A multi-sensor fusion adaptive detection method for stone picking machine and intelligent stone picking machine

Abstract

The present invention belongs to the field of agricultural machinery, and relates to a multi-sensor fusion adaptive detection method for a stone picker and an intelligent stone picker. The detection data of several sensors on the stone picker are collected, and a pre-trained transfer learning model in different environments is constructed. The transfer learning model is trained based on multi-task learning and adaptive gradient pruning, and the hyperparameters of the transfer learning model are adjusted through Bayesian optimization to obtain the optimal object recognition model under the current soil environment; based on a heuristic search algorithm, the parameter configuration of the sensor and the optimal object recognition model is automatically adjusted in combination with detection accuracy and energy consumption indicators; previous detection data are collected, and a clustering algorithm is used to determine whether there is a new pattern in the object detection result. If so, the optimal strategy of the object recognition model is generated through deep reinforcement learning and an environmental reward mechanism. Accurate detection, classification and efficient cleaning of stones and construction waste in reclaimed land and agricultural land are achieved, and manual labor is reduced.

Description

Technical Field

[0001] The invention relates to a multi-sensor fusion adaptive detection method and an intelligent stone picker suitable for a stone picker in reclaimed land and agricultural land, belonging to the technical field of agricultural machinery and equipment. Background Art

[0002] In the process of development and use of reclaimed land and agricultural land, a large amount of construction waste and stones are often encountered. These wastes not only seriously affect the soil quality, but also hinder the planting and growth of crops. The existing technology mainly relies on manual picking or ordinary stone picking machines for processing. A stone picker is a mechanical device used to clean up stones in places such as farmland and construction sites. It mainly separates the stones by screening the stones in the soil to facilitate subsequent agricultural or construction engineering operations. It is usually driven by a power transmission device, and the soil and stones are fed into the machine together through a drum screen. The stones are separated and collected in a specific container, while the soil continues to be used for agriculture or other purposes. However, it has problems such as low efficiency, high labor intensity, and incomplete treatment. Ordinary stone pickers are usually unable to accurately identify and separate stones and construction waste of different sizes, which can easily damage the soil and further reduce the utilization value of the land. Therefore, there is an urgent need for a stone picking device that can be efficient, accurate, and environmentally friendly to solve the shortcomings of the existing technology. Summary of the invention

[0003] In view of the above problems, the purpose of the present invention is to provide a multi-sensor fusion adaptive detection method and an intelligent stone picker suitable for reclaimed land and agricultural land, so as to realize the accurate detection, classification and efficient cleaning of stones and construction waste in reclaimed land and agricultural land, thereby improving soil quality, increasing the cultivability of land and reducing manual labor intensity.

[0004] To achieve the above-mentioned purpose, the present invention proposes the following technical scheme: a multi-sensor fusion adaptive detection method for a stone picker, comprising the following steps: collecting detection data of several sensors on the stone picker, and calibrating the time and space coordinates of the sensor detection data through the timestamp and position information; inputting the calibrated detection data into a neural network model to obtain a high-order feature vector of the object to be collected; constructing a pre-trained transfer learning model in different environments, training the transfer learning model based on multi-task learning and adaptive gradient pruning, and adjusting the hyperparameters of the transfer learning model through Bayesian optimization to obtain an optimal object recognition model under the current soil environment; inputting the high-order feature vector into the optimal object recognition model under the current soil environment to obtain the detection result and detection accuracy of the object to be collected; based on the detection accuracy and energy consumption indicators, automatically adjusting the parameter configuration of the sensor and the optimal object recognition model through a heuristic search algorithm; collecting previous detection data, and judging whether there is a new pattern in the object detection result through a clustering algorithm, if so, generating the optimal strategy of the object recognition model through deep reinforcement learning and environmental reward mechanism, thereby further optimizing the optimal object recognition model during the detection process.

[0005] Furthermore, the method for calibrating the time and space coordinates of the detection data of the sensor is: simultaneously obtaining the corresponding timestamps of the detection data of several sensors, using a time-space coupling algorithm to model the motion trajectory and environment of the stone picker, obtaining the real-time position information of the stone picker through the established model, and calibrating the time and space coordinates of the sensor's detection data in real time through a central signal processor.

