Multifunctional automotive upholstery skip car

Abstract

The multifunctional automobile interior trim part skip car comprises a car frame composed of four supporting columns and a bottom plate, a plurality of layers of net-shaped storage racks are arranged on the car frame, and four wheels are arranged at the bottom of the car frame; a load sensor is arranged on a bottom plate of the frame, a displayer and an energy device are arranged on a supporting column of the frame, an algorithm control system is arranged below the displayer, a visual sensor is arranged on the top of the frame, a laser radar is arranged on a beam on the top of the frame, and a motor is arranged on the outer side of the bottom plate of the frame. An IMU (inertial measurement unit) is arranged in the center of the inner side of a bottom plate of the frame, and a blowing cleaning system is further arranged on a cross beam at the top of the frame and comprises an air outlet and an air blower. Automatic cleaning of the skip car is achieved through various sensors, an algorithm control system and an air blowing cleaning system, and the problem of product scrapping caused by smudginess and improper transportation of the skip car is solved.

Description

Technical Field

[0001] The invention relates to the technical field of cleaning control of a material cart, and in particular to a multifunctional material cart for automobile interior decoration parts. Background Art

[0002] At present, the existing material cart adopts a sliding sleeve and spring structure, has a telescopic function, and can save space. However, this design can only carry the product and has no other functions.

[0003] For example, patent CN214566839U "A folding material rack for easy storage" discloses a folding material rack for easy storage, including two fixed frames, one side of the fixed frame is provided with four installation grooves in a rectangular array, a connecting rod is provided between the inner cavities of the two installation grooves, one end of the connecting rod is provided with a fixed block, one side of the fixed block is symmetrically provided with a first limit rod, a limit groove is provided on the inner wall of one side of the inner cavity of the installation groove, a second limit rod is provided in the inner cavity of the limit groove, a spring is slidably sleeved on the outer wall of the second limit rod, a clamping block is provided at one end of the spring and the second limit rod, and three bearing rods are provided between the inner walls on both sides of the fixed frame.

[0004] It can be seen that the current material cart design function is too single and cannot provide other functions, making it difficult to solve the problem of product scrapping caused by dirty material carts and improper transportation. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a multifunctional automobile interior trim material cart, the material cart comprising: a frame consisting of four support columns and a bottom plate, the frame being provided with a multi-layer mesh storage rack and having four wheels at the bottom; a load sensor being provided on the bottom plate of the frame, a display and an energy device being provided on the support columns of the frame, an algorithm control system being provided below the display, a visual sensor being provided on the top of the frame, a laser radar being provided on the top crossbeam of the frame, a motor being provided on the outer side of the bottom plate of the frame, an IMU inertial measurement unit being provided at the center of the inner side of the bottom plate of the frame, and an air blowing cleaning system being provided on the top crossbeam of the frame, the air blowing cleaning system comprising an air outlet and a blower; the algorithm control system is connected to the load sensor, the display, the visual sensor, the laser radar, the motor, the IMU inertial measurement unit and the air blowing cleaning system;

[0006] The load sensor is used to measure the load weight of the trolley and send it to the algorithm control system;

[0007] The display is used to provide a user interaction interface;

[0008] The energy device is used to provide power support for all devices on the skip;

[0009] The algorithm control system is used to perform dirt detection and adaptive speed control on the material vehicle according to the vehicle-mounted data collected by each sensor;

[0010] The visual sensor is used to detect the contamination level of each shelf and send it to the algorithm control system;

[0011] The laser radar is used to measure ground undulations and obstacles and send them to the algorithm control system;

[0012] The motor is used to provide power to the wheels of the trolley under the control of the algorithm control system;

[0013] The IMU inertial measurement unit is used to measure the slope and bump conditions and send them to the algorithm control system;

[0014] The air blowing cleaning system is used to clean the material vehicle by the gas sprayed from the blower and the air outlet under the control of the algorithm control system.

[0015] Preferably, the visual sensor detects dirt in the material car in combination with the computer vision technology and deep learning technology of the algorithm control system;

[0016] Different types of dirt on the skip are identified by training a deep learning model, and the training process of the deep learning model includes:

[0017] (1) Data Collection

[0018] Collecting a data set as training samples, marking the location and category of stains for each image in the data set, and using image enhancement technology to generate more training samples, the data set includes images of various types of stains inside the car and images under various environmental conditions, the various types of stains include dust, mud, and wet stains, and the various environmental conditions include different lighting, angles, and stain coverage;

[0019] (2) Data preprocessing

[0020] Before model training, the size of the images in the data set is adjusted to a fixed size required by the deep learning model, the pixel values ​​of the images are standardized, and the data set is divided into a training set, a validation set, and a test set;

[0021] (3) Selecting the model architecture

[0022] Use traditional convolutional neural networks, modern deep networks, and object detection models;

[0023] (4) Model training

[0024] The training process includes initializing network parameters, selecting loss functions and optimization algorithms, forward propagation and back propagation, and hyperparameter tuning;

[0025] (5) Model evaluation and testing

[0026] After training, the model is evaluated using the test set, and the evaluation indicators include accuracy, recall, precision, F1-score and IoU;

[0027] (6) Model optimization and deployment

[0028] After the model training is completed and verified by the test set, the model is optimized to improve the reasoning speed and deployed based on the hardware environment to ensure that the model has sufficient computing power for real-time detection;

[0029] (7) Continuous Improvement

[0030] The data set and the optimization model are continuously updated through online learning and incremental learning.

