Intelligent car washer and vehicle cooperative sensing communication method based on 5G communication

Abstract

The invention discloses an intelligent car washer and vehicle collaborative perception communication method based on 5G communication, and relates to the field of communication, and the method comprises the steps: obtaining visual data based on a 5G heterogeneous data synchronization link, and obtaining vehicle end inertial measurement unit data of a vehicle; extracting visual features in the visual data; determining an initial posture of the vehicle entering the vehicle washing area based on the visual data and the vehicle end inertial measurement unit data; constructing a dynamic virtual coupling model based on vehicle-end inertial measurement unit data; in response to received real-time acceleration data and angular velocity data uploaded by the vehicle through a 5G heterogeneous data synchronization link, driving the vehicle in the dynamic virtual coupling model to move, and dynamically locking an intelligent vehicle washing machine coordinate system and a vehicle coordinate system based on visual features; and a dynamically locked coordinate system is determined, and the intelligent car washing machine is driven to wash the car according to the dynamically locked coordinate system. The safety and the quality stability of cleaning operation can be effectively guaranteed.

Description

Technical Field

[0001] This application relates to the field of communications, and more particularly to a method for intelligent car wash machines and vehicle collaborative sensing and communication based on 5G communication. Background Technology

[0002] With the continuous growth of car ownership and consumers' increasing demands for the efficiency and quality of vehicle cleaning services, automated car wash technology has become an important development direction in the automotive aftermarket service sector. Intelligent car wash machines, by integrating robotic arms, high-pressure spray systems, and sensor arrays, can achieve automated vehicle cleaning. However, in actual car washing processes, vehicles are not absolutely stationary but exhibit various forms of micro-movements, including natural micro-movements caused by the handbrake not being engaged, periodic vibrations generated when being dragged by the conveyor belt, and low-speed creeping during automatic parking. Although these micro-movements are small in magnitude, they cause continuous changes in the relative position between the vehicle surface and the end effector of the car wash machine, posing a serious challenge to the accuracy of the cleaning operation.

[0003] In existing technologies, intelligent car wash machines primarily rely on external visual sensors (such as cameras) to perceive the spatial position and attitude information of vehicles. A typical technical solution involves deploying multiple fixed cameras in the car wash area, extracting vehicle contour features through image recognition algorithms, and planning the motion trajectory of the robotic arm accordingly. However, the car wash operation site contains a large amount of high-density water mist, cleaning foam, and the high-speed movement of the robotic arm. These factors cause severe medium obstruction and multipath interference to optical sensors. Water mist particles scatter and refract light, resulting in blurred images captured by the camera; cleaning foam adheres to the lens surface or floats in the field of view, forming large visual blind spots; and the rapid movement of the robotic arm produces motion blur and dynamic occlusion in the image. These interference factors make it difficult for external visual sensors to continuously and stably obtain accurate vehicle position information.

[0004] Therefore, when a vehicle makes a slight movement, its spatial coordinates relative to the car wash machine will drift in real time. Due to the severely reduced sensing capabilities of external sensors in water mist and foam environments, the car wash machine cannot capture changes in the vehicle's posture in time. This leads to a decrease in the spatial alignment accuracy between the end effector of the robotic arm and the vehicle surface, resulting in the nozzle being too close or too far from the vehicle body. If the distance is too close, it will cause damage to the paint from the high-pressure water jet; if the distance is too far, the cleaning effect will be poor.

[0005] In summary, in car wash environments with complex media such as water mist and foam, existing perception solutions based on external visual sensors cannot effectively address the real-time coordinate system drift caused by minor vehicle movements, thus affecting the safety and quality stability of the cleaning operation. Summary of the Invention

[0006] This application provides a method for intelligent car wash machine and vehicle collaborative sensing and communication based on 5G communication, which is used to ensure the safety and quality stability of the cleaning operation in car wash environment with complex media such as water mist and foam.

[0007] To achieve the above objectives, the embodiments of this application adopt the following technical solutions: Firstly, a collaborative sensing and communication method between a smart car wash machine and a vehicle based on 5G communication is provided, applied to a car wash system. The car wash system includes a smart car wash machine and a vehicle, and the smart car wash machine includes a camera. The method includes: Construct a 5G heterogeneous data synchronization link between the smart car wash machine and the vehicle, and based on the 5G heterogeneous data synchronization link, acquire visual data through the camera and acquire data from the vehicle's inertial measurement unit. Extracting visual features from visual data; Based on visual data and vehicle-mounted inertial measurement unit data, the initial attitude of the vehicle entering the car wash area is determined; A dynamic virtual coupling model is constructed based on data from the vehicle-mounted inertial measurement unit. In response to receiving real-time acceleration and angular velocity data uploaded by the vehicle through the 5G heterogeneous data synchronization link, the vehicle in the dynamic virtual coupling model is driven to move, and the coordinate system of the smart car wash machine is dynamically locked to the vehicle coordinate system based on visual features. A dynamically locked coordinate system is determined, and the intelligent car wash machine is driven to perform cleaning operations on the vehicle according to the dynamically locked coordinate system.

[0008] In one possible implementation of the first aspect, the intelligent car wash machine further includes a robotic arm equipped with a nozzle and a servo motor. Based on data from the vehicle-side inertial measurement unit, a dynamic virtual coupling model is constructed, including: The vehicle's attitude change is calculated based on data from the vehicle-mounted inertial measurement unit. Calculate the displacement of each preset feature point on the vehicle surface based on the attitude change; The robotic arm drives a servo motor to perform compensating movements in the same direction and with equal amount of displacement to keep the relative position of the nozzle and the vehicle surface constant.

[0009] In another possible implementation of the first aspect, the vehicle-side inertial measurement unit data includes roll angle, pitch angle, and yaw angle, and the displacement of each preset feature point on the vehicle surface is calculated based on the attitude change, including: The static bounding box model of the vehicle is determined based on visual features, and the vehicle size data in the static bounding box model is obtained. Based on the roll angle, pitch angle, yaw angle and vehicle size data, calculate the three-dimensional spatial displacement vector of each preset feature point on the vehicle surface; By mapping the three-dimensional spatial displacement vector to the coordinate system of the intelligent car wash machine, the displacement of each preset feature point on the vehicle surface is obtained.

[0010] In another possible implementation of the first aspect, after the robotic arm drives the servo motor to perform a compensating motion of the same direction and amount based on the displacement, it further includes: In response to the audio signal reflected from the impact of high-pressure water flow on the vehicle body, the echo delay and spectral attenuation rate of the audio signal are determined. The actual physical distance between the nozzle and the vehicle body is calculated based on the echo delay and spectral attenuation rate. The actual physical distance is compared with the theoretical distance calculated based on data from the vehicle-mounted inertial measurement unit to obtain the distance deviation; An error correction vector is generated based on the distance deviation, and the spatial coordinates of the static bounding box model are adjusted based on the error correction vector to eliminate the integral drift error of the vehicle-side inertial measurement unit.

[0011] In another possible implementation of the first aspect, the actual physical distance between the nozzle and the vehicle body is calculated based on the echo delay and spectral attenuation rate, including: Extract the acoustic signature features produced by water flow hitting different materials on the car body from the audio signal; Determine the vehicle body material type of the current cleaning area based on voiceprint characteristics; Load the sound wave propagation speed and energy attenuation coefficient corresponding to the vehicle body material type from the preset acoustic parameter database; The actual physical distance between the nozzle and the vehicle body is calculated based on the echo delay, sound wave propagation speed, and energy attenuation coefficient.

[0012] In another possible implementation of the first aspect, the intelligent car wash machine performs the washing operation on the vehicle based on a dynamically locked coordinate system, including: Real-time monitoring of the real-time latency value of the 5G heterogeneous data synchronization link; When the real-time delay value is lower than the first preset threshold, the drive robotic arm maintains a high-rigidity mode and performs high-pressure rinsing or close-fitting brushing. When the real-time delay value is higher than the second preset threshold, the proportional-integral-derivative parameters of the robotic arm joint are determined based on the real-time delay value, and the stiffness coefficient and damping coefficient are reduced based on the proportional-integral-derivative parameters, so that the robotic arm enters the flexible following mode. In flexible following mode, the robotic arm is driven to perform a passive retreat mode based on the vehicle's current contact physical data to prevent scratches on the paint.

