Automatic obstacle avoidance method and system for cleaning vehicle
Through real-time obstacle feature data processing and policy network optimization, combined with the improved adversarial network GGAN model, the intelligent cleaning vehicle automatic obstacle avoidance method solves the problems of poor stability and high energy consumption, achieving low carbon energy saving and efficient obstacle avoidance.
Patent Information
- Application Number
- CN202411821597.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-11
- Publication Date
- 2025-04-01
AI Technical Summary
Smart cleaning vehicles have poor stability and high energy consumption during obstacle avoidance, making it difficult to effectively avoid static and dynamic obstacles, resulting in increased collision and calculation costs.
A clean vehicle automatic obstacle avoidance method is adopted. By obtaining real-time obstacle characteristic data, using the target model to identify obstacles and predict action, combining the policy network to form a set of obstacle avoidance strategies for intelligent cleaning vehicles, and outputting the best strategy selection under low-carbon energy-saving conditions through the improved adversarial network GGAN model to achieve automatic obstacle avoidance.
It improves the stability and low-carbon energy-saving performance of smart cleaning vehicles in the obstacle avoidance process, reduces collision risks and calculation costs, and realizes efficient, stable and green and low-carbon obstacle avoidance control of smart cleaning vehicles.
Smart Images
Figure CN119739157A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of obstacle avoidance planning, and more specifically, to an automatic obstacle avoidance method and system for a cleaning vehicle. Background Art
[0002] As the construction of civilized cities gradually improves, urban cleanliness has also become an important part of civilized city construction. With the increase in labor costs, intelligent automatic cleaning vehicles have become emerging maintainers of urban cleanliness. Intelligent automatic cleaning vehicles not only improve cleaning efficiency and achieve uninterrupted and continuous cleaning, but also reduce labor expenditures.
[0003] However, during the automatic cleaning process of the intelligent cleaning vehicle, path planning is the core technology of the intelligent cleaning vehicle's autonomous navigation. The intelligent cleaning vehicle acts autonomously according to a pre-set path. However, movement in a real environment is usually uncertain, and random obstacles often appear. In the actual movement of the intelligent cleaning vehicle, there are not only static obstacles but also dynamic obstacles. If the intelligent cleaning vehicle does not have obstacle avoidance behavior in a random obstacle environment, it is very easy to produce a collision result. In the process of obstacle avoidance path planning, if the obstacle avoidance behavior and obstacle avoidance connection in the planned global path are not considered after avoiding the obstacle, the computational cost of re-planning the path and driving will be increased, and even incomplete obstacle avoidance will occur, as well as unstable driving and poor driving quality of the cleaning vehicle during the obstacle avoidance process. Therefore, in the obstacle avoidance process, not only the control parameters of the cleaning vehicle should be considered, but also the movement behavior of the obstacles should be analyzed. Therefore, it is necessary to plan the applicability of the path of the intelligent cleaning vehicle during the obstacle avoidance process. Summary of the invention
[0004] In view of the problems of poor stability and high energy consumption in the obstacle avoidance process of the intelligent cleaning vehicle in the prior art, the purpose of the present invention is to provide an automatic obstacle avoidance method and system for a cleaning vehicle, so as to achieve low-carbon and stable driving of the cleaning vehicle during the automatic obstacle avoidance process.
[0005] To achieve the above object, the present invention provides the following technical solution: an automatic obstacle avoidance method for a cleaning vehicle, the automatic obstacle avoidance method for a cleaning vehicle comprising the following steps: Step S1: Obtain feature data of the obstacle at the current moment and construct a real-time state feature sequence; the trained target model identifies the obstacle and predicts the action according to the real-time state feature sequence; Step S2: Obtain the historical ultimate feature sequence output by the target model, collect the historical obstacle avoidance data of the smart cleaning vehicle corresponding to the historical ultimate feature sequence, establish a four-tuple set, use the four-tuple set to train the strategy network, and input the current motion state of the smart cleaning vehicle into the trained strategy network to obtain the output strategy set; Step S3: input the strategy set into the improved adversarial network GGAN model, and output the optimal strategy selection of the smart cleaning vehicle under the constraints of low-carbon energy-saving conditions; Step S4: Select and control the intelligent cleaning vehicle to automatically avoid obstacles according to the output optimal strategy.