[0031] Preferably, the model training includes:

[0032] The network parameters are initialized as follows: the weights in the network need to be randomly initialized, and the initialization methods include Gaussian distribution initialization, Xavier initialization and He initialization;

[0033] The selected loss function includes: the loss function is used to measure the gap between the prediction result and the true label. For the dirt detection task, if it is a classification task, the cross entropy loss function is used to calculate the error between the model prediction category and the true category; if it is a regression task, including predicting the area or coverage of the stain, the mean square error is used; if it is a multi-task, that is, classification and positioning are performed simultaneously, a weighted loss function is used to weight the classification and regression losses;

[0034] The selected optimization algorithm includes: an optimization algorithm for adjusting weights in the network to minimize the loss function, the optimization algorithm includes: a stochastic gradient descent suitable for large-scale data sets, and an Adam optimizer that combines the advantages of momentum and adaptive learning rate for complex deep learning tasks;

[0035] The forward propagation includes: passing the input image through each layer of the network to finally obtain the predicted output;

[0036] The back propagation includes: calculating the gradient of the loss function for each weight, and updating the weight by gradient descent; during the training process, batch gradient descent is used, that is, a small batch of samples is used for each update instead of all training samples;

[0037] The hyperparameter tuning includes: evaluating the model performance through the validation set, and adjusting the hyperparameters according to the evaluation results, including the learning rate, batch size, and number of network layers, and finding the optimal hyperparameter combination through cross-validation or grid search.

[0038] Preferably, the selection model architecture includes:

[0039] Use traditional convolutional neural networks to handle simple image classification tasks;

[0040] Use modern deep networks to handle complex image recognition tasks, capable of extracting deeper features;

[0041] Use object detection models for tasks involving localizing and detecting the location of stains;

[0042] The evaluation indicators for model evaluation and testing include:

[0043] For classification tasks, accuracy is used to measure the proportion of samples that the model predicts correctly;

[0044] Use recall and precision to measure the model's ability to correctly identify stains. Recall focuses on the proportion of stains detected, while precision focuses on whether the detected stains are real stains.

[0045] For tasks with imbalanced categories, use the F1-score metric that takes both recall and precision into account;

[0046] For the object detection task, IoU is used to evaluate the overlap between the detection box and the actual stain area.

[0047] Preferably, the visual sensor detects dirt in the material car in combination with the computer vision technology and deep learning technology of the algorithm control system, including:

[0048] The visual sensor collects images inside the material car;

[0049] The algorithm control system uses computer vision technology to convert the image collected by the visual sensor into a color space, converts the RGB mode into a grayscale / HSV mode, removes noise through Gaussian filtering, and enhances the contrast of stains;

[0050] The algorithm control system uses a pre-trained convolutional neural network, a modern deep network and a target detection model to classify dirt to obtain the dirt category and pollution degree. The training data set consists of different types of dirt, including dust, mud and particulate matter. When the pollution degree, that is, the dirt coverage area> the preset threshold, the blowing cleaning system is started.

[0051] Preferably, the air outlet includes a plurality of high-pressure air nozzles evenly arranged on the top of the frame, and the air nozzles can adjust the angle according to the detection results;

[0052] The blowing cleaning system includes three wind speed settings: low wind speed, medium wind speed and high wind speed. The low wind speed is used for low-level cleaning, suitable for light dust, the medium wind speed is used for medium-level cleaning, suitable for particulate matter, and the high wind speed is used for high-level cleaning, suitable for wet dirt.

[0053] The algorithm control system collects the dirt type and dirt coverage area in the material car, and based on the dirt type and dirt coverage area, uses a fuzzy control algorithm to control the wind force and wind nozzle direction of the air blowing cleaning system.

[0054] Preferably, the algorithm control system performs adaptive speed control on the material vehicle, including:

[0055] The load sensor measures the load weight of the truck; the laser radar measures ground undulations and obstacles; the IMU inertial measurement unit is equipped with an accelerometer and a gyroscope to measure slope and bumps;

[0056] The algorithm control system calculates the speed adjustment value using a fuzzy control algorithm based on the vehicle-mounted data collected by the load sensor, the laser radar and the IMU inertial measurement unit, including load weight, slope, and ground smoothness, and controls the motor according to the speed adjustment value.

[0057] Preferably, the algorithm control system uses reinforcement learning to optimize the speed regulation strategy of the material vehicle, specifically uses a DQN model to train the speed regulation strategy, uses experience replay and target network to improve training stability, and continuously optimizes the strategy during training to make the vehicle speed regulation more intelligent.

[0058] Preferably, the implementation process of the DQN model includes:

[0059] (1) Environment settings

[0060] For the state space, the percentage of load is used to indicate the proportion of the weight of the materials in the carriage to the maximum load. The angle of slope is used to indicate the slope of the road section where the vehicle is currently located, which affects the acceleration and braking of the vehicle. The degree of bumpiness is used to indicate the degree of bumpiness of the current road surface conditions of the vehicle, which affects the stability of the vehicle speed.

[0061] The state space is defined as the vector S:

[0062] S = [load, slope, bumpiness]

[0063] For the action space, at each moment, the shuttle can choose the following three actions: acceleration, i.e. +am / s, maintenance, i.e. 0m / s, deceleration, i.e. -am / s, where a is a positive number;

[0064] The action space is defined as the vector A:

[0065] A=[+am / s,0m / s,-am / s]

[0066] (2) Designing the Reward Function

[0067] The reward function is the core of reinforcement learning and is used to guide the material truck to optimize its decision. According to the task requirements, the reward function is designed as follows: smooth driving, when the bumpiness is minimized, it means that the vehicle speed is stable, and a +m reward is given; speed optimization, when there is no unnecessary deceleration, it means that the vehicle speed has been effectively optimized, and a +n reward is given; avoid speeding, if the vehicle speed is too fast and there is a risk of tilting or material dumping, a -m reward is given; where m and n are both positive numbers;

[0068] The reward function is expressed as:

[0069]

[0070] (3) State Transfer

[0071] At each step, after the skip performs an action, the system moves to the next state based on the current state and the selected action, which is usually determined by the dynamics of the environment;