[0013] In another possible implementation of the first aspect, the proportional-integral-derivative (PID) parameters of the robotic arm joints are determined based on real-time delay values, and the stiffness and damping coefficients are reduced based on these PID parameters to enable the robotic arm to enter a flexible following mode, including: Calculate the amount by which the real-time delay value exceeds the second preset threshold; The proportional gain attenuation factor is calculated based on the delay excess, where the larger the delay excess, the smaller the proportional gain attenuation factor. Multiply the preset reference proportional gain of the robotic arm joint by the proportional gain attenuation factor to obtain the dynamically adjusted proportional gain parameter. Calculate the integral time constant amplification factor based on the delay excess; the larger the delay excess, the larger the integral time constant amplification factor. The dynamic-adjusted integral parameters are obtained by multiplying the reference integral time constant of the robotic arm joint by the integral time constant amplification factor. The differential gain suppression factor is calculated based on the delay excess, where the larger the delay excess, the smaller the differential gain suppression factor. The differential parameters are obtained by multiplying the baseline differential gain of the robotic arm joint by the differential gain suppression factor. Based on the dynamically adjusted proportional gain parameters, integral parameters, and differential parameters, calculate the reduction ratio of stiffness coefficient and damping coefficient. The actual stiffness coefficient in the flexible following mode is obtained by multiplying the reference stiffness coefficient of the robotic arm joint by the stiffness coefficient reduction ratio. The actual damping coefficient in the flexible following mode is obtained by multiplying the reference damping coefficient of the robotic arm joint by the damping coefficient reduction ratio.

[0014] In another possible implementation of the first aspect, the current contact physical data includes contact force data when the robotic arm contacts the vehicle body. In flexible following mode, based on the vehicle's current contact physical data, the robotic arm is driven to execute a passive yielding mode to prevent scratches to the paint, including: Determine whether the contact force data exceeds the safety threshold; If the contact force data exceeds the safety threshold, calculate the direction vector of the contact force; The robotic arm is controlled to retract in the opposite direction of the contact force vector until the contact force data is below the safety threshold.

[0015] Secondly, this application provides an intelligent car wash machine, comprising: Camera; The memory is configured to store instructions; and The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the aforementioned 5G-based intelligent car wash machine and vehicle collaborative perception and communication method.

[0016] Thirdly, this application provides a car wash system, comprising: vehicle; Intelligent car wash machines connect to vehicles.

[0017] By leveraging the low latency and high reliability of 5G networks, a real-time data synchronization link was established between the vehicle's inertial measurement unit and the car wash machine, achieving millisecond-level transmission of vehicle motion information. By fusing visual and inertial data, the initial attitude of the vehicle was accurately determined, and a dynamic virtual coupling model was constructed, enabling the car wash robotic arm to synchronously compensate for the vehicle's minute movements in real time. A dynamic coordinate system locking mechanism ensures that the car wash machine always performs cleaning operations in the correct spatial position, avoiding nozzle-vehicle distance deviations caused by minor vehicle movements. Adaptive stiffness adjustment and passive yielding mechanisms further enhance system safety, effectively preventing vehicle damage even in the event of communication delays or sudden contact. This method significantly improves the operational accuracy and reliability of automated car washes, reduces reliance on external visual sensors, and provides an innovative solution for the application of intelligent car wash technology in complex environments.

[0018] Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0019] Figure 1 A flowchart illustrating a 5G-enabled smart car wash machine and vehicle collaborative sensing and communication method provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of an intelligent car wash machine provided in an embodiment of this application; Figure 3 This is a schematic diagram of a car wash system provided in an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0021] It should be noted that if the embodiments of this application involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0022] Furthermore, if the embodiments of this application involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0023] Figure 1 This illustration schematically depicts a flowchart of a 5G-based intelligent car wash machine and vehicle collaborative sensing and communication method according to an embodiment of this application. Figure 1 As shown in the figure, this application provides a method for collaborative perception and communication between a smart car wash machine and a vehicle based on 5G communication, which is applied to a car wash system. The car wash system includes a smart car wash machine and a vehicle. The smart car wash machine includes a camera. The method may include the following steps.

[0024] S110. Construct a 5G heterogeneous data synchronization link between the smart car wash machine and the vehicle, and based on the 5G heterogeneous data synchronization link, acquire visual data through the camera and acquire data from the vehicle's inertial measurement unit. S120. Extract visual features from visual data; S130. Based on visual data and vehicle-side inertial measurement unit data, determine the initial attitude of the vehicle as it enters the car wash area; S140. Based on the data from the vehicle-mounted inertial measurement unit, a dynamic virtual coupling model is constructed; S150, in response to receiving real-time acceleration data and angular velocity data uploaded by the vehicle through the 5G heterogeneous data synchronization link, drive the vehicle movement in the dynamic virtual coupling model, and dynamically lock the coordinate system of the intelligent car wash machine and the vehicle coordinate system based on visual features. S160. Determine the dynamically locked coordinate system and drive the intelligent car wash machine to perform cleaning operations on the vehicle according to the dynamically locked coordinate system.

[0025] In this embodiment, before establishing the 5G heterogeneous data synchronization link between the smart car wash machine and the vehicle, the following steps are also included: In response to a vehicle entry signal, the intelligent car wash machine sends a telemetry request command, wherein the vehicle responds to the telemetry request command by activating the chassis controller local area network bus gateway and locking the data ports of the inertial measurement unit and the suspension travel sensor. A 5G slicing channel is constructed for transmitting timing data of the inertial measurement unit and the suspension travel sensor.

[0026] The specific implementation mechanism of 5G network slicing is as follows: (1) Vehicle communication module configuration: The vehicle is equipped with a 5G-V2X communication module; the module supports Sub-6GHz band (n78:3.5GHz) and C-V2XPC5 direct communication; it integrates automotive-grade security chips and supports USIM / eSIM dual-mode authentication. (2) When the vehicle enters the car wash area, it sends a slice request message to the 5G core network (5GC) through the cellular network; the request message adopts the Network Slice Selection Assistance Information (NSSAI) format, which includes S-NSSAI (Single Network Slice Selection Assistance Information): SST=2 (URLLC type), SD=0x000001 (car wash service identifier), and the quality of service (QoS) requirements are: end-to-end latency ≤10ms, packet loss rate ≤10 -5, bandwidth ≥ 5Mbps, the network slice selection function of the 5G core network selects the corresponding network slice instance according to the request; the slice instance includes dedicated user plane function (UPF), session management function (SMF) and access and mobility management function (AMF). (3) A dedicated protocol data unit (PDU) session is established between the vehicle and the car wash machine; PDU session type: IPv4, allocated dedicated IP address range 192.168.100.0 / 24; QoS flow identifier (QFI): 5 (corresponding to 5QI=82, URLLC service); guaranteed bit rate (GFBR): uplink 3Mbps, downlink 2Mbps. (4) The precise time protocol is used to realize the clock synchronization between the vehicle and the car wash machine; the 5G base station is used as the PTP master clock, and the synchronization signal block (SSB) is sent down through the air interface; the vehicle module and the car wash machine module are used as slave clocks, and clock calibration is realized through bidirectional timestamp exchange. (5) IMU data is encapsulated using a custom application layer protocol: Protocol header (8 bytes): Magic number (2B) + Version number (1B) + Data type (1B) + Timestamp (4B), Data payload (48 bytes): Acceleration (3×4B) + Angular velocity (3×4B) + Attitude angle (3×4B) + Sequence number (4B) + CRC check (4B), Total packet length 56 bytes. At a transmission frequency of 100Hz, the uplink bandwidth requirement is: 56×8×100=44.8kbps. UDP protocol is used for transmission to reduce latency. Packet loss detection and retransmission are implemented at the application layer: The receiver maintains a sequence number window and detects packet loss; when the packet loss rate is <1%, no retransmission is performed, relying on Kalman filter interpolation; when the packet loss rate is ≥1%, a request is made to retransmit the most recent 10 frames of data. (6) Configure traffic policies through the policy control function of the 5G core network: the car wash slice traffic is completely isolated from other services (such as eMBB) at the UPF level; reserve dedicated radio resource blocks (PRB) for the car wash slice, accounting for 15-20%; when the network is congested, prioritize the QoS of the car wash slice, and reduce the speed of other services. (7) After the cleaning operation is completed, the vehicle sends a PDU session release request; the 5G core network reclaims the allocated IP address and radio resources; the slice instance remains active for the next vehicle to use.

[0027] During the startup phase of the car wash system, a highly reliable communication link must first be established between the intelligent car wash machine and the vehicle. When a vehicle enters the car wash area, the entrance detection sensor of the intelligent car wash machine triggers a vehicle entry signal, which activates the car wash machine's 5G communication module. The car wash machine immediately broadcasts a telemetry request command to the vehicle via the 5G network. This command is encapsulated using the V2X (Vehicle-to-Everything) protocol and includes the car wash machine's device identifier, the requested data type, and the communication encryption key. After receiving the telemetry request command, the vehicle's onboard communication unit verifies the command's legitimacy. Once it confirms that the command originates from an authorized car wash service device, it automatically activates the chassis controller's local area network (LAN) bus gateway. This gateway acts as a bridge connecting the vehicle's internal CAN bus with the external communication system. By activating the gateway, the data ports of the vehicle's inertial measurement unit (IMU) and suspension travel sensors are activated and locked in a dedicated transmission mode. The IMU can measure the vehicle's three-axis acceleration, three-axis angular velocity, and attitude angle information in real time. The sampling frequency of these data is typically set to 100Hz to ensure that even minute changes in vehicle motion can be captured. The suspension travel sensor monitors the suspension compression of the four wheels of the vehicle to determine the load distribution and changes in vehicle height.