[0006] Preferably, the training of the target model in step S1 includes: Using the convolutional neural network CNN as the framework, an external connection layer is constructed. The external connection layer includes a CRFs conditional random field, which is denoted as a CRF layer. The category feature sequence extracted by the convolutional neural network CNN is used as the input of the CRF layer. The CRF layer will use the category feature sequence to perform action prediction inference; Obtain a sample set, divide the data in the sample set into multiple batches, each batch includes multiple data groups, each data group includes feature data of multiple categories of obstacles in multiple historical cycle times, and construct a historical state feature sequence of each category of obstacles in each cycle time; The data in the sample set is input into the initial target model in batches for training, and the forward propagation is performed to calculate the loss function. The back propagation calculates the gradient of each parameter of the initial target model, and the SGD optimizer uses the gradient to update the parameters of the initial target detection model, that is, to complete an iteration. After each batch is trained, the loss function value is calculated. When the loss function value reaches the threshold, the training is stopped to obtain the final target model. The loss function of the target model is set to , represents the number of data groups in each batch, Indicates the time period label of each obstacle in each data group, It is the actual ultimate characteristic sequence of the obstacle; is the ultimate characteristic sequence of obstacles predicted by the model; The real-time state feature sequence of the obstacle at the current moment is imported into the trained target model, and the category and action prediction of the obstacle are analyzed to obtain the real-time ultimate feature sequence.
[0007] Preferably, the characteristic data of the obstacle includes location information data , speed information data and size information data ; Historical cycle time Discretization to obtain the location information data of obstacles at each moment , speed information data and size information data , the location information data Use binary encoding, the encoding length is p; the speed information data Normalize and map to a real number; convert the size information data Use binary encoding, the encoding length is m; Constructing historical state feature sequences ,Will Input target model, , and Indicates the cycle time Location information data corresponding to the moment , speed information data and size information data , , and within the cycle time, represents the initial moment of the obstacle, Indicates the ultimate moment of the obstacle and outputs the conjugate feature sequence , Indicates the category of the obstacle; right Apply CRFs conditional random fields to obtain the ultimate feature sequence of obstacles , Indicates the obstacle's motion prediction data.
[0008] Preferably, the historical obstacle avoidance data in S2 includes the start and end parameter control data of the cleaning vehicle , Multi-dimensional reward data and conditional data ; Start and end parameter control data Including initial parameter control data and result parameter control data, the initial parameter control data packet and the result parameter control data packet respectively include the control parameters of each device state of the cleaning vehicle at the beginning and end of obstacle avoidance; Quadruple Set Including the ultimate characteristic sequence of history , start and end parameter control data , Multi-dimensional reward data and conditional data , The multi-dimensional reward data includes obstacle avoidance quality coefficient, safety coefficient and time consumption coefficient; Multi-dimensional reward data ; In the formula, is the first-order derivative of the driving curve of the smart cleaning vehicle, is the second-order derivative of the driving curve of the intelligent cleaning vehicle, The time required to complete obstacle avoidance, Edge track for smart cleaning vehicles Interference coefficient with obstacles; The strategy set includes multiple groups of strategy selections, each group of strategy selections includes a parameter control guidance and an obstacle avoidance path for an intelligent cleaning vehicle; the parameter control guidance is used to guide the parameters of the equipment control of the intelligent cleaning vehicle during the obstacle avoidance process, and the obstacle avoidance path is the path that the intelligent cleaning vehicle travels during the obstacle avoidance process.
[0009] Preferably, the strategy network training process includes: Construct a policy network framework. The policy network uses a combination of support vector machine (SVM) and radial basis function (RBF) kernel. Set the quaternary set as the input of the policy network and perform standardization. Set the policy set as the output. Set the radial basis function (RBF) kernel as a Gaussian function. Adjust the parameters of the RBF kernel through cross-validation. Construct a training set, which includes multiple sets of four-tuple sets and corresponding policy sets. Input the data in the training set into the policy network in batches, and continuously adjust the network parameters through iterative training until all batches in the training set are input, and obtain the optimized policy network. The motion state of the smart cleaning vehicle at the current moment is used to establish a four-tuple set at the current moment and input it into the trained strategy network to obtain the output strategy set.
[0010] Preferably, the GGAN model construction in S3 includes: The GGAN model sets a two-layer adversarial network, namely the GAN1 adversarial network and the GAN2 adversarial network. The GAN1 adversarial network is used to analyze the correlation characteristics between the strategy set containing multiple data and the optimal strategy selection. The GAN1 adversarial network sets the generator structure to adopt a multi-input neural network structure and sets the constraint layer. Construct an experimental set, which includes multiple groups of strategy sets and the corresponding optimal strategy selections. Input the data in the experimental set into the GGAN model in batches. The input of the GGAN model is the strategy set in the experimental set, and the output is the predicted optimal strategy selection. Use the function value of the diversity control loss function as a supervisory signal, and update the parameters in the framework network through back propagation. When the function value of the mixed loss function is minimized and does not change, a trained GGAN model is obtained.
[0011] Preferably, the constraint layer is formed by a constraint function Limit the generation of the best strategy selection and set the constraint function for: ; is the power efficiency coefficient, Indicates the quality of the smart cleaning vehicle, , and represents the correlation coefficient, Indicates the operating temperature of the power equipment.
