Constricted space intelligent gangue selection robot heterogeneous sensor fusion detection method

Abstract

The invention relates to a fusion detection method for heterogeneous sensors of an intelligent gangue selection robot in a limited space, and belongs to the field of fusion of technical heterogeneous sensors. VOC time sequence signals of a bionic electronic nose array, laser radar cloud data and pose information of an inertial measurement unit are synchronously collected, and the VOC time sequence signals are denoised through a generative adversarial network model; voxel filtering and outlier elimination are carried out on the laser radar cloud data; the VOC concentration gradient is predicted based on CFD simulation, a three-dimensional environment map is constructed in combination with laser radar cloud data and inertial measurement unit poses, and a smell concentration thermodynamic diagram is marked; and extracting robust odor fingerprints from the denoised VOC time sequence signals, aligning feature distribution of different mining areas, and updating the classification model. According to the scheme, the gangue selection robot is guided to preferentially explore a hidden area with low VOC concentration and high CO2 accumulation through the odor concentration gradient, and the detection capacity of non-dominant gangue piles is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of heterogeneous sensor fusion, and in particular to a heterogeneous sensor fusion detection method for an intelligent waste sorting robot in a confined space. Background Art

[0002] Gangue sorting robots are designed to address the challenge of gangue removal during raw coal production and transportation, effectively achieving clean coal utilization and green production. Gangue sorting robots are primarily categorized as cable-driven, Cartesian, and parallel robots. Among parallel gangue sorting robots, the Delta robot offers advantages such as high structural rigidity and high speed, meeting the demands of gangue sorting.

[0003] The heterogeneous sensors of traditional gangue sorting robots rely on single modes such as vision (such as cameras) and weight detection. They are easily disturbed by complex mine environments such as dust and insufficient light, and cannot directly perceive the volatile organic compound (VOC) characteristics released by gangue. Their ability to identify low-concentration, hidden accumulations of gangue is limited. Traditional methods miss detection due to the lack of gas sensing capabilities, resulting in low sorting efficiency, which needs to be improved. Summary of the Invention

[0004] The purpose of the present invention is to provide a heterogeneous sensor fusion detection method for an intelligent waste sorting robot in a confined space to solve the problems raised in the above background technology.

[0005] The technical solution of the present invention is a heterogeneous sensor fusion detection method for an intelligent waste sorting robot in a confined space, comprising the following steps:

[0006] S1: Start the bionic electronic nose array, lidar, inertial measurement unit and tactile sensor, perform baseline drift correction on the bionic electronic nose array, and align the coordinate system of the lidar;

[0007] S2: Synchronously collects the VOC time series signals of the bionic electronic nose array, LiDAR cloud data, and the pose information of the inertial measurement unit. It uses a generative adversarial network model to denoise the VOC time series signals, performs voxel filtering and outlier removal on the LiDAR cloud data, and fuses the IMU and LiDAR cloud data to estimate the robot's pose.

[0008] S3: Predict VOC concentration gradients based on CFD simulation, build a 3D environmental map by combining lidar cloud data and robot pose information from an inertial measurement unit, and annotate the odor concentration heat map;

[0009] S4: Build a classification model, extract robust odor fingerprints from the denoised VOC time series signals, align the characteristic distributions of different mining areas, and update the classification model;

[0010] S5: Generate a gangue sorting task queue based on the odor concentration heat map, establish a deep reinforcement learning model to plan the robot's gripping path, and adjust the gripping arm's motion based on real-time feedback from the tactile sensor to update the odor concentration heat map in real time.

[0011] S6: Collect grasping results and feed them back to the classification model for incremental learning, and allocate unexplored areas by sharing multi-robot task progress;

[0012] S7: Stop the operation according to the preset termination conditions, which include that the probability of gangue in the odor concentration thermodynamic map is lower than 0.1 or the preset operation time threshold is reached;

[0013] S8: Store to the classification model; classification model parameters and sorting log, and generate a sorting efficiency analysis report.

[0014] As a further optimization of this technical solution, the baseline drift correction of the bionic electronic nose array in S1 is achieved by the following steps:

[0015] S1a: Collect the original signal of the bionic electronic nose array in a VOC-free environment and calculate the baseline offset;

[0016] S1b: A polynomial fitting algorithm is used to generate a compensation curve based on the original signal of the bionic electronic nose array and the calculated baseline offset, and the data output of the bionic electronic nose array is corrected in real time.

