot crushing method. The handheld industrial silicon ingot crushing device comprises two core components: an AR monitoring and display device and the handheld crushing device itself. The two components achieve bidirectional data transmission via a high-interference-resistant data cable. Specifically:

[0133] The AR monitoring and display device 1 includes a control module, an image perception module, a power supply module, a display module, and a communication module. The image perception module includes a binocular camera and a depth sensor, used to scan the work area and acquire scene images and depth information. The control module includes an image processor, a path planning module, and a power supply module, receiving data from the image perception module for target silicon ingot identification, size determination, path planning, and generation of crushing parameter commands. The display module is primarily a semi-transparent, semi-reflective silicon-based micro-organic light-emitting diode (OLED) display screen, used to present the target silicon ingot's locking frame, impact point guidance prompts, and crushing progress information within the wearer's field of vision. The power supply module supports entering a low-power mode after the operation is completed. The communication module supports two transmission methods: Bluetooth pairing and data cable transmission.

[0134] The handheld crushing device 2 includes a communication module, a crushing execution module, a drive module, a gripping and shock-absorbing module, and a sensing module. The communication module receives crushing parameter commands from the AR monitoring and display device 1. The crushing execution module is primarily a detachable tungsten carbide alloy impact breaker, with the hammer type and size adjustable according to actual construction needs. The drive module includes a servo driver, which drives the breaker to perform crushing operations at a predetermined frequency and amplitude according to the crushing parameter commands received from the AR monitoring and display device 1. The gripping and shock-absorbing module uses a non-slip rubber sleeve on the handle and internal shock-absorbing springs to buffer high-frequency vibrations during the crushing process, reducing operational discomfort.

[0135] The above embodiments are merely illustrative of the implementation methods of the present invention, but should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A handheld industrial silicon ingot crushing method based on visual recognition and intelligent path optimization, characterized in that, The handheld industrial silicon ingot crushing method includes the following steps: The first step is for the operator to wear AR monitoring and display equipment and carry a handheld crushing equipment into the work area to collect scene images and depth data of each target silicon block in the work area. The work area is scanned using binocular cameras and depth sensors on AR monitoring and display devices to generate scene images and pixel point cloud data. ,in Let be the pixel coordinates of a single pixel in the pixel cloud with respect to the image captured by the binocular camera. This is the distance from the actual location of the pixel to the optical center of the binocular camera, i.e., the depth data; The second step is to perform target silicon block size analysis and surface local curvature analysis within the scanning area based on the scene image obtained in the first step. The scene image is denoised and partitioned to identify and separate the target silicon block and its background; pixel point cloud data of the pixel region where the target silicon block is located are extracted sequentially. The minimum circumscribed cuboid model size is analyzed to determine whether the target silicon block size exceeds the standard; the local curvature of the target silicon ingot surface is calculated to determine the selection of the breaking point. The third step is to analyze the surface equivalent hardness value of the target silicon block and determine the cracks. The fourth step is to display the locking frame of the excessive silicon block in the field of view of the AR monitoring and display device, and generate the breakage path and candidate impact points through comprehensive optimization of the measurement strategy to provide a breakage warning. The fifth step involves generating amplitude and frequency execution parameters based on the crack surface energy and success probability, transmitting them to the handheld crushing device, and having the operator perform the operation according to the execution parameters. Step 6: When the size of the target silicon blocks within the viewing area meets the requirements, the AR monitoring and display equipment generates a breakage report, and the power module enters low-power mode.