[0012] Preferably, the GAN2 adversarial network is used to generate an adversarial loss function, introduce a diversity loss, and set the diversity control loss function to ; in, To combat the loss function, To improve the Euclidean distance loss function, To adjust the parameters.
[0013] Preferably, the improved Euclidean distance loss function is ; in, is the number of samples generated, , Indicates that the feature dimension is And the labels are , Samples of Adversarial Loss Function for: ; in, is the objective function of the entire generative adversarial network, where is the discriminator network, is the generator network; is the reconstruction loss, which measures the optimal strategy choice for the predictions generated by the generator and the true best strategy selection The reconstruction error between is a metric loss that measures the optimal strategy choice for the predictions generated by the generator With its output The measurement error between is a generative adversarial loss used to make the predicted optimal policy choices produced by the generator more difficult for the discriminator to distinguish.
[0014] Preferably, the automatic obstacle avoidance system of the cleaning vehicle comprises: A sample storage module is used to store feature data and historical obstacle avoidance data of multiple categories of obstacles within a historical period; A data acquisition module, which is used to collect characteristic data of obstacles at the current moment and the motion state of the intelligent cleaning vehicle at the current moment; Data processing module, including target model, policy network and adversarial network GGAN model. The target model is used for obstacle recognition and action prediction. The policy network is used to form the smart cleaning vehicle strategy set. The adversarial network GGAN model is used to output the best strategy selection in the strategy set. The adjustment module selects the best strategy to automatically avoid obstacles for the smart cleaning vehicle.
[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. In the present invention, the target model realizes obstacle type recognition and action prediction according to the real-time state feature sequence, realizes the division of dynamic obstacles and static obstacles, and provides data support for the obstacle avoidance behavior of the intelligent cleaning vehicle through the action prediction of the obstacles. The strategy network is used to form a strategy set for the obstacle avoidance control of the intelligent cleaning vehicle, and all automatic obstacle avoidance behaviors completed during the obstacle avoidance process are provided, thereby improving the quality of the optimal obstacle avoidance behavior that can be selected.
[0016] 2. In the present invention, the computational occupancy of the loss function constraint is reduced by improving the double-layer adversarial network in the adversarial network GGAN model. By setting the constraint layer in low-carbon energy saving, the optimal strategy selection of the intelligent cleaning vehicle is output in the strategy set under the constraint condition, and the optimal operational control of the obstacle avoidance behavior of the cleaning vehicle is achieved under low-carbon energy saving conditions, thereby realizing good, stable, green and low-carbon obstacle avoidance control management of the intelligent cleaning vehicle. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 A structural schematic diagram of the steps of an automatic obstacle avoidance method for a cleaning vehicle proposed by the present invention; Figure 2 It is a schematic diagram of the system structure in the present invention. DETAILED DESCRIPTION
[0018] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.
[0019] Embodiment 1 Reference Figure 1 and Figure 2 Embodiment 1 further illustrates an automatic obstacle avoidance method and system for a cleaning vehicle proposed in the present invention.
[0020] As the construction of civilized cities gradually improves, urban cleanliness has also become an important part of civilized city construction. With the increase in labor costs, intelligent automatic cleaning vehicles have become emerging maintainers of urban cleanliness. Intelligent automatic cleaning vehicles not only improve cleaning efficiency and achieve uninterrupted and continuous cleaning, but also reduce labor expenditures.
[0021] However, during the automatic cleaning process of the intelligent cleaning vehicle, path planning is the core technology of the intelligent cleaning vehicle's autonomous navigation. The intelligent cleaning vehicle acts autonomously according to a pre-set path. However, movement in a real environment is usually uncertain, and random obstacles often appear. In the actual movement of the intelligent cleaning vehicle, there are not only static obstacles but also dynamic obstacles. If the intelligent cleaning vehicle does not have obstacle avoidance behavior in a random obstacle environment, it is very easy to cause a collision. In the process of obstacle avoidance path planning, if the obstacle avoidance behavior and obstacle avoidance connection in the planned global path after avoiding the obstacle are not considered, the computational cost of re-planning the path and driving will be increased, and even incomplete obstacle avoidance will occur. Therefore, in the process of obstacle avoidance, not only the control parameters of the cleaning vehicle should be considered, but also the movement behavior of the obstacle should be analyzed. Therefore, it is necessary to plan the applicability of the path of the intelligent cleaning vehicle in the obstacle avoidance process.
