Intelligent control method for loading building
By installing lidar, millimeter-wave radar and visual cameras on the loading building and combining them with long-short-term memory network training models, the problems of loading operations requiring human experience and fault tolerance were solved, and loading accuracy and efficiency were improved.
Patent Information
- Application Number
- CN202511206409.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-27
- Publication Date
- 2025-10-03
AI Technical Summary
Existing loading operations require high levels of operator skill and experience, as well as high fault tolerance in the loading process, resulting in low loading efficiency.
LiDAR, millimeter-wave radar, and visual cameras are installed at the entrance and exit of the loading building. An initial model is constructed and the inference model is trained through a long short-term memory network. These sensors are used to obtain real-time information on the carriage position, speed, and material volume, and generate chute control signals to dynamically adjust the degree of opening and closing.
Through neural network training and data fusion, the loading process is automatically controlled, which improves loading accuracy and efficiency and reduces dependence on operators.
Smart Images
Figure CN120736288A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of railway freight transportation, and in particular to an intelligent control method for a loading building. Background Art
[0002] The loading building typically consists of a steel tower, a top belt level, a buffer bunker level, a metering bunker level, and a ground equipment level. The steel tower spans the railway. Materials conveyed from the top belt level are temporarily stored in the buffer bunker. When a train is ready to be loaded, the materials are conveyed to the metering bunker via gates below the buffer bunker. Weighing sensors are located below each of the four support points of the metering bunker to measure the weight of the materials delivered to the metering bunker. Once the loading weight is reached, the buffer bunker gates close, and the weighed materials are lowered into the train cars, completing the loading operation. Simultaneously, the weight of each load is stored in the loading building system's database to generate the final loading data sheet.
[0003] The existing loading process is that the operator in the loading building operates the circulating weight-bearing button to weigh the materials to be loaded into the train car; when a car starts to pass under the chute of the ground equipment layer, the circulating weight-bearing button is pressed to open the gate under the quantitative bin and put the measured materials into the chute; the flow gate on the chute is opened to load the materials into the car; during the loading process, the opening and closing degree of the flow gate is continuously adjusted according to the loading experience, and the weight of the remaining material in the quantitative bin is observed. While ensuring the uniformity of the material loading into the car, it is necessary to ensure that all the materials can be loaded into it when the car completely passes under the chute. When the car completely passes under the chute, the opening degree of the flow gate needs to be opened to the maximum to ensure that all the materials are loaded into the car; before the next car enters under the chute, the flow gate is closed to the appropriate position. This requires operators to maintain a high level of concentration throughout the entire train loading process. Each loading operation requires them to constantly monitor the uniformity of the materials within each car, while also observing the weight of the materials within the metering bin. Based on their experience, they must constantly adjust the opening and closing of the flow gates to ensure even loading of the materials into the cars. While constantly adjusting the flow gates, they must also press the cycle load button before loading each car to weigh the materials to be loaded into the next car. Therefore, operators in the loading hall not only require a high level of operational skills and experience, but also have a high tolerance for errors. This leads to long operation times and heavy workloads, resulting in low loading efficiency. Summary of the Invention
[0004] The technical problem to be solved by the present invention is that the existing loading operation has high requirements on the operating level and operating experience of the operator, and has high requirements on the fault tolerance rate of the loading process.
[0005] In order to solve the above problem, the present invention provides a control method capable of improving the loading accuracy of a loading building, and the specific solution is:
[0006] An intelligent control method for a loading building comprises the following steps:
[0007] S1. Pre-install a laser radar at the loading building exit near the chute, install a millimeter-wave radar on the top beams of the loading building exit and entrance, and install a visual camera on the top of the loading building entrance and exit;
[0008] S2. Build an initial model, obtain historical loading data, train the initial model based on the historical loading data, and obtain an inference model;
[0009] S3. The millimeter waves emitted by the millimeter-wave radar are used to obtain the car's position and speed in real time. The laser beam of the lidar is used to obtain the three-dimensional point cloud data of the car below the loading building, and a three-dimensional point cloud model is generated. The material point cloud in the three-dimensional point cloud model is extracted to calculate the volume of the material in the car. The visual camera is used to obtain the position of the blanking hole below the loading building in real time, and the car's position, car speed and blanking hole position are integrated to determine the position and posture of the car relative to the blanking hole.