2. The handheld industrial silicon ingot crushing method based on visual recognition and intelligent path optimization according to claim 1, characterized in that, The second step is specifically as follows: Step 2.1: Remove noise points in the scene image by Gaussian filtering, and then segment the scene image by Euclidean clustering algorithm to separate and identify the target silicon ingot from the background; Step 2.2: Extract pixel point cloud data of the pixel region where the target silicon block is located. This is transformed into three-dimensional spatial coordinates under a unified main viewpoint coordinate system, as shown in formula (1). The main field of view coordinate system is based on the optical center of the binocular camera as the origin, with the x-axis extending horizontally to the right along the lens of the AR monitoring and interactive device, the y-axis extending vertically upward along the lens, and the z-axis extending outward perpendicular to the lens. （1） In the formula: The target silicon block's three-dimensional spatial coordinates are in the main field-of-view coordinate system; The coordinates of the center of the scene image; The horizontal focal length of the binocular camera. The focal length is in the vertical direction; Let be a rotation matrix. These are translation vectors, and both are built-in parameters of the binocular camera. The superscript T indicates transpose; If the target silicon block size If the size is greater than 200mm, the target silicon ingot is determined to be out of standard, thus providing a basis for the identification of out-of-standard silicon blocks in subsequent AR monitoring and display equipment; Step 2.3, obtaining 3D spatial coordinates based on pixel point cloud data correction Construct the minimum bounding box model of the target silicon block; use formula (2) to calculate the covariance matrix of the pixel point cloud data, use formula (3) to perform eigenvalue decomposition on the covariance matrix, use formula (4) to calculate the vertex coordinate difference of the minimum bounding box model in the axial direction, and output the actual size of the target silicon block. （2） （3） （4） In the formula: N represents the total number of pixel point cloud data; Here, p represents the covariance matrix; p represents the three-dimensional spatial coordinates of a pixel. The target silicon block centroid is considered to be a homogeneous material; the superscript T indicates transpose. The principal axis eigenvector; For eigenvalues; This represents the minimum circumscribed cuboid size. These correspond to the length, width, and height of the target silicon block, respectively. Determine the size of the target silicon block based on the minimum circumscribed cuboid model size. Does it exceed the size threshold? Step 2.4, obtaining 3D spatial coordinates based on pixel point cloud data correction By fitting the three-dimensional surface equation, a second-order term matrix H is obtained, and the eigenvalues ​​are solved. , ; （5） （6） （7） （8） in, Let each represent a three-dimensional spatial coordinate variable of the three-dimensional surface equation. Substitute the three-dimensional spatial coordinates... It can fit the coefficients of each term in the three-dimensional surface equation in formula (5). H is a second-order term matrix; and The maximum and minimum eigenvalues, respectively; and These represent the maximum and minimum local surface curvature of a pixel, respectively.

3. The handheld industrial silicon ingot crushing method based on visual recognition and intelligent path optimization according to claim 2, characterized in that, The third step is specifically as follows: Step 3.1: Perform grayscale processing on the scene image obtained from denoising and partitioning in Step 2.1, and use local adaptive thresholding to determine the cracks on the surface of the target silicon ingot; Step 3.2: Extract the contrast within each scene image partition. Entropy Two texture roughness features are input into a pre-defined convolutional neural network (CNN) model, which outputs the surface equivalent hardness value of each region of the silicon ingot. ; like It is divided into high-hardness zones; if It is divided into a medium hardness zone; if It is divided into a low-hardness zone.

4. The handheld industrial silicon ingot crushing method based on visual recognition and intelligent path optimization according to claim 3, characterized in that, In the third step mentioned above: Step 3.1 specifically refers to: For each pixel corresponding to the identified target silicon block, take all pixels of the entire target silicon block. Where the index n represents the total number of pixels, the local mean is calculated. Standard deviation Local adaptive threshold for this pixel And perform analysis to determine whether a pixel is a crack; the calculation formula is as follows: （9） （10） （11） in, This represents the grayscale dynamic range constant, used to normalize the grayscale variation range; express The grayscale value of the pixel at the location; This represents the sensitivity coefficient, used to adjust the degree of response of the threshold to changes in contrast, and is typically between 0.2 and 0.