[0006] Furthermore, the detection data of the sensor is preprocessed, and the preprocessing method is as follows: the detection data of the sensor is processed in parallel by a multi-core processor; the original signal is decomposed into multiple frequency bands by frequency domain decomposition, each frequency band is subjected to independent wavelet transform and filtering operations, the weight coefficient of each frequency band is adjusted, and the denoised signal is resynthesized; in the process of normalizing the detection data, the mean and equation of the real-time detection data are estimated, and the normalization function is adjusted according to the current detection data distribution; and the normalization parameters are dynamically updated by a Gaussian fitting algorithm.

[0007] Furthermore, if it cannot be determined whether the detection data of the sensor can be used to predict the test results, the detection data of the sensor is input into the Bayesian network model to establish a probabilistic relationship between the detection data of the sensor and the detection result, and the detection result is obtained through the maximum a posteriori probability MAP combined with fuzzy logic.

[0008] Furthermore, the multi-task learning adopts a multi-task learning network, the top-level network branches of the multi-task learning network are used for different tasks, and the bottom-level network is shared by various tasks. Through the joint loss function, each task is weighed and optimized; the adaptive gradient pruning is to prune according to the gradient change and the importance of the parameters in each iteration cycle of the transfer learning model training to remove redundant parameters; the hyperparameters include the learning rate, regularization parameter and the number of layers. After adjusting the hyperparameters, the performance of the adjusted transfer learning model is predicted by the Gaussian process.

[0009] Furthermore, the parameter configuration of the automatic adjustment sensor and the optimal object recognition model adopts an adaptive adjustment method of a dual feedback loop, wherein the dual feedback loop includes a fast feedback loop and a slow feedback loop, wherein the fast feedback loop triggers local adjustment through a detected error signal; and the slow feedback loop adjusts the global parameters or working mode of the sensor by analyzing long-term detection data.

[0010] Furthermore, the method for generating the optimal strategy of the object recognition model through deep reinforcement learning and environmental reward mechanism is as follows: in the meta-learning network, the transfer learning model is trained through multiple tasks to obtain meta-parameters, and when the stone picker enters different environments, the transfer learning model is adjusted in real time by adjusting the meta-parameters; the previous detection data is input into the online clustering algorithm model to analyze whether there are new data points in the detection data, and if so, it is identified whether the cluster to which the new data point belongs is an abnormal point, and if not, it is determined that there is a new pattern; a deep Q network model is constructed, so that the detection process is regarded as a state-action-reward cycle, the detection strategy of the transfer learning model is adjusted through real-time feedback, and the effective detection behavior is strengthened through environmental reward signals, so as to gradually converge to obtain the optimal strategy. The present invention also discloses an intelligent stone picker, comprising: an object acquisition module, a transmission module, a classification module, a sensor module, a data processing module and a power supply module; the object acquisition module is used to grab objects according to the object size category obtained by the data processing module; the transmission module is used to transmit the object to the classification module, which can adjust the transmission speed and transmission angle according to the size of the object; the classification module includes screens of different sizes for separating objects of different categories; the sensor module includes a front sensor of the stone picker and an internal sensor of the stone picker, and the front sensor of the stone picker is used to detect the position and characteristics of stones and construction waste in the soil. The internal sensors of the stone picker are used to monitor the operating status of each key component in real time. Once an abnormal situation is detected, an alarm is automatically sounded and a reminder is sent to the user. The data processing module adopts the multi-sensor fusion adaptive detection method described in any of the above items to scan the soil in real time, and accurately identify stones and construction waste in the soil by fusing data from different sensors, and generate a three-dimensional position map. The power supply module adopts hybrid power supply, including solar panels, lithium battery packs and auxiliary batteries. Solar panels are used when the sun is sufficient, and lithium battery packs are used at other times. The mechanical energy generated by the stone picker is converted into electrical energy and stored in the auxiliary battery.

[0011] Furthermore, the chassis of the stone picker is tracked or wheeled, and the object collection module, transmission module, classification module, sensor module, data processing module and power supply module are fixed on the chassis through a bracket; the chassis is also provided with a GPS and an inertial navigation system, and the GPS and the inertial navigation system are connected to the data processing module to realize autonomous working path planning.