[0010] S4. Input the carriage position, the carriage position and posture relative to the material drop port, and the material volume into the inference model to generate a chute control signal. The opening and closing degree of the loading building chute is controlled according to the chute control signal.
[0011] Compared with the prior art, the present invention adopting the above technical solution has the following beneficial effects:
[0012] The present invention collects a large amount of historical loading data as well as data collected by millimeter-wave radar, laser radar, and visual cameras to train a recurrent neural network. The neural network of the initial model learns the complex mapping relationship between loading operations and various influencing factors under different working conditions, such as different vehicle models, different material properties, and different weather conditions, thereby generating an inference model. During the loading process, the inference model is used to control the loading. According to the detection data input in real time by the laser radar, millimeter-wave radar, and visual camera, combined with the inference logic of the inference model, a chute control signal is output to dynamically adjust the opening and closing degree of the chute. For example, when it is detected that the car is about to be filled, the control automatically reduces the loading speed to avoid overloading. When it is detected that the car position is slightly deviated, the control adjusts the discharge direction of the chute in real time to ensure that the material falls accurately into the car.
[0013] Preferably, a further technical solution of the present invention is:
[0014] Specifically, S2 includes: selecting a long short-term memory network as the neural network architecture of the initial model, dividing the historical loading data into a training set, a validation set, and a test set, training the initial model with the training set to obtain an inference model, using the validation set to monitor the training process of the initial model, and using the test set to test the trained inference model.
[0015] The training process of S2 is specifically as follows:
[0016] S201. Calculate the error of the output layer of the initial model using the loss function. The loss function is:
[0017] ;
[0018] Where N is the number of training samples;
[0019] S202. Calculate the gradient of the loss function with respect to the weights and biases of the initial model through the back-propagation algorithm:
[0020] ;
[0021] ;
[0022] Where T is the length of the sequence, is the output value, is hidden state, is the weight matrix, is the bias vector;
[0023] S203: Update the weights and biases of the initial model through an optimization algorithm to obtain an inference model.
[0024] The optimization algorithm uses the gradient descent method with momentum, and the formula is:
[0025] ;
[0026] ;
[0027] in is the speed variable, is the momentum factor, which ranges from [0,1]. is the learning rate.
[0028] The other parameters include loading quality, visual loading status, chute opening degree, material type, weather information and car model. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Figure 1 It is a structural diagram of an embodiment of the present invention; DETAILED DESCRIPTION
[0030] The present invention will be further described below with reference to the embodiments, the purpose of which is only to provide a better understanding of the content of the present invention. Therefore, the examples given do not limit the scope of protection of the present invention.
[0031] Referring to the accompanying drawings, an embodiment of the present invention discloses an intelligent control method for a loading building, comprising the following steps:
[0032] S1. Pre-install a laser radar at the loading building exit near the chute, install a millimeter-wave radar on the top beams of the loading building exit and entrance, and install a visual camera on the top of the loading building entrance and exit;
[0033] S2. Build an initial model, obtain historical loading data, train the initial model based on the historical loading data, and obtain an inference model;
[0034] S3. Use the laser beam of the LiDAR to obtain the three-dimensional point cloud data of the carriage under the loading building, generate a three-dimensional point cloud model, extract the material point cloud in the three-dimensional point cloud model, and calculate the volume of the material in the carriage; use the millimeter waves emitted by the millimeter wave radar to obtain the carriage position and speed in real time; use the visual camera to obtain the position of the blanking hole under the loading building in real time, and fuse the carriage position, carriage speed and blanking hole position to determine the position and posture of the carriage relative to the blanking hole;
[0035] S4. Input the carriage position, the carriage position and posture relative to the material drop port, and the material volume into the inference model, generate a chute control signal, and send it to the loading control system. The loading control system controls the opening and closing degree of the loading building chute according to the chute control signal.
