Method and system for detecting material level of unmanned intelligent loading

Abstract

The invention discloses a material level detection method and system for unmanned intelligent loading, and relates to the technical field of material level detection, and the method comprises the steps: obtaining the current weather condition, the vehicle type and the vehicle hopper size, carrying out the collection through a sensor or manual registration or a range radar, and measuring the information redundancy and value density in different data sources through information entropy, automatically judging whether multi-sensor redundancy input is needed or not; according to the material level detection method and system for unmanned intelligent loading, the importance and redundancy degree of all data sources in environment and vehicle information can be dynamically evaluated, whether multi-sensor cooperative collection needs to be started or not is automatically decided, and the adaptive capacity of the system to different working conditions is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of material level detection, and in particular to a method and system for material level detection in unmanned intelligent loading. Background Art

[0002] In automated loading scenarios for bulk materials, such as mines, ports, and power plants, unmanned intelligent loading technology is gaining widespread adoption to improve operational efficiency and reduce labor costs. Traditional loading material level detection methods rely primarily on manual observation, fixed-point radar, or simple camera systems monitoring the vehicle bed.

[0003] Such methods often suffer from low precision, weak dynamic adaptability, and the inability to adjust adaptively to vehicle type and loading environment. This can easily lead to underloading or overloading during the loading process, affecting the accuracy of material measurement and the safety of subsequent transportation. In recent years, with the development of deep learning, image recognition, and multi-source information fusion technologies, intelligent perception and decision-making systems based on artificial intelligence have provided new solutions for unmanned loading. However, most current systems still fail to achieve the joint modeling of complex working conditions such as weather, dust, and humidity with variables such as loading speed and truck bed structure, and lack the ability to dynamically adjust loading strategies based on different vehicle types. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for material level detection in unmanned intelligent loading to solve the problems existing in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for material level detection in unmanned intelligent loading, comprising:

[0006] S1. Obtain the current weather conditions, vehicle type, and truck bed size using sensors, manual registration, or ranging radar. Information entropy is used to measure the information redundancy and value density in different data sources, automatically determining whether multi-sensor redundant input is required.

[0007] S2. Use a ranging radar to measure the material level. During the loading process, capture images of the truck bed and perform image recognition on the images. Analyze the current dust concentration based on the image recognition results. Determine the dynamic compensation value of the current ranging radar data based on the humidity and dust concentration. Measure the degree of dust interference in the truck bed image using image entropy. Use image entropy as a weighting factor to adjust the compensation ratio of the ranging radar.

[0008] S3. Use the changes in bucket material level during the loading process of different vehicles as a dataset for deep learning, thereby establishing a functional relationship between the material level changes of different vehicles and the loading speed and loading position. When training the material level change model for different vehicle types, use the KL divergence to measure the difference between the predicted distribution and the actual material level change distribution as the loss function;

[0009] S4. During the loading process, the vehicle is moved to change the loading position and displace the material level. The material level changes in different buckets are determined based on the trained functional relationship, and the loading speed at the next moment and whether to stop loading are determined accordingly. The material accumulation caused by the change in loading position is modeled, the direction and result of the resultant force at each loading point are calculated, and the material flow trend is analyzed. When the direction of the resultant force is aligned with the center of the bucket, the loading is judged to be uniform. Otherwise, the loading trajectory is corrected by adjusting the position and direction of the vehicle.

[0010] Preferably, it also includes S5, ending the loading when the initial height of the reset material level is close to the vehicle's allowable loading height threshold, and correcting the parameters of the loading function based on the loading result, locally approximating the function curve between the current material level and the height threshold using Newton interpolation, estimating when the threshold will be reached, and issuing an early warning. If the final result deviates from the prediction, the error quotient is updated using error feedback to correct the loading function model.

[0011] Preferably, it also includes S6, adding the relationship between the change in material level during the loading process and the change in vehicle position to the deep learning training set, further training and optimizing the relationship between the loading speed and function of the buckets of different vehicle models, and mapping different vehicle model numbers, material level characteristics, and position combinations to different training channels through hash mapping.

[0012] Preferably, when training the material level change model for different vehicle types in S3, the KL divergence is used to measure the difference between the predicted distribution and the actual material level change distribution. The specific formula is:

[0013] ;

[0014] in, is the KL divergence, which indicates the degree of difference between the model prediction and the actual observation. The actual bucket material level change data is in the The probability of a state, The material level distribution predicted for the network is The probability of a state, It is the index of the material level distribution interval.