[0012] Furthermore, an edge computing chip is embedded in the data processing module for preprocessing and preliminary analysis of the detection data of the sensor, and uploading the preliminary analysis results to the cloud server for global reasoning and decision-making; the data processing module includes a touch screen operation interface, which can set fine-grained feedback, set different detection priorities, adjust the objective function and parameter settings of the model, and support remote control.

[0013] The technical solution of the present invention has at least the following technical effects or advantages: 1. The stone picker in the present invention can realize accurate detection, classification and efficient cleaning of stones and construction waste in reclaimed land and agricultural land, thereby improving soil quality, increasing the cultivability of land, and reducing manual labor intensity; by recycling the mechanical energy generated during the use of the stone picker, the energy utilization efficiency of the stone picker is further improved.

[0014] 2. The present invention adopts a spatiotemporal coupling algorithm, which dynamically adjusts the acquisition frequency and the synchronization strategy between sensors according to the moving speed and direction of the trebuchet to avoid data deviation caused by trebuchet vibration or terrain changes.

[0015] 3. The present invention uses multi-task learning (MTL) to optimize multiple related tasks at the same time. Through the shared representation method, the relevant information of these tasks can be learned at the same time to improve the overall performance. In the model training process, the adaptive gradient pruning technology is used to dynamically adjust the model parameters to adapt to different computing resources and detection requirements. This method reduces the complexity of the model by pruning redundant parameters, while maintaining or even improving the detection accuracy of the model.

[0016] 4. The present invention adopts edge computing architecture to delegate data processing and some model reasoning tasks to the local device, thereby reducing data transmission delays and improving real-time performance. Edge computing nodes can independently complete data preprocessing, feature extraction, and preliminary classification tasks, and then send the processing results to the cloud for complex analysis.

[0017] 5. The present invention introduces a meta-learning framework, which enables the model to quickly adapt to data changes in new environments. Through meta-learning, the model can extract meta-knowledge from previous detection experience and apply it to the rapid learning process in new environments, greatly reducing the dependence on new data. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flow chart of a multi-sensor fusion adaptive detection method for a stone picker in one embodiment of the present invention; Figure 2 It is a schematic diagram of a multi-sensor fusion adaptive detection method for a stone picker in one embodiment of the present invention; Figure 3 is a schematic diagram of a method for obtaining a high-order feature vector of an object to be collected in one embodiment of the present invention; Figure 4 It is a structural schematic diagram of a stone picking machine in one embodiment of the present invention; Figure 5 It is a schematic diagram of a multi-sensor fusion adaptive detection system for a stone picker in one embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is described in detail through specific embodiments. However, it should be understood that the provision of specific embodiments is only for a better understanding of the present invention, and they should not be understood as limitations of the present invention. In the description of the present invention, it should be understood that the terms used are only for the purpose of description and cannot be understood as indicating or implying relative importance.

[0020] In order to solve the problems existing in the prior art that ordinary stone pickers are usually unable to accurately identify and separate stones and construction waste of different sizes, are prone to damage the soil, and further reduce the utilization value of the land, the present invention proposes a multi-sensor fusion adaptive detection method and an intelligent stone picker suitable for reclamation land and agricultural land, which integrates a high-performance processor and a data fusion algorithm module, is responsible for real-time processing of sensor data, and generates accurate target position and size information. The control algorithm based on machine learning can automatically adjust the working parameters of the equipment according to different soil conditions and working environments, such as sensor sensitivity, mechanical arm operation mode, conveyor belt speed, etc., through adaptive control online learning, continuously accumulate operation data, and optimize its own working logic. Equipped with a touch screen operation interface, users can directly set the working mode and view the equipment status on the device. At the same time, the device supports remote control, and users can remotely monitor and operate the stone picker through mobile devices (such as smartphones or tablets). Remote operation includes functions such as start / stop, path planning, and real-time adjustment. It can realize accurate detection, classification and efficient cleaning of stones and construction waste in reclaimed land and agricultural land through a multi-sensor fusion system and an adaptive detection method, thereby improving soil quality, increasing the cultivability of land, and reducing manual labor intensity. The scheme of the present invention is described in detail below through embodiments.