[0028] After the data port is locked, the smart car wash machine negotiates with the vehicle to establish a 5G slicing channel. This channel is a virtual private channel in the 5G network, featuring independent bandwidth guarantees and ultra-low latency. Through network slicing technology, the timing data of the inertial measurement unit and suspension travel sensors can be isolated from other network traffic, avoiding data transmission delays caused by network congestion. The end-to-end latency of the 5G slicing channel is typically controlled within 5 milliseconds. After the 5G heterogeneous data synchronization link is established, the smart car wash machine begins to collect visual data through its high-resolution cameras. These cameras are typically mounted on the top and sides of the car wash area, allowing observation of the vehicle from multiple angles. The visual data includes the vehicle's outline image, color information, surface texture, etc., which is collected as a video stream at a frame rate of 30fps. Simultaneously, the vehicle continuously uploads data from its inertial measurement unit via the 5G slicing channel, including real-time acceleration vectors, angular velocity vectors, and attitude angles calculated through sensor fusion algorithms. This dual-channel data acquisition mechanism achieves synergy between visual perception and inertial perception, laying the data foundation for subsequent vehicle attitude recognition and motion tracking. Leveraging the high bandwidth of 5G networks, visual and inertial data can be synchronously transmitted to the edge computing unit of the car wash machine for real-time processing, effectively avoiding the perception lag problem caused by data transmission delays in traditional solutions.

[0029] After acquiring the visual data, it is necessary to extract key information that characterizes the spatial features of the vehicle from the original image. The visual feature extraction process first preprocesses the image captured by the camera, including noise reduction, contrast enhancement, and distortion correction. Since water mist and foam exist in the car wash environment, image quality is affected to some extent; therefore, an adaptive filtering algorithm is needed to remove noise interference.

[0030] After preprocessing, a deep learning object detection algorithm is used to identify vehicle regions in the image. This algorithm, based on a convolutional neural network architecture, can accurately locate the vehicle's bounding box in complex backgrounds. Once the vehicle is detected, its geometric features are further extracted, including its length, width, height, and the positions of key points on the vehicle body. Key points include the four corners of the vehicle, door edges, window outlines, and rearview mirror positions. Through image fusion from multiple viewpoint cameras, a 3D spatial model of the vehicle can be reconstructed. Visual features also include the texture features of the vehicle surface, such as the reflectivity of the paint and the direction of the body lines. This texture information helps in accurately locating the target area for subsequent cleaning operations. To improve the robustness of feature extraction, the algorithm constructs a multi-scale feature pyramid for the vehicle's appearance, extracting and fusing features at different resolution levels. The extracted visual features are encoded as feature vectors, which contain comprehensive information about the vehicle's position, pose, and size in image space. The feature vectors are typically 512-dimensional or 1024-dimensional, fully expressing the vehicle's visual attributes. These visual features are not only used to determine the vehicle's initial pose but also serve as a reference for coordinate system locking during subsequent dynamic tracking. By extracting visual features, the raw pixel-level image data is transformed into structured semantic information, which greatly reduces the computational complexity of subsequent processing and improves the resistance to environmental interference.

[0031] The initial attitude of the vehicle entering the car wash area is determined, defining its initial position and orientation within the car wash machine's coordinate system. Determining the initial attitude requires fusing visual data and data from the vehicle's inertial measurement unit (IMU) to achieve multi-sensor collaborative localization. Based on extracted visual features, vehicle feature points in image space are mapped to the car wash machine's 3D world coordinate system through perspective transformation and camera calibration parameters. This process requires knowledge of the camera's intrinsic and extrinsic parameter matrices. The intrinsic matrix describes the camera's focal length and optical center position, while the extrinsic matrix describes the camera's rotation and translation relative to the car wash machine's coordinate system. By solving the perspective projection equations, the coordinates of the vehicle's key points in 3D space can be calculated.

[0032] However, relying solely on visual data suffers from inaccurate depth estimation, especially under water mist interference, where decreased image clarity leads to increased feature point localization errors. Therefore, data from the vehicle-mounted inertial measurement unit (IMU) is introduced for fusion correction. The IMU provides attitude angle data including roll, pitch, and yaw angles, which describe the vehicle's spatial attitude relative to the horizontal plane and geographic north. When fusing the IMU's attitude angles with visual data, an extended Kalman filter algorithm is employed, which can provide optimal estimation of the vehicle's state even in the presence of measurement noise.

[0033] Specifically, the vehicle's position and attitude are treated as state variables, while visual observations and inertial measurements are treated as observables. Through an iterative process of prediction and updates, the system gradually converges to the true initial attitude. The fused initial attitude includes the three-dimensional coordinates of the vehicle's center point in the car wash machine's coordinate system. and the vehicle's three attitude angles These six parameters correspond to roll, pitch, and yaw, respectively. They comprehensively describe the vehicle's spatial attitude, providing accurate initial conditions for the subsequent construction of a dynamic virtual coupled model. The accuracy of the initial attitude determination directly affects the alignment accuracy of subsequent cleaning operations. Through multi-sensor fusion, the attitude estimation error can be controlled within the centimeter and degree range, meeting the requirements of high-precision cleaning operations.

[0034] The dynamic virtual coupling model can realize the synchronization of car wash machine and vehicle movement. This model establishes a real-time mapping relationship between vehicle movement and car wash robotic arm movement at the software level.

[0035] The model is built based on data from the vehicle's inertial measurement unit (IMU), which reflects real-time changes in the vehicle's motion. Based on the acceleration and angular velocity data continuously uploaded by the IMU, the vehicle's attitude changes are calculated through numerical integration. Acceleration data is integrated once to obtain the velocity change and twice to obtain the position change; angular velocity data is integrated to obtain the attitude angle change. Because the integration process accumulates sensor measurement errors, drift may occur after prolonged operation; therefore, periodic correction using visual data is necessary.

[0036] After obtaining the vehicle's attitude change, the displacement of each preset feature point on the vehicle surface is further calculated. These preset feature points are several key locations selected on the vehicle surface, such as the center of the roof, door handles, and wheel hubs. Changes in the position of these points directly affect the target position of the end effector of the car wash robotic arm. When calculating the feature point displacement, the rigid body motion characteristics of the vehicle must be considered; that is, any motion of the vehicle can be decomposed into a combination of translation and rotation. The translation part is directly given by the position change, while the rotation part requires transforming the coordinates of the feature points in the vehicle coordinate system to the car wash machine coordinate system using a rotation matrix. The rotation matrix is ​​constructed from three attitude angles through Euler angle transformation. After calculating the displacement of each feature point, the car wash machine's robotic arm control system drives the servo motor to perform compensating movements in the same direction and with equal magnitude based on these displacements.

[0037] Specifically, if the vehicle moves forward The robotic arm moves forward accordingly as the distance is measured. If the vehicle is involved in an accident As the vehicle tilts, the end effector of the robotic arm adjusts its angle accordingly. The goal of this compensating motion is to maintain a constant relative position between the nozzle and the vehicle surface; that is, the nozzle's position remains unchanged within the vehicle's coordinate system. Through a dynamic virtual coupling model, the vehicle's motion is transmitted to the car wash machine in real time, achieving a virtual rigid connection between the two in space. This coupling mechanism requires no physical contact, relying entirely on data communication and motion control, thus avoiding the complexity and failure risks associated with mechanical connections. The model's real-time performance depends on the low latency of 5G communication, ensuring that vehicle motion data is transmitted to the car wash machine within milliseconds, allowing the compensating motion to occur almost synchronously with the vehicle's movement.

[0038] During the washing process, the vehicle continuously undergoes minute movements, which are captured and uploaded in real time by the vehicle's inertial measurement unit. When the car wash machine receives real-time acceleration and angular velocity data uploaded by the vehicle via a 5G heterogeneous data synchronization link, it immediately updates the vehicle's motion state in the dynamic virtual coupling model. The acceleration data reflects the vehicle's linear acceleration in three spatial directions, while the angular velocity data reflects the vehicle's angular velocity around three rotational axes. This data is received in time-series format, with a sampling frequency of 100Hz, meaning it is updated every 10 milliseconds.

[0039] Upon receiving new data, the vehicle state variables in the virtual coupled model are updated, including the vehicle's position, velocity, attitude angles, and angular velocity. State updates are calculated recursively using kinematic equations. For example, the position update formula is the current position plus the product of velocity and time step; the velocity update is the current velocity plus the product of acceleration and time step; and the attitude angle update is achieved by adding the product of angular velocity and time step to the current attitude angle.