[0015] Preferably, in S4, the material accumulation caused by the change in the charging position is modeled, and the specific formula for calculating the resultant force direction and resultant force result of each charging point is:

[0016] ;

[0017] in, represents the sum of the accumulation trend vectors of all charging points, Indicates the Material accumulation trend vector at each loading location, Indicates the index number of the loading position point, Indicates the total number of charging points.

[0018] Preferably, the S2 also includes performing a time sequence consistency check on the truck bed images of multiple consecutive time frames in the image recognition stage to eliminate the interference of instantaneous occlusion or abnormal lighting on the image entropy calculation.

[0019] Preferably, the S3 also includes introducing a model memory mechanism during the deep learning training process to identify and memorize the recurring loading behavior patterns of specific vehicle models, thereby reducing the amount of real-time calculations.

[0020] Preferably, the step S4 further includes performing intelligent path filtering on the relative displacement between the current loading position and the previous loading point, and only retaining the displacement direction that can bring material level optimization benefits.

[0021] Preferably, in S6, different vehicle model numbers, material level characteristics, and location combinations are mapped to different training channels through hashing. The specific formula is:

[0022] ;

[0023] in, represents the input feature encoding, Indicates the total number of hash buckets, Indicates the training subset number to which the input sample is assigned, Represents the modulo operator.

[0024] A material level detection system for unmanned intelligent loading is used to implement the steps of a material level detection method for unmanned intelligent loading, and the system includes:

[0025] The information collection module is used to obtain weather, vehicle type, and truck bed size information. It uses sensors, manual registration, or ranging radar to collect data and determines whether to enable multi-sensor redundant input based on information entropy.

[0026] Image and radar fusion module, used to obtain truck bed images and identify dust concentration during the loading process, and dynamically compensate the ranging radar data by combining humidity and image entropy;

[0027] The model training module is used to build a deep learning model based on the material level changes during the loading process of different vehicles, and optimize the functional relationship between the material level and the loading speed and loading position through KL divergence;

[0028] The control decision module is used to determine the current material level status based on the model, control the vehicle position adjustment and update the loading speed, analyze the material flow uniformity according to the direction of the resultant force at each loading point, and dynamically correct the loading trajectory.

[0029] It can be seen from the above technical solution that the present invention has the following beneficial effects:

[0030] The method and system for material level detection of unmanned intelligent loading, by introducing an information entropy evaluation mechanism, can dynamically evaluate the importance and redundancy of each data source in the environment and vehicle information, automatically decide whether to enable multi-sensor collaborative acquisition, effectively improve the system's adaptability to different working conditions, integrate image recognition and ranging radar, combine dust concentration and image entropy weight to dynamically compensate radar data, improve the system's ranging accuracy in complex environments such as dust and humidity, avoid material level estimation deviation caused by single perception error, establish a mapping relationship between material level change and loading speed and position through deep learning, and optimize the prediction model with KL divergence, which can It can accurately reflect the differences in loading behaviors of multi-type and multi-structure truck beds, improve the generalization ability of loading strategies, adjust vehicle position and loading speed according to real-time judgment results, and analyze material flow trends through combined force modeling to ensure uniform distribution of materials in the truck bed, avoid overloading or uneven loading due to loading offset, and improve loading quality and transportation safety. Compared with the traditional method that relies on fixed-point radar or manual observation, the present invention realizes the intelligent fusion of perception-judgment-control closed loop, and has stronger autonomous loading decision-making ability. It is suitable for the deployment of unmanned loading systems in bulk material scenarios such as mines, ports, and power plants, and has good practicality and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Flow chart of the method of the present invention;

[0032] Figure 2 This is a system connection diagram of the present invention. DETAILED DESCRIPTION

[0033] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0034] like Figure 1 As shown, the present invention provides a technical solution: a method for material level detection in unmanned intelligent loading, comprising:

[0035] S1. Obtain the current weather conditions, vehicle type, and truck bed size using sensors, manual registration, or ranging radar. Information entropy is used to measure the information redundancy and value density in different data sources, automatically determining whether multi-sensor redundant input is required.