[0022] The intelligent cleaning vehicle will pre-set the working range or working path during the working process. In the pre-set working area, due to the uncertainty of the environment, different static obstacles and dynamic obstacles will randomly appear in the pre-set working area. In the process of obstacle avoidance, the optimal path for safe and collision-free operation of the intelligent cleaning vehicle is the result of comprehensive analysis of multiple factors. In order to achieve the planning of the optimal path, in the obstacle avoidance behavior of dynamic obstacles, it is necessary to consider not only the control parameters of the intelligent cleaning vehicle equipment and the adaptive identification of obstacles, but also the control parameters of the cleaning vehicle equipment include the maximum driving speed and the turning angle range. The adaptive identification of obstacles includes the dynamic and static states, categories and motion behaviors of obstacles. For the path planning problem of mobile robots in a random obstacle environment, the main purpose of this study is to plan the optimal obstacle avoidance motion path for the intelligent cleaning vehicle from the starting point to the end point of the working area, and when random obstacles appear locally, the intelligent cleaning vehicle can avoid these obstacles and return to the planned optimal path.
[0023] This paper establishes a two-layer planning for the obstacle avoidance operation scheduling problem of intelligent cleaning vehicles. The upper layer considers minimizing costs and minimizing energy consumption and carbon emissions, that is, the obstacle avoidance scheduling control of the intelligent cleaning vehicle, and the lower layer is the optimal motion state, that is, the equipment control operation quality of the intelligent cleaning vehicle during the obstacle avoidance process. The optimal obstacle avoidance behavior is achieved through optimal path planning, and the motion conflict between the intelligent cleaning vehicle and obstacles is resolved through the optimal obstacle avoidance behavior. The operation scheduling and path planning are modeled as upper and lower double-layer models respectively; low-carbon energy-saving factors are added to achieve the easy handling and practicality of the obstacle avoidance behavior.
[0024] The obstacle avoidance scheduling control of the intelligent cleaning vehicle includes the dynamic output of various motion parameters of the intelligent cleaning vehicle in the obstacle avoidance process based on the time window in the initial state. The motion parameters include driving speed and obstacle avoidance angle. The driving speed is limited by the maximum driving speed of the intelligent cleaning vehicle, and the obstacle avoidance angle is affected by the wheelbase between the driving wheels, the position of the driving wheels, and the relative active curvature between the obstacle and the intelligent cleaning vehicle.
[0025] The shortest moving distance can also be formalized as finding an optimal path from the starting point to the end point in a discrete or continuous state space, and the optimal path will not cause vehicle collisions and deadlocks to avoid operation interruptions.
[0026] An automatic obstacle avoidance method for a cleaning vehicle, the automatic obstacle avoidance method for a cleaning vehicle comprising the following steps: Step S1: Obtain feature data of the obstacle at the current moment and construct a real-time state feature sequence; the trained target model identifies the obstacle and predicts the action according to the real-time state feature sequence; The training of the target model includes: Using the convolutional neural network CNN as the framework, an external connection layer is constructed. The external connection layer includes a CRFs conditional random field, which is denoted as a CRF layer. The category feature sequence extracted by the convolutional neural network CNN is used as the input of the CRF layer. The CRF layer will use the category feature sequence to perform action prediction inference; Obtain a sample set, divide the data in the sample set into multiple batches, each batch includes multiple data groups, each data group includes feature data of multiple categories of obstacles in multiple historical cycle times, and construct a historical state feature sequence of each category of obstacles in each cycle time, The characteristic data of the obstacle includes location information data , speed information data and size information data , Historical cycle time Discretization to obtain the location information data of obstacles at each moment , speed information data and size information data , the location information data Use binary encoding, the encoding length is p; the speed information data Normalize and map to a real number; convert the size information data Use binary encoding, the encoding length is m; Constructing historical state feature sequences ,Will Input the target model; the target model uses the forward calculation of the convolutional layer and the fully connected layer of the convolutional neural network CNN, updates the learning rate, adds the ReLU activation function between the convolutional layer and the fully connected layer, adds the Dropout layer after the pooling layer and before the fully connected layer, and sets the Dropout layer to 0.5. Set the convolution kernel size to 3x3, the number to 64, the stride to 1, and the padding to the same. , and Indicates the cycle time Location information data corresponding to the moment , speed information data and size information data , , and within the cycle time, represents the initial moment of the obstacle, Indicates the ultimate moment of the obstacle and outputs the conjugate feature sequence , Indicates the category of obstacles, such as trash cans, pedestrians, cars, and others. right Apply CRFs conditional random fields, update learning parameters through maximum likelihood estimation or gradient descent, and obtain the ultimate feature sequence of obstacles. , Predicted motion data representing obstacles, such as stationary, accelerated motion, uniform motion, and others; The loss function of the target model is set to , represents the number of data groups in each batch, Indicates the time period label of each obstacle in each data group, It is the actual ultimate characteristic sequence of the obstacle; is the ultimate characteristic sequence of obstacles predicted by the model; The data in the sample set is input into the initial target model in batches for training, and the forward propagation is performed to calculate the loss function. The back propagation calculates the gradient of each parameter of the initial target model, and the SGD optimizer uses the gradient to update the parameters of the initial target detection model, that is, to complete an iteration. After each batch is trained, the loss function value is calculated. When the loss function value reaches the threshold, the training is stopped to obtain the final target model.