[0072] (4) Q function update rules

[0073] The core of the DQN model is to learn the Q value function Q(s) through a neural network. t ,a t ; θ), where θ represents the parameters of the Q network, which are continuously adjusted through back propagation, with the goal of minimizing the following loss function:

[0074]

[0075] Among them, a represents the learning rate, which controls the update step size; r t Represents the reward at the current moment; s t represents the current state; at represents the current action; γ represents the discount factor, which determines the weight of future rewards; Q(s t ,a t ; θ) represents the Q value of the action selected in the current state; θ - Represents the parameters of the target network, which is used to stabilize the training process; Represents the maximum Q value in the next state, which is used to estimate the optimal future return;

[0076] (5) Experience Replay

[0077] The DQN model adopts an experience replay mechanism to improve training efficiency and stability by storing and reusing the state, action, reward and next state of the car. The experience generated by the interaction between the car and the environment is stored in the experience replay pool, and then small batches of data are randomly sampled from the experience replay pool for training to break the time correlation of the data and increase sample utilization.

[0078] (6) Target network

[0079] The DQN model uses a target network, which is a delayed version of the Q network. The parameters of the target network are updated only at certain training steps, while the parameters of the Q network are updated at each step, avoiding drastic fluctuations in the Q value and improving the stability of training.

[0080] Preferably, the training process of the DQN model includes a training phase and a reasoning and control phase;

[0081] The training phase includes:

[0082] (1) Initialization

[0083] Initialize the parameters θ of the Q network and θ of the target network - , initialize the experience replay pool;

[0084] (2) Training process

[0085] At each time step t, the agent takes the current state s t Select an action t , the selection strategy adopts the ε-greedy strategy, that is, a random action is selected with probability ∈, and the action with the maximum Q value is selected at other times;

[0086] After executing the action, the system returns the reward r t and the next state s t +1;

[0087] Will (s t ,a t ,r t ,s t +1) Deposit into the experience replay pool;

[0088] Randomly sample batches of data from the experience replay pool to train the Q network and update the parameters θ of the Q network;

[0089] At certain intervals, the parameters of the Q network are copied to the target network.

[0090] (3) Strategy Update

[0091] Each time the Q network is updated, the Q value of the current state and action is compared with the target Q value, and the parameters of the Q network are updated so that the network can learn a more appropriate strategy;

[0092] The reasoning and control phase includes:

[0093] The load sensor, the laser radar and the IMU inertial measurement unit collect the vehicle data of the material truck, including load weight, slope and ground stability.

[0094] Using the current vehicle data, predict the current state s t ;

[0095] The current state s t Input the trained DQN model to calculate and select the optimal action, including acceleration, deceleration or maintaining speed;

[0096] According to the selected action, the motor is controlled to perform corresponding acceleration or deceleration.

[0097] Compared with the prior art, the present invention has the following beneficial effects:

[0098] The present invention measures the load weight of the material truck through a load sensor, detects the pollution degree of each layer of the storage rack through a visual sensor, measures the ground undulations and obstacles through a laser radar, and measures the slope and bumps through an IMU inertial measurement unit.

[0099] The present invention uses an algorithm control system to perform dirt detection and adaptive speed control on the material truck according to the on-board data collected by various sensors and laser radars.

[0100] The present invention uses a blowing cleaning system to clean the material cart using the ejected gas under the control of an algorithm control system. BRIEF DESCRIPTION OF THE DRAWINGS

[0101] Figure 1 The present invention provides a schematic structural diagram of a multifunctional automobile interior trim material trolley. DETAILED DESCRIPTION

[0102] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0103] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0104] like Figure 1As shown, an embodiment of the present invention provides a multifunctional automobile interior trim material cart, including: a frame composed of four support columns and a bottom plate, a multi-layer mesh storage rack is provided on the frame, and four wheels are provided at the bottom; a load sensor 100 is provided on the bottom plate of the frame, a display 200 and an energy device 300 are provided on the support columns of the frame, an algorithm control system 400 is provided below the display 200, a visual sensor 500 is provided on the top of the frame, a laser radar 600 is provided on the top crossbeam of the frame, a motor 700 is provided on the outer side of the bottom plate of the frame, an IMU inertial measurement unit 800 is provided in the center of the inner side of the bottom plate of the frame, and a blowing cleaning system is also provided on the top crossbeam of the frame, and the blowing cleaning system includes an air outlet 900 and a blower 1000; the algorithm control system 400 is connected with the load sensor 100, the display 200, the visual sensor 500, the laser radar 600, the motor 700, the IMU inertial measurement unit 800 and the blowing cleaning system;

[0105] The load sensor 100 is used to measure the load weight of the material truck and send it to the algorithm control system 400;

[0106] Display 200, used to provide a user interaction interface;

[0107] Energy device 300, used to provide power support for all devices on the skip;

[0108] An algorithm control system 400 for performing dirt detection and adaptive speed control on the material vehicle based on the vehicle-mounted data collected by various sensors;

[0109] Visual sensor 500, used to detect the contamination level of each shelf and send it to algorithm control system 400;

[0110] Laser radar 600, used to measure ground undulations and obstacles and send them to the algorithm control system 400;

[0111] The motor 700 is used to provide power to the wheels of the feeder vehicle under the control of the algorithm control system 400;

[0112] IMU inertial measurement unit 800, used to measure slope and bump conditions and send them to the algorithm control system 400;

[0113] The air blowing cleaning system is used to clean the material cart by the gas sprayed by the blower 1000 and the air outlet 900 under the control of the algorithm control system 400.