[0036] S2. Use a ranging radar to measure the material level. During the loading process, capture images of the truck bed and perform image recognition on the images. Analyze the current dust concentration based on the image recognition results. Determine the dynamic compensation value of the current ranging radar data based on the humidity and dust concentration. Measure the degree of dust interference in the truck bed image using image entropy. Use image entropy as a weighting factor to adjust the compensation ratio of the ranging radar.

[0037] S3. Use the changes in bucket material level during the loading process of different vehicles as a dataset for deep learning, thereby establishing a functional relationship between the material level changes of different vehicles and the loading speed and loading position. When training the material level change model for different vehicle types, use the KL divergence to measure the difference between the predicted distribution and the actual material level change distribution as the loss function;

[0038] S4. During the loading process, the vehicle is moved to change the loading position and displace the material level. The material level changes in different buckets are determined based on the trained functional relationship, and the loading speed at the next moment and whether to stop loading are determined accordingly. The material accumulation caused by the change in loading position is modeled, the direction and result of the resultant force at each loading point are calculated, and the material flow trend is analyzed. When the direction of the resultant force is aligned with the center of the bucket, the loading is judged to be uniform. Otherwise, the loading trajectory is corrected by adjusting the position and direction of the vehicle.

[0039] This implementation establishes an unmanned intelligent loading system that integrates environmental perception, adaptive compensation, image recognition, and intelligent decision-making to achieve real-time and accurate monitoring and dynamic adjustment of the loading status of materials in the truck bed. The overall system is built on the basis of multimodal data fusion and deep learning. The key technical paths are as follows:

[0040] Environment and initial parameter perception stage:

[0041] First, environmental sensors are used to capture key variables such as weather parameters (such as temperature, humidity, and wind speed) and vehicle information (model number, truck bed dimensions) in the current loading environment. Data collection can be automated (e.g., using temperature and humidity sensors or millimeter-wave radar) or manually entered. Information entropy is used to calculate the redundancy and effective information density of different input sources. When multi-source data is highly redundant and the confidence level of a single sensor is insufficient, the system automatically activates a multi-sensor redundancy mechanism to enhance the robustness and robustness of the input data.

[0042] Real-time material level detection and image information fusion stage:

[0043] During the loading process, a ranging radar is used to continuously and dynamically scan the material height (i.e., material level) within the truck bed. Simultaneously, an industrial camera mounted above or to the side of the loading equipment captures intermittent or continuous images of the truck bed's entirety. The system uses image recognition algorithms (such as YOLO and Mask R-CNN) to detect dust distribution characteristics, changes in the pile's outline, and illumination offsets within the truck bed. Combined with current humidity data, the system then assesses the degree to which dust interferes with the radar's ranging signal reflection intensity.

[0044] Based on this, image information is quantified as an image entropy metric, reflecting the impact of interfering elements such as dust on material level identification. Based on image entropy, the system sets a dynamic weighting coefficient to adjust the compensation ratio of the ranging radar output signal to offset signal distortion in areas with high dust concentrations. The compensation strategy can be dynamically calculated using a linear regression model or polynomial fitting method.

[0045] Charging process modeling and learning phase:

[0046] To enhance intelligent decision-making capabilities, the system collects historical data on bucket level changes during loading processes, creating a dataset encompassing multiple vehicle types and loading conditions. Sequential modeling networks, such as LSTM and Transformer, are used to model bucket level variations over time and position, training the functional mapping between level, speed, and position.

[0047] The KL divergence is introduced in the loss function to measure the difference between the material level distribution predicted by the model and the actual observed material level distribution. This indicator reflects the degree of fit of the model to the actual charging trend and helps to adjust the model structure parameters and learning rate strategy.

[0048] Intelligent path adjustment and accumulation modeling stage:

[0049] Based on the trained model, the system analyzes the material level distribution and spatial uniformity within the loading bucket in real time during the loading process. If the pile shifts or the material level is abnormally high, the system controls the vehicle to make small movements (forward, backward, or lateral) to change the loading point.

[0050] The system also uses physical simulation methods (such as discrete element methods (DEM)) to model the material flow paths and accumulation trends caused by varying loading point positions. This model calculates the direction of the local and total force vectors generated at each loading point and compares them to the vehicle bed's geometric center. If the force direction approaches the centerline, the system determines that the loading is uniform; otherwise, fine-tuning the vehicle's trajectory is corrected.