[0017] The S3 specifically includes the following steps:

[0018] S31: CFD simulation establishes a gas diffusion model based on predefined VOC release rates from waste rock and mine environmental parameters. CFD simulation calculates and predicts the model-predicted values ​​of VOC concentration distribution at different spatial locations. The VOC concentrations measured by the bionic electronic nose array are compared with the model-predicted values ​​to calibrate candidate waste rock areas. A pre-set odor fingerprint library stores the VOC composition and concentration characteristics of waste rock released from different mining areas.

[0019] S32: constructing a three-dimensional geometric map using the lidar cloud data and the pose information of the inertial measurement unit, and superimposing the concentration features as voxel attributes onto the three-dimensional geometric map to form a dynamic three-dimensional map integrating the odor concentration;

[0020] S33: A particle filter algorithm is used to update the probability distribution of waste rock in the dynamic three-dimensional map in real time, and an A-Star algorithm is used to plan the robot navigation path, with priority given to exploring areas with low VOC concentration and high CO2 accumulation.

[0021] The conditions for calibration of the candidate area of ​​gangue in S31 are:

[0022] When the VOC concentration measured by the bionic electronic nose array is lower than the preset low threshold and the CO2 concentration in the candidate waste rock area is higher than the preset high threshold, it is determined to be a candidate waste rock area;

[0023] When the deviation between the VOC concentration measured by the bionic electronic nose array and the model prediction value exceeds 20%, the S4 is triggered and is specifically implemented by the following steps:

[0024] S41: Construct a synthetic dataset containing dust and humidity noise, and use an adversarial network model to generate denoised VOC time series signals;

[0025] S42: An LSTM-Transformer hybrid model is used to extract odor fingerprints from the VOC time series signals collected by the bionic electronic nose array. The time series features output by the LSTM-Transformer hybrid model are dynamically weighted through the attention mechanism to generate an odor fingerprint vector, which is finally mapped to the voxel-level probability value of the gangue candidate area.

[0026] The S5 is specifically implemented by the following steps:

[0027] S51: Based on the probability value of the gangue candidate area and the distance to the target area, a sorting priority queue is generated through the Q-learning algorithm, where the area with a probability value higher than a preset threshold is marked as a high priority;

[0028] S52: Generate optimized paths for robot arm joint angles and end-point velocities using a deep reinforcement learning model, and perform collision detection based on LiDAR cloud data;

[0029] S53: When the tactile sensor detects abnormal pressure distribution, the robot arm joint path is triggered to replan and the odor concentration heat map is updated.

[0030] The S6 also has the following steps:

[0031] S61: If the robotic arm fails to grasp continuously for more than a preset number of times, step S2 is triggered to rescan the target area;

[0032] S62: Allocate multi-robot collaborative tasks through the ant colony optimization algorithm to cover unexplored areas and reduce duplication of work.

[0033] The preset operation time threshold in S7 is dynamically calculated based on the area of ​​the operation area.

[0034] The S7 also includes equipment maintenance steps, specifically including:

[0035] S71: Clean the filter of the bionic electronic nose array and remove dust attachment using compressed air;

[0036] S72: Scan the joint gap of the robotic arm through the laser radar to detect whether the wear exceeds the preset threshold;

[0037] S73: If the wear of the end effector of the robot arm is detected to be ≥0.5mm, a warning signal is triggered and a maintenance log is recorded.

[0038] The present invention provides a heterogeneous sensor fusion detection method for an intelligent waste sorting robot in a confined space through improvement. Compared with the prior art, the present invention has the following improvements and advantages:

[0039] 1. This invention uses CFD simulation to predict VOC concentration gradients, combines measured data with real-time corrections to the gas diffusion model, and annotates the probability distribution of waste rock. Compared to traditional approaches that rely solely on visual features (such as color and shape), this approach uses odor concentration gradients to guide waste rock sorting robots to prioritize hidden areas with low VOC concentrations but high CO2 accumulation, significantly improving their detection capabilities for non-visible waste rock piles.

[0040] 2. Using an LSTM-Transformer hybrid model to extract odor fingerprints: LSTM captures temporal dependencies, while Transformer extracts global features through a self-attention mechanism to generate a 12-dimensional robust odor fingerprint vector. Compared with traditional frequency domain analysis methods such as Fourier transform, this model can more accurately characterize dynamically changing VOC signal characteristics.

[0041] 3. Targeting low-VOC concentration environments in mines, the system improves low-concentration signal detection accuracy through baseline drift correction of the bionic electronic nose array. Combined with a particle filter algorithm, it updates the probability distribution of waste rock in a three-dimensional map in real time, achieving millimeter-level positioning in confined areas. The operating time threshold is dynamically calculated based on the area, and integrated equipment maintenance mechanisms (such as cleaning the electronic nose filter and detecting wear on the robotic arm) ensure long-term stable operation in dusty and humid environments.