[0114] In the embodiment of the present invention, the visual sensor 500 detects dirt in the material truck by combining the computer vision technology (OpenCV) and deep learning technology (CNN) of the algorithm control system 400;

[0115] The deep learning model is trained to identify different types of dirt on the skip. The training process of the deep learning model includes:

[0116] (1) Data Collection

[0117] Collect a dataset as training samples, annotate the location and category of stains for each image in the dataset, and use image enhancement technology to generate more training samples. The dataset includes images of various stains inside the car and images under various environmental conditions. Various stains include dust, mud, and wet stains. Various environmental conditions include different lighting, angles, and stain coverage.

[0118] First, in order to train a deep learning model, a large and representative dataset must be collected. These datasets contain images of various types of stains inside the car, including dust, mud, wet stains, etc. The dataset includes images under various environmental conditions, such as different lighting, angles, and stain coverage.

[0119] Image labeling: Each image needs to be clearly labeled with the location and type of stains (such as dust, mud, wet stains, etc.). This can be done manually or using semi-automatic annotation tools.

[0120] Image enhancement: To improve the robustness of the model, image enhancement techniques (such as rotation, scaling, cropping, brightness adjustment, etc.) can be used to generate more training samples.

[0121] (2) Data preprocessing

[0122] Before model training, the size of the images in the dataset was adjusted to the fixed size required by the deep learning model, the pixel values ​​of the images were standardized, and the dataset was divided into training set, validation set, and test set;

[0123] Before neural network training, image data needs to be preprocessed. The preprocessing steps include:

[0124] Image resizing: All input images are resized to a fixed size required by the neural network (e.g. 224x224 pixels) to ensure consistency in the input layer.

[0125] Normalization: Normalize the pixel values ​​of the image to the range [0,1], which helps speed up the training process and improves convergence rate.

[0126] Data Split: Divide the dataset into training set, validation set and test set. We divide it into 80% of the data for training, 10% for validation and 10% for final testing of the model’s performance.

[0127] (3) Selecting the model architecture

[0128] Use traditional convolutional neural networks, modern deep networks, and object detection models;

[0129] Choosing a model architecture involves:

[0130] Use traditional convolutional neural networks (LeNet) to handle simple image classification tasks;

[0131] Use modern deep networks (ResNet) to handle complex image recognition tasks, capable of extracting deeper features;

[0132] An object detection model (YOLO) is used to handle tasks involving localization and detection of stain locations (e.g., the distribution of dust).

[0133] (4) Model training

[0134] The training process includes initializing network parameters, selecting loss functions and optimization algorithms, forward propagation and back propagation, and hyperparameter tuning;

[0135] Initializing network parameters includes: the weights in the network need to be randomly initialized, and the initialization methods include Gaussian distribution initialization, Xavier initialization, and He initialization;

[0136] The loss function is selected as follows: the loss function is used to measure the gap between the predicted result and the true label. For the dirt detection task, if it is a classification task, the cross entropy loss function is used to calculate the error between the model prediction category and the true category; if it is a regression task (such as predicting the area or coverage of the stain), including predicting the area or coverage of the stain, the mean square error (MSE) is used; if it is a multi-task, that is, classification and positioning are performed simultaneously, the weighted loss function is used to weight the classification and regression losses;

[0137] Select optimization algorithms, including: Optimization algorithms are used to adjust the weights in the network to minimize the loss function. Optimization algorithms include: Stochastic Gradient Descent (SGD) for large-scale datasets, and Adam optimizer, which combines the advantages of momentum and adaptive learning rate for complex deep learning tasks;

[0138] Forward propagation includes: passing the input image through each layer of the network to finally get the predicted output;

[0139] Backpropagation includes: calculating the gradient of the loss function for each weight and updating the weights by gradient descent; during training, batch gradient descent is used, that is, each update uses a small batch of samples instead of all training samples;

[0140] Hyperparameter tuning includes: evaluating model performance through a validation set, adjusting hyperparameters based on the evaluation results, including learning rate, batch size, and number of network layers, and finding the optimal hyperparameter combination through cross-validation or grid search.

[0141] (5) Model evaluation and testing

[0142] After training, the model is evaluated using the test set. The evaluation indicators include accuracy, recall, precision, F1-score, and IoU (Intersection over Union).

[0143] The evaluation indicators for model evaluation and testing include:

[0144] For classification tasks, accuracy is used to measure the proportion of samples that the model predicts correctly;

[0145] Use recall and precision to measure the model's ability to correctly identify stains. Recall focuses on the proportion of stains detected, while precision focuses on whether the detected stains are real stains.

[0146] For tasks with imbalanced categories, use the F1-score metric that takes both recall and precision into account;

[0147] For the object detection task, IoU is used to evaluate the overlap between the detection box and the actual stain area.

[0148] (6) Model optimization and deployment

[0149] After the model training is completed and verified by the test set, the model is optimized to improve the inference speed and deployed based on the hardware environment to ensure that the model has sufficient computing power for real-time detection;

[0150] Once the training is completed and verified by the test set, the model can be optimized (such as quantization, pruning) to improve the inference speed and finally deployed to the actual application. During the deployment process, the support of the hardware environment needs to be considered to ensure that the model has sufficient computing power for real-time detection.

[0151] (7) Continuous Improvement

[0152] Continuously update data sets and optimize models through online learning and incremental learning.

[0153] Dirt detection is a long-term project. As the cabin environment changes and new types of stains emerge, it is necessary to continuously update the data set and optimize the model. The accuracy and stability of the system can be continuously improved through online learning, incremental learning, etc.

[0154] In the embodiment of the present invention, the visual sensor 500 detects dirt in the material truck by combining the computer vision technology and deep learning technology of the algorithm control system 400, including:

[0155] The visual sensor collects images inside the 500-meter material car;

[0156] The algorithm control system 400 uses computer vision technology to convert the image collected by the visual sensor 500 into a color space, converts the RGB mode into a grayscale / HSV mode, removes noise through Gaussian filtering, and enhances the contrast of the stain;

[0157] The algorithm control system 400 uses pre-trained convolutional neural networks, modern deep networks and target detection models to classify dirt and obtain dirt categories and pollution levels (0-100%). The training data set consists of different types of dirt, including dust, mud and particulate matter. When the pollution level, that is, the dirt coverage area > the preset threshold (30%), the blowing cleaning system is started.