5. When the gray value of a pixel When the grayscale value is high, the pixel is determined to be a crack region; when the grayscale value is low, the pixel is determined to be a crack region. When this happens, the pixel is determined to be a non-cracked region; In step 3.2, contrast Entropy The calculation is shown in formulas (12) and (13), and the convolutional neural network model CNN is shown in formulas (14) and (15); （12） （13） （14） （15） In the formula: For the first Contrast of the partitions; represents the probability of the gray-level co-occurrence matrix of this partition; i and j are gray-level indices, ranging from 0 to 255; For the first The entropy value of the partition; To prevent zero constant, take ; This is the surface equivalent hardness value; This refers to the Convolutional Neural Network (CNN) model. The input feature vector; This is the normal roughness of the partition; The convolutional neural network (CNN) model comprises two convolutional layers, two pooling layers, two fully connected layers, and an output layer. The convolutional layers extract the surface texture features of the target silicon block, and the pooling layers remove local acquisition noise. The fully connected layers map the surface texture features after removing local acquisition noise to the equivalent surface hardness value. The output layer uses linear regression to predict the equivalent surface hardness value of each partition. The prediction method of the output layer is shown in formula (16). （16） In the formula, To output the predicted surface equivalent hardness value, a real number output, ranging from 0 to 10; These are the feature vectors of the fully connected layer. Output layer weights; superscript Indicates transpose; This is a bias term.

5. The handheld industrial silicon ingot crushing method based on visual recognition and intelligent path optimization according to claim 4, characterized in that, The fourth step is specifically as follows: Step 4.1: Based on the excessive silicon blocks screened out by the size threshold in Step 2.3, a red rectangular locking frame matching the outline of the excessive silicon blocks, i.e. the minimum circumscribed cuboid model, is generated in the AR monitoring and display device; if some areas of the target silicon block are outside the field of view and it is impossible to determine whether its size exceeds the standard, then an arrow is displayed on the edge of the locking frame of the AR monitoring and display device to guide the viewer to adjust the field of view to see the complete target silicon block. Step 4.2: Based on the calculation, analysis and identification of the target silicon ingot size, local surface curvature, surface equivalent hardness value and cracks completed in steps 2 and 3, a comprehensive optimization strategy considering multiple factors is adopted to generate the fracture path, as shown in formula (17): （17） In the formula: Prioritize based on overall criteria; For the benefit of sizing; Surface equivalent hardness value Weighted terms; For crack weighting; These are safety penalties, and are mandatory conditions that must be met. Weights for each item; Step 4.3, based on the comprehensive priority of candidate hitting points Sorting conditions, plan path nodes, and generate Node sequence; arrows appear in the AR monitoring and display equipment to guide the crushing operation in sequence.

6. The handheld industrial silicon ingot crushing method based on visual recognition and intelligent path optimization according to claim 5, characterized in that, In the fourth step mentioned above: The benefits of sizing As shown in formula (18): （18） In the formula: This represents the minimum circumscribed cuboid side length of the current target silicon block. In order to be in Updated edge length predicted by the local fracture model after impact; The hardness weighting term is the surface equivalent hardness value; The approximate calculation of the crack weighting term takes into account both the crack and the local curvature, as shown in formula (19): （19） In the formula: the clip function is a numerically limited function; Confidence level; These are curvature weighting coefficients; The maximum principal curvature; The safety penalty item considers three points: (1) the distance between the hammer of the crushing device and personnel or other non-target silicon block objects is less than 100mm; (2) the elevation angle between the hammer and the ground is less than 30°; (3) if the amplitude and frequency parameters exceed the upper limit of the equipment, the device will be triggered to stop, and the penalty item will be infinitely large.

7. The handheld industrial silicon ingot crushing method based on visual recognition and intelligent path optimization according to claim 6, characterized in that, The fifth step is specifically as follows: Step 5.1: Based on the surface equivalent hardness value obtained in step 3.2, estimate the crack energy required for the candidate impact point, as shown in formula (20): （20） In the formula: Candidate hitting points Required crack energy; The critical energy release rate; The local equivalent thickness is measured in real time by the minimum circumscribed cuboid model of the target silicon ingot; This is the equivalent crack length; Energy utilization efficiency; This is the hardness scaling factor; The surface equivalent hardness value of the partition r where the candidate impact point is located, obtained from step 3.