[0042] 4. This invention is based on multi-physics field coupling modeling, integrating the gas diffusion model and the robot kinematic model. It introduces the VOC concentration gradient factor through the A-Star algorithm cost function, and prioritizes exploring the boundary areas of waste rock where the odor concentration changes significantly. Compared with traditional path planning that only considers distance or energy consumption, it significantly improves the detection efficiency in complex terrain. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The present invention will be further explained below in conjunction with the accompanying drawings and Examples:

[0044] Figure 1 It is a flow chart of the heterogeneous sensor fusion detection method of the confined space intelligent waste sorting robot of the present invention. DETAILED DESCRIPTION

[0045] The following will be combined with the Figure 1The present invention is described in detail, and the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0046] The present invention constructs an intelligent gangue sorting system suitable for complex confined spaces. The core solution is to achieve high-precision real-time detection, positioning and efficient sorting of gangue through multi-sensor fusion and intelligent algorithms in a strong noise and dynamically changing mine environment.

[0047] Example 1:

[0048] The present invention provides a heterogeneous sensor fusion detection method for an intelligent waste sorting robot in a confined space by improving Figure 1 As shown, the following steps are included:

[0049] S1: Start the bionic electronic nose array, lidar, inertial measurement unit and tactile sensor, perform baseline drift correction on the bionic electronic nose array, and align the coordinate system of the lidar; the tactile sensor enables the robot to achieve reliable operation in extreme environments such as mine dust and insufficient light through accurate feedback of tactile data. Tactile sensors are existing technology and will not be described in detail here.

[0050] S2: Synchronously collects the VOC time series signals of the bionic electronic nose array, LiDAR cloud data, and the pose information of the inertial measurement unit. It uses a generative adversarial network model to denoise the VOC time series signals, performs voxel filtering and outlier removal on the LiDAR cloud data, and fuses the IMU and LiDAR cloud data to estimate the robot's pose.

[0051] The Generative Adversarial Network (GAN) model consists of a generator and a discriminator. The generator receives a noisy VOC time series signal and outputs a denoised signal; the discriminator determines whether the signal is truly noise-free. Through adversarial training, the generator gradually improves its denoising capabilities, ultimately outputting a VOC signal with a signal-to-noise ratio of 30dB. The inertial measurement unit provides high-frequency but drift-prone pose data, while the lidar provides high-precision but low-frequency spatial information. The Kalman filter fuses these two data sets to estimate the robot's pose in real time.

[0052] S3: Predict VOC concentration gradients based on CFD simulation, build a 3D environmental map by combining lidar cloud data and robot pose information from an inertial measurement unit, and annotate the odor concentration heat map;

[0053] S4: Build a classification model to extract robust odor fingerprints from the denoised VOC time series signal, align the feature distributions of different mining areas, and update the classification model. Align the feature distributions of different mining areas through domain adversarial neural networks and update the classification model. Domain adversarial neural networks are used in transfer learning to align the feature distributions of different mining areas (source domain and target domain) and reduce domain differences.

[0054] By inputting mine environmental parameters (wind speed, temperature, and VOC release rate of waste rock), the VOC concentration gradient distribution is predicted and the model is calibrated by comparing with the measured data.

[0055] S5: Generate a gangue sorting task queue based on the odor concentration heat map, establish a deep reinforcement learning model to plan the robot's gripping path, and adjust the gripping arm's motion based on real-time feedback from the tactile sensor to update the odor concentration heat map in real time.

[0056] S6: Collect grasping results and feed them back to the classification model for incremental learning, and allocate unexplored areas by sharing multi-robot task progress;

[0057] The A-Star algorithm is a heuristic path search algorithm used for robot navigation path planning. It introduces an odor concentration gradient factor based on the traditional A-Star algorithm to balance path distance and waste rock detection efficiency, giving priority to areas with significant VOC concentration gradients (such as waste rock boundaries) to improve exploration efficiency. The cost function is:

[0058] g(n): actual cost (path length, energy consumption)

[0059] h(n): heuristic cost (Euclidean distance to the goal);

[0060] VOC concentration gradient (guiding the robot to preferentially explore stone boundaries).

[0061] β: Weight coefficient (balance between path efficiency and detection sensitivity).

[0062] Odor gradient factor: Add the VOC concentration change rate to g(n) and give priority to exploring areas with significant concentration gradients (such as gangue boundaries).

[0063] The technical effect is that hidden areas can be explored first: the waste rock boundary can be identified by concentration gradient. Even when the VOC concentration is below the preset threshold, the waste rock pile can still be located by the gradient change (for example, the gradient of -0.1ppm / m in the 0.2ppm concentration area indicates the existence of a boundary);

[0064] It can also improve detection efficiency: compared with traditional path planning that only considers distance, detection efficiency in complex terrain is improved by more than 30%.