[0027] Import the real-time state feature sequence of the obstacle at the current moment into the trained target model, analyze the category and action prediction of the obstacle, and obtain the real-time ultimate feature sequence; A high-definition camera is installed on the smart cleaning vehicle to detect the road environment within the vehicle body. The high-definition camera acquires the characteristic data of obstacles at a discrete time (per k seconds / g frames) within the cycle time and transmits it to the cloud processing platform. The cloud processing platform uses the target model to perform obstacle recognition analysis and motion prediction. Step S2: Obtain the historical ultimate feature sequence output by the target model, collect the historical obstacle avoidance data of the smart cleaning vehicle corresponding to the historical ultimate feature sequence, establish a four-tuple set, use the four-tuple set to train the strategy network, and input the current motion state of the smart cleaning vehicle into the trained strategy network to obtain the output strategy set. Historical obstacle avoidance data including the start and end parameter control data of the cleaning vehicle , Multi-dimensional reward data and conditional data ; Start and end parameter control data Including initial parameter control data and result parameter control data, the initial parameter control data packet and the result parameter control data packet respectively include the control parameters of each device state of the cleaning vehicle at the beginning and end of obstacle avoidance; Quadruple Set Including the ultimate characteristic sequence of history , start and end parameter control data , Multi-dimensional reward data and conditional data , The multi-dimensional reward data includes obstacle avoidance quality coefficient, safety coefficient and time consumption coefficient; Multi-dimensional reward data ; In the formula, is the first-order derivative of the driving curve of the intelligent cleaning vehicle, is the second-order derivative of the driving curve of the intelligent cleaning vehicle, The time required to complete obstacle avoidance, Edge track for smart cleaning vehicles Interference coefficient with obstacles; Among them, the driving curve of the intelligent cleaning vehicle is obtained by GPS, the time required to complete obstacle avoidance is obtained by the timer; the interference coefficient is obtained by radar and camera.
[0028] The larger the multi-dimensional reward data value is, the higher the reward coefficient is, that is, the better the obstacle avoidance behavior of the smart cleaning vehicle is; Conditional data Including the maximum driving speed and angular velocity of the smart cleaning vehicle, the radar and camera occlusion and blind spot avoidance range, and the obstacle avoidance restricted activity area; The strategy set includes multiple groups of strategy selections, each of which includes a parameter control guide and an obstacle avoidance path for the smart cleaning vehicle; the parameter control guide is used to guide the parameters of the equipment control of the smart cleaning vehicle during the obstacle avoidance process, that is, to regulate the driving speed, turning angle and other related motion changes of the smart cleaning vehicle, and the obstacle avoidance path is the path that the smart cleaning vehicle travels during the obstacle avoidance process; The policy network training process includes: The policy network framework is constructed. The policy network adopts the combination of support vector machine (SVM) and radial basis function (RBF) kernel. The four-tuple set is set as the input of the policy network and standardized. The policy set is taken as the output. The radial basis function (RBF) kernel is set as the Gaussian function, and the parameters of the RBF kernel are adjusted through cross-validation.
[0029] Construct a training set, which includes multiple sets of four-tuple sets and corresponding policy sets. Input the data in the training set into the policy network in batches, and continuously adjust the network parameters through iterative training until all batches in the training set are input, thus obtaining the optimized policy network.
[0030] The motion state of the smart cleaning vehicle at the current moment is used to establish a four-tuple set at the current moment and input it into the trained strategy network to obtain the output strategy set.
[0031] In this embodiment, the target model realizes obstacle type recognition and action prediction based on the real-time state feature sequence, realizes the division of dynamic obstacles and static obstacles, and provides data support for the obstacle avoidance behavior of the intelligent cleaning vehicle through obstacle action prediction. The strategy network is used to form a strategy set for obstacle avoidance control of the intelligent cleaning vehicle, and all automatic obstacle avoidance behaviors completed during the obstacle avoidance process are provided, thereby improving the quality of the optimal obstacle avoidance behavior that can be selected.
[0032] Embodiment 2 Reference Figure 1 and Figure 2 , Example 2 further illustrates an automatic obstacle avoidance method and system for a cleaning vehicle proposed by the present invention.
[0033] A method for automatic obstacle avoidance of a cleaning vehicle comprises step S3: inputting a strategy set into an improved adversarial network GGAN model, and outputting an optimal strategy selection of an intelligent cleaning vehicle under the constraints of low-carbon energy-saving conditions.