[0036] In this embodiment, S2 specifically includes: selecting a long short-term memory network as the neural network architecture of the initial model, dividing the historical loading data into a training set, a validation set, and a test set, training the initial model with the training set to obtain an inference model, using the validation set to monitor the training process of the initial model, and using the test set to test the trained inference model.
[0037] In this embodiment, a high-precision laser radar is selected as the laser radar, which is irradiated vertically downward. The FOV of the radar must be able to cover the material falling area below the chute discharge port, so as to ensure the real-time calculation of the mass of the loaded material in the car during the loading process. During installation, the laser radar is firmly fixed on the upper steel beam structure of the loading building near the chute discharge port to ensure that it can still work stably in a long-term vibration environment. The laser radar periodically emits a laser beam. When the laser encounters the material, it will be reflected. According to the time difference between the laser emission and reception, the distance between each point on the surface of the material in the car and the radar is accurately calculated, thereby constructing a three-dimensional point cloud model of the material. Extract the material point cloud, and accurately calculate the volume of the loaded material through the voxel algorithm. The voxel algorithm calculates the volume as follows:
[0038] 1. Determine the voxel grid range
[0039] Calculate the bounding box: Find the extreme values of the material point cloud on the x, y, and z axes, and determine the minimum three-dimensional space occupied by the point cloud:
[0040] ;
[0041] ;
[0042] ;
[0043] The length, width, and height of the bounding box are:
[0044] ;
[0045] ;
[0046] ;
[0047] 2. Voxel grid division
[0048] Set voxel size: Select voxel side length d=0.05 (unit: meter) based on the accuracy requirement. This setting meets the material volume measurement requirements.
[0049] Calculate voxel grid parameters:
[0050] Number of voxels along the X axis (length):
[0051] ( To round up to ensure the entire range is covered)
[0052] Number of voxels along the Y axis (width):
[0053] ;
[0054] Number of voxels along the Z axis (height):
[0055] ;
[0056] Volume of a single voxel:
[0057] ;
[0058] 3. Voxel occupancy judgment
[0059] Voxel coordinate mapping: for each material point , calculate the voxel index to which it belongs :
[0060] ;
[0061] ;
[0062] ;in To round down, The position of the corresponding voxel in the grid.
[0063] Deduplication statistics: Record all voxel indices occupied by the point cloud through a hash table or a three-dimensional array (to avoid repeated counting) to obtain the total number of occupied voxels .
[0064] 4. Volume calculation
[0065] Total volume formula:
[0066] * = ;
[0067] At the same time, the density information of different materials is stored in advance, and combined with the material volume, the material mass m is obtained in real time, and the mass data is sent to the loading control system.
[0068] In this embodiment, when installing millimeter-wave radars on the top beams of the loading building's entrance and exit, their tilt angles are adjusted so that their beams cover the train cars traveling on the tracks. The millimeter-wave radars utilize the Doppler effect to transmit millimeter waves and receive echoes reflected from the cars, acquiring real-time information about the car's position p, speed, and other parameters. To ensure data accuracy, on-site calibration is required after installation. This is done by simulating car movements at different speeds and positions to correct the millimeter-wave radar's measurements. The car's position, speed, and other parameters acquired by the millimeter-wave radars are transmitted to the loading control system.
[0069] In this embodiment, high-resolution visual cameras are installed at the top center of the entrance and exit of the loading building to ensure that their field of view can cover at least one entire train car area and the blanking port. The loading control system is equipped with a server image processing unit. The server image processing unit uses advanced image recognition algorithms to identify and detect the outline of a single car and the blanking port, thereby determining the position and posture of the car relative to the blanking port. At the same time, the image data obtained by the visual camera is fused with the data of the millimeter-wave radar. The specific fusion method is to input the car position, the position and posture of the car relative to the blanking port, the material volume and other parameters into the inference model. The inference model assigns corresponding weights to different types of data based on the data characteristics and accuracy of the visual camera and millimeter-wave radar, and then fuses the data of the two through a weighted average algorithm to generate more accurate car position and chute control signals.
[0070] The weighted average algorithm is as follows:
[0071] The visual signal is set as V. The visual system determines its value by identifying the relative position of the detection carriage and the blanking port target in the image. When the visual detection carriage is below the blanking port, the value of V is 1, otherwise it is 0, indicating that blanking is prohibited. The millimeter-wave radar signal is R. The millimeter-wave radar determines its value by detecting the radar position of the front and rear ends of the carriage and the blanking port. When the blanking port is located between the front and rear ends, its value is 1, otherwise it is 0. By testing multiple carriages, the accuracy of the visual signal is obtained. and radar signal accuracy For example, if 100 carriages are tested and the visual signal is correct for 92 carriages and wrong for 8 carriages, then its accuracy is 0.92; if the millimeter-wave radar is correct for 99 carriages, then its accuracy is 0.99;
[0072] By weighted average
[0073]
[0074]
[0075] Score range , set the threshold to 0.99. When S≥0.99, loading is possible, and when S<0.99, loading is not possible.
[0076] The other parameters include loading mass, visual loading status, chute opening and closing degree, material type, weather information and car model; the volume v is obtained by laser radar, and the loading mass m is obtained based on the material density, the car position p is obtained by millimeter wave radar, and the position and posture c of the drop port are obtained by vision, which are integrated together to form a (m, p, c) visual loading signal, and then the current chute opening and closing degree s, material type k, weather information w and car model u are integrated to generate a chute control signal I (m, p, c, s, k, w, u), which is input into the loading control system. The loading control system controls the opening and closing degree of the chute according to the chute control signal.
[0077] In this embodiment, step S2 specifically includes: collecting historical loading data, including car model, material type, weather information, car location, loading mass, and chute opening and closing degree. The collected data is preprocessed, including data cleaning and normalization. Data cleaning is used to remove obviously erroneous or abnormal data; normalization is used to unify values in different ranges to the same scale to facilitate initial model processing; the longer the data collection time, the better, and it is recommended to collect data from 100 trains. The more data, the higher the accuracy of the trained initial model; a long short-term memory network is selected as the neural network architecture of the initial model, and the preprocessed historical loading data is divided into a training set, a validation set, and a test set according to a certain ratio. The initial model is trained with the training set to obtain an inference model, the validation set is used to monitor the training process of the initial model, and the trained inference model is tested with the test set.
[0078] In this embodiment, the specific S2 is:
[0079] S201. During the training process, we define a loss function To measure the predicted value and the true value The difference between them is used to calculate the error of the output layer of the initial model through the loss function. The loss function is:
[0080] ;
[0081] Where N is the number of training samples;
[0082] S202. Calculate the gradient of the loss function with respect to the weights and biases of the initial model through the back-propagation algorithm:
[0083] ;
[0084] ;
[0085] in, is the length of the sequence, is the output value, is hidden state, is the weight matrix, is the bias vector of the hidden layer;
[0086] S203. After obtaining the gradient, the weights and biases of the initial model are updated through an optimization algorithm to obtain an inference model. A common optimization algorithm is the gradient descent method, and its update formula is:
[0087] , , , ;
[0088] in, is the learning rate, which controls the step size of each update.
[0089] In this embodiment, the optimization algorithm adopts the gradient descent method with momentum, and the formula is:
[0090] ;
[0091] ;
[0092] in is the speed variable, is the momentum factor, and its value is between [0,1].
[0093] The principle is to expand the long short-term memory network in the time dimension, similar to a multi-layer feedforward neural network, and then calculate the gradient according to the traditional back propagation algorithm; assuming that at time step t, the input is , the hidden state is , the output is The hidden state update formula of the long short-term memory network is usually: ,in is the activation function, is the weight matrix input to the hidden layer, is the weight matrix from hidden layer to hidden layer, and the output calculation is: ,in, is the activation function of the output layer, is the bias vector of the output layer; by continuously performing forward propagation to calculate the output, backpropagation to calculate the gradient, and using the optimization algorithm to update the weights and biases, the long short-term memory network gradually learns the complex mapping relationship between loading operations and various influencing factors under different working conditions; the validation set is used to monitor the training process to avoid overfitting of the network; overfitting means that the initial model performs very well on the training data and can perfectly fit various details in the training data, including noise and some special, non-generally representative features, but performs poorly on new, unseen data (test data) and lacks generalization ability, thus obtaining a well-trained inference model.
[0094] The inference model in the previous step is deployed to the server, and then the chute control signal I (m, p, c, s, k, w, u) from the millimeter wave radar, lidar and visual camera fusion during the loading process is received. The inference model infers the opening and closing degree of the chute at the next moment, whether there is an abnormality, etc. based on the trained parameters; the chute control signal at the next moment , where s is the chute opening degree at the next moment, and e is whether the loading is abnormal. The loading control system sends these parameters to the PLC for loading control. For example, if the inference model determines that the car is about to be full, it sends a command to the loading control system to reduce the chute opening degree and automatically reduce the loading speed to prevent overloading. If the inference model determines that the loading mass is too small, the loading control system increases the chute opening degree and the loading speed to prevent underloading. If the inference model determines that the loading mass is too small, the loading control system activates chute vibration control to prevent material from sticking to the chute.
[0095] The present invention collects a large amount of historical loading data and data collected by real-time sensors to train a recurrent neural network. The neural network of the initial model learns the complex mapping relationship between loading operations and various influencing factors under different working conditions, such as different vehicle models, different material properties, and different weather conditions, thereby generating an inference model. During the loading process, the inference model is used to control the loading process. Based on the detection data input in real time by the lidar, millimeter-wave radar, and visual camera, combined with the inference logic of the inference model, a chute control signal is output to dynamically adjust the opening and closing degree of the chute. For example, when it detects that the car is about to be filled, the system will automatically reduce the loading speed to avoid overloading. When it detects that the car position is slightly deviated, the system will adjust the discharge direction of the chute in real time to ensure that the material falls accurately into the car.
[0096] The above description is only a preferred embodiment of the present invention and does not limit the scope of the present invention. Any equivalent changes made using the contents of the present invention specification and its drawings are included in the scope of the present invention.
Claims
1. A method for intelligent control of a loading building, characterized in that: The following steps are involved: S1. Pre-install a laser radar at the loading building exit near the chute, install a millimeter-wave radar on the top beams of the loading building exit and entrance, and install a visual camera on the top of the loading building entrance and exit; S2. Build an initial model, obtain historical loading data, train the initial model based on the historical loading data, and obtain an inference model; S3. The millimeter waves emitted by the millimeter-wave radar are used to obtain the car's position and speed in real time. The laser beam of the lidar is used to obtain the three-dimensional point cloud data of the car below the loading building, and a three-dimensional point cloud model is generated. The material point cloud in the three-dimensional point cloud model is extracted to calculate the volume of the material in the car. The visual camera is used to obtain the position of the blanking hole below the loading building in real time, and the car's position, car speed and blanking hole position are integrated to determine the position and posture of the car relative to the blanking hole. S4. Input the carriage position, the carriage position and posture relative to the material drop port, the material volume, and other parameters into the inference model to generate a chute control signal. The opening and closing degree of the loading building chute is controlled according to the chute control signal.
2. The intelligent control method for a loading building according to claim 1, characterized in that: Specifically, S2 includes: selecting a long short-term memory network as the neural network architecture of the initial model, dividing the historical loading data into a training set, a validation set, and a test set, training the initial model with the training set to obtain an inference model, using the validation set to monitor the training process of the initial model, and using the test set to test the trained inference model.
3. The intelligent control method for a loading building according to claim 2, characterized in that: The training process of S2 is specifically as follows: S201. Calculate the error of the output layer of the initial model using the loss function. The loss function is: ； Where N is the number of training samples; S202. Calculate the gradient of the loss function with respect to the weights and biases of the initial model through the back-propagation algorithm: ； ； Where T is the length of the sequence, is the output value, is hidden state, is the weight matrix, is the bias vector; S203: Update the weights and biases of the initial model through an optimization algorithm to obtain an inference model.
4. The intelligent control method for a loading building according to claim 2, characterized in that: The optimization algorithm uses the gradient descent method with momentum, and the formula is: ； ； in is the speed variable, is the momentum factor, which ranges from [0,1]. is the learning rate.
5. The intelligent control method for a loading building according to claim 1, characterized in that: The other parameters include loading mass, visual loading signals, chute opening, material type, weather information and car model.