[0065] The Q-value update formula of Q-learning is used to generate a priority queue for sorting tasks, optimize the order of gangue sorting through reinforcement learning, balance the sorting success rate with mobile energy consumption, and dynamically adjust task priorities (regions with probability values ​​greater than 0.8 are given priority).

[0066] Q value update:

[0067]

[0068] Q(s,a): The Q-value (expected cumulative reward) of performing action a in state s.

[0069] s: state, defined as the probability value of gangue in the odor map and the distance between the target and the robot.

[0070] a: Action, that is, selecting the next gangue area to be sorted.

[0071] α: Learning rate, which controls the update rate of the old Q value with new information.

[0072] R: Reward function, the formula is sorting success rate moving energy consumption R = sorting success rate - λ·

[0073] Mobile energy consumption (λ is the energy consumption weight coefficient; the document does not specify the value, but it reflects the goal of "minimizing energy consumption").

[0074] γ: Discount factor used to balance the importance of current rewards and future rewards (0≤γ≤1).

[0075] s': The next state after executing action a.

[0076] a': all possible actions in the next state s', Indicates the maximum Q value of the next state.

[0077] Task queue generation: Dynamically adjust the priority based on the Q value, give priority to areas with probability value > 0.8, and minimize mobile energy consumption.

[0078] The LSTM-Transformer hybrid model refers to: LSTM: It is good at capturing temporal dependencies and is used to extract local temporal features of VOC signals.

[0079] The practical significance lies in that LSTM captures the temporal dependency of VOC signals, and Transformer extracts global features through the self-attention mechanism to generate a 12-dimensional odor fingerprint vector.

[0080] S7: Stop the operation according to the preset termination conditions, which include that the probability of gangue in the odor concentration thermodynamic map is lower than 0.1 or the preset operation time threshold is reached;

[0081] S8: Store to the classification model and generate a sorting efficiency analysis report.

[0082] In this embodiment, after the robot is started, it will first perform system initialization and sensor calibration, turn on the bionic electronic nose array, lidar, inertial measurement unit, tactile sensor and robotic arm control system, correct the baseline drift of the bionic electronic nose array, align the lidar coordinate system, load the confined space map framework and set the initial parameters of the gas diffusion model, establish low-latency communication between the edge computing unit and the cloud and robotic arm controller, and output the calibrated bionic electronic nose array, lidar, inertial measurement unit, tactile sensor data stream and initial environmental parameter configuration.

[0083] Entering the stage of real-time multimodal data acquisition and preprocessing, the bionic electronic nose array collects VOC concentration time series signals at a frequency of ≥100Hz, the lidar generates high-precision point cloud data at 10Hz, and the inertial measurement unit provides robot posture information at 100Hz. These data are then preprocessed, and the odor signal is denoised using a pre-trained GAN model. The lidar cloud data is optimized through voxel filtering and outlier removal. The inertial measurement unit and lidar cloud data are then fused and the robot's precise posture is estimated through Kalman filtering. The denoised VOC signal, optimized point cloud, and high-precision posture information are output.

[0084] Dynamic odor SLAM and map construction, based on CFD simulation and real-time environmental parameters, predicts the VOC concentration gradient distribution, compares the predicted results with the measured data of the bionic electronic nose array to dynamically correct the model parameters, and then inputs the lidar cloud data and the inertial measurement unit posture into ORB-SLAM3 to build a three-dimensional environmental map, and annotates the odor concentration heat map in the map, uses the particle filter algorithm to update the odor map in real time, combines Gaussian process regression to optimize hotspot positioning, marks explored and unexplored areas, and uses the A-Star algorithm to plan the robot navigation path, outputting a dynamic three-dimensional map and robot navigation path that integrates odor concentration. In this process, environmental parameters are collected and preprocessed in real time, temperature and humidity sensors, and barometers are deployed to collect environmental parameters, and the bionic electronic nose array collects VOC raw signals at 10Hz and synchronously records the lidar cloud data timestamp.

[0085] The particle filter algorithm is used to update the probability distribution of gangue in real time. The particle weight calculation is based on the following formula:

[0086]

[0087] in, The observation likelihood function dynamically adjusts weights based on the deviation between measured VOC concentrations and CFD-predicted values. Particle clusters with weights greater than 0.8 are marked as high priority. A particle filter algorithm fuses multi-source data (VOC concentration and lidar reflectivity) to improve the accuracy of waste rock positioning, surpassing the error of traditional visual positioning. When changes in the mine ventilation system cause altered VOC diffusion patterns, the model automatically corrects the path to avoid missed detections caused by static planning.