[0040] While updating the vehicle's state, it's necessary to dynamically lock the intelligent car wash machine's coordinate system with the vehicle's coordinate system based on visual features. Dynamic locking means calculating the transformation relationship between the two coordinate systems in real time, ensuring the vehicle's position and attitude are accurately represented in the car wash machine's coordinate system. The locking process is achieved through a coordinate transformation matrix, composed of a rotation matrix and a translation vector. The rotation matrix is ​​calculated based on the vehicle's attitude angles, and the translation vector is calculated based on the vehicle's center point position. Whenever the vehicle's state is updated, the coordinate transformation matrix is ​​also updated to ensure the two coordinate systems maintain correct alignment. Visual features act as a reference calibration in dynamic locking; due to integral drift in the inertial measurement unit, the calculated position and attitude may deviate from the true values ​​after prolonged operation.

[0041] By periodically comparing the vehicle's position extracted from visual features with the position calculated by inertial measurements, drift errors can be detected and corrected. The correction method involves calculating the deviation between the two and then feeding back the deviation proportionally into the state estimation, causing the estimated value to gradually approach the true value. This dynamic locking mechanism, combining the high frequency of inertial measurement with the high precision of visual measurement, can both quickly respond to the instantaneous movement of the vehicle and maintain long-term positioning accuracy. Through dynamic locking, the car wash machine can grasp the precise pose of the vehicle in its own coordinate system in real time, providing an accurate target reference for the motion control of the robotic arm.

[0042] After dynamically locking the coordinate system, the car wash machine obtains the vehicle's real-time pose information within its own coordinate system, at which point it can drive the robotic arm and other actuators to begin the cleaning operation. The cleaning operation is performed based on the dynamically locked coordinate system, meaning that the robotic arm's motion planning and control are all completed within this coordinate system. First, based on the vehicle's size and shape, a cleaning path is planned, defining a series of spatial locations the nozzles need to traverse. Path planning considers both the completeness and efficiency of the cleaning coverage, typically employing either raster scanning or helical scanning modes.

[0043] After path planning is completed, the robotic arm's motion controller transforms the path points from the vehicle coordinate system to the car wash machine coordinate system based on the current coordinate transformation matrix, obtaining the target position that the robotic arm needs to reach. The robotic arm then uses inverse kinematics calculations to convert the target position of the end effector into the target angle of each joint, and then drives the servo motors to move the joints to the target angle. During the washing process, because the vehicle may continuously undergo slight movements, the coordinate transformation matrix is ​​updated in real time, and the target position of the robotic arm is dynamically adjusted accordingly. This dynamic adjustment is continuous, ensuring that the nozzle is always aligned with the target cleaning area on the vehicle surface.

[0044] To ensure cleaning quality and safety, the car wash machine also monitors the latency of the 5G heterogeneous data synchronization link in real time. When the latency is low, vehicle motion data can be transmitted promptly, allowing the robotic arm to maintain a high-stiffness mode and perform operations requiring precise control, such as high-pressure washing or close-fitting brushing. In high-stiffness mode, the joint stiffness and damping coefficients of the robotic arm are set high, enabling it to quickly and accurately reach the target position and remain stable. When the latency increases, the timeliness of data transmission decreases. If the high-stiffness mode is maintained, the movement of the robotic arm may become asynchronous with the actual movement of the vehicle, increasing the risk of collision. In this case, the car wash machine dynamically adjusts the control parameters of the robotic arm based on the latency value, reducing the stiffness and damping coefficients, allowing the robotic arm to enter a flexible following mode.

[0045] Specifically, in flexible following mode, the robotic arm's response speed decreases, but its adaptability to external disturbances is enhanced. Even if there is a delay in the vehicle's motion data, the robotic arm can absorb positional deviations through flexible deformation, avoiding damage to the vehicle. Furthermore, the car wash machine is equipped with contact force sensors to monitor the contact force between the robotic arm and the vehicle body in real time. When the contact force exceeds a safety threshold, the control system immediately calculates the direction of the contact force and drives the robotic arm to retreat in the opposite direction until the contact force is reduced to a safe range. This passive retreat mechanism provides a final safety guarantee for the washing operation, preventing scratches to the paint or impacts to the vehicle body even in extreme situations. By performing the washing operation based on a dynamically locked coordinate system, the car wash machine achieves real-time tracking and compensation for the vehicle's minute movements, ensuring the accuracy and safety of the washing operation.

[0046] This embodiment leverages the low latency and high reliability of 5G networks to establish a real-time data synchronization link between the vehicle's inertial measurement unit and the car wash machine, achieving millisecond-level transmission of vehicle motion information. By fusing visual and inertial data, the initial attitude of the vehicle is accurately determined, and a dynamic virtual coupling model is constructed, enabling the car wash robotic arm to synchronously compensate for the vehicle's minute movements in real time. A dynamic coordinate system locking mechanism ensures that the car wash machine always performs cleaning operations in the correct spatial position, avoiding nozzle-vehicle distance deviations caused by minor vehicle movements. Adaptive stiffness adjustment and passive yielding mechanisms further enhance system safety, effectively preventing vehicle damage even in the event of communication delays or sudden contact. This method significantly improves the operational accuracy and reliability of automated car washing, reduces reliance on external visual sensors, and provides an innovative solution for the application of intelligent car wash technology in complex environments.

[0047] In one embodiment of this invention, the intelligent car wash machine further includes a robotic arm equipped with a nozzle and a servo motor. Based on data from the vehicle-side inertial measurement unit, a dynamic virtual coupling model is constructed, including the following steps: S210. Calculate the vehicle's attitude change based on the data from the vehicle-side inertial measurement unit. S220. Calculate the displacement of each preset feature point on the vehicle surface based on the attitude change. S230: The robotic arm drives the servo motor to perform compensating movements in the same direction and with equal amount based on the displacement, so as to keep the relative position of the nozzle and the vehicle surface constant.

[0048] In the process of constructing the dynamic virtual coupling model, the first step is to calculate the vehicle's attitude change based on the data from the vehicle-mounted inertial measurement unit. The vehicle-mounted inertial measurement unit continuously collects the vehicle's three-axis acceleration and three-axis angular velocity data at a sampling frequency of 100Hz. This raw data is transmitted in real time to the edge computing unit of the car wash machine via a 5G heterogeneous data synchronization link.

[0049] The attitude change is calculated using a quaternion integration method, which effectively avoids the gimbal lock problem in Euler angle representation. Specifically, the angular velocity vector output by the inertial measurement unit is constructed as a quaternion differential equation, and the change in attitude quaternions is obtained through numerical integration. Within each sampling period, the angular velocity vector... Multiply the quaternion by the current attitude to obtain the rate of change of the quaternion, then multiply by the time step of 0.01 seconds, and add it to the current quaternion to complete one attitude update.

[0050] To suppress accumulated errors during the integration process, a complementary filtering algorithm is employed to correct the attitude estimation. This algorithm weights and fuses the high-frequency angular velocity integral results from the inertial measurement unit (IMU) with the attitude angles calculated by the low-frequency accelerometer. Leveraging the accelerometer's ability to accurately measure the direction of gravity during static or uniform motion, it corrects for drift components in the attitude angles. The fusion weights are adaptively adjusted based on the vehicle's motion state. When the vehicle is detected to be in violent motion, the accelerometer's weight is reduced to avoid interference from motion acceleration on gravity measurements; when the vehicle is relatively stationary, the accelerometer's weight is increased to fully utilize its long-term stability.

[0051] By combining quaternion integration and complementary filtering, the change in the vehicle's roll angle relative to its initial attitude can be calculated in real time. Pitch angle change and yaw angle change Simultaneously, a second integration is performed on the triaxial acceleration data to obtain the change in position of the vehicle's center point. During integration, the acceleration needs to be transformed from the vehicle coordinate system to the world coordinate system to remove the influence of gravitational acceleration before integration. Because acceleration measurements are noisy, double integration can cause position errors to diverge rapidly. Therefore, at regular intervals, visual data is used to perform zero-speed correction or position reset on the position estimate.

[0052] After obtaining the vehicle's attitude changes, it is necessary to further calculate the displacement of each preset feature point on the vehicle surface. First, a static bounding box model of the vehicle is determined based on visual features. This model is a cuboid that tightly wraps around the vehicle's shape, and the vehicle's length, width, and height are extracted using a visual recognition algorithm. In the bounding box model, preset feature points are defined at key locations on the vehicle surface, such as the four corners of the roof, the center of the doors, the centers of the front and rear bumpers, and the centers of the wheel hubs. Typically, 20 to 30 feature points are selected to cover the entire vehicle body. Each feature point has fixed three-dimensional coordinates in the vehicle coordinate system.