[0051] In the process of material level detection, this embodiment effectively improves the quality assessment and dynamic compensation capabilities of sensor data by introducing information entropy and image entropy quantification mechanisms, and enhances the system's adaptability to dust interference and humidity changes in harsh environments; by using deep learning to construct a material level change function model, the prediction accuracy of charging trends for different vehicle models is significantly improved; combined with the combined force modeling of material accumulation, automatic identification and real-time correction of charging uniformity are achieved, effectively reducing safety hazards and material waste caused by unbalanced loading; the overall system can complete high-precision, intelligent charging monitoring and control without human intervention, and has high practicality and promotion value.

[0052] It also includes S5, ending loading when the initial height of the reset material level is close to the vehicle's allowable loading height threshold, and correcting the parameters of the loading function based on the loading result, locally approximating the function curve between the current material level and the height threshold using Newton interpolation, estimating when the threshold will be reached, and issuing an early warning. If the final result deviates from the prediction, the error quotient is updated using error feedback to correct the loading function model.

[0053] After completing the initial loading phase, the system continuously monitors the real-time material level and compares it with the set upper limit for safe loading of the vehicle. When the system detects that the current material level is approaching the preset threshold, it immediately triggers the loading termination prediction mechanism. The system calls the established loading function relationship (derived from a historical deep learning model or a real-time fitting function), selects the loading data points for the current time period, and uses Newton interpolation to construct a curve and estimate the slope in a local area of ​​the material level change function. It predicts the material level change trend in the short term in the future, and combines time calculation to estimate the time point when the threshold is reached, sending an early warning signal to the control system or operation interface.

[0054] If the actual loading endpoint level deviates from the Newton interpolation prediction, the system automatically calculates the error and uses it as feedback to update the interpolation spread quotient table, dynamically correcting the approximate structure of the local interpolation function. Simultaneously, the system backtracks and adjusts the parameters of the original loading function model based on the direction and magnitude of the error, improving prediction accuracy and response speed in similar loading environments in the future.

[0055] By introducing a local interpolation prediction and error feedback mechanism when the material level approaches the safe loading height threshold, it can not only effectively prevent overflow or safety hazards caused by overloading, but also improve the accuracy and response sensitivity of loading end control; in addition, the adaptive correction capability of the function model improves the system's learning efficiency and generalization ability in dynamic and complex environments, making the intelligent loading process more stable, efficient and controllable.

[0056] It also includes S6, adding the relationship between the material level change and the vehicle position during the loading process to the deep learning training set, further training and optimizing the loading speed and function relationship of the truck beds of different vehicle models, and mapping different vehicle model numbers, material level characteristics, and position combinations to different training channels through hash mapping.

[0057] The existing deep learning training system, built on the relationship between material level and loading speed, further incorporates vehicle position information during the loading process, such as longitude and latitude coordinates, steering angle, and the bucket's rotational position. These spatiotemporal position parameters are combined with real-time material level change data to form a model and added as new feature vectors to the training set. These composite features not only reflect the dynamic changes in the loading path but also reveal the loading efficiency characteristics of specific vehicle models in specific position structures.

[0058] To improve the parallelism and feature differentiation of the training process, the system combines the vehicle model number, material level characteristics, and vehicle location status into a hash mapping function. Using a distributed encoding strategy, these are mapped to multiple independent training channels or neural network branches, enabling separate training of loading behaviors for different vehicle models in a semantic space. Each channel can independently optimize its specific model parameters, resulting in a model with stronger recognition and generalization capabilities.

[0059] During training, a neural network structure based on an attention mechanism or a gating unit is used, which enables the model to automatically capture the key dependency between material level and position, and grasp the changing trend during the loading process through time series modeling, thereby realizing refined modeling of loading modes for different vehicles.

[0060] This implementation introduces the joint features of vehicle spatial position and material level changes, which not only improves the model's adaptability to complex loading scenarios, but also enhances its ability to characterize the loading efficiency of different vehicle models. Through hash mapping channel training, it significantly improves the modeling efficiency and generalization accuracy under multi-vehicle conditions, allowing the same system to maintain high performance when dealing with different types of vehicles and working conditions.