[0034] The GGAN model uses an improved GAN adversarial network to analyze the associated features of a strategy set containing multiple data and the best strategy selection, and generates a new GGAN model based on the associated features obtained through training. The improved adversarial network in the GGAN model consists of a generator and a discriminator, and generates realistic data through adversarial training. GGAN model construction includes: The GGAN model sets up a two-layer adversarial network, namely the GAN1 adversarial network and the GAN2 adversarial network, and the adversarial network consists of a generator and a discriminator. The GAN1 adversarial network is used to analyze the associated features of a strategy set containing multiple data and the optimal strategy selection, and the GAN2 adversarial network is used to generate an adversarial loss function. The GAN1 adversarial network sets the generator structure to adopt a multi-input neural network structure to receive multiple data of the strategy set as input. According to the strategy selection in the strategy set, two feature groups (a smart cleaning vehicle parameter control guidance and an obstacle avoidance path) are included. The number of data of the smart cleaning vehicle parameter control guidance is set to n1, and the number of data of the obstacle avoidance path is set to n2. Then, the input dimension of the generator is set to n1+n2.
[0035] GAN1 adversarial network sets a constraint layer, which is implemented by constraint function Limit the generation of the best strategy selection and set the constraint function for, ; is the power efficiency coefficient, Indicates the quality of the smart cleaning vehicle, , and represents the correlation coefficient, Indicates the operating temperature of the power equipment.
[0036] Diversity control is usually achieved by designing a loss function. In generating the GAN1 adversarial network, a diversity loss is introduced, which encourages the generator to generate diverse samples. The design of the diversity loss can be based on the difference between the generated samples.
[0037] Set the diversity control loss function to be.
[0038] ; in, To combat the loss function, To improve the Euclidean distance loss function, To adjust the parameters, ; in, is the number of samples generated, , Indicates that the feature dimension is And the labels are , of the sample.
[0039] Adversarial Loss Function for: ; in, is the objective function of the entire generative adversarial network, where is the discriminator network, is the generator network; is the reconstruction loss, which measures the optimal strategy choice for the predictions generated by the generator and the true best strategy selection The reconstruction error between is a metric loss that measures the optimal strategy choice for the predictions generated by the generator With its output The measurement error between is a generative adversarial loss used to make the predicted optimal policy choices produced by the generator more difficult for the discriminator to distinguish.
[0040] This formula reflects the game process in which the generator seeks to generate the best strategy choice that is close to reality, while the discriminator seeks to accurately distinguish the truth from the false.
[0041] Construct an experimental set, which includes multiple groups of strategy sets and the corresponding optimal strategy selections. Input the data in the experimental set into the GGAN model in batches. The input of the GGAN model is the strategy set in the experimental set, and the output is the predicted optimal strategy selection. Use the function value of the diversity control loss function as a supervisory signal, and update the parameters in the framework network through back propagation. When the function value of the mixed loss function is minimized and does not change, a trained GGAN model is obtained.
[0042] Step S4: Select and control the intelligent cleaning vehicle to automatically avoid obstacles according to the output optimal strategy.
[0043] The intelligent cleaning vehicle dynamically adjusts the parameters of each device according to the parameter control guidance in the optimal strategy selection. The adjustment method can be PID control, and the obstacle avoidance route is moved according to the obstacle avoidance path to realize automatic obstacle avoidance of the intelligent cleaning vehicle.
[0044] In this embodiment, the computational occupancy of the loss function constraint is reduced by improving the double-layer adversarial network in the adversarial network GGAN model. By setting the constraint layer in low-carbon energy saving, the optimal strategy selection of the intelligent cleaning vehicle is output in the strategy set under the constraint condition, and the optimal operational control of the obstacle avoidance behavior of the cleaning vehicle is achieved under low-carbon energy saving conditions, thereby realizing good, stable, green and low-carbon obstacle avoidance control management of the intelligent cleaning vehicle.
[0045] Embodiment 3 Reference Figure 1 and Figure 2 , Example 3 further illustrates an automatic obstacle avoidance method and system for a cleaning vehicle proposed by the present invention.
[0046] An automatic obstacle avoidance method for a cleaning vehicle, the automatic obstacle avoidance method for a cleaning vehicle comprising the following steps: Step S1: Obtain feature data of the obstacle at the current moment and construct a real-time state feature sequence; the trained target model identifies the obstacle and predicts the action according to the real-time state feature sequence; Step S2: Obtain the historical ultimate feature sequence output by the target model, collect the historical obstacle avoidance data of the smart cleaning vehicle corresponding to the historical ultimate feature sequence, establish a four-tuple set, use the four-tuple set to train the strategy network, and input the current motion state of the smart cleaning vehicle into the trained strategy network to obtain the output strategy set; Step S3: input the strategy set into the improved adversarial network GGAN model, and output the optimal strategy selection of the smart cleaning vehicle under the constraints of low-carbon energy-saving conditions; Step S4: Select and control the intelligent cleaning vehicle to automatically avoid obstacles according to the output optimal strategy.