[0021] Embodiment 1 This embodiment discloses a multi-sensor fusion adaptive detection method for a stone picking machine. Figure 1 , Figure 2 As shown, the following steps are included: S1 collects the detection data of several sensors on the stone picker, and calibrates the time and space coordinates of the sensor's detection data through timestamps and position information.

[0022] The method for calibrating the time and space coordinates of the sensor detection data is as follows: the detection data of several sensors are simultaneously obtained with their corresponding timestamps, that is, a high-precision timestamp module is installed on each sensor, and the motion trajectory and environment of the stone picker are modeled using a spatiotemporal coupling algorithm. The real-time position information of the stone picker is obtained through the established model to ensure the precise alignment of the data of different sensors in the time and space dimensions, and the time and space coordinates of the sensor detection data are calibrated in real time through the central signal processor. The spatiotemporal coupling algorithm dynamically adjusts the acquisition frequency and the synchronization strategy between sensors according to the moving speed and direction of the trebuchet, and generates a spatiotemporal aligned data set for subsequent processing to avoid data deviations caused by the vibration of the trebuchet or terrain changes.

[0023] S2 inputs the calibrated detection data into the neural network model to obtain the high-order feature vector of the object to be collected.

[0024] The sensor detection data is preprocessed by combining multi-dimensional spectrum analysis, such as short-time Fourier transform (STFT) with wavelet transform, and introducing noise filtering process to distinguish and eliminate noise in different frequency bands. The original signal is decomposed into multiple frequency bands through frequency domain decomposition technology, and each frequency band is independently denoised, and then the denoised signal is resynthesized to ensure that the key features are retained while removing the noise.

[0025] The detection data of the sensor is processed in parallel by a multi-core processor; the original signal is decomposed into multiple frequency bands through frequency domain decomposition, and independent wavelet transform and filtering operations are performed on each frequency band. The weight coefficient of each frequency band is adjusted, and the denoised signal is re-synthesized to achieve efficient filtering of different types of noise, thereby improving the purity of the signal and detection accuracy.

[0026] The adaptive Gaussian normalization method is introduced to dynamically adjust the normalization parameters according to the real-time data distribution, so that the data distribution is closer to the Gaussian distribution. In the process of normalizing the detection data, the mean and equation of the real-time detection data are estimated, and the normalization function is adjusted according to the current detection data distribution; the normalization parameters are dynamically updated through the Gaussian fitting algorithm to ensure that the original information of the data is retained to the greatest extent during the normalization process, making the data normalization process more adaptable, especially when the data distribution changes, it can be adjusted quickly.

[0027] A multi-level feature extraction mechanism is used to further extract advanced features (such as complex pattern recognition and multi-sensor collaborative features) based on primary feature extraction (such as texture, shape, density, etc.). The primary features of each sensor are convolved and pooled layer by layer using a convolutional neural network (CNN) to generate a multi-dimensional feature map. Then, the multi-sensor data is jointly analyzed through the feature fusion module to extract comprehensive features with stronger discriminative power, which are input into the model training to extract high-order feature vectors that can characterize the characteristics of the stone.

[0028] If it is uncertain whether the sensor's detection data can be used to predict the test results, the sensor's detection data is input into the Bayesian network model to establish a probabilistic relationship between the sensor's detection data and the detection results. The detection results are obtained by combining the maximum a posteriori probability MAP with fuzzy logic. In the real-time detection process, a fuzzy reasoning system based on the Bayesian network is introduced to deal with the uncertainty and ambiguity in the sensor data. The Bayesian network can give a more robust detection result based on the probability distribution of the sensor input, and improve the reliability and stability of the detection, especially when the data is incomplete or conflicting.

[0029] S3 builds pre-trained transfer learning models in different environments, trains the transfer learning models based on multi-task learning and adaptive gradient pruning, and adjusts the hyperparameters of the transfer learning models through Bayesian optimization to obtain the optimal object recognition model under the current soil environment.

[0030] In order to cope with the detection tasks in different soil environments, the domain adaptation technology is used to transfer the pre-trained models in different environments to make them more adaptable in the new target domain (specific soil environment). The model's learning ability for unlabeled data is enhanced through self-supervised learning.

[0031] The initial transfer learning model is pre-trained on a public dataset, and then self-supervised learning is performed on unlabeled data in the target domain to gradually adjust the model parameters to adapt it to the target soil environment. Virtual samples are generated by a generative adversarial network (GAN) to enhance the generalization ability of the model.