[0061] When training the material level change model for different vehicle models in S3, the KL divergence is used to measure the difference between the predicted distribution and the actual material level change distribution. The specific formula is:

[0062] ;

[0063] in, is the KL divergence, which indicates the degree of difference between the model prediction and the actual observation. The actual bucket material level change data is in the The probability of a state, The material level distribution predicted for the network is The probability of a state, It is the index of the material level distribution interval.

[0064] To achieve high-precision fitting of the material level prediction model, this implementation incorporates KL divergence as a loss function during the model training phase. After collecting extensive data on material level changes during the loading process for various vehicle models, the system first discretizes the continuous material level into several state intervals and calculates the frequency of occurrence in each state to form a true probability distribution. Simultaneously, the constructed neural network model takes the loading speed, time series, and vehicle parameters as input and outputs a predicted material level distribution. KL divergence measures the degree of deviation between the two and serves as the objective function during training.

[0065] Through the backpropagation mechanism, if the model output deviates significantly from the true distribution, the corresponding KL divergence value increases. The optimizer adjusts the network weight parameters accordingly, causing the predicted distribution to gradually approach the actual material level variation pattern. This process is repeated iteratively until the model converges, ensuring that the model accurately represents the dynamic characteristics of vehicle material level changes.

[0066] Introducing KL divergence as a loss function strengthens the model's ability to capture the true distribution characteristics of material levels. This allows the neural network to not only learn overall trends but also focus on micro-distribution characteristics, effectively improving the model's prediction accuracy and robustness. This allows for intelligent judgment and real-time feedback across different vehicle models and complex loading paths.

[0067] In S4, the material accumulation caused by the change of the loading position is modeled, and the specific formula for calculating the resultant force direction and resultant force result of each loading point is:

[0068] ;

[0069] in, represents the sum of the accumulation trend vectors of all charging points, Indicates the Material accumulation trend vector at each loading location, Indicates the index number of the loading position point, Indicates the total number of charging points.

[0070] During loading, dynamic adjustments to the vehicle's position (such as forward, backward, and turning) can cause spatial shifts in the loading point, altering the direction and distribution of material accumulation within the truck bed. To accurately determine the uniformity of the current loading state and guide subsequent loading strategies, the system models the accumulation trend of each loading point in three dimensions.

[0071] Accumulation trend vector at each charging point It can be calculated from the material flow direction at that location, the mass loaded per unit time, the stacking inclination and the inertial direction. It is usually obtained by modeling based on the center of mass offset or the discrete element method (DEM) simulation. On this basis, the trend vectors of all loading points are summed to obtain the total resultant force direction. .

[0072] like If it is consistent with the center axis of the truck bed, it means that the loading is uniform; if it deviates, the system will identify it as an overloaded state and control the vehicle's movement direction or loading arm angle accordingly to correct the loading trajectory and realize dynamic control of the material accumulation trend.

[0073] By introducing the loading trend resultant force vector model, the system can perceive the loading deviation trend in real time during the loading process, avoid the center of gravity deviation of the vehicle body caused by unilateral loading, and improve the vehicle driving stability; at the same time, it also improves the spatial utilization efficiency and dynamic control capabilities of loading, providing key decision-making basis for the unmanned intelligent loading process.

[0074] S2 also includes temporal consistency verification of the truck bed images of multiple consecutive time frames during the image recognition stage to eliminate the interference of instantaneous occlusion or abnormal lighting on the image entropy calculation.

[0075] In the material level detection system, the bucket image is an important basis for evaluating dust interference and distance compensation accuracy. To improve the accuracy of image entropy calculation, the system introduces a time sequence consistency check mechanism during the image recognition stage. The specific operation process is as follows:

[0076] A sequence of truck bed images is continuously acquired at a set frame rate (e.g., 5 frames per second). Image preprocessing modules (including grayscale conversion, normalization, and edge enhancement) ensure the consistency of the input images. The system then compares pixel regions at the same location across multiple time frames for features such as brightness, contrast, and texture boundaries. Dynamic time warping or sliding window correlation analysis is used to determine the consistency of visual features within the sequence.

[0077] If a frame exhibits significant deviations (e.g., a sudden increase in brightness or the sudden appearance of an obstructing object's outline), the system marks it as an abnormal frame and removes it from the image entropy calculation to prevent transient interference from misleading the dust weight correction coefficient. Furthermore, within each time window, only image frames that have passed the consistency check are retained for subsequent weighted averaging of image entropy, ensuring that the compensation mechanism is based on stable and reliable image features.

[0078] By introducing temporal consistency analysis of image frames, image fluctuations caused by non-structural factors such as instantaneous occlusion, strong light exposure, and dynamic interference can be effectively filtered out, thereby improving the robustness of image entropy in actual working conditions. This in turn makes the ranging radar compensation mechanism more accurate, avoids material level misjudgments caused by abnormal images, and improves the stability and environmental adaptability of the entire loading system.

[0079] S3 also includes the introduction of a model memory mechanism during deep learning training to identify and memorize recurring loading behavior patterns of specific vehicle models, reducing the amount of real-time calculations.

[0080] In practice, some vehicle models exhibit highly repetitive loading behavior patterns under similar operating conditions, such as stable loading speeds, consistent bucket structures, and similar force responses. To improve system processing efficiency, this implementation embeds a model memory mechanism within the deep learning training architecture. By identifying and storing the loading behavior characteristics of typical vehicle models, this enables rapid response and low-computational-load inference for repetitive scenarios.

[0081] During the implementation phase, the system first constructs a model memory unit (such as Key-ValueMemory) during the training phase. Key features such as vehicle number, loading path characteristics, and material level trends are used as search keys. The corresponding trained loading prediction model parameters or state vectors are cached as values ​​in the model memory. Subsequently, when the system identifies a newly entered vehicle whose basic parameters closely match a key in the known model (for example, if the Euclidean distance or cosine similarity is within a threshold), the full neural network forward propagation process is skipped, and the matching value is directly used for rapid inference and prediction.

[0082] Furthermore, to prevent memory redundancy and failure, a memory update strategy is employed: parameters of frequently used memory cells are dynamically refreshed, and cells that have been unused for a long time are recycled to ensure optimal allocation of memory resources. This mechanism is particularly suitable for fixed vehicle models or company-owned fleets that perform frequently repetitive tasks, improving inference efficiency and reducing hardware consumption.

[0083] By introducing a model memory mechanism, the system can identify and reuse historically high-frequency loading patterns, avoiding repeated model training and high-cost inference. This effectively reduces real-time computing load, improves model response speed and resource utilization efficiency, and offers significant advantages in edge computing environments. It also enhances the system's accuracy and stability when handling known vehicle models.

[0084] S4 also includes intelligent path filtering of the relative displacement between the current loading position and the previous loading point, retaining only the displacement directions that can bring material level optimization benefits.

[0085] During unmanned intelligent loading, adjustments to the vehicle or loading device position directly affect the material accumulation structure and level distribution in the bucket. To avoid ineffective movement or uneven loading due to improper path planning, this implementation introduces an "intelligent path filtering" mechanism during the S4 loading adjustment phase. This mechanism analyzes and filters the relative displacement directions between consecutive loading points, allowing only those paths that positively optimize the material level distribution to be executed.

[0086] First, after each loading action is completed, the system records the current loading position coordinates (such as the relative X and Y displacement values ​​relative to the center of the truck bed) and the corresponding material level changes (such as local accumulation height, slope distribution, and central symmetry indicators). Before proceeding to the next loading action, the system calculates the displacement vector between the current position and the previous position and establishes a dynamic prediction model based on the current material level distribution to assess whether continuing loading in this direction will effectively improve material level center deviation, reduce edge accumulation, and improve overall flatness.

[0087] If the prediction results indicate a direction that will optimize material level (e.g., the direction of the resultant force after simulated accumulation approaches the center of gravity of the bucket), the path is retained as a candidate. Otherwise, it is considered an inefficient path and filtered out to avoid wasting resources and worsening accumulation. Ultimately, based on a multi-objective optimization strategy, the system selects the optimal direction from several candidate paths to perform displacement adjustments and loading operations.

[0088] By introducing a path screening mechanism, the loading path is transformed from "blind traversal" to "target-driven", effectively improving loading efficiency and stacking uniformity, reducing ineffective actions and energy consumption during the loading process; at the same time, it improves the system's decision-making intelligence level and resource scheduling efficiency, which is especially suitable for large-scale mining trucks or engineering vehicles with high requirements for load balance.

[0089] In S6, different vehicle model numbers, material level characteristics, and location combinations are mapped to different training channels through hashing. The specific formula is:

[0090] ;

[0091] in, represents the input feature encoding, Indicates the total number of hash buckets, Indicates the training subset number to which the input sample is assigned, Represents the modulo operator.

[0092] When processing large-scale, multi-type vehicle material level data, directly feeding all samples into the same training network channel can easily lead to model overfitting, reduced generalization ability, and training resource conflicts. Therefore, this implementation introduces a hash mapping mechanism in the S6 stage to project multi-dimensional feature combinations into different training channels for distributed modeling.

[0093] The specific method is to encode the feature group of each input sample (such as vehicle model number, loading posture characteristics, real-time material level change) into an integer value , through the hash function Map it to numbers 0 to The total number of hash buckets It can be set according to the system hardware capabilities and data scale. For example, when deployed in a GPU cluster environment, it can be set to the number of parallel channels.

[0094] This hashing mechanism distributes samples with similar features to the same channel for local modeling. Different channels maintain their own model weights without interfering with each other, enabling highly parallel and highly discriminative feature learning. Furthermore, the uniformity of the hash function helps ensure a balanced distribution of training samples across subsets, preventing model failure due to sparse channel data.

[0095] Using hash maps to group training channels can significantly improve training efficiency and the scalability of the model structure, reduce resource conflicts and memory usage, and enhance the model's generalization and portability in multi-vehicle and multi-location material level scenarios, making it suitable for large-scale deployment and accelerated training on heterogeneous platforms.

[0096] like Figure 2 As shown, a material level detection system for unmanned intelligent loading is also provided, which is used to implement the steps of the material level detection method for unmanned intelligent loading. The system includes:

[0097] The information collection module is used to obtain weather, vehicle type, and truck bed size information. It uses sensors, manual registration, or ranging radar to collect data and determines whether to enable multi-sensor redundant input based on information entropy.

[0098] Image and radar fusion module, used to obtain truck bed images and identify dust concentration during the loading process, and dynamically compensate the ranging radar data by combining humidity and image entropy;

[0099] The model training module is used to build a deep learning model based on the material level changes during the loading process of different vehicles, and optimize the functional relationship between the material level and the loading speed and loading position through KL divergence;

[0100] The control decision module is used to determine the current material level status based on the model, control the vehicle position adjustment and update the loading speed, analyze the material flow uniformity according to the direction of the resultant force at each loading point, and dynamically correct the loading trajectory.

[0101] This system utilizes a modular structure to achieve a closed-loop process of multi-source heterogeneous information perception, fusion analysis, predictive modeling, and intelligent control. By guiding input optimization and compensation strategies using information entropy and image entropy, combined with deep model training driven by KL divergence, material level detection achieves high accuracy and robustness while also enabling real-time response to dynamic loading conditions. The control decision module, centered around a combined force model, analyzes material accumulation trends and adaptively adjusts the path to ensure safe, efficient, and uniform loading operations.

[0102] The system can realize full-process automation from data collection, intelligent analysis, model prediction to real-time decision-making, significantly improving the accuracy of material level detection and the intelligence level of loading operations; its modular design is suitable for distributed deployment, facilitating its promotion and application in large and complex operating environments such as mining areas and ports; the system has learning and optimization capabilities, and can continuously iterate to improve performance and meet the needs of long-term high-intensity unmanned operations.

[0103] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for material level detection in unmanned intelligent loading, characterized in that: include: S1. Obtain the current weather conditions, vehicle type, and truck bed size using sensors, manual registration, or ranging radar. Information entropy is used to measure the information redundancy and value density in different data sources, automatically determining whether multi-sensor redundant input is required. S2. Use a ranging radar to measure the material level. During the loading process, capture images of the truck bed and perform image recognition on the images. Analyze the current dust concentration based on the image recognition results. Determine the dynamic compensation value of the current ranging radar data based on the humidity and dust concentration. Measure the degree of dust interference in the truck bed image using image entropy. Use image entropy as a weighting factor to adjust the compensation ratio of the ranging radar. S3. Use the changes in bucket material level during the loading process of different vehicles as a dataset for deep learning, thereby establishing a functional relationship between the material level changes of different vehicles and the loading speed and loading position. When training the material level change model for different vehicle types, use the KL divergence to measure the difference between the predicted distribution and the actual material level change distribution as the loss function; S4. During the loading process, the vehicle is moved to change the loading position and displace the material level. The material level changes in different buckets are determined based on the trained functional relationship, and the loading speed at the next moment and whether to stop loading are determined accordingly. The material accumulation caused by the change in loading position is modeled, the direction and result of the resultant force at each loading point are calculated, and the material flow trend is analyzed. When the direction of the resultant force is aligned with the center of the bucket, the loading is judged to be uniform. Otherwise, the loading trajectory is corrected by adjusting the position and direction of the vehicle.

2. The method for material level detection in unmanned intelligent loading according to claim 1, characterized in that: It also includes S5, ending loading when the initial height of the reset material level is close to the vehicle's allowable loading height threshold, and correcting the parameters of the loading function based on the loading result, locally approximating the function curve between the current material level and the height threshold using Newton interpolation, estimating when the threshold will be reached, and issuing an early warning. If the final result deviates from the prediction, the error quotient is updated using error feedback to correct the loading function model.

3. The method for material level detection in unmanned intelligent loading according to claim 2, characterized in that: It also includes S6, adding the relationship between the material level change and the vehicle position during the loading process to the deep learning training set, further training and optimizing the loading speed and function relationship of the truck beds of different vehicle models, and mapping different vehicle model numbers, material level characteristics, and position combinations to different training channels through hash mapping.

4. The method for material level detection in unmanned intelligent loading according to claim 1 is characterized in that: In S3, when training the material level change model for different vehicle models, the KL divergence is used to measure the difference between the predicted distribution and the actual material level change distribution. The specific formula is: ； in, is the KL divergence, which indicates the degree of difference between the model prediction and the actual observation. The actual bucket material level change data is in the The probability of a state, The material level distribution predicted for the network is The probability of a state, It is the index of the material level distribution interval.

5. The method for material level detection in unmanned intelligent loading according to claim 1 is characterized in that: In S4, the material accumulation caused by the change in the charging position is modeled, and the specific formula for calculating the resultant force direction and resultant force result of each charging point is: ； in, represents the sum of the accumulation trend vectors of all charging points, Indicates the Material accumulation trend vector at each loading location, Indicates the index number of the loading position point, Indicates the total number of charging points.

6. The method for material level detection in unmanned intelligent loading according to claim 1 is characterized in that: The S2 also includes performing a time sequence consistency check on the truck bed images of multiple consecutive time frames in the image recognition stage to eliminate the interference of instantaneous occlusion or abnormal lighting on the image entropy calculation.

7. The method for material level detection in unmanned intelligent loading according to claim 1, characterized in that: The S3 also includes introducing a model memory mechanism during the deep learning training process to identify and memorize recurring loading behavior patterns of specific vehicle models, thereby reducing the amount of real-time calculations.

8. The method for material level detection in unmanned intelligent loading according to claim 1, characterized in that: The S4 also includes performing intelligent path filtering on the relative displacement between the current loading position and the previous loading point, and only retaining the displacement direction that can bring material level optimization benefits.

9. The method for material level detection in unmanned intelligent loading according to claim 1, characterized in that: In S6, different vehicle model numbers, material level characteristics, and location combinations are mapped to different training channels through hashing. The specific formula is: ； in, represents the input feature encoding, Indicates the total number of hash buckets, Indicates the training subset number to which the input sample is assigned, Represents the modulo operator.

10. A material level detection system for unmanned intelligent loading, used to implement the steps of the material level detection method for unmanned intelligent loading according to claims 1-9, characterized in that: The system comprises: The information collection module is used to obtain weather, vehicle type, and truck bed size information. It uses sensors, manual registration, or ranging radar to collect data and determines whether to enable multi-sensor redundant input based on information entropy. Image and radar fusion module, used to obtain truck bed images and identify dust concentration during the loading process, and dynamically compensate the ranging radar data by combining humidity and image entropy; The model training module is used to build a deep learning model based on the material level changes during the loading process of different vehicles, and optimize the functional relationship between the material level and the loading speed and loading position through KL divergence; The control decision module is used to determine the current material level status based on the model, control the vehicle position adjustment and update the loading speed, analyze the material flow uniformity according to the direction of the resultant force at each loading point, and dynamically correct the loading trajectory.

Images