[0158] In the embodiment of the present invention, the air outlet 900 includes a plurality of high-pressure air nozzles evenly arranged on the top of the frame, and the air nozzles can adjust the angle according to the detection results;

[0159] The blowing cleaning system includes three wind speed adjustments: low wind speed, medium wind speed and high wind speed. The low wind speed is used for low-level cleaning and is suitable for light dust. The medium wind speed is used for medium-level cleaning and is suitable for particulate matter. The high wind speed is used for high-level cleaning and is suitable for wet dirt.

[0160] The algorithm control system 400 collects the dirt type and dirt coverage area in the material truck, and based on the dirt type and dirt coverage area, uses a fuzzy control algorithm to control the wind force and wind nozzle direction of the air blowing cleaning system.

[0161] Input variables: dirt type (dust / particles / wet dirt), dirt coverage area (0-100%).

[0162] Output variables: wind force (low / medium / high), wind duration (short / medium / long).

[0163] The fuzzy rules are shown in Table 1:

[0164] Table 1

[0165] Dirt Type Dirt coverage area (%) Wind force Blow dry time dust 0-30％ Low short dust 30-60％ middle middle Particles 30-60％ high middle Wet dirt 60-100％ high long

[0166] In the embodiment of the present invention, the algorithm control system 400 performs adaptive speed control on the material vehicle, including:

[0167] The load sensor 100 measures the load weight of the material truck; the laser radar 600 measures ground undulations and obstacles; the IMU inertial measurement unit 800 is equipped with an accelerometer and a gyroscope to measure slope and bumps;

[0168] The algorithm control system 400 calculates the speed adjustment value using a fuzzy control algorithm based on the vehicle data collected by the load sensor 100, the laser radar 600 and the IMU inertial measurement unit 800, including the load weight, slope, and ground smoothness, and controls the motor 700 according to the speed adjustment value.

[0169] 1. Sensor input

[0170] Load sensor (measures the weight of the carriage load).

[0171] IMU inertial measurement unit (accelerometer + gyroscope, measuring slope and bumps).

[0172] LiDAR (to measure ground relief and obstacles).

[0173] 2. Fuzzy control algorithm

[0174] Input variables: load weight (low / medium / high), slope (small / medium / steep), surface smoothness (smooth / medium / bumpy).

[0175] Output variable: vehicle speed adjustment (slow down / maintain / accelerate).

[0176] The fuzzy rules are shown in Table 2:

[0177] Table 2

[0178] Load slope Ground status Speed ​​adjustment Low Small smooth accelerate Low Small Bumping maintain Low steep Bumping slow down middle Small smooth maintain high Small Bumping slow down high steep Bumping slow down

[0179] In the embodiment of the present invention, the algorithm control system 400 adopts reinforcement learning (RL) to optimize the speed regulation strategy of the material vehicle, specifically adopts the DQN model to train the speed regulation strategy, uses experience replay and target network to improve the training stability, and continuously optimizes the strategy during the training process to make the vehicle speed regulation more intelligent.

[0180] In the embodiment of the present invention, the implementation process of the DQN model includes:

[0181] (1) Environment settings

[0182] For the state space, the percentage of load (0%-100%) is used to indicate the proportion of the weight of the materials in the carriage to the maximum load, the angle of slope (0°-30°) is used to indicate the slope of the road section where the vehicle is currently located, which affects the acceleration and braking of the vehicle, and the degree of bumpiness (0%-100%) is used to indicate the degree of bumpiness of the current road surface conditions of the vehicle, such as the roughness of the road, which affects the stability of the vehicle speed;

[0183] The state space is defined as the vector S:

[0184] S = [load, slope, bumpiness]

[0185] For the action space, at each moment, the material vehicle can choose the following three actions: acceleration, which is +am / s, maintenance, which is 0m / s, and deceleration, which is -am / s, where a is a positive number;

[0186] The action space is defined as the vector A:

[0187] A=[+am / s,0m / s,-am / s]

[0188] Assume that acceleration (+1m / s) increases the speed; maintenance (0m / s) maintains the current speed; deceleration (-1m / s) reduces the speed. The action space is defined as: A = [+1m / s, 0m / s, -1m / s].

[0189] (2) Designing the Reward Function

[0190] The reward function is the core of reinforcement learning and is used to guide the material truck to optimize its decision. According to the task requirements, the reward function is designed as follows: smooth driving, when the bumpiness is minimized, it means that the vehicle speed is stable, and a +m reward is given; speed optimization, when there is no unnecessary deceleration, it means that the vehicle speed has been effectively optimized, and a +n reward is given; avoid speeding, if the vehicle speed is too fast and there is a risk of tilting or material dumping, a -m reward is given; where m and n are both positive numbers;

[0191] The reward function is expressed as:

[0192]

[0193] Specifically, according to the task requirements, the reward function is designed as follows:

[0194] Bumpiness Minimization: When the bumpiness is minimized, it means the vehicle is running smoothly, and a +1 point reward is given.

[0195] Speed ​​Optimization: If there is no unnecessary deceleration, the vehicle's speed has been effectively optimized, and a +0.5 point bonus is given.

[0196] Avoid Over-speeding: If the vehicle is driving too fast and there is a risk of tilting or material dumping, a -1 point bonus will be given.