[0047] The automatic obstacle avoidance system for the cleaning vehicle comprises: A sample storage module is used to store feature data and historical obstacle avoidance data of multiple categories of obstacles within a historical period; A data acquisition module, which is used to collect characteristic data of obstacles at the current moment and the motion state of the intelligent cleaning vehicle at the current moment; Data processing module, including target model, policy network and adversarial network GGAN model. The target model is used for obstacle recognition and action prediction. The policy network is used to form the smart cleaning vehicle strategy set. The adversarial network GGAN model is used to output the best strategy selection in the strategy set. The adjustment module selects the best strategy to automatically avoid obstacles for the smart cleaning vehicle.
[0048] In addition, according to the implementation of the present application, the process described in the figure of the automatic obstacle avoidance method for a cleaning vehicle can be implemented as a computer software program. For example, the present application provides a non-transitory machine-readable storage medium, which stores machine-readable instructions, and the machine-readable instructions can be executed by a processor to execute instructions corresponding to the method steps provided in the present application. Of course, the architecture shown in the figure of the automatic obstacle avoidance method for a cleaning vehicle is only exemplary. When implementing different devices, adaptive selection or adjustment can be made according to actual needs.
[0049] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.
[0050] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technical users in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.
Claims
1. A method for automatic obstacle avoidance of a cleaning vehicle, characterized in that: The automatic obstacle avoidance method for a cleaning vehicle comprises the following steps: Step S1: Obtain feature data of the obstacle at the current moment and construct a real-time state feature sequence; the trained target model identifies the obstacle and predicts the action according to the real-time state feature sequence; Step S2: Obtain the historical ultimate feature sequence output by the target model, collect the historical obstacle avoidance data of the smart cleaning vehicle corresponding to the historical ultimate feature sequence, establish a four-tuple set, use the four-tuple set to train the strategy network, and input the current motion state of the smart cleaning vehicle into the trained strategy network to obtain the output strategy set; Step S3: input the strategy set into the improved adversarial network GGAN model, and output the optimal strategy selection of the smart cleaning vehicle under the constraints of low-carbon energy-saving conditions; Step S4: Select and control the intelligent cleaning vehicle to automatically avoid obstacles according to the output optimal strategy.
2. The automatic obstacle avoidance method for a cleaning vehicle according to claim 1, characterized in that: The training of the target model in step S1 includes: Using the convolutional neural network CNN as the framework, an external connection layer is constructed. The external connection layer includes a CRFs conditional random field, which is denoted as a CRF layer. The category feature sequence extracted by the convolutional neural network CNN is used as the input of the CRF layer. The CRF layer will use the category feature sequence to perform action prediction inference; Obtain a sample set, divide the data in the sample set into multiple batches, each batch includes multiple data groups, each data group includes feature data of multiple categories of obstacles in multiple historical cycle times, and construct a historical state feature sequence of each category of obstacles in each cycle time; The data in the sample set is input into the initial target model in batches for training, and the forward propagation is performed to calculate the loss function. The back propagation calculates the gradient of each parameter of the initial target model, and the SGD optimizer uses the gradient to update the parameters of the initial target detection model, that is, to complete an iteration. After each batch is trained, the loss function value is calculated. When the loss function value reaches the threshold, the training is stopped to obtain the final target model. The loss function of the target model is set as: ， represents the number of data groups in each batch, Indicates the time period number of each obstacle in each data set, It is the actual ultimate characteristic sequence of the obstacle; is the ultimate characteristic sequence of obstacles predicted by the model; The real-time state feature sequence of the obstacle at the current moment is imported into the trained target model, and the category and action prediction of the obstacle are analyzed to obtain the real-time ultimate feature sequence.
3. The automatic obstacle avoidance method for a cleaning vehicle according to claim 2, characterized in that: The characteristic data of the obstacle includes location information data , speed information data and size information data ; Historical cycle time Discretization to obtain the location information data of obstacles at each moment , speed information data and size information data , the location information data Use binary encoding, the encoding length is p; Speed information data Normalize and map to a real number; convert the size information data Use binary encoding, the encoding length is m; Constructing historical state feature sequences ,Will Input target model, , and Indicates the cycle time Location information data corresponding to the moment , speed information data and size information data , , and within the cycle time, represents the initial moment of the obstacle, Indicates the ultimate moment of the obstacle and outputs the conjugate feature sequence , Indicates the category of the obstacle; right Apply CRFs conditional random fields to obtain the ultimate feature sequence of obstacles , Indicates the obstacle's motion prediction data.
4. The automatic obstacle avoidance method for a cleaning vehicle according to claim 3, characterized in that: The historical obstacle avoidance data in S2 includes the start and end parameter control data of the cleaning vehicle , Multi-dimensional reward data and conditional data ; Start and end parameter control data Including initial parameter control data and result parameter control data, the initial parameter control data packet and the result parameter control data packet respectively include the control parameters of each device state of the cleaning vehicle at the beginning and end of obstacle avoidance; Quadruple Set Including the ultimate characteristic sequence of history , start and end parameter control data , Multi-dimensional reward data and conditional data , The multi-dimensional reward data includes obstacle avoidance quality coefficient, safety coefficient and time consumption coefficient; Multi-dimensional reward data ; In the formula, is the first-order derivative of the driving curve of the intelligent cleaning vehicle, is the second-order derivative of the driving curve of the intelligent cleaning vehicle, The time required to complete obstacle avoidance, Edge track for smart cleaning vehicles Interference coefficient with obstacles; The strategy set includes multiple groups of strategy selections, each group of strategy selections includes a parameter control guidance and an obstacle avoidance path for an intelligent cleaning vehicle; the parameter control guidance is used to guide the parameters of the equipment control of the intelligent cleaning vehicle during the obstacle avoidance process, and the obstacle avoidance path is the path that the intelligent cleaning vehicle travels during the obstacle avoidance process.
5. The automatic obstacle avoidance method for a cleaning vehicle according to claim 4, characterized in that: The strategy network training process includes: Construct a policy network framework. The policy network uses a combination of support vector machine (SVM) and radial basis function (RBF) kernel. Set the quaternary set as the input of the policy network and perform standardization. Set the policy set as the output. Set the radial basis function (RBF) kernel as a Gaussian function. Adjust the parameters of the RBF kernel through cross-validation. Construct a training set, which includes multiple sets of four-tuple sets and corresponding policy sets. Input the data in the training set into the policy network in batches, and continuously adjust the network parameters through iterative training until all batches in the training set are input, and obtain the optimized policy network. The motion state of the smart cleaning vehicle at the current moment is used to establish a four-tuple set at the current moment and input it into the trained strategy network to obtain the output strategy set.
6. The automatic obstacle avoidance method for a cleaning vehicle according to claim 5, characterized in that: The GGAN model construction in S3 includes: The GGAN model sets a two-layer adversarial network, namely the GAN1 adversarial network and the GAN2 adversarial network. The GAN1 adversarial network is used to analyze the correlation characteristics between the strategy set containing multiple data and the optimal strategy selection. The GAN1 adversarial network sets the generator structure to adopt a multi-input neural network structure and sets the constraint layer. Construct an experimental set, which includes multiple groups of strategy sets and the corresponding optimal strategy selections. Input the data in the experimental set into the GGAN model in batches. The input of the GGAN model is the strategy set in the experimental set, and the output is the predicted optimal strategy selection. Use the function value of the diversity control loss function as a supervisory signal, and update the parameters in the framework network through back propagation. When the function value of the mixed loss function is minimized and does not change, a trained GGAN model is obtained.
7. The automatic obstacle avoidance method for a cleaning vehicle according to claim 6, characterized in that: The constraint layer is defined by the constraint function Limit the generation of the best strategy selection and set the constraint function for: ； is the power efficiency coefficient, Indicates the quality of the smart cleaning vehicle, , and represents the correlation coefficient, Indicates the operating temperature of the power equipment.
8. The automatic obstacle avoidance method for a cleaning vehicle according to claim 6, characterized in that: The GAN2 adversarial network is used to generate adversarial loss functions, introduce diversity loss, and set the diversity control loss function as: ； in, To combat the loss function, To improve the Euclidean distance loss function, To adjust the parameters.
9. The automatic obstacle avoidance method for a cleaning vehicle according to claim 6, characterized in that: The improved Euclidean distance loss function is: ； in, is the number of samples generated, , Indicates that the feature dimension is And the labels are , Samples of Adversarial Loss Function for: ； in, is the objective function of the entire generative adversarial network, where is the discriminator network, is the generator network; is the reconstruction loss, which measures the optimal strategy choice for the predictions generated by the generator and the true best strategy selection The reconstruction error between is a metric loss that measures the optimal strategy choice for the predictions generated by the generator With its output The measurement error between is a generative adversarial loss used to make the predicted optimal policy choices produced by the generator more difficult for the discriminator to distinguish.
10. An automatic obstacle avoidance system for a cleaning vehicle, based on an automatic obstacle avoidance method for a cleaning vehicle according to any one of claims 1 to 9, characterized in that: The automatic obstacle avoidance system for the cleaning vehicle comprises: A sample storage module is used to store feature data and historical obstacle avoidance data of multiple categories of obstacles within a historical period; A data acquisition module, which is used to collect characteristic data of obstacles at the current moment and the motion state of the intelligent cleaning vehicle at the current moment; Data processing module, including target model, policy network and adversarial network GGAN model. The target model is used for obstacle recognition and action prediction. The policy network is used to form the smart cleaning vehicle strategy set. The adversarial network GGAN model is used to output the best strategy selection in the strategy set. The adjustment module selects the best strategy to automatically avoid obstacles for the smart cleaning vehicle.