2. ; Based on the expected success rate of breaking The target energy required for a single strike is calculated as shown in formula (21): （21） In the formula: The target energy for a single strike; The crack surface energy requirement is calculated using formula (20); Candidate hitting points; The expected success rate of breaking is set to 0.6-0.9; The shape parameter of the silicon ingot is 1 to 3; Step 5.2, considering the equivalent impact model, the crack energy required at the candidate impact point is mapped to the frequency and amplitude required by the handheld crusher, as shown in formula (22): （22） In the formula: The expected target energy for a single strike is obtained by formula (21); For kinetic energy transfer efficiency, we take 0.6; For the hammerhead mass; It represents the displacement amplitude; For frequency; Step 5.3: After starting the handheld crushing device, the handheld crushing device drives the breaker hammer to carry out impact according to the predetermined control parameters. During the impact process, the AR monitoring and display equipment continuously monitors the crushing progress and dynamically adjusts the path and impact plan.

8. A handheld industrial silicon ingot crushing method based on visual recognition and intelligent path optimization according to claim 7, characterized in that, In the fifth step mentioned above: In step 5.2, the displacement amplitude A and frequency f are selected by choosing a fixed, small recommended frequency. ,correspond Execution amplitude of the striking point With execution frequency Select using the following three steps: (1) Use the recommended frequency The amplitude of the action is calculated according to formula (23). (2) If the solution is obtained Upper limit of amplitude range Then take the execution frequency. (3) If the obtained execution amplitude Upper limit of amplitude range Then Substituting into formula (24), we obtain the execution frequency. ; （23） （24） In the formula: The specific amplitude of a single strike; The specific execution frequency of a single strike; This is the upper limit of the amplitude, which is adjusted according to the driving and damping systems used. This is the upper limit of the frequency, which is also adjusted according to the drive and damping systems used. In step 5.3, the AR monitoring and display device acquires 10 frames of images of the target silicon block impact area per second, updates the pixel point cloud data of the target area once according to steps 2.2 and 2.3, identifies and calculates the real-time size of the target silicon block, and determines whether it is qualified. Based on the actual crushing situation of the target silicon block on site, the crushing path and crushing point are dynamically adjusted according to step 4.

2. During the operation, the scene image, impact point, impact time, number of impacts, execution amplitude, and execution frequency are automatically recorded in the process log.

9. A handheld industrial silicon ingot crushing method based on visual recognition and intelligent path optimization according to claim 8, characterized in that, The sixth step is specifically as follows: When the AR monitoring and display device measures the size of all target silicon blocks in the field of view to be within the qualified range of less than 200mm for three consecutive times at 500ms intervals, it automatically determines that the operation is completed and stops the operation of the handheld crushing device. Then, key data is extracted from the process log and a breakage report is output in a preset format.

10. A handheld industrial silicon ingot crushing device based on visual recognition and intelligent path optimization, characterized in that, The handheld industrial silicon ingot crushing method described in any one of claims 1-9 is implemented using a handheld industrial silicon ingot crushing device. The handheld industrial silicon ingot crushing device comprises two core components: an AR monitoring and display device and a handheld crushing device, which transmit data bidirectionally. Specifically: The AR monitoring and display device includes a control module, an image perception module, a power supply module, a display module, and a communication module. The image perception module includes a binocular camera and a depth sensor for scanning the work area and acquiring scene images and depth information. The control module includes an image processor, a path planning module, and a power supply module, which receives data from the image perception module and is used for target silicon block identification, size determination, path planning, and generation of crushing parameter instructions. The handheld crushing device includes a communication module, a crushing execution module, a drive module, a gripping and shock-absorbing module, and a sensing module. The communication module is used to receive crushing parameter commands sent by the AR monitoring and display device. The crushing execution module is mainly a breaker hammer. The drive module includes a servo driver, which drives the breaker hammer to perform crushing operations at a predetermined frequency and amplitude according to the crushing parameter commands received by the AR monitoring and display device. The gripping and shock-absorbing module is equipped with shock-absorbing springs to buffer vibrations during the crushing process.

Images