[0088] Considering the low concentration of VOC released by gangue, the bionic electronic nose array was switched to high-gain mode to improve low-concentration detection accuracy. CFD gas diffusion modeling was performed based on the characteristics of gangue. Environmental parameters, gangue VOC release rate, and prior knowledge of gangue distribution were input. The k-ε turbulence model was used to simulate gas diffusion and mark possible gangue areas with low VOC concentration and high CO2 concentration.

[0089] By calibrating the measured data with the model's waste rock characteristics, a model correction is triggered when the measured concentration of the bionic electronic nose array deviates by more than 20% from the CFD predicted value. Areas with low VOC concentrations but high CO2 concentrations are calibrated as candidate waste rock piles. The waste rock VOC release rate is corrected and the waste rock odor fingerprint library is updated. Multimodal SLAM fusion and semantic annotation are performed. A 3D map with an accuracy of ±5cm is constructed using the LOAM algorithm and ORB-SLAM3. Coal and waste rock piles are marked, and the standardized VOC concentration is superimposed on the map as a voxel attribute to mark the coal and waste rock candidate areas. A particle filter based on the waste rock distribution is updated to evenly distribute the particle set in the unexplored area.

[0090] Particles are weighted based on the probability of gangue predicted by CFD. If the measured VOC concentration at the particle location is less than 0.3 and the lidar detects a low-reflectivity surface, the weight is multiplied by 2. Particles with a weight less than 0.1 are excluded. The exploration focuses on areas with high gangue probability, such as roadway corners, for dynamic path planning and hotspot tracking. An odor concentration gradient factor is added to the A-Star algorithm cost function of the A-Star algorithm planning robot, prioritizing the path with the fastest VOC concentration decrease. Obstacle avoidance constraints are set, and a dynamic 3D map integrating odor concentration and a robot navigation path containing waypoints is output.

[0091] Next, anti-interference feature extraction and transfer learning are performed. The denoised VOC signal is input into an LSTM-Transformer hybrid model. A robust odor fingerprint is extracted through an attention mechanism. A domain adversarial neural network is used to align the feature distributions of the new mining area with those of the source mining area. Under the federated learning framework, parameters from each mining area are uploaded to the cloud to aggregate the global model. Based on the transfer learning model, the waste rock probability is output and the odor map classification labels are updated. The waste rock probability distribution map and the optimized classification model are output.

[0092] Based on the probability and distance of gangue in the odor map, the sorting task queue is generated in combination with the Q-learning algorithm. The target position is input into the deep reinforcement learning (DDPG) model to generate the optimal grasping path. The path is previewed and obstacles are avoided. When the robotic arm grasps, the pressure distribution is fed back in real time through the tactile sensor. If abnormal pressure is detected, the path replanning is triggered, and the robotic arm grasping path and real-time tactile feedback signal are output.

[0093] The system then enters the closed-loop execution and dynamic optimization phase. The robotic arm grasps the target waste rock according to the planned path and places it in the designated area. The odor map is updated in real time, and the collected grasping results are fed back to the transfer learning model to incrementally update the classification weights. If three consecutive grasping failures occur, the SLAM module is triggered to rescan the area. The updated odor map and task progress are shared via blockchain. A swarm intelligence algorithm is used to assign unexplored areas to idle robots, and the sorted waste rock pile and the optimized global map and model are output.

[0094] When the probability of gangue in the odor map is less than 0.1 or the preset operation time threshold is reached, the task is terminated, the complete SLAM map, classification model parameters and sorting log are stored, and a sorting efficiency report is generated. It is also necessary to clean the sensor surface and check the wear status of the robotic arm, and output the archive data package and sorting efficiency analysis report.

[0095] Example 2: This example is an explanation of Example 1. Specifically,

[0096] The baseline drift correction of the bionic electronic nose array in S1 is achieved by the following steps:

[0097] S1a: Collect the original signal of the bionic electronic nose array in a VOC-free environment and calculate the baseline offset;

[0098] S1b: A polynomial fitting algorithm is used to generate a compensation curve based on the original signal of the bionic electronic nose array and the calculated baseline offset, and the data output of the bionic electronic nose array is corrected in real time.

[0099] As a further optimization of this technical solution, S3 specifically includes the following steps:

[0100] S31: CFD simulation establishes a gas diffusion model based on predefined VOC release rates from waste rock and mine environmental parameters. CFD simulation calculates and predicts the model-predicted values ​​of VOC concentration distribution at different spatial locations. The VOC concentrations measured by the bionic electronic nose array are compared with the model-predicted values ​​to calibrate candidate waste rock areas. A pre-set odor fingerprint library stores the VOC composition and concentration characteristics of waste rock released from different mining areas.