[0032] Multi-task learning (MTL) optimizes multiple related tasks at the same time, such as stone positioning, shape classification, soil structure analysis, etc. Through the shared representation method, it can simultaneously learn the relevant information of these tasks and improve the overall performance. Multi-task learning (MTL) uses a multi-task learning network. The top-level network branches of the multi-task learning network are used for different tasks, and the bottom-level network is shared by various tasks. Through the joint loss function, each task is weighed and optimized to ensure that the model can perform well on multiple tasks.

[0033] Adaptive gradient pruning is to perform pruning according to gradient changes and parameter importance in each iteration cycle of transfer learning model training to remove redundant parameters. The pruned transfer learning model is then fine-tuned to continue training to restore the performance before pruning. It can significantly reduce the consumption of computing resources while ensuring performance.

[0034] The hyperparameters of the transfer learning model are adjusted through Bayesian Optimization. Bayesian Optimization intelligently searches the parameter space to find the optimal hyperparameter combination, thereby improving the overall performance of the model. Among them, the hyperparameters include learning rate, regularization parameter and number of layers. After the hyperparameters are adjusted, the performance of the adjusted transfer learning model is predicted through Gaussian Process, so as to efficiently find the optimal configuration.

[0035] S4 inputs the high-order feature vector into the optimal object recognition model under the current soil environment to obtain the detection results and detection accuracy of the objects to be collected.

[0036] Based on the detection accuracy and energy consumption indicators, S5 automatically adjusts the parameter configuration of the sensor and the optimal object recognition model through a heuristic search algorithm.

[0037] The adaptive heuristic parameter tuner dynamically adjusts the parameter configuration of sensors and models based on heuristic search algorithms (such as genetic algorithms and simulated annealing algorithms) to cope with complex and changing soil environments. The tuner automatically optimizes parameters by monitoring detection accuracy and energy consumption indicators in real time. During the real-time detection process, the tuner continuously evaluates the effect of the current parameter settings and generates new parameter combinations using heuristic search algorithms. Through the test and evaluation cycle, the optimal parameter configuration is gradually found to ensure the best balance between detection accuracy and energy consumption.

[0038] The automatic adjustment of the sensor and the parameter configuration of the optimal object recognition model adopts an adaptive adjustment method of dual feedback loops. The dual feedback loops include a fast feedback loop and a slow feedback loop. The fast feedback loop triggers local adjustment through the detected error signal. The fast feedback loop performs instant correction on the real-time detection results. For example, when an obvious error is detected, it immediately triggers re-detection. The slow feedback loop adjusts the global parameters or working mode of the sensor by analyzing long-term detection data to cope with continuously changing environmental conditions. The slow feedback loop is used to detect adjustments to the overall trend, such as adjusting the global configuration parameters of the sensor through long-term monitoring.

[0039] S6 collects previous detection data and uses a clustering algorithm to determine whether there are new patterns in the object detection results. If so, it generates the optimal strategy for the object recognition model through deep reinforcement learning and environmental reward mechanism, thereby further optimizing the optimal object recognition model during the detection process.

[0040] The method to generate the optimal strategy for the object recognition model through deep reinforcement learning and environmental reward mechanism is: In the meta-learning network, the transfer learning model is trained through multiple tasks to obtain meta-parameters. When the stone picker enters different environments, the transfer learning model is adjusted in real time by adjusting the meta-parameters, so that it can perform well even with a small amount of new data.

[0041] Previous test data is input into the online clustering algorithm model to analyze whether there are new data points in the test data. If so, it identifies whether the cluster to which the new data point belongs is an outlier. If not, it determines that there are new patterns and fine-tunes the model based on these new findings. The online clustering algorithm can update the cluster center and adjust the classification boundaries in real time, so that the model always maintains sensitivity to the data distribution.

[0042] In the process of self-learning, a deep Q network DQN model is constructed, so that the detection process is regarded as a state-action-reward cycle. The detection strategy of the transfer learning model is adjusted through real-time feedback, and effective detection behavior is strengthened through environmental reward signals, so as to gradually converge to the optimal strategy. The model searches for the optimal strategy by experimenting with different parameter settings and detection paths, and continuously optimizes its own performance through reinforcement learning.