[0053] When the vehicle changes attitude, these feature points will undergo spatial displacement as the vehicle's rigid body moves. Calculating this displacement requires considering both translation and rotation. The translation component is directly given by the change in position of the vehicle's center point; that is, all feature points are superimposed with the same translation vector. The rotational part needs to be calculated using a rotational transformation matrix, which consists of the change in the roll angle. Pitch angle change and yaw angle change It is constructed.

[0054] For coordinates in the vehicle coordinate system For the i-th feature point, first subtract the vehicle center point coordinates from its coordinates to obtain its position vector relative to the center. Then, multiply this vector by a rotation matrix to obtain the rotated relative position vector. Add this to the new position of the vehicle center to obtain the new coordinates of the feature point in the world coordinate system. The difference between the old and new coordinates is the three-dimensional spatial displacement vector of the feature point. Since the car wash robotic arm moves in the car wash machine coordinate system, it is necessary to map the displacement vector of the feature point from the world coordinate system to the car wash machine coordinate system. The mapping process is achieved through a coordinate transformation matrix, which is determined by the calibration program during system initialization.

[0055] To eliminate the integral drift error of the vehicle-mounted inertial measurement unit, the dynamic virtual coupling model integrates the following triple feedback calibration loop: (1) Zero-speed correction circuit: The vehicle's motion status is monitored in real time, and it is determined to be stationary when the following conditions are met: the amplitude of the three-axis acceleration is close to gravity: 9.6 < ||a|| < 10.0 m / s². 2 The triaxial angular velocity is close to zero: ||ω|| < 0.01 rad / s; the duration of the above conditions is > 0.5 seconds.

[0056] Once a stationary state is detected, a zero-velocity correction operation is performed: the velocity estimate v is forcibly set to zero. Update the gyroscope bias estimate; reset the cumulative error of the accelerometer integrator.

[0057] Zero-speed correction can effectively suppress short-term drift in velocity and position, but it cannot correct long-term drift in attitude angle.

[0058] (2) Visual relocation loop: Every 10 seconds or when the cumulative displacement exceeds 50cm, the visual relocalization process is triggered: Step 1: The camera acquires an image of the current vehicle and extracts vehicle contour feature points (such as the centers of the four wheel hubs, door handles, etc.); Step 2: The coordinates of the feature points are transformed from the image coordinate system to the world coordinate system to obtain the visually measured vehicle position p. v Step 3: Calculate the deviation between visual position and inertial position: Step 4: If the deviation exceeds the threshold (||Δp||>2cm), then the inertial integral is considered to have drift. Step 5: Use a Kalman filter to fuse visual and inertial data. The Kalman gain K is dynamically adjusted based on the confidence level of the visual measurement: when there is less water mist and the image is clear (image contrast > 0.6), K = 0.7, mainly trusting the visual measurement; when there is more water mist and the image is blurry (image contrast < 0.4), K = 0.3, reducing the visual weight; Step 6: merging the positions p f The spatial coordinates of the static bounding box model are updated as a new reference location. Visual relocalization can correct position drift, but its reliability decreases in cases of severe water mist.

[0059] (3) Acoustic distance calibration loop (low-frequency calibration, continuous monitoring): During the cleaning process, the distance is calibrated by continuously using the audio signal generated by the high-pressure water jet hitting the vehicle body.

[0060] The advantage of acoustic calibration lies in its immunity to optical obstruction from water mist and foam, providing reliable distance feedback even in the event of visual failure. The specific calibration process is as follows: Step 1: The acoustic sensor measures the actual physical distance d between the nozzle and the vehicle body. a (Accuracy ±2mm); Step 2: Calculate the theoretical distance d based on inertial data. I (Based on vehicle pose and robotic arm forward kinematics); Step 3: Calculate distance deviation: Step 4: If the deviation exceeds the threshold (|Δd|>10mm), then trigger error correction; Step 5: Generate a correction vector according to the deviation direction and adjust the coordinates of the static bounding box model; Step 6: Adopt a progressive correction strategy and apply the correction amount in 5 control cycles to avoid abrupt changes.

[0061] The three calibration methods mentioned above have complementary characteristics. Zero-speed correction has a fast response (real-time) but is only effective when stationary; visual repositioning has high accuracy but fails under water mist; acoustic calibration is robust (not affected by optical obstruction) but is only available when sprayed with water.

[0062] This embodiment can dynamically select a calibration strategy based on the current environmental conditions and sensor status: before the cleaning operation (when the vehicle is stationary), it mainly relies on zero-speed correction + visual repositioning; during the cleaning operation (when the water mist is severe), it mainly relies on acoustic calibration; and during the cleaning interval (when the nozzle moves and the water spraying is paused), it restores visual repositioning.

[0063] Through the synergistic effect of triple calibration, the cumulative positional error after long-term operation (30 minutes) is controlled within ±8mm, meeting the accuracy requirements of high-pressure washing (requirement ±10mm) and close-fitting brushing (requirement ±5mm).

[0064] The robotic arm drives a servo motor to perform compensating movements in the same direction and with equal magnitude based on the calculated displacement of feature points, thus maintaining a constant relative position between the nozzle and the vehicle surface. The robotic arm is typically a six-degree-of-freedom tandem arm, with cleaning actuators such as a high-pressure nozzle and a rotating brush mounted at its end. Each joint is equipped with a high-precision servo motor, enabling precise control with an angular resolution of 0.01 degrees.

[0065] The execution of compensating motion first requires inverse kinematics solving, which involves calculating the rotation angles of each joint of the robotic arm based on the target position the nozzle needs to reach. Inverse kinematics solving employs analytical or numerical iteration methods. For a six-DOF robotic arm, there are typically multiple solutions for joint angles, and the optimal solution must be selected based on constraints such as joint limits and singularity avoidance. After obtaining the target joint angles, the servo motor controller calculates the motor's driving torque based on the deviation between the current joint angles and the target angles.

[0066] The control algorithm employs a three-loop cascaded control structure of position, velocity, and torque. The outermost position loop calculates the desired velocity based on the angular deviation, the middle velocity loop calculates the desired torque based on the velocity deviation, and the innermost torque loop achieves the desired torque output through current control. The unidirectional and equal-quantity characteristic of the compensating motion is reflected in the fact that when the vehicle moves or rotates in a certain direction, the end effector of the robotic arm also moves at the same speed and direction, ensuring that the nozzle always remains in the predetermined cleaning position when viewed in the vehicle coordinate system.

[0067] Specifically, when the vehicle moves slowly forward due to the handbrake not being engaged, the robotic arm moves forward synchronously; when the vehicle lowers due to suspension compression, the robotic arm lowers synchronously. This real-time tracking and compensation control frequency is synchronized with the data update frequency of the inertial measurement unit, reaching 100Hz. During the execution of the compensation motion, the encoder of the servo motor provides feedback on the actual joint angle, forming a closed-loop control that eliminates the effects of transmission backlash and load disturbances. Through precise compensation motion control, the distance between the nozzle and the vehicle surface can be stably maintained at the preset optimal cleaning distance, typically 8 to 12 centimeters.

[0068] To further improve the real-time performance and accuracy of the compensation motion, this embodiment also introduces a feedforward compensation mechanism based on Kalman prediction. Specifically, the vehicle's motion state is first modeled as a linear dynamic system: ; Where the state vector Includes position, velocity, attitude angle, and angular velocity; the system matrix A describes the natural evolution of the state, such as the velocity integral being the position; control input. For the vehicle's acceleration and angular acceleration; This is process noise, used to simulate the randomness of vehicle micro-movements.

[0069] Based on 10 historical frames of IMU data, the Kalman filter estimates the current state. and its covariance P k ; Then predict the state for the next 5-10 frames (50-100ms): ; The predicted time domain n is dynamically adjusted according to the 5G link delay: when the delay is <10ms, n=5 (predict 50ms); when the delay is 10-25ms, n=delay value (ms) / 10 (predict to compensated delay); when the delay is >25ms, n=10 (predict 100ms, enter flexible mode).

[0070] Based on the predicted vehicle pose Calculate the target position of the end effector (nozzle) of the robotic arm in advance.

[0071] Let the coordinates of the target feature point being cleaned in the vehicle coordinate system {V} be... (For example, the coordinates of the center point of the car door in the vehicle coordinate system are...) (Unit: meters)

[0072] Step 1: Based on the predicted vehicle pose, construct the transformation matrix from the vehicle coordinate system to the world coordinate system: ; in: It is a 3×3 rotation matrix, constructed from the predicted Euler angles (pitch φ, pitch θ, yaw ψ), i.e. , The predicted position of the vehicle's center in the world coordinate system (3×1 vector). Step 2: Transform the target feature points from the vehicle coordinate system to the world coordinate system: ; Step 3: In the world coordinate system, calculate the expected offset vector of the nozzle relative to the feature point. The nozzle should be maintained at a certain distance d (usually 8-12 cm) outside the normal direction of the feature point. V Defined in the vehicle coordinate system (e.g., the external normal of the door is...). ).

[0073] Transform the normal vector to the world coordinate system: ; Calculate the target position of the nozzle: ; First item This refers to rotating the "feature points + offset" from the vehicle coordinate system to the world coordinate system; the second item... By adding the world coordinates of the vehicle's center, a translation transformation is completed. When the vehicle tilts / pitches / yaws, the rotation matrix R will change accordingly to ensure that the nozzle is always aligned with the correct position on the vehicle's surface.

[0074] Step 4: Convert the target position into joint angle commands using inverse kinematics: ; The inverse kinematics considers not only position constraints but also nozzle attitude constraints (the spray direction should be along...). ).

[0075] For example, assuming the vehicle's initial position No rotation (R=I); the target feature point is the center of the right door. The desired nozzle distance d = 0.1m, normal direction (Pointing outwards); then the target position of the nozzle is: ; When the vehicle yaws ψ=10°, the rotation matrix This makes the position of the feature point in the world coordinate system become The normal vector becomes The nozzle target position becomes .

[0076] The robotic arm moves to a new target position in advance through predictive compensation, maintaining a constant relative position with the vehicle body.

[0077] The target position is converted into joint angle commands through inverse kinematics and sent to the servo motors in advance. ; When the actual IMU data arrives, calculate the prediction error: ; Adaptive prediction is achieved by updating the parameters of the Kalman filter (process noise covariance Q and measurement noise covariance R) based on the prediction error. If the prediction error exceeds a threshold (||e^(-Q / R)), the Kalman filter is updated accordingly. predict If the value is greater than 5mm, a quick correction is triggered: the robotic arm is switched to speed control mode and moved to the correct position at the maximum speed (200mm / s), with a correction time of <0.5s.

[0078] Through predictive compensation, the robotic arm's movement can be initiated 50-100ms in advance, offsetting the effects of 5G communication latency and robotic arm response latency.

[0079] Actual test data shows that without predictive compensation, the peak relative position error between the nozzle and the vehicle body can reach 15mm (when the vehicle suddenly moves slightly); with predictive compensation, the peak relative position error is reduced to less than 5mm, and the root mean square error is reduced from 8mm to 3mm.

[0080] The predictive compensation mechanism and the triple calibration mechanism work together to form the core technical architecture of the dynamic virtual coupling model, realizing complete closed-loop control of measurement, calibration, prediction, and compensation.

[0081] This implementation achieves real-time and precise compensation for vehicle micro-motions by constructing a dynamic virtual coupling model based on data from the vehicle's inertial measurement unit. The method utilizes quaternion integration and complementary filtering algorithms to accurately calculate the vehicle's attitude changes, effectively suppressing the integral drift error of the inertial sensor. By establishing a static bounding box model of the vehicle and defining preset feature points, the overall motion of the vehicle is decomposed into the spatial displacements of each feature point, providing a clear compensation target for the robotic arm. The servo motor-driven, same-direction, equal-quantity compensation motion ensures that the nozzle maintains a constant relative position to the vehicle surface. This dynamic coupling mechanism requires no physical connection and relies entirely on data communication and motion control.

[0082] In one embodiment of this invention, the vehicle-side inertial measurement unit data includes roll angle, pitch angle, and yaw angle. The displacement of each preset feature point on the vehicle surface is calculated based on the attitude change, including the following steps: S310. Determine the static bounding box model of the vehicle based on visual features, and obtain the vehicle size data in the static bounding box model. S320. Based on the roll angle, pitch angle, yaw angle and vehicle size data, calculate the three-dimensional spatial displacement vector of each preset feature point on the vehicle surface; S330. Map the three-dimensional spatial displacement vector to the coordinate system of the intelligent car wash machine to obtain the displacement of each preset feature point on the vehicle surface.

[0083] In this embodiment, the visual data includes still images of the vehicle. Determining the static bounding box model of the vehicle based on visual features includes: The center points of the four wheel hubs of a vehicle in a still image are identified as reference feature points; In response to the vehicle size data uploaded by the vehicle via the 5G heterogeneous data synchronization link, the static bounding box model is constructed based on the reference feature points. The vehicle size data includes vehicle length, vehicle width, vehicle height, and wheelbase.

[0084] Before calculating the displacement of each preset feature point on the vehicle surface, a static bounding box model of the vehicle needs to be established as a geometric reference. The static bounding box model is a three-dimensional cuboid frame that tightly wraps around the shape of the vehicle. The establishment of this model depends on the accurate identification of visual features and the acquisition of vehicle size data.

[0085] When a vehicle enters the car wash area and remains relatively stationary, the smart car wash machine's camera captures a still image of the vehicle. After image acquisition, a deep learning object detection algorithm is used to analyze the image, focusing on identifying the center points of the vehicle's four wheel hubs as reference feature points. The selection of wheel hub center points has significant engineering implications because the wheel hubs are rigid connection points of the vehicle's chassis, their positions are relatively stable, and they are easily visually identifiable. The identification process employs a circular object detector using the Hough circle detection algorithm, which can accurately locate the center coordinates of the four wheel hubs against complex backgrounds. Through image fusion from multi-view cameras and the principle of triangulation, the three-dimensional spatial coordinates of the four wheel hub center points in the world coordinate system can be calculated.

[0086] After identifying the reference feature points, the vehicle actively uploads its dimensions via a 5G heterogeneous data synchronization link. This data is stored in the vehicle's electronic control unit, including precise geometric parameters such as vehicle length, width, height, and wheelbase. Vehicle length refers to the distance from the frontmost point of the front bumper to the rearmost point of the rear bumper; vehicle width refers to the maximum width between the outer edges of the side mirrors; vehicle height refers to the vertical distance from the ground to the highest point of the roof; and wheelbase refers to the horizontal distance from the center of the front axle to the center of the rear axle.

[0087] Based on the spatial coordinates of the four wheel hub center points and vehicle dimensions, a static body kit model can be accurately constructed. The construction process begins by determining the vehicle's longitudinal axis based on the line connecting the front and rear wheel hub center points; this axis defines the vehicle's forward direction. Then, the vehicle's lateral axis is determined based on the line connecting the left and right wheel hub center points; this axis is perpendicular to the longitudinal axis. The relative positions of the front and rear wheel hubs are verified using wheelbase data, and the relative positions of the left and right wheel hubs are verified using track data. Using the plane containing the four wheel hub center points as the base, the top surface of the body kit is determined by extending upwards based on the vehicle's height data. Finally, based on the vehicle's length and width data, and using the wheel hub center points as references, the six sides of the body kit are determined by expanding forward, backward, left, and right.

[0088] The resulting static bounding box model fully describes the vehicle's spatial footprint. The origin of the model's coordinate system is typically set at the vehicle's geometric center or rear axle center, with the coordinate axes aligned with the vehicle's longitudinal, lateral, and vertical directions. The static bounding box model not only provides overall vehicle dimensions but also offers a unified coordinate reference framework for defining predefined feature points.

[0089] After establishing the static bounding box model, the three-dimensional spatial displacement vectors of each preset feature point on the vehicle surface need to be calculated based on the roll angle, pitch angle, yaw angle, and vehicle size data provided by the vehicle-side inertial measurement unit. These preset feature points are a series of key location points predefined in the static bounding box model. These points are distributed in important areas of the vehicle surface and are used to guide the cleaning path planning of the car wash robotic arm. Typical feature points include the four corners of the roof, the centers of the front and rear windshields, the door handle positions, the centers of the front and rear bumpers, and the fender positions above the wheel hubs. Typically, 25 to 35 feature points are defined on a standard sedan.

[0090] Each feature point has fixed three-dimensional coordinates in the vehicle coordinate system, which are determined based on the size data of the static bounding box model and the geometric features of the standard vehicle model. When the vehicle's attitude changes, the vehicle-side inertial measurement unit outputs the roll angle, pitch angle, and yaw angle in real time. The roll angle reflects the vehicle's rotation about the longitudinal axis, the pitch angle reflects the vehicle's rotation about the lateral axis, and the yaw angle reflects the vehicle's rotation about the vertical axis.

[0091] The calculation of the three-dimensional spatial displacement vector employs the rigid body kinematics transformation method. This method treats the vehicle as a rigid body, and the motion of any point within it can be decomposed into translational motion along with the overall vehicle and rotational motion about the vehicle's center. For coordinates... For the i-th feature point, first calculate its position vector relative to the vehicle center. ,in Let the coordinates be the vehicle center coordinates. Then, construct a rotation transformation matrix, which is built from the three attitude angles through a ZYX Euler angle rotation sequence.

[0092] position vector Multiplying by the rotation matrix on the left yields the rotated position vector. The new coordinates of the feature point after rotation are the vehicle center coordinates plus the rotated position vector. The three-dimensional spatial displacement vector is the difference between the new coordinates and the original coordinates, expressed as... This displacement vector contains components in the X, Y, and Z directions, fully describing the trajectory of the feature point in space. Since the attitude changes of the vehicle are usually small, the roll and pitch angles are generally within ±5 degrees, and the yaw angle is within ±10 degrees. Therefore, the amplitude of the displacement vector is usually in the centimeter range.

[0093] After obtaining the three-dimensional spatial displacement vectors of each preset feature point in the world coordinate system or vehicle coordinate system, these displacement vectors need to be mapped to the intelligent car wash machine coordinate system so that the robotic arm can perform compensating motion based on the mapped displacement. The intelligent car wash machine coordinate system is a fixed coordinate system established with the car wash machine base as the origin, and the motion planning and control of the robotic arm are all carried out in this coordinate system.

[0094] The essence of coordinate system mapping is coordinate transformation, which requires determining the spatial relationship between the world coordinate system and the car wash machine coordinate system. This relationship is determined during the installation and commissioning phase of the car wash machine through a system calibration program. The calibration process typically uses precision measuring equipment such as a calibration plate or laser tracker to measure the position of the origin of the car wash machine coordinate system in the world coordinate system, as well as the rotational relationship between the coordinate axes of the two coordinate systems. The calibration result is represented as a 4×4 homogeneous transformation matrix, which contains a 3×3 rotation submatrix and a 3×1 translation vector.

[0095] The coordinate mapping calculation process involves representing the three-dimensional spatial displacement vector of the feature point in homogeneous coordinate form, that is, adding a component 1 to the three-dimensional vector to form a four-dimensional vector. Then, the four-dimensional vector is multiplied on the left by the homogeneous transformation matrix to obtain the displacement vector in the car wash coordinate system. Since the displacement vector is a relative quantity rather than an absolute position, only rotation transformation is needed during the transformation process, and translation transformation is not required. Specifically, the fourth component of the homogeneous coordinates of the displacement vector is set to 0, so that the translation vector part has no effect in the matrix multiplication, and only the rotation matrix affects the displacement vector.

[0096] The displacement vector obtained after transformation is the displacement of the feature point in the car wash machine coordinate system. The three components of this displacement correspond to the X, Y, and Z axes of the car wash machine coordinate system, respectively. After receiving these displacement data, the motion controller of the robotic arm uses them as the target position increment of the end effector. It then calculates the required rotation angle increment of each joint through inverse kinematics and drives the servo motor to perform compensating motion.

[0097] The accuracy of coordinate mapping directly affects the precision of compensated motion. Errors in the mapping can cause a mismatch between the robotic arm's movement direction or amplitude and the vehicle's actual movement. Therefore, during system operation, coordinate system calibration verification needs to be performed periodically. This involves moving the robotic arm to a known position and comparing the deviation between the actual and theoretical positions to evaluate the accuracy of the coordinate mapping. If the deviation exceeds a threshold, the calibration procedure needs to be re-executed to update the homogeneous transformation matrix.

[0098] This implementation constructs a static bounding box model based on visual features and vehicle size data, providing a precise geometric reference framework for defining vehicle surface feature points. This method utilizes the wheel hub center point as a reference feature point, combined with actively uploaded vehicle size data, to achieve rapid and accurate establishment of the vehicle's geometric model. By calculating the three-dimensional spatial displacement vectors of each preset feature point based on roll angle, pitch angle, yaw angle, and vehicle size data, the overall attitude change of the vehicle is accurately decomposed into the spatial motion of each surface point. The rigid body kinematics transformation method ensures the theoretical accuracy of the displacement vector calculation, realistically reflecting the impact of vehicle micro-motions on the positions of each feature point. By mapping the three-dimensional spatial displacement vectors to the car wash machine coordinate system, a direct correlation between vehicle motion information and robotic arm control commands is achieved, eliminating control deviations caused by coordinate system differences.

[0099] In one embodiment of this invention, after the robotic arm drives the servo motor to perform a compensating motion of the same direction and amount based on the displacement, the following steps are also included: S410, In response to receiving an audio signal reflected from a high-pressure water jet hitting the vehicle body, determine the echo delay and spectral attenuation rate of the audio signal; S420. Calculate the actual physical distance between the nozzle and the vehicle body based on the echo delay and spectral attenuation rate. S430. Compare the actual physical distance with the theoretical distance calculated based on the data from the vehicle-mounted inertial measurement unit to obtain the distance deviation; S440. Generate an error correction vector based on the distance deviation, and adjust the spatial coordinates of the static bounding box model based on the error correction vector to eliminate the integral drift error of the vehicle-end inertial measurement unit.

[0100] After the robotic arm performs compensating movements and begins the cleaning operation, high-pressure water jets are ejected from the nozzles, striking the vehicle's surface at a speed of 15 to 25 meters per second. The moment the water jets hit the vehicle, they generate a strong acoustic signal containing rich physical information that can be used to calculate the actual distance between the nozzles and the vehicle.

[0101] Intelligent car wash machines are equipped with highly sensitive acoustic sensor arrays installed at the joints or base of their robotic arms. These sensors utilize MEMS microphone technology, with a frequency response range covering 20Hz to 20kHz, enabling them to capture full-frequency audio signals generated by the impact of water flow. When high-pressure water flows onto the car body, the sound waves generated at the impact point propagate outwards in the form of spherical waves. Part of these sound waves travel directly to the acoustic sensors, forming direct sound signals; the other part is reflected off the car body surface and reaches the sensors, forming reflected sound signals or echoes.

[0102] The audio signal received by the acoustic sensor is a superposition of direct sound and echo. Signal processing algorithms can separate the echo component and extract its characteristic parameters. Echo delay refers to the time difference between the moment the water hits the vehicle body and the moment the echo arrives at the sensor; this delay reflects the length of the sound wave's propagation path. The delay is measured using a cross-correlation algorithm. The received signal is cross-correlated with a known water-impact sound template. The peak position of the cross-correlation function corresponds to the arrival time of the echo, and the difference between the peak position and the impact time is the echo delay.

[0103] To improve the accuracy of time delay measurement, oversampling technology was used to increase the sampling rate of the audio signal from the standard 48kHz to 192kHz, achieving a time delay resolution of approximately 5 microseconds, corresponding to a distance resolution of approximately 1.7 millimeters. Spectral attenuation rate refers to the degree of energy attenuation of the echo signal relative to the direct sound signal in the frequency domain. This parameter reflects the energy loss of sound waves during propagation and reflection. The calculation of spectral attenuation rate first involves performing Fast Fourier Transform on both the direct sound signal and the echo signal to obtain their respective spectral distributions. Then, the energy ratio of the two spectra at each frequency point is calculated, and the logarithm is taken to obtain the attenuation rate curve.

[0104] Sound waves of different frequencies attenuate at different rates when propagating in air, with higher frequencies attenuating faster. Therefore, the spectral attenuation rate curve shows an increasing trend with frequency. By fitting the slope of the attenuation rate curve, a comprehensive spectral attenuation rate parameter can be extracted. This parameter is related not only to the propagation distance but also to the characteristics of the propagation medium. In a car wash environment, the water mist in the air increases the attenuation of sound waves.

[0105] In practical implementation, the acoustic sensor continuously acquires audio signals at a frequency of 1 kHz, generating one frame of audio data every 1 millisecond, and calculating echo delay and spectral attenuation rate in real time. These parameters are smoothed using a sliding window averaging filter to eliminate transient noise interference. Through acoustic analysis of the audio signals, a distance sensing channel independent of visual and inertial sensors is established, unaffected by optical obstruction from water mist and foam.

[0106] After obtaining the echo delay and spectral attenuation rate of the audio signal, it is necessary to calculate the actual physical distance between the nozzle and the car body based on these acoustic parameters. The basic principle of distance calculation is the time-distance relationship of sound wave propagation, that is, distance equals the speed of sound multiplied by the propagation time. However, in a car wash environment, the propagation path of sound waves is not a simple straight line, but a complex path involving the nozzle to the point of impact on the car body and then reflection to the sensor.

[0107] To simplify calculations, it is assumed that the nozzle, impact point, and sensor are approximately collinear. This is generally reasonable in the geometry of the robotic arm, as the sensor is mounted close to the nozzle. Under this assumption, the echo delay corresponds to a propagation distance approximately twice the distance from the nozzle to the vehicle body. The speed of sound needs to be considered in relation to ambient temperature and humidity. Under standard atmospheric conditions, the speed of sound is approximately 343 meters per second, but in a car wash environment, the speed of sound will deviate due to the presence of water mist and temperature variations. By installing temperature and humidity sensors on the car wash machine to measure environmental parameters in real time, the speed of sound value is corrected using empirical formulas.

[0108] Substituting the corrected sound velocity into the distance calculation formula yields a preliminary estimate of the distance between the nozzle and the vehicle body. However, this estimate still contains errors, primarily due to the additional attenuation of sound waves in the water mist medium and the multipath propagation effect. Therefore, a spectral attenuation rate parameter is introduced for correction. The spectral attenuation rate reflects the degree of sound wave energy loss. By establishing an empirical model relating the attenuation rate to the propagation distance and water mist concentration, the distance estimate can be compensated for. Empirical models typically employ polynomial fitting or neural network fitting, and are trained using extensive experimental data during the system debugging phase.

[0109] To further improve ranging accuracy, the acoustic signature of water impacting different materials on the vehicle body can be extracted from the audio signal. Different parts of the vehicle body are made of different materials; metal panels, plastic bumpers, and rubber seals, among others, produce different sound spectrum characteristics when impacted by water. Through spectrum analysis and machine learning classification algorithms, the type of vehicle body material in the current cleaning area can be identified. Different materials have different sound wave reflection characteristics; metal surfaces have a higher sound reflection coefficient, while plastic surfaces have a higher sound absorption coefficient.

[0110] The system loads and identifies the sound wave propagation speed and energy attenuation coefficient corresponding to the material type from a pre-set acoustic parameter database. These parameters are the acoustic properties of the material determined through offline experiments. The material-related acoustic parameters are then substituted into the distance calculation model to obtain distance calculation results optimized for the current material.

[0111] After obtaining the actual physical distance between the nozzle and the vehicle body using acoustic methods, this measurement needs to be compared with the theoretical distance calculated based on data from the vehicle-mounted inertial measurement unit (IMU) to assess the cumulative error of the inertial navigation system. The theoretical distance refers to the expected distance between the nozzle and the vehicle body calculated via forward kinematics based on the vehicle's initial attitude, motion data output by the IMU, and motion commands from the robotic arm.

[0112] The theoretical distance calculation process first obtains the coordinates of the current cleaning target point in the vehicle coordinate system from the static bounding box model. These coordinates are predefined feature point positions. Then, based on the vehicle's current attitude obtained by integration from the vehicle-side inertial measurement unit, the feature point coordinates are transformed from the vehicle coordinate system to the world coordinate system, and further transformed to the car wash machine coordinate system. Simultaneously, based on the encoder feedback angles of each joint of the robotic arm, the current position of the nozzle in the car wash machine coordinate system is calculated through forward kinematics. The theoretical distance is the Euclidean distance between the nozzle position and the target point position.

[0113] This theoretical distance represents the distance that the nozzle should maintain between the vehicle body under ideal conditions, i.e., when the sensor has no error and the motion control is completely precise. However, in actual operation, the acceleration and angular velocity measurements of the vehicle-side inertial measurement unit have zero bias, noise, and scale factor errors. These errors accumulate and are amplified after integration, causing the calculated vehicle attitude to gradually deviate from the actual attitude, thus resulting in a deviation between the theoretical and actual distances.

[0114] Distance deviation is calculated by subtracting the theoretical distance from the actual physical distance. The sign of the deviation value reflects the direction of the error: a positive deviation indicates that the actual distance is greater than the theoretical distance, meaning that the vehicle's actual position is farther than the position calculated by inertia; a negative deviation indicates the opposite. The magnitude of the deviation value reflects the severity of the error. When the absolute value of the deviation exceeds a preset threshold, it indicates that the integral drift of the inertial measurement unit has reached a level that requires correction.

[0115] In practical implementation, a deviation threshold of 10 mm is set. When a deviation exceeding this threshold is detected, the error correction process is triggered. To avoid accidental triggering due to random measurement noise, a strategy of continuous multiple detections is adopted. Only when the deviation of five consecutive measurements exceeds the threshold is a systematic error confirmed and correction initiated. Statistical analysis of distance deviation can also be used to assess the long-term stability of the system. By recording the trend of deviation changes over time, sensor aging or drift of calibration parameters can be identified.

[0116] After determining the distance deviation, an error correction vector needs to be generated based on the deviation value. This vector is then used to adjust the spatial coordinates of the static bounding box model, thereby eliminating the integral drift error of the vehicle-mounted inertial measurement unit. Generating the error correction vector requires converting the scalar distance deviation into a vector correction in three-dimensional space. This conversion depends on the relative positional relationship between the nozzle and the target point.

[0117] First, calculate the unit direction vector pointing from the nozzle to the target point. This vector is obtained by subtracting the nozzle coordinates from the target point coordinates and then normalizing. The error correction vector is the unit direction vector multiplied by the distance deviation value. This vector points in the direction that needs correction, and its magnitude is equal to the distance that needs correction. For example, if the distance deviation is positive 10 millimeters, meaning the actual position of the target point is 10 millimeters farther than the theoretical position, then the error correction vector points away from the nozzle and has a length of 10 millimeters.

[0118] The error correction vector, expressed in the car wash machine coordinate system, needs further transformation to the vehicle coordinate system to adjust the static bounding box model. The coordinate transformation is achieved using the inverse transformation matrix from the car wash machine coordinate system to the vehicle coordinate system. Multiplying the error correction vector by the inverse transformation matrix on the left yields the correction vector in the vehicle coordinate system. Adjusting the static bounding box model involves adding the correction vector to the coordinates of all preset feature points in the model, achieving an overall translation.

[0119] To avoid introducing new oscillations or instabilities during the correction process, a gradual correction strategy is adopted. This means that instead of applying the entire correction amount to the model all at once, it is applied gradually over multiple control cycles. Specifically, the error correction vector is multiplied by a gain coefficient less than 1, such as 0.2, and then this reduced correction amount is applied in each control cycle. After five cycles, the deviation is completely eliminated. This gradual correction method can smoothly adjust the system state and avoid the shocks caused by sudden changes.

[0120] The correction process also needs to consider the consistency of corrections for multiple feature points. When distance measurements are performed on multiple feature points simultaneously, inconsistent deviation values ​​may be obtained. In this case, multiple correction vectors need to be fused. The fusion method uses a weighted average, with weights determined based on the signal-to-noise ratio or measurement confidence of each measurement point. Measurement points with higher signal-to-noise ratios are assigned greater weights. The resulting composite correction vector is used as the final adjustment and applied to the static bounding box model.

[0121] The offline calibration process of the acoustic parameter database includes the following steps: (1) Sample collection. Specifically, in the anechoic laboratory, three types of standard material samples, namely glass (4-6mm thick tempered glass), metal (0.8-1.2mm steel plate + 30-50μm car paint) and tire (rubber hardness 60-70 Shore A), are subjected to impact tests at 1cm intervals within a range of 5-20cm using high-pressure water flow with a pressure of 8-15MPa and a flow rate of 15-25L / min. (2) Multi-parameter synchronous measurement. Specifically, a laser rangefinder with an accuracy of 0.1mm is used to measure the actual distance as the reference value; a sound level meter (frequency range 20Hz-20kHz, accuracy 0.1dB) is used to measure the echo signal intensity; a high-speed microphone array (sampling rate 192kHz) is used to record the complete acoustic signal; and a temperature and humidity sensor (accuracy ±0.1℃ / ±1%RH) is used to record environmental parameters. (3) Parameter extraction and modeling. Specifically, for each set of test data, the echo delay τ is extracted using a cross-correlation algorithm with an accuracy of 5μs; the energy ratio E between the direct sound and the echo is calculated. e / E d Establish an attenuation model: ; Where R is the material reflection coefficient, α is the overall attenuation coefficient (unit: Np / m), and d is the propagation distance.

[0122] The characteristic parameters of each material were obtained by fitting using the least squares method: glass: R=0.92±0.03, α=0.15±0.02Np / m; metal: R=0.75±0.05, α=0.20±0.03Np / m; tire: R=0.35±0.08, α=0.35±0.05Np / m.

[0123] (4) Construct an environmental correction model, specifically, establish correction formulas for sound speed and temperature / humidity: Where T is the temperature in Celsius and H is the relative humidity (%).

[0124] (5) Based on the measured echo delay τ and energy attenuation ratio Solve the system of equations: , ; The corrected distance d is obtained through iterative solution. c .

[0125] To address the strong noise interference in the car wash environment, the following signal processing strategies are adopted: (1) Adaptive noise cancellation, i.e., when the water pump is started but not spraying water, 30 seconds of background noise is collected to establish a noise template N(f); the real-time signal is passed through an adaptive filter: , where β is the adaptive coefficient, dynamically adjusted according to the frequency band signal-to-noise ratio (0.6-0.9). (2) Short-time Fourier transform (window length 25ms, overlap rate 50%) is used to generate the time spectrum; the water flow impact event is manifested in the time spectrