Crane suspension arm anti-collision early warning method and system
By using millimeter wave radar to acquire 3D point cloud data and combining with PointNet++ model for wire identification, obstacle identification problem in complex environments is solved, and the safety and efficiency of crane operations are improved.
Patent Information
- Application Number
- CN202510182823.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-19
- Publication Date
- 2025-06-27
Smart Images
Figure CN120208098A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of crane safety monitoring, and particularly relates to a method and system for anti-collision warning of a crane boom. Background Art
[0002] Currently, when a crane operates in a complex environment, there is a relatively high risk of the crane boom colliding with surrounding obstacles (especially high-voltage wires); existing technologies mainly rely on manual monitoring or traditional sensors to detect obstacles to prevent collision accidents. However, in blind spots of vision or complex and changeable environments, these existing measures are difficult to ensure operation safety.
[0003] Using a sensor to directly measure distance, although its system architecture is simple and implementation is convenient, it is only limited to measuring the distance information of a single obstacle. In a complex and changeable environment, this method is difficult to comprehensively capture the detailed details of the surrounding environment. In addition, it lacks the ability to comprehensively identify the characteristics of obstacles, especially the effective identification of fine objects such as wires. Summary of the Invention
[0004] The present invention solves the problem that existing technologies rely on manual monitoring or traditional sensors to detect obstacles, and proposes a method and system for anti-collision warning of a crane boom, which significantly improves the safety and efficiency of crane operation and reduces the risk of accidents.
[0005] To achieve the above object, the present invention adopts the following technical solutions: A method for anti-collision warning of a crane boom, comprising the following steps: S1, a radar scans to obtain 3D point cloud data and forms a radar reflection signal sequence, and preprocesses the radar reflection signal sequence to obtain a processed time-domain signal; S2, based on the PointNet++ model, segment the point cloud data, identify the wire point cloud data, and calculate the distance of the wire point cloud data relative to the radar; S3, determine whether there is a collision risk according to the distance of the wire point cloud data relative to the radar. If so, send a warning signal; if not, the crane continues to work.
[0006] In this technical solution, a millimeter-wave radar is used to scan environmental data to obtain high-precision 3D point cloud data, and a deep learning algorithm is used to analyze and process the data to detect obstacles around the crane boom in real time. The system can accurately calculate the relative distance between the crane boom and obstacles (especially wires), and automatically issue a warning or control the shutdown when detecting a potential collision risk, thereby significantly improving the safety and efficiency of crane operation.
[0007] The present invention is further set as: The radar scans to obtain 3D point cloud data of the environment and forms a radar reflection signal sequence, specifically including: The millimeter-wave radar installed at the top of the crane scans to obtain 3D point cloud data of the environment, and combines the reflection intensity S of each point in the 3D point cloud data i and the timestamp t i to form a reflection intensity value based on each time point, denoted as the radar reflection signal sequence.
[0008] The radar scans to obtain 3D point cloud data of the environment and forms a radar reflection signal sequence, specifically including: The millimeter-wave radar installed at the top of the crane scans to obtain 3D point cloud data of the environment, and combines the reflection intensity and timestamp of each point in the 3D point cloud data to form a reflection intensity value based on each time point, denoted as the radar reflection signal sequence.
[0009] The present invention is further configured as: the preprocessing of the radar reflection signal sequence to obtain the processed time-domain signal specifically includes: S11, performing a fast Fourier transform on the radar reflection signal sequence; S12, suppressing high-frequency components through a low-pass filter; multiplying with the result of the Fourier transform to obtain the filtered frequency-domain signal; S13, performing an inverse Fourier transform on the filtered frequency-domain signal to obtain the denoised time-domain signal, and updating it into the point cloud data.
[0010] In this technical solution, first, the Fourier transform is used to complete the conversion from the time domain to the frequency domain. Then, a low-pass filter is designed to suppress unnecessary high-frequency components and remove some high-frequency noises; multiplying with the result of the Fourier transform to obtain the filtered frequency-domain signal; finally, an inverse Fourier transform is performed to obtain the processed time-domain signal, which is updated into the point cloud data and used as the input of the PointNet++ model.
[0011] The present invention is further configured as: the step S2 includes a local feature learning process and a global feature aggregation process. The local feature learning process uses k-nearest neighbors to select the local neighborhood for each point, and aggregates the features of this neighborhood through a multi-layer perceptron.
[0012] In this technical solution, in the feature learning process of each layer, the model extracts local features according to each point and the points in its neighborhood.
[0013] The present invention is further configured as: the global feature aggregation process aggregates all features into a global feature representation, and assigns a class label to each point through a fully connected layer; during the training process, through the labeled wire point data, the model learns how to segment and identify the area where the wire is located, and uses the cross-entropy loss function to optimize the network parameters.
[0014] In this technical solution, local features are aggregated layer by layer upward and finally global features are formed. Through the max pooling operation, all features are aggregated into a global feature representation, thereby providing a basis for higher-level feature learning.
[0015] The present invention is further configured as: the step S3 includes: Set the position where the radar is located as the origin coordinate, calculate the distance of the wire point cloud data relative to the position where the radar is located, compare this distance with a preset safety threshold. If the distance is less than the safety threshold, there is a collision risk, trigger the anti-collision mechanism and send out a warning signal; if the distance is greater than the safety threshold, the crane continues to work.
[0016] In this technical solution, the calculated relative distance is compared with the safety threshold to further determine whether to issue a collision warning.
[0017] The present invention is further configured as: the 3D point cloud data includes the three-dimensional coordinates of each point in the point cloud, the timestamp of the radar echo corresponding to the point, and the reflection intensity of the corresponding point.
[0018] In this technical solution, the timestamp of the radar echo corresponding to the 3D point cloud data can reflect the time from the emission to the reception of the radar wave and can be used for subsequent distance calculation.
[0019] The present invention is further configured as: the step S3 further includes: during the working process of the crane, continuously monitor it throughout the whole process.
[0020] In this technical solution, the crane boom is monitored throughout the whole process to provide sufficient safety guarantee.
[0021] A crane boom anti-collision warning system applicable to the above-mentioned crane boom anti-collision warning method includes: An acquisition module that uses a millimeter-wave radar to obtain environmental point cloud data; A preprocessing module that preprocesses the radar reflection signal sequence to obtain a processed time-domain signal; A deep learning module that uses the PointNet++ model for 3D point cloud segmentation and recognition; A warning module that performs real-time warning according to the distance of the wire point cloud data relative to the radar.
[0022] The crane boom anti-collision warning system of this technical solution mainly includes an acquisition module, a preprocessing module, a deep learning module, and a warning module. The 3D point cloud data is acquired through the acquisition module, data filtering and feature extraction are performed through the preprocessing module, model training and detection are performed through the deep learning module, and real-time monitoring and warning are realized through the warning module.
[0023] The present invention can bring the following beneficial effects: A crane boom anti-collision warning method according to the present invention obtains environmental data through millimeter-wave radar scanning, performs 3D modeling, real-time detects obstacles around the boom, accurately calculates the relative distance, and automatically issues a warning or controls the shutdown when detecting a potential collision risk, thereby significantly improving the safety and efficiency of crane operation. Brief Description of the Drawings
[0024] Figure 1 is a schematic flow chart of a crane boom anti-collision warning method of the present application.
[0025] Figure 2 is a schematic framework diagram of a crane boom anti-collision warning system of the present application. Detailed Embodiments
[0026] To make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only the best embodiments of the present invention, which are only used to explain the present invention and do not limit the protection scope of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.
[0027] Embodiment 1 The application scope of crane operation is constantly expanding, especially showing irreplaceable value in key areas of construction sites and infrastructure construction; however, when the crane operates near high-voltage wires or dense power lines, the risk of collision between the boom and the wires increases significantly; once a collision occurs, not only may the crane suffer losses, but it may also trigger power accidents, resulting in serious casualties and property losses; currently, the industry mainly uses manual monitoring or traditional sensors to detect obstacles in order to prevent collision accidents; however, in blind spots of vision or complex and changeable environments, these existing measures often cannot ensure operation safety.
[0028] In the prior art, the method of directly measuring the distance using sensors, although its system architecture is simple and easy to implement, is only limited to measuring the distance information of a single obstacle; in complex and changeable scenarios, the above method is difficult to comprehensively capture the detailed details of the surrounding environment; in addition, the above method lacks the ability to comprehensively identify the characteristics of obstacles, especially the effective identification of fine objects such as wires.
[0029] To this end, this embodiment proposes a method for anti-collision warning of a crane boom, which can track the relative position of the boom and the wire in real time, has a dynamic warning function and can achieve automatic control; greatly improving the safety of crane operation and providing a solid protection barrier for construction workers.
[0030] A method for anti-collision warning of a crane boom in this embodiment refers to Figure 1 , and mainly includes the following several steps.
[0031] Step S1, use radar scanning to obtain 3D point cloud data. After obtaining the 3D point cloud data, recombine the data to obtain a radar reflection signal sequence, and then preprocess the radar reflection signal sequence to finally obtain a denoised time-domain signal.
[0032] For the above step S1, regarding its acquisition process, specifically: use a millimeter-wave radar installed at the top of the crane to scan and obtain 3D point cloud data of its surrounding environment, and combine the reflection intensity S i and timestamp t i of each point in the 3D point cloud data to obtain a reflection intensity value based on each time point, that is, a radar reflection signal sequence.
[0033] In the above technical solution, use a millimeter-wave radar to scan and obtain corresponding 3D point cloud data. After the scanning is completed, reorganize the 3D point cloud data, combine the reflection intensity and timestamp of each point in the point cloud to form a time series for subsequent frequency-domain analysis.
[0034] In this embodiment, the millimeter-wave radar is specifically a millimeter-wave radar in the 77GHz band, which is specifically installed at the top of the crane to facilitate obtaining 3D point cloud data of the surrounding environment of the crane.
[0035] The above 3D point cloud data mainly includes the three-dimensional coordinates of each point in the point cloud data, the timestamp of the radar echo corresponding to the point, and the reflection intensity of the corresponding point.
[0036] In this technical solution, the timestamp of the radar echo corresponding to the 3D point cloud data can reflect the time from the radar wave emission to reception and can be used for subsequent distance calculation.
[0037] More specifically, the 3D point cloud data of the surrounding environment is specifically expressed as: P i ={(x i , y i , z i , t i , S i )|=1,2,...,N} Among them, (x i, y i , z i ) are the three-dimensional coordinates of each point in the point cloud; t i represents the timestamp of the radar echo of point i, which reflects the time from the transmission to the reception of the radar wave and is mainly for subsequent distance calculation; S i represents the reflection intensity of point i.
[0038] Since the scanning signal of the radar is a time-domain signal, the reflection intensity S of each point i and the timestamp t i combined can form a time series, that is, the reflection intensity value at each time point is S(t) = [S1, S2,..., S N .
[0039] For the above process of preprocessing the radar reflection signal sequence to obtain the processed time-domain signal, it specifically includes the following several sub-steps.
[0040] Step S11, perform frequency-domain analysis on the formed radar reflection signal sequence using the fast Fourier transform.
[0041] Step S12, use a low-pass filter to suppress unnecessary high-frequency components and remove some high-frequency noises; by multiplying it with the result of the Fourier transform, a filtered frequency-domain signal can be formed.
[0042] Step S13, perform an inverse Fourier transform on the filtered frequency-domain signal. After completing the Fourier transform and filtering process, a denoised time-domain signal can be obtained and updated into the point cloud data.
[0043] In the above technical solution, first use the Fourier transform to complete the conversion from the time domain to the frequency domain. Then, design a low-pass filter to suppress unnecessary high-frequency components and remove some high-frequency noises; obtain a filtered frequency-domain signal by multiplying with the result of the Fourier transform; finally, perform an inverse Fourier transform to obtain the processed time-domain signal, update it into the point cloud data, and use it as the input of the PointNet++ model.
[0044] For the above steps S11 to S13, more specifically, it includes the following process.
[0045] For step S11, it uses the Fourier transform to perform frequency-domain analysis on the radar reflection signal sequence S(t), thereby removing noise and enhancing the effective signal. First, perform the Fourier transform on it to complete the conversion from the time domain to the frequency domain, and its formula is specifically as follows: where X(f) is the representation of the signal in the frequency domain; S iis the signal strength at the i-th time point; t i is the timestamp at the i-th time point; f is the frequency in the Fourier transform.
[0046] Next, for step S12, by designing a low-pass filter to suppress unnecessary high-frequency components and remove some high-frequency noise, the specific formula is: where f cut is the cut-off frequency of the set low-pass filter. Subsequently, by multiplying with the result X(f) of the Fourier transform, the filtered frequency-domain signal is obtained, and the specific formula is: X filtered (f) = X(f) · H(f).
[0047] For step S13, finally, by performing the inverse Fourier transform (IFFT) on the filtered frequency-domain signal, the denoised time-domain signal can be obtained, and the calculation formula is specifically: After completing the Fourier transform and filtering, the processed time-domain signal S filtered (t) can be obtained. These signals represent the reflection intensity of each point in the radar scan, and they are updated into the point cloud data, expressed as: P filtered = {(y i , y i , z i , t i , S filtered，i ) | i = 1, 2, …, N), and it is used as the input of PointNet++.
[0048] Step S2, according to the PointNet++ model, segment the point cloud data, identify the wire point cloud data, and finally calculate the distance of the wire point cloud data relative to the radar.
[0049] PointNet++ is developed on the basis of PointNet. By introducing a hierarchical feature learning mechanism, it can process point cloud data with local structures. Its main process is divided into two stages: the local feature learning process and the global feature aggregation process.
[0050] For the local feature learning process, in PointNet++, local feature learning is mainly achieved through neighborhood aggregation. In the feature learning process of each layer, the model extracts local features based on each point and its neighboring points. The k-nearest neighbors are used for each point p i = (x i , y i , z i , S filtered，i) Select its local neighborhood N(p i ), and perform feature aggregation on this neighborhood through a multi-layer perceptron. For the neighborhood N(p i ) of each point, its local feature representation (the feature of the l-th layer) can be calculated by the following formula: Where, is the local feature of the i-th point at the l-th layer; MLP (l) represents the multi-layer perceptron of the l-th layer, which is used to exchange the features of neighboring points.
[0051] For global feature aggregation, it includes feature aggregation and global feature learning. The local features are aggregated layer by layer upward, and finally a global feature is formed; through the max-pooling operation, all features are aggregated into a global feature representation, which provides a basis for higher-level feature learning.
[0052] The global feature representation at the l-th layer is calculated by the following formula: Based on the comprehensive representation of the global feature and the local feature, a class label, such as a wire or a boom, is assigned to each point through a fully connected layer. Wires are usually slender objects in the point cloud. PointNet++ can accurately identify wire points and classify them through fine segmentation of the point cloud.
[0053] During the training process, through the labeled wire point data, the model learns how to segment and identify the area where the wire is located, and uses the cross-entropy loss function to optimize the network parameters, as shown in the following formula: Where, y i，c is the true label (0 or 1) of the i-th point, indicating whether it belongs to the class c (such as a wire); p i，c is the probability that the model predicts the i-th point belongs to the class c, and C is the number of classes.
[0054] The segmented wire point cloud data can be expressed as: P line = { {(x i , y i , z i ) | i ∈ [1, n]} Where, n is the number of points in the wire point cloud. The position of the radar is the origin coordinate, and the distance between it and all points in the wire point cloud is calculated, specifically referring to the following formula:
[0055] Step S3: Determine whether there is a collision risk based on the distance between the wire point cloud data and the radar. If there is a corresponding collision risk, a warning signal will be issued; if there is no corresponding collision, the crane will continue to work.
[0056] More specifically, in the above steps, set the position of the radar as the origin coordinate, and then calculate the distance between the wire point cloud data and the position of the radar. After calculating the above distance, compare the above distance with the safety threshold. If the above distance is smaller than the safety threshold, it is considered that there is a collision risk, and the anti-collision mechanism can be immediately triggered and a warning signal will be issued. If the above distance is larger than the safety threshold, the crane will work normally.
[0057] In the above technical solution, compare the calculated relative distance with the safety threshold to further determine whether to issue a collision warning.
[0058] Once the above distance d min is obtained, it needs to be compared with the pre-set safety threshold d safe . If d min is less than the safety threshold d safe , it means that there is a collision risk, and the system should trigger the anti-collision mechanism, issue a warning signal, or take other control measures, such as stopping the boom operation or adjusting the boom position. Otherwise, the operation can continue.
[0059] After step S3, after sending the corresponding warning signal, the signal is displayed through the mobile app and the crane operator is reminded through the walkie-talkie.
[0060] During the operation of the crane, the whole process of the crane will be continuously monitored and warned.
[0061] In this technical solution, a millimeter-wave radar is used to scan the environmental data to obtain high-precision 3D point cloud data, and deep learning algorithms are used to analyze and process the data to detect obstacles around the boom in real time. The system can accurately calculate the relative distance between the boom and the obstacles (especially wires), and automatically issue a warning or control the shutdown when detecting potential collision risks, thus significantly improving the safety and efficiency of crane operations.
[0062] Embodiment 2 This embodiment proposes a method for preventing collision warning of the crane boom, referring to Figure 1 , which mainly includes the following several steps.
[0063] Step S1: Obtain 3D point cloud data through radar scanning. After obtaining the 3D point cloud data, reorganize the data to obtain a radar reflection signal sequence, and then preprocess the radar reflection signal sequence to finally obtain a denoised time-domain signal.
[0064] For the above step S1, regarding its acquisition process, specifically: Use a millimeter-wave radar installed at the top of the crane to scan and obtain 3D point cloud data of its surrounding environment, and combine the reflection intensity S i and the timestamp t i of each point in the 3D point cloud data to obtain the reflection intensity value based on each time point, that is, the radar reflection signal sequence.
[0065] In the above technical solution, use a millimeter-wave radar to scan and obtain the corresponding 3D point cloud data. After the scanning is completed, reorganize the 3D point cloud data, combine the reflection intensity and timestamp of each point in the point cloud to form a time series for subsequent frequency-domain analysis.
[0066] In this embodiment, the millimeter-wave radar is specifically a millimeter-wave radar in the 77GHz frequency band, which is specifically installed at the top of the crane to facilitate obtaining 3D point cloud data of the surrounding environment of the crane.
[0067] The above 3D point cloud data mainly includes the three-dimensional coordinates of each point in the point cloud data, the timestamp of the radar echo corresponding to the point, and the reflection intensity of the corresponding point.
[0068] In this technical solution, the timestamp of the radar echo corresponding to the 3D point cloud data can reflect the time from the radar wave emission to reception and can be used for subsequent distance calculation.
[0069] More specifically, the 3D point cloud data of the surrounding environment is specifically expressed as: P i ={(x i , y i , z i , t i , S i )|i = 1, 2,..., N} where (x i , y i , z i ) are the three-dimensional coordinates of each point in the point cloud; t i represents the timestamp of the radar echo of point i, which reflects the time from the radar wave emission to reception, mainly for subsequent distance calculation; S i represents the reflection intensity of point i.
[0070] Since the scanning signal of the radar is a time-domain signal, the reflection intensity S of each point iand the timestamp t i Combined, a time series can be formed, where the reflected intensity value at each time point is S(t) = [S1, S2,..., S N .
[0071] For the above process of preprocessing the radar reflection signal sequence to obtain the processed time-domain signal, it specifically includes the following several sub-steps.
[0072] Step S11: Perform frequency-domain analysis on the formed radar reflection signal sequence using the fast Fourier transform.
[0073] Step S12: Use a low-pass filter to suppress unnecessary high-frequency components and remove some high-frequency noises; by multiplying it with the result of the Fourier transform, a filtered frequency-domain signal can be formed.
[0074] Step S13: Perform an inverse Fourier transform on the filtered frequency-domain signal. After completing the Fourier transform and filtering processes, a denoised time-domain signal can be obtained and updated into the point cloud data.
[0075] In the above technical solution, first, the Fourier transform is used to complete the conversion from the time domain to the frequency domain. Then, a low-pass filter is designed to suppress unnecessary high-frequency components and remove some high-frequency noises; the filtered frequency-domain signal is obtained by multiplying it with the result of the Fourier transform; finally, an inverse Fourier transform is performed to obtain the processed time-domain signal, which is updated into the point cloud data and used as the input of the PointNet++ model.
[0076] For steps S11 to S13 above, more specifically, it includes the following process.
[0077] For step S11, it uses the Fourier transform to perform frequency-domain analysis on the radar reflection signal sequence S(t), thereby removing noises and enhancing the effective signal. First, perform the Fourier transform on it to complete the conversion from the time domain to the frequency domain, and its formula is specifically as follows: where X(f) is the representation of the signal in the frequency domain; S i is the signal intensity at the i-th time point; t i is the timestamp at the i-th time point; f is the frequency in the Fourier transform.
[0078] Immediately afterwards, for step S12, by designing a low-pass filter to suppress unnecessary high-frequency components and remove some high-frequency noises, the specific formula is: where f cutis the set cut-off frequency of the low-pass filter. Subsequently, by multiplying with the result X(f) of the Fourier transform, the filtered frequency-domain signal is obtained. The specific formula is: X filtered (f) = X(f) · H(f).
[0079] For step S13, finally, by performing the inverse Fourier transform (IFFT) on the filtered frequency-domain signal, the denoised time-domain signal can be obtained. The specific calculation formula is: After completing the Fourier transform and filtering, the processed time-domain signal S filtered (t) can be obtained. These signals represent the reflection intensity of each point scanned by the radar, and updating them into the point cloud data is expressed as: P filtered = {(x i , y i , z i , t i , S filtered，i ) | i = 1, 2,..., N}, and it is used as the input of PointNet++.
[0080] Step S2: Segment the point cloud data according to the PointNet++ model, identify the wire point cloud data, and finally calculate the distance of the wire point cloud data relative to the radar.
[0081] PointNet++ is developed on the basis of PointNet. By introducing a hierarchical feature learning mechanism, it can process point cloud data with local structures. Its main process is divided into two stages: the local feature learning process and the global feature aggregation process.
[0082] For the local feature learning process, in PointNet++, local feature learning is mainly achieved through neighborhood aggregation. In the feature learning process of each layer, the model extracts local features based on each point and its neighboring points. Using k-nearest neighbors, for each point p i = (x i , y i , z i , S filtered，i ), its local neighborhood N(p i ) is selected, and the feature aggregation of this neighborhood is performed through a multi-layer perceptron. For the neighborhood N(p i ) of each point, its local feature representation (the feature of the l-th layer) can be calculated by the following formula: Among them, is the local feature of the i-th point in the l-th layer; MLP (l)The multi-layer perceptron representing the l-th layer is used to exchange the features of neighboring points.
[0083] For global feature aggregation, it includes feature aggregation and global feature learning. The local features are aggregated layer by layer upward to finally form global features. Through the max-pooling operation, all features are aggregated into a global feature representation, thus providing a basis for higher-level feature learning.
[0084] The global feature representation at the l-th layer is calculated by the following formula: Based on the comprehensive representation of global features and local features, a class label is assigned to each point through a fully connected layer, such as a wire or a boom. Wires are usually slender objects in the point cloud. PointNet++ can accurately identify wire points and classify them through fine segmentation of the point cloud.
[0085] During the training process, through the labeled wire point data, the model learns how to segment and identify the area where the wire is located, and uses the cross-entropy loss function to optimize the network parameters, as shown in the following formula: where, y i，c is the true label (0 or 1) of the i-th point, indicating whether it belongs to class c (such as a wire); p i，c is the probability that the model predicts the i-th point belongs to class c, and C is the number of classes.
[0086] The segmented wire point cloud data can be expressed as: P line ={(x i , y i , z i )|i∈[1,n]} where, n is the number of points in the wire point cloud. The position of the radar is the origin coordinate, and the distance between it and all points in the wire point cloud is calculated, specifically referring to the following formula:
[0087] Step S3, determine whether there is a collision risk according to the distance of the wire point cloud data relative to the radar. If there is a corresponding collision risk, then a warning signal will be issued; if there is no corresponding collision, then the crane will continue to work.
[0088] More specifically, in the above steps, the position where the radar is located is set as the origin coordinate, and then the distance of the wire point cloud data relative to the position where the radar is located is calculated. After calculating the above distance, the above distance is compared with the safety threshold. If the above distance is smaller than the safety threshold, it is considered that there is a collision risk, and the anti-collision mechanism can be immediately triggered and a warning signal can be issued. If the above distance is larger than the safety threshold, the crane will work normally.
[0089] On this basis, referring to Figure 2 , this embodiment also proposes a crane boom anti-collision warning system, which includes an acquisition module, a preprocessing module, a deep learning module, and a warning module. The acquisition module is connected to the preprocessing module, the preprocessing module is connected to the deep learning module, and the deep learning module is connected to the warning module.
[0090] For the acquisition module, its main function is: to obtain 3D point cloud data about the surrounding environment according to a preset millimeter-wave radar.
[0091] The acquisition module mainly includes an acquisition unit, a sending unit, and a backup unit. The acquisition unit is respectively connected to the sending unit and the backup unit. The acquisition unit is the above-mentioned millimeter-wave radar. After the millimeter-wave radar completes the acquisition, on the one hand, the acquired 3D point cloud data is sent to the preprocessing unit through the sending unit, and the sending unit can be any form of communication device; on the other hand, the above 3D point cloud data is backed up through the backup unit, and when the sending fails or the communication is interrupted, the data can be restored and viewed through the backup unit.
[0092] For the preprocessing module, its main function is: to preprocess the radar reflection signal sequence to obtain the processed time-domain signal.
[0093] The preprocessing module mainly includes a filtering unit and an extraction unit. The filtering unit is connected to the extraction unit. After performing Fourier transform on the above 3D point cloud data, the transformed data is filtered in the filtering unit, and features are extracted from the filtered data in the extraction unit.
[0094] For the deep learning module, its main function is: to perform 3D point cloud segmentation and recognition using the PointNet++ model.
[0095] The deep learning module mainly includes a training unit, a model detection unit, and a calculation unit. The training unit can train the model according to the data of the preprocessing module and PointNet++; the model detection unit can perform model detection; the calculation unit can calculate the distance of the wire point cloud data relative to the radar.
[0096] For the early warning module, its main function is to conduct real-time early warning based on the distance between the wire point cloud data and the radar.
[0097] The early warning unit mainly includes a data acquisition unit, a distance judgment unit, and an execution unit. The data acquisition unit can obtain data from the deep learning module. The distance judgment unit can compare and analyze the distance between the wire point cloud data and the radar with the safety threshold. The execution unit can trigger the anti-collision mechanism and send early warning signals.
[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit them. Although the present invention has been described in detail with reference to the embodiments, those of ordinary skill in the art should understand that any modification or equivalent replacement of the technical solutions of the present invention does not depart from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.
Claims
1. A crane boom anti-collision warning method, characterized in that: The following steps are involved: S1, radar scanning acquires 3D point cloud data and forms a radar reflection signal sequence, and preprocesses the radar reflection signal sequence to obtain a processed time domain signal; S2, based on the PointNet++ model, segments the point cloud data, identifies the wire point cloud data and calculates the distance of the wire point cloud data relative to the radar; S3, determines whether there is a collision risk based on the distance of the wire point cloud data relative to the radar. If so, a warning signal is issued; if not, the crane continues to work.
2. A crane boom anti-collision warning method according to claim 1, characterized in that: The radar scanning acquires 3D point cloud data of the environment and forms a radar reflection signal sequence, specifically including: The millimeter-wave radar installed on the top of the crane scans and obtains the 3D point cloud data of the environment, and calculates the reflection intensity S of each point in the 3D point cloud data. i and timestamp t i The reflection intensity values at each time point are combined and recorded as radar reflection signal sequence.
3. A crane boom anti-collision warning method according to claim 2, characterized in that: The preprocessing of the radar reflection signal sequence to obtain the processed time domain signal specifically includes: S11, performing fast Fourier transform on the radar reflection signal sequence; S12, suppressing high frequency components by a low-pass filter; and obtaining a filtered frequency domain signal by multiplying the result of Fourier transform; S13, performing inverse Fourier transform on the filtered frequency domain signal to obtain a denoised time domain signal, and updating it to the point cloud data.
4. A crane boom anti-collision warning method according to claim 1, 2 or 3, characterized in that: The step S2 includes a local feature learning process and a global feature aggregation process. The local feature learning process uses k nearest neighbors to select a local neighborhood for each point, and performs feature aggregation on the neighborhood through a multi-layer perceptron.
5. A crane boom anti-collision warning method according to claim 4, characterized in that: The global feature aggregation process aggregates all features into a global feature representation, and assigns a category label to each point through a fully connected layer; during the training process, the model learns how to segment and identify the areas where the wires are located through labeled wire point data, and uses a cross entropy loss function to optimize network parameters.
6. A crane boom anti-collision warning method according to claim 1, 2 or 3, characterized in that: The step S3 comprises: Set the radar location as the origin coordinate, calculate the distance of the wire point cloud data relative to the radar location, and compare the distance with the preset safety threshold. If the distance is less than the safety threshold, there is a collision risk, triggering the anti-collision mechanism and issuing a warning signal; if the distance is greater than the safety threshold, the crane continues to work.
7. A crane boom anti-collision warning method according to claim 1, characterized in that: The 3D point cloud data includes the three-dimensional coordinates of each point in the point cloud, the timestamp of the radar echo of the corresponding point, and the reflection intensity of the corresponding point.
8. The crane boom anti-collision warning method according to claim 6, characterized in that: The step S3 also includes: during the operation of the crane, continuously monitoring the entire process.
9. A crane boom anti-collision warning method according to claim 6, characterized in that: The step S3 also includes: after sending the warning signal, it can be displayed through the mobile phone APP and the crane operator can be reminded through the intercom.
10. A crane boom anti-collision warning system, applicable to a crane boom anti-collision warning method according to any one of claims 1 to 9, characterized in that: include: The acquisition module uses millimeter-wave radar to obtain environmental point cloud data; A preprocessing module preprocesses the radar reflection signal sequence to obtain a processed time domain signal; Deep learning module, using PointNet++ model for 3D point cloud segmentation and recognition; The early warning module provides real-time early warning based on the distance of the wire point cloud data relative to the radar.