[0101] S32: constructing a three-dimensional geometric map using the lidar cloud data and the pose information of the inertial measurement unit, and superimposing the concentration features as voxel attributes onto the three-dimensional geometric map to form a dynamic three-dimensional map integrating the odor concentration;

[0102] S33: A particle filter algorithm is used to update the probability distribution of waste rock in the dynamic three-dimensional map in real time, and an A-Star algorithm is used to plan the robot navigation path, with priority given to exploring areas with low VOC concentration and high CO2 accumulation.

[0103] Particle filter algorithm formula:

[0104]

[0105] Parameter explanation: The state of the i-th particle (position + gangue probability); w: particle weight; t:

[0106] Time step.

[0107] Observation likelihood (based on VOC concentration and lidar reflectivity).

[0108] It is used to update the probability distribution of gangue in the 3D map in real time. The areas with particle weight > 0.1 are marked as high-probability candidate areas, with a positioning accuracy of ±0.5m.

[0109] Polynomial fitting is a mathematical method that approximates data points using a polynomial function. Its goal is to find a set of polynomial coefficients that minimizes the error relative to the given data points (usually using the least squares method). For baseline drift correction of the bionic electronic nose array, polynomial fitting is used to model the baseline offset of the sensor in a VOC-free environment and generate a compensation curve to eliminate drift noise.

[0110] In Example 2, a cubic polynomial is used for dynamic compensation:

[0111] y=a0+a1x+a2x 2 +a3x 3

[0112] Parameter explanation:

[0113] y: sensor output value after compensation.

[0114] x: time or sensor raw output sequence.

[0115] a0,a1,a2,a3: Polynomial coefficients, determined by fitting the baseline shift data.

[0116] The practical significance lies in the fact that the cubic polynomial can flexibly fit the nonlinear trend of the sensor baseline drift

[0117] (such as slow exponential decay or periodic fluctuations), while avoiding the overfitting problem that may be caused by high-order polynomials.

[0118] In Example 2, the system uses a cubic polynomial fitting algorithm to dynamically compensate for baseline drift in the bionic electronic nose array. Specifically, in a closed, VOC-free environment, the bionic electronic nose array continuously samples raw signals at a 10Hz sampling frequency for 10 minutes. The standard deviation of the baseline drift (typical value ≤ 0.8mV) is calculated to determine the cubic polynomial coefficients. This compensation curve is then generated to correct the bionic electronic nose array output in real time.

[0119] Example 3, this example is an explanation of Example 2:

[0120] The conditions for calibration of the candidate area of ​​gangue in S31 are:

[0121] When the VOC concentration measured by the bionic electronic nose array is lower than the preset low threshold and the CO2 concentration in the candidate waste rock area is higher than the preset high threshold, it is determined to be a candidate waste rock area;

[0122] When the deviation between the VOC concentration measured by the bionic electronic nose array and the model prediction value exceeds 20%, the dynamic correction of the gangue VOC release rate and the odor fingerprint library is triggered.

[0123] The conditions for calibration of the candidate area for waste rock in Example 3 are achieved through multi-sensor data fusion. The preset low threshold value of VOC concentration is 0.3ppm (based on the typical release characteristics of waste rock), and the high threshold value of CO2 concentration is 1000ppm (reflecting the oxygen-deficient environment in the waste rock accumulation area). When the VOC concentration measured by the bionic electronic nose array is lower than 0.3ppm and the CO2 concentration is higher than 1000ppm, it is determined to be a candidate area for waste rock; if the deviation between the measured VOC concentration and the predicted value of the CFD model exceeds 20% (such as the predicted value 1.0ppm, the measured value ≤0.8ppm or ≥1.2ppm), the dynamic correction mechanism is triggered. The correction process includes adjusting the model parameters of the waste rock VOC release rate (such as correcting the release coefficient from 0.05g / s to 0.04g / s) and updating the odor fingerprint library.

[0124] Example 4: This example is an explanation of Example 1. S4 is specifically implemented by the following steps:

[0125] S41: Construct a synthetic dataset containing dust and humidity noise, and use an adversarial network model to generate denoised VOC time series signals;

[0126] S42: An LSTM-Transformer hybrid model is used to extract odor fingerprints from the VOC time series signals collected by the bionic electronic nose array. The time series features output by the LSTM-Transformer hybrid model are dynamically weighted through the attention mechanism to generate an odor fingerprint vector, which is finally mapped to the voxel-level probability value of the gangue candidate area.