[0197] In summary, the reward function can be expressed as:

[0198]

[0199] (3) State Transfer

[0200] At each step, after the vehicle performs an action, the system moves to the next state based on the current state and the selected action, which is usually determined by the dynamics of the environment; for example, the acceleration or deceleration of the vehicle will be affected by the current load, slope, and bumpiness.

[0201] (4) Q function update rules

[0202] Q-learning is a model-free reinforcement learning algorithm that guides the decision-making of an agent by continuously estimating the state-action value function Q(s,a). The Deep Q Network (DQN) uses a neural network to approximate this value function, allowing it to handle complex, high-dimensional state spaces.

[0203] The core of the DQN model is to learn the Q value function Q(s) through a neural network. t ,a t ; θ), where θ represents the parameters of the Q network, which are continuously adjusted through back propagation, with the goal of minimizing the following loss function:

[0204]

[0205] Among them, a represents the learning rate, which controls the update step size; r t Represents the reward at the current moment; s t represents the current state; at represents the current action; γ represents the discount factor, which determines the weight of future rewards; Q(s t ,a t ; θ) represents the Q value of the action selected in the current state; θ - Represents the parameters of the target network, which is used to stabilize the training process; Represents the maximum Q value in the next state, which is used to estimate the optimal future return;

[0206] (5) Experience Replay

[0207] The DQN model uses an experience replay mechanism to improve training efficiency and stability by storing and reusing the state, action, reward, and next state of the car. The experience generated by the interaction between the car and the environment is stored in the experience replay pool. Then, small batches of data are randomly sampled from the experience replay pool for training to break the time correlation of the data and increase sample utilization.

[0208] (6) Target Network

[0209] To avoid instability during training, the DQN model uses a target network, which is a delayed version of the Q network. The parameters of the target network are updated only at certain training steps, while the parameters of the Q network are updated at every step. This mechanism effectively avoids drastic fluctuations in the Q value and improves the stability of training.

[0210] In the embodiment of the present invention, the training process of the DQN model includes a training phase and a reasoning and control phase;

[0211] The training phase includes:

[0212] (1) Initialization

[0213] Initialize the parameters θ of the Q network and θ of the target network - , initialize the experience replay pool;

[0214] (2) Training process

[0215] At each time step t, the agent takes the current state s t Select an action t , the selection strategy adopts the epsilon-greedy strategy, that is, a random action is selected with probability ∈, and the action with the maximum Q value is selected at other times;

[0216] After executing the action, the system returns the reward r t and the next state s t +1;

[0217] Will (s t ,a t ,r t ,s t +1) Deposit into the experience replay pool;

[0218] Randomly sample batches of data from the experience replay pool to train the Q network and update the parameters θ of the Q network;

[0219] At certain intervals, the parameters of the Q network are copied to the target network.

[0220] (3) Strategy Update

[0221] Each time the Q network is updated, the Q value of the current state and action is compared with the target Q value, and the parameters of the Q network are updated so that the network can learn a more appropriate strategy;

[0222] The reasoning and control phase includes:

[0223] The load sensor 100, the laser radar 600 and the IMU inertial measurement unit 800 collect the vehicle data of the material truck, including the load weight, slope and ground stability.

[0224] Using the current vehicle data, predict the current state s t ;

[0225] The current state s t Input the trained DQN model to calculate and select the optimal action, including acceleration, deceleration or maintaining speed;

[0226] According to the selected action, the motor 700 is controlled to perform corresponding acceleration or deceleration.

[0227] Compared with the prior art, the advantages of the present invention are:

[0228] In order to effectively detect the dirtiness inside the car, the present invention measures the load weight of the material cart by a load sensor, detects the degree of contamination of each layer of the storage rack by a visual sensor, measures ground undulations and obstacles by a lidar, and measures slope and bumps by an IMU inertial measurement unit.

[0229] In order to effectively detect the dirtiness of the material truck and realize adaptive speed control, the present invention uses an algorithm control system to perform dirtiness detection and adaptive speed control on the material truck according to the on-board data collected by each sensor and laser radar. The algorithm control system selects different network models for processing for different tasks, and selecting a suitable deep learning architecture is the key to success. In the dirtiness detection task, convolutional neural network (CNN) is a common choice because CNN performs well in image classification and target detection tasks. The architecture selected by the present invention includes: traditional convolutional neural network, which is suitable for simpler image classification tasks; modern deep network, which is suitable for complex image recognition tasks and can extract deeper features; target detection model, if the task involves locating and detecting the location of stains, the target detection model YOLO can be used. The algorithm control system uses reinforcement learning (RL) to optimize the speed adjustment strategy of the material truck, specifically uses the DQN model to train the speed adjustment strategy, uses experience playback and target network to improve training stability, and continuously optimizes the strategy during training to make the vehicle speed adjustment more intelligent.

[0230] The present invention uses a blowing cleaning system to clean the material cart with the ejected gas under the control of an algorithm control system. The blowing cleaning system is provided with three wind speed adjustments: low wind speed, medium wind speed and high wind speed. The low wind speed is used for low-level cleaning and is suitable for light dust, the medium wind speed is used for medium-level cleaning and is suitable for particulate matter, and the high wind speed is used for high-level cleaning and is suitable for wet dirt.

[0231] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A multifunctional automobile interior parts material trolley, characterized in that: The material cart comprises: a frame consisting of four support columns and a bottom plate, the frame is provided with a multi-layer mesh storage rack, and the bottom is provided with four wheels; a load sensor (100) is provided on the bottom plate of the frame, a display (200) and an energy device (300) are provided on the support columns of the frame, an algorithm control system (400) is provided below the display (200), a visual sensor (500) is provided on the top of the frame, a laser radar (600) is provided on the cross beam at the top of the frame, and a motor (70) is provided on the outer side of the bottom plate of the frame. 0), an IMU inertial measurement unit (800) is provided at the center of the inner side of the bottom plate of the frame, and a blowing cleaning system is also provided on the top crossbeam of the frame, and the blowing cleaning system includes an air outlet (900) and a blower (1000); the algorithm control system (400) is connected with the load sensor (100), the display (200), the visual sensor (500), the laser radar (600), the motor (700), the IMU inertial measurement unit (800) and the blowing cleaning system; The load sensor (100) is used to measure the load weight of the material truck and send it to the algorithm control system (400); The display (200) is used to provide a user interaction interface; The energy device (300) is used to provide power support for all devices on the skip; The algorithm control system (400) is used to perform dirt detection and adaptive speed control on the material vehicle according to the vehicle-borne data collected by each sensor; The visual sensor (500) is used to detect the contamination level of each layer of the storage rack and send it to the algorithm control system (400); The laser radar (600) is used to measure ground undulations and obstacles and send them to the algorithm control system (400); The motor (700) is used to provide power to the wheels of the trolley under the control of the algorithm control system (400); The IMU inertial measurement unit (800) is used to measure the slope and bump conditions and send them to the algorithm control system (400); The air blowing cleaning system is used to clean the material vehicle by means of the gas ejected by the blower (1000) and the air outlet (900) under the control of the algorithm control system (400).

2. The multifunctional automobile interior trim material trolley according to claim 1, characterized in that: The visual sensor (500) detects dirt in the material vehicle in combination with the computer vision technology and deep learning technology of the algorithm control system (400); Different types of dirt on the skip are identified by training a deep learning model, and the training process of the deep learning model includes: (1) Data Collection Collecting a data set as training samples, marking the location and category of stains for each image in the data set, and using image enhancement technology to generate more training samples, the data set includes images of various types of stains inside the car and images under various environmental conditions, the various types of stains include dust, mud, and wet stains, and the various environmental conditions include different lighting, angles, and stain coverage; (2) Data preprocessing Before model training, the size of the images in the data set is adjusted to a fixed size required by the deep learning model, the pixel values ​​of the images are standardized, and the data set is divided into a training set, a validation set, and a test set; (3) Selecting the model architecture Use traditional convolutional neural networks, modern deep networks, and object detection models; (4) Model training The training process includes initializing network parameters, selecting loss functions and optimization algorithms, forward propagation and back propagation, and hyperparameter tuning; (5) Model evaluation and testing After training, the model is evaluated using the test set, and the evaluation indicators include accuracy, recall, precision, F1-score and IoU; (6) Model optimization and deployment After the model training is completed and verified by the test set, the model is optimized to improve the reasoning speed and deployed based on the hardware environment to ensure that the model has sufficient computing power for real-time detection; (7) Continuous Improvement The data set and the optimization model are continuously updated through online learning and incremental learning.

3. The multifunctional automobile interior trim material trolley according to claim 2, characterized in that: The model training includes: The network parameters are initialized as follows: the weights in the network need to be randomly initialized, and the initialization methods include Gaussian distribution initialization, Xavier initialization and He initialization; The selected loss function includes: the loss function is used to measure the gap between the prediction result and the true label. For the dirt detection task, if it is a classification task, the cross entropy loss function is used to calculate the error between the model prediction category and the true category; if it is a regression task, including predicting the area or coverage of the stain, the mean square error is used; if it is a multi-task, that is, classification and positioning are performed simultaneously, a weighted loss function is used to weight the classification and regression losses; The selected optimization algorithm includes: an optimization algorithm for adjusting weights in the network to minimize the loss function, the optimization algorithm includes: a stochastic gradient descent suitable for large-scale data sets, and an Adam optimizer that combines the advantages of momentum and adaptive learning rate for complex deep learning tasks; The forward propagation includes: passing the input image through each layer of the network to finally obtain the predicted output; The back propagation includes: calculating the gradient of the loss function for each weight, and updating the weight by gradient descent; during the training process, batch gradient descent is used, that is, a small batch of samples is used for each update instead of all training samples; The hyperparameter tuning includes: evaluating the model performance through the validation set, and adjusting the hyperparameters according to the evaluation results, including the learning rate, batch size, and number of network layers, and finding the optimal hyperparameter combination through cross-validation or grid search.

4. The multifunctional automobile interior trim material trolley according to claim 2, characterized in that: The selection model architecture includes: Use traditional convolutional neural networks to handle simple image classification tasks; Use modern deep networks to handle complex image recognition tasks, capable of extracting deeper features; Use object detection models for tasks involving localizing and detecting the location of stains; The evaluation indicators for model evaluation and testing include: For classification tasks, accuracy is used to measure the proportion of samples that the model predicts correctly; Use recall and precision to measure the model's ability to correctly identify stains. Recall focuses on the proportion of stains detected, while precision focuses on whether the detected stains are real stains. For tasks with imbalanced categories, use the F1-score metric that takes both recall and precision into account; For the object detection task, IoU is used to evaluate the overlap between the detection box and the actual stain area.

5. The multifunctional automobile interior trim material trolley according to claim 2, characterized in that: The visual sensor (500), in combination with the computer vision technology and deep learning technology of the algorithm control system (400), detects the dirt in the material vehicle, including: The visual sensor collects (500) an image inside the material vehicle; The algorithm control system (400) uses computer vision technology to convert the image collected by the visual sensor (500) into a color space, converts the RGB mode into a grayscale / HSV mode, removes noise through Gaussian filtering, and enhances the contrast of stains; The algorithm control system (400) uses a pre-trained convolutional neural network, a modern deep network and a target detection model to classify dirt to obtain dirt categories and pollution levels. The training data set consists of different types of dirt, including dust, mud and particulate matter. When the pollution level, i.e., the dirt coverage area> a preset threshold, the air blowing cleaning system is started.

6. The multifunctional automobile interior trim material trolley according to claim 1, characterized in that: The air outlet (900) comprises a plurality of high-pressure air nozzles evenly arranged on the top of the frame, and the air nozzles can adjust their angles according to the detection results; The blowing cleaning system includes three wind speed settings: low wind speed, medium wind speed and high wind speed. The low wind speed is used for low-level cleaning, suitable for light dust, the medium wind speed is used for medium-level cleaning, suitable for particulate matter, and the high wind speed is used for high-level cleaning, suitable for wet dirt. The algorithm control system (400) collects the type of dirt and the area covered by the dirt in the material truck, and based on the type of dirt and the area covered by the dirt, uses a fuzzy control algorithm to control the wind force and the direction of the wind nozzle of the blowing cleaning system.

7. The multifunctional automobile interior trim material trolley according to claim 1, characterized in that: The algorithm control system (400) performs adaptive speed control on the material vehicle, including: The load sensor (100) measures the load weight of the truck; the laser radar (600) measures ground undulations and obstacles; the IMU inertial measurement unit (800) is provided with an accelerometer and a gyroscope to measure slope and bumps; The algorithm control system (400) calculates a speed adjustment value using a fuzzy control algorithm based on vehicle-mounted data collected by the load sensor (100), the laser radar (600) and the IMU inertial measurement unit (800), including load weight, slope and ground smoothness, and controls the motor (700) based on the speed adjustment value.

8. The multifunctional automobile interior trim material trolley according to claim 7, characterized in that: The algorithm control system (400) uses reinforcement learning to optimize the speed regulation strategy of the material vehicle, specifically uses a DQN model to train the speed regulation strategy, uses experience replay and a target network to improve training stability, and continuously optimizes the strategy during the training process, making the vehicle speed regulation more intelligent.

9. The multifunctional automobile interior trim material trolley according to claim 8, characterized in that: The implementation process of the DQN model includes: (1) Environment settings For the state space, the percentage of load is used to indicate the proportion of the weight of the materials in the carriage to the maximum load, the angle of slope is used to indicate the slope of the road section where the vehicle is currently located, which affects the acceleration and braking of the vehicle, and the degree of bumpiness is used to indicate the degree of bumpiness of the current road surface conditions of the vehicle, which affects the stability of the vehicle speed; The state space is defined as the vector S: S = [load, slope, bumpiness] For the action space, at each moment, the shuttle can choose the following three actions: acceleration, i.e. +am / s, maintenance, i.e. 0m / s, deceleration, i.e. -am / s, where a is a positive number; The action space is defined as the vector A: A=[+am / s,0m / s,-am / s] (2) Designing the Reward Function The reward function is the core of reinforcement learning and is used to guide the material truck to optimize its decision. According to the task requirements, the reward function is designed as follows: smooth driving, when the bumpiness is minimized, it means that the vehicle speed is stable, and a +m reward is given; speed optimization, when there is no unnecessary deceleration, it means that the vehicle speed has been effectively optimized, and a +n reward is given; avoid speeding, if the vehicle speed is too fast and there is a risk of tilting or material dumping, a -m reward is given; where m and n are both positive numbers; The reward function is expressed as: (3) State Transfer At each step, after the skip performs an action, the system moves to the next state based on the current state and the selected action, which is usually determined by the dynamics of the environment; (4) Q function update rules The core of the DQN model is to learn the Q value function Q(s) through a neural network. t ,a t ; θ), where θ represents the parameters of the Q network, which are continuously adjusted through back propagation, with the goal of minimizing the following loss function: Among them, a represents the learning rate, which controls the update step size; r t Represents the reward at the current moment; s t represents the current state; at represents the current action; γ represents the discount factor, which determines the weight of future rewards; Q(s t ,a t ; θ) represents the Q value of the action selected in the current state; θ - Represents the parameters of the target network, which is used to stabilize the training process; Represents the maximum Q value in the next state, which is used to estimate the optimal future return; (5) Experience Replay The DQN model adopts an experience replay mechanism to improve training efficiency and stability by storing and reusing the state, action, reward and next state of the car. The experience generated by the interaction between the car and the environment is stored in the experience replay pool, and then small batches of data are randomly sampled from the experience replay pool for training to break the time correlation of the data and increase sample utilization. (6) Target network The DQN model uses a target network, which is a delayed version of the Q network. The parameters of the target network are updated only at certain training steps, while the parameters of the Q network are updated at each step, avoiding drastic fluctuations in the Q value and improving the stability of training.

10. The multifunctional automobile interior trim material trolley according to claim 9, characterized in that: The training process of the DQN model includes the training phase and the reasoning and control phase; The training phase includes: (1) Initialization Initialize the parameters θ of the Q network and θ of the target network - , initialize the experience replay pool; (2) Training process At each time step t, the agent takes the current state s t Select an action t , the selection strategy adopts the ε-greedy strategy, that is, a random action is selected with probability ∈, and the action with the maximum Q value is selected at other times; After executing the action, the system returns the reward r t and the next state s t +1; Will (s t ,a t ,r t ,s t +1) stored in the experience replay pool; Randomly sample batches of data from the experience replay pool to train the Q network and update the parameters θ of the Q network; At certain intervals, the parameters of the Q network are copied to the target network; (3) Strategy Update Each time the Q network is updated, the Q value of the current state and action is compared with the target Q value, and the parameters of the Q network are updated so that the network can learn a more appropriate strategy; The reasoning and control phase includes: The load sensor (100), the laser radar (600) and the IMU inertial measurement unit (800) are used to collect the vehicle data of the material truck, including the load weight, slope and ground stability. Using the current vehicle data, predict the current state s t ; The current state s t Input the trained DQN model to calculate and select the optimal action, including acceleration, deceleration or maintaining speed; According to the selected action, the motor (700) is controlled to perform corresponding acceleration or deceleration.

Images