[0043] Embodiment 2 Based on the same inventive concept, this embodiment discloses an intelligent stone picking machine, such as Figure 4 , Figure 5 As shown, it includes: an object collection module 1, a transmission module 2, a classification module 3, a sensor module, a data processing module, a power supply module, a chassis and a protective shell.

[0044] The object acquisition module 1 is used to grab objects according to the object size category obtained by the data processing module; the object acquisition module 1 may include a mechanical arm and a grabbing tool. In this embodiment, the mechanical arm is a highly flexible mechanical arm with adjustable length and has six degrees of freedom of movement. The grabbing tool may be a multifunctional clamp, a vacuum suction cup or a manipulator, etc. The specific grabbing tool to be used is determined according to the object to be grabbed, wherein the opening of the clamp and the manipulator can be determined according to the object to be grabbed, and the force of the clamping object can also be adjusted according to actual needs, and the vacuum suction cup can also be selected according to the object to be grabbed. The grabbing method is automatically adjusted according to the size and shape of the target object, so that the target object can be efficiently located and extracted.

[0045] The transmission module 2 is arranged between the object collection module 1 and the classification module 3, and is used to transmit the object to the classification module 3. It can also be used to transmit a certain type of classified objects. It can adjust the transmission speed and transmission angle according to the size of the object to ensure that the object reaches the classification module 3 smoothly.

[0046] The classification module 3 includes replaceable screens of different sizes for separating objects of different categories, and can classify objects according to their sizes, and is equipped with a classification collection box to store different types of waste. The screening device can also be equipped with a vibrating screen or a rotary screen to enhance the classification effect.

[0047] The sensor module includes an ultrasonic sensor, an infrared sensor, a geomagnetic sensor and a radar sensor. The above sensors are set at the front of the stone picker. The ultrasonic sensor is used to detect the density difference of underground stones to provide location information. The infrared sensor is used to detect the temperature change of the soil surface and identify the thermal characteristics of the stone. The geomagnetic sensor is used to detect underground magnetic objects, such as iron-containing stones, and provide magnetic field change data. The radar sensor detects stones and obstacles under the ground through radar waves and generates a depth image. The camera and lidar are used for sensing environmental perception and navigation, real-time monitoring of the surrounding environment, and identification of potential obstacles or dangerous areas. The obstacle avoidance system can adjust the direction of the equipment in time or stop working to ensure safe operation. Multiple status monitoring sensors are set inside the stone picker to monitor the operating status of each key component in real time (such as battery power, mechanical arm position, sensor working status, etc.). Once an abnormal situation is detected, the system will automatically alarm and send a reminder to the user. The sensor module can learn and pick up stones by identifying indicators such as the size, shape, and porosity of objects. Stones and bricks have different colors, shapes, and porosities, while construction waste plastic sheets, steel bars, etc. have no pores and are mostly in the form of sheets and columns. The above distinctions can be used to accurately determine the category of construction waste in the data processing module so that it can be processed separately.

[0048] The data processing module adopts any of the above multi-sensor fusion adaptive detection methods to scan the soil in real time, accurately identify stones and construction waste in the soil by fusing data from different sensors, and generate a three-dimensional location map.

[0049] The data processing module is embedded with an edge computing chip, which is used to pre-process and preliminarily analyze the sensor's detection data, and upload the preliminary results of the analysis to the cloud server for global reasoning and decision-making. The data processing module includes a touch screen operation interface that can set fine-grained feedback, such as giving priority to detection speed or accuracy, or optimizing for specific environments, allowing users to provide fine-grained feedback based on actual needs. By setting different detection priorities (such as speed, accuracy, energy consumption, etc.), adjusting the model's objective function and parameter settings, users can directly set the working mode and view the status of the stone picker on the stone picker. By learning the user's preferences, the model's working mode is automatically adjusted to meet personalized needs.

[0050] At the same time, the data processing module supports remote control, and users can remotely monitor and operate the stone picker through mobile devices such as smartphones or tablets. Remote operation includes functions such as start / stop, path planning, and real-time adjustment.

[0051] The power supply module adopts hybrid power supply, including solar panels, lithium battery packs and auxiliary batteries. The solar panels are set on the top of the protective shell. The solar panels are used when there is sufficient sunshine, and the lithium battery packs are used at other times, such as on cloudy days or at night, to ensure that the stone picker can work continuously and convert the mechanical energy generated by the stone picker (such as the mechanical energy generated by the movement and vibration of the stone picker) into electrical energy, which is stored in the auxiliary battery to support some functions of the stone picker such as sensors and data processing modules, thereby further improving the energy utilization efficiency of the stone picker.

[0052] The chassis of the stone picker is crawler or wheeled, and the object collection module 1, transmission module 2, classification module 3, sensor module, data processing module and power supply module are fixed on the chassis through a bracket. The specific choice of crawler or wheeled chassis is determined according to the specific working environment to ensure that the equipment can adapt to various complex terrains, such as soft soil, slopes or stone-dense areas. The chassis is equipped with a shock absorber to reduce the vibration of the equipment during driving and protect the soil structure. Among them, the bracket is made of high-strength lightweight alloy or composite material, which can reduce the weight of the equipment while ensuring strength, and facilitate movement on different terrains. GPS and inertial navigation systems are also provided on the chassis. GPS and inertial navigation systems are connected to the data processing module to realize autonomous work path planning, avoid obstacles, and ensure full coverage operations. The inertial navigation system combines the detection data of the sensor to adjust the path in real time during the operation to improve work efficiency and accuracy.

[0053] The protective housing is used to place the transmission module 2, the classification module 3, the sensor module, the data processing module and the power supply module, and protect the above modules. The protective housing is made of corrosion-resistant materials and has a high protection level, which can effectively prevent dust, moisture and other environmental factors from damaging the internal electronic components.

[0054] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0055] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0056] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0057] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0058] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit them. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation of the present invention can still be modified or replaced by equivalents, and any modification or equivalent replacement that does not deviate from the spirit and scope of the present invention should be included in the protection scope of the claims of the present invention. The above content is only a specific implementation of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A multi-sensor fusion adaptive detection method for a stone picker, characterized in that: The following steps are involved: Collect detection data of several sensors on the stone picker, and calibrate the time and space coordinates of the detection data of the sensors through the timestamp and position information; Inputting the calibrated detection data into a neural network model to obtain a high-order feature vector of the object to be collected; Constructing pre-trained transfer learning models in different environments, training the transfer learning models based on multi-task learning and adaptive gradient pruning, and adjusting the hyperparameters of the transfer learning models through Bayesian optimization to obtain the optimal object recognition model under the current soil environment; Inputting the high-order feature vector into the optimal object recognition model under the current soil environment to obtain the detection result and detection accuracy of the object to be collected; Based on the detection accuracy and energy consumption indicators, the parameter configuration of the sensor and the optimal object recognition model is automatically adjusted through a heuristic search algorithm; Collect previous detection data, and use a clustering algorithm to determine whether there is a new pattern in the object detection results. If so, generate the optimal strategy for the object recognition model through deep reinforcement learning and environmental reward mechanism, so as to further optimize the optimal object recognition model during the detection process.

2. The multi-sensor fusion adaptive detection method for a stone picker according to claim 1, characterized in that: The method for calibrating the time and space coordinates of the detection data of the sensor is: simultaneously obtaining the corresponding timestamps of the detection data of several sensors, using a time-space coupling algorithm to model the motion trajectory and environment of the stone picker, obtaining the real-time position information of the stone picker through the established model, and calibrating the time and space coordinates of the sensor's detection data in real time through a central signal processor.

3. The multi-sensor fusion adaptive detection method for a stone picker according to claim 2, characterized in that: The detection data of the sensor is preprocessed, and the preprocessing method is: Process the sensor detection data in parallel through multi-core processors; Through frequency domain decomposition, the original signal is decomposed into multiple frequency bands, and independent wavelet transform and filtering operations are performed on each frequency band, and the weight coefficient of each frequency band is adjusted, and then the denoised signal is resynthesized; In the process of normalizing the detection data, the mean and equation of the real-time detection data are estimated, and the normalization function is adjusted according to the current detection data distribution; the normalization parameters are dynamically updated through the Gaussian fitting algorithm.

4. The multi-sensor fusion adaptive detection method for a stone picker according to claim 2, characterized in that: If it is not certain whether the detection data of the sensor can be used to predict the test results, the detection data of the sensor is input into the Bayesian network model to establish a probabilistic relationship between the detection data of the sensor and the detection result, and the detection result is obtained through the maximum a posteriori probability MAP combined with fuzzy logic.

5. The multi-sensor fusion adaptive detection method for a stone picker according to claim 1, characterized in that: The multi-task learning adopts a multi-task learning network, the top network branches of the multi-task learning network are used for different tasks, and the bottom network is shared by various tasks, and each task is weighed and optimized through a joint loss function; The adaptive gradient pruning is to perform pruning according to the gradient change and the importance of the parameters in each iteration cycle of the transfer learning model training to remove redundant parameters; The hyperparameters include a learning rate, a regularization parameter, and the number of layers. After adjusting the hyperparameters, the performance of the adjusted transfer learning model is predicted by a Gaussian process.

6. The multi-sensor fusion adaptive detection method for a stone picker according to claim 1, characterized in that: The parameter configuration of the automatic adjustment sensor and the optimal object recognition model adopts an adaptive adjustment method of a dual feedback loop, wherein the dual feedback loop includes a fast feedback loop and a slow feedback loop, wherein the fast feedback loop triggers local adjustment through a detected error signal; and the slow feedback loop adjusts the global parameters or working mode of the sensor by analyzing long-term detection data.

7. The multi-sensor fusion adaptive detection method for a stone picker according to claim 1, characterized in that: The method for generating the optimal strategy of the object recognition model through deep reinforcement learning and environmental reward mechanism is: In the meta-learning network, the transfer learning model is trained through multiple tasks to obtain meta-parameters. When the stone picker enters different environments, the transfer learning model is adjusted in real time by adjusting the meta-parameters. The previous detection data is input into the online clustering algorithm model to analyze whether there are new data points in the detection data. If so, it is identified whether the cluster to which the new data point belongs is an outlier. If not, it is determined that there is a new pattern. A deep Q-network model is constructed to regard the detection process as a state-action-reward cycle. The detection strategy of the transfer learning model is adjusted through real-time feedback, and effective detection behavior is strengthened through environmental reward signals, so as to gradually converge to the optimal strategy.

8. An intelligent stone picking machine, characterized in that: include: Object acquisition module, transmission module, classification module, sensor module, data processing module and power supply module; The object acquisition module is used to grab objects according to the object size category obtained by the data processing module; The transmission module is used to transmit the object to the classification module, and can adjust the transmission speed and transmission angle according to the size of the object; The classification module includes screens of different sizes for separating objects of different categories; The sensor module includes a front sensor of the stone picker and an internal sensor of the stone picker. The front sensor of the stone picker is used to detect the position and characteristics of stones and construction waste in the soil. The internal sensor of the stone picker is used to monitor the operating status of each key component in real time. Once an abnormal situation is detected, an alarm is automatically sounded and a reminder is sent to the user. The data processing module adopts the multi-sensor fusion adaptive detection method according to any one of claims 1 to 7, and is used to scan the soil in real time, accurately identify stones and construction waste in the soil by fusing data from different sensors, and generate a three-dimensional location map; The power supply module adopts hybrid power supply, including solar panels, lithium battery packs and auxiliary batteries. Solar panels are used when there is sufficient sunshine, and lithium battery packs are used at other times. The mechanical energy generated by the stone picker is converted into electrical energy and stored in the auxiliary battery.

9. The intelligent stone picking machine according to claim 8, characterized in that: The chassis of the stone picker is tracked or wheeled, and the object collection module, transmission module, classification module, sensor module, data processing module and power supply module are fixed on the chassis through a bracket; the chassis is also provided with a GPS and an inertial navigation system, and the GPS and the inertial navigation system are connected to the data processing module to realize autonomous working path planning.

10. The intelligent stone picking machine according to claim 8, characterized in that: The data processing module is embedded with an edge computing chip for preprocessing and preliminary analysis of the detection data of the sensor, and uploading the preliminary analysis results to the cloud server for global reasoning and decision-making; the data processing module includes a touch screen operation interface, which can set fine-grained feedback, set different detection priorities, adjust the objective function and parameter settings of the model, and support remote control.

Images