[0127] Example 5: This example is an explanation of Example 1. S5 is specifically implemented by the following steps:

[0128] S51: Based on the probability value of the gangue candidate area and the distance to the target area, a sorting priority queue is generated through the Q-learning algorithm, where the area with a probability value higher than a preset threshold is marked as a high priority;

[0129] S52: Generate optimized paths for robot arm joint angles and end-point velocities using a deep reinforcement learning model, and perform collision detection based on LiDAR cloud data;

[0130] S53: When the tactile sensor detects abnormal pressure distribution, the robot arm joint path is triggered to replan and the odor concentration heat map is updated.

[0131] During the robotic arm's grasping process, the tactile sensor continuously monitors the pressure distribution. If abnormal pressure is detected (such as local pressure is too high / too low, or pressure distribution is uneven), it indicates that the grasping posture or force is inappropriate, triggering the path replanning algorithm to adjust the robotic arm's joint angle and terminal speed in real time to ensure stable grasping.

[0132] The S6 also has the following steps:

[0133] S61: If the robotic arm fails to grasp continuously for more than a preset number of times, step S2 is triggered to rescan the target area;

[0134] S62: Allocate multi-robot collaborative tasks through the ant colony optimization algorithm to cover unexplored areas and reduce duplication of work.

[0135] Example 6: This example is an explanation of Example 1. The preset operation time threshold in S7 is dynamically calculated based on the area of ​​the operation area.

[0136] As one of the criteria for grasping failure: If the robotic arm fails to grasp continuously for more than a preset number of times, combined with the pressure data of the tactile sensor, the dynamic odor SLAM module is triggered to rescan the target area and correct the gangue positioning information

[0137] The S7 also includes equipment maintenance steps, specifically including:

[0138] S71: Clean the filter of the bionic electronic nose array and remove dust attachment using compressed air;

[0139] S72: Scan the joint gap of the robotic arm through the laser radar to detect whether the wear exceeds the preset threshold;

[0140] S73: If the wear of the end effector of the robot arm is detected to be ≥0.5mm, a warning signal is triggered and a maintenance log is recorded.

[0141] The high-speed impact of compressed air is used to blow away dust particles attached to the filter, restoring the filter's air permeability and sensor sensitivity. The compressed air nozzle is positioned on the robot, with the jet direction of the compressed air nozzle directed toward the filter of the bionic electronic nose array, and the pressure is set to 0.5 MPa.

[0142] It should be noted that all calculation formulas in this application document utilize, including but not limited to, regression analysis within machine learning algorithms to deeply analyze the collected parameters and identify their natural trends and interrelationships. Professional software, such as Python's Scikit-learn library or the R language, is used to automatically generate mathematical models that match the data. Model performance is then objectively evaluated through methods such as cross-validation, combined with continuous feedback and optimization to ensure that the created formulas truly reflect the inherent laws of the data, thereby guaranteeing their validity and accuracy, and ensuring that the calculation process complies with the constraints of natural laws rather than being based on artificially set rules.

[0143] The technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk or optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods of various embodiments of the present invention.

[0144] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0145] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

[0146] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A heterogeneous sensor fusion detection method for an intelligent waste sorting robot in a confined space, characterized by: The following steps are involved: S1: Start the bionic electronic nose array, lidar, inertial measurement unit and tactile sensor, perform baseline drift correction on the bionic electronic nose array, and align the coordinate system of the lidar; S2: Synchronously collects the VOC time series signals of the bionic electronic nose array, LiDAR cloud data, and the pose information of the inertial measurement unit. It uses a generative adversarial network model to denoise the VOC time series signals, performs voxel filtering and outlier removal on the LiDAR cloud data, and fuses the IMU and LiDAR cloud data to estimate the robot's pose. S3: Predict VOC concentration gradients based on CFD simulation, build a 3D environmental map by combining lidar cloud data and robot pose information from an inertial measurement unit, and annotate the odor concentration heat map; S4: Build a classification model, extract robust odor fingerprints from the denoised VOC time series signals, align the characteristic distributions of different mining areas, and update the classification model; S5: Generate a gangue sorting task queue based on the odor concentration heat map, establish a deep reinforcement learning model to plan the robot's gripping path, and adjust the gripping arm's motion based on real-time feedback from the tactile sensor to update the odor concentration heat map in real time. S6: Collect grasping results and feed them back to the classification model for incremental learning, and allocate unexplored areas by sharing multi-robot task progress; S7: Stop the operation according to the preset termination conditions, which include that the probability of gangue in the odor concentration thermodynamic map is lower than 0.1 or the preset operation time threshold is reached; S8: Store to the classification model.

2. The heterogeneous sensor fusion detection method for an intelligent waste sorting robot in a confined space according to claim 1 is characterized in that: The baseline drift correction of the bionic electronic nose array in S1 is achieved by the following steps: S1a: Collect the original signal of the bionic electronic nose array in a VOC-free environment and calculate the baseline offset; S1b: A polynomial fitting algorithm is used to generate a compensation curve based on the original signal of the bionic electronic nose array and the calculated baseline offset, and the data output of the bionic electronic nose array is corrected in real time.

3. The heterogeneous sensor fusion detection method for an intelligent waste sorting robot in a confined space according to claim 1 is characterized in that: The S3 specifically includes the following steps: S31: CFD simulation establishes a gas diffusion model based on predefined VOC release rates from waste rock and mine environmental parameters. CFD simulation calculates and predicts the model-predicted values ​​of VOC concentration distribution at different spatial locations. The VOC concentrations measured by the bionic electronic nose array are compared with the model-predicted values ​​to calibrate candidate waste rock areas. A pre-set odor fingerprint library stores the VOC composition and concentration characteristics of waste rock released from different mining areas. S32: constructing a three-dimensional geometric map using the lidar cloud data and the pose information of the inertial measurement unit, and superimposing the concentration features as voxel attributes onto the three-dimensional geometric map to form a dynamic three-dimensional map integrating the odor concentration; S33: A particle filter algorithm is used to update the probability distribution of waste rock in the dynamic three-dimensional map in real time, and an A-Star algorithm is used to plan the robot navigation path, with priority given to exploring areas with low VOC concentration and high CO2 accumulation.

4. The heterogeneous sensor fusion detection method for an intelligent waste sorting robot in a confined space according to claim 3 is characterized in that: The conditions for calibration of the candidate area of ​​gangue in S31 are: When the VOC concentration measured by the bionic electronic nose array is lower than the preset low threshold and the CO2 concentration in the candidate waste rock area is higher than the preset high threshold, it is determined to be a candidate waste rock area; When the deviation between the VOC concentration measured by the bionic electronic nose array and the model prediction value exceeds 20%, the dynamic correction of the gangue VOC release rate and the odor fingerprint library is triggered.

5. The heterogeneous sensor fusion detection method for an intelligent waste sorting robot in a confined space according to claim 1 is characterized in that: The S4 is specifically implemented by the following steps: S41: Construct a synthetic dataset containing dust and humidity noise, and use an adversarial network model to generate denoised VOC time series signals; S42: An LSTM-Transformer hybrid model is used to extract odor fingerprints from the VOC time series signals collected by the bionic electronic nose array. The time series features output by the LSTM-Transformer hybrid model are dynamically weighted through the attention mechanism to generate an odor fingerprint vector, which is finally mapped to the voxel-level probability value of the gangue candidate area.

6. The heterogeneous sensor fusion detection method for an intelligent waste sorting robot in a confined space according to claim 1 is characterized in that: The S5 is specifically implemented by the following steps: S51: Based on the probability value of the gangue candidate area and the distance to the target area, a sorting priority queue is generated through the Q-learning algorithm, where the area with a probability value higher than a preset threshold is marked as a high priority; S52: Generate optimized paths for robot arm joint angles and end-point velocities using a deep reinforcement learning model, and perform collision detection based on LiDAR cloud data; S53: When the tactile sensor detects abnormal pressure distribution, the robot arm joint path is triggered to replan and the odor concentration heat map is updated.

7. The heterogeneous sensor fusion detection method for an intelligent waste sorting robot in a confined space according to claim 1 is characterized in that: The S6 also has the following steps: S61: If the robotic arm fails to grasp continuously for more than a preset number of times, step S2 is triggered to rescan the target area; S62: Allocate multi-robot collaborative tasks through the ant colony optimization algorithm to cover unexplored areas and reduce duplication of work.

8. The heterogeneous sensor fusion detection method for an intelligent waste sorting robot in a confined space according to claim 1 is characterized in that: The preset operation time threshold in S7 is dynamically calculated based on the area of ​​the operation area.

9. The heterogeneous sensor fusion detection method for an intelligent waste sorting robot in a confined space according to claim 1 is characterized in that: The S7 also includes equipment maintenance steps, specifically including: S71: Clean the filter of the bionic electronic nose array and remove dust attachment using compressed air; S72: Scan the joint gap of the robotic arm through the laser radar to detect whether the wear exceeds the preset threshold; S73: If the wear of the end effector of the robot arm is detected to be ≥0.5mm, a warning signal is triggered and a maintenance log is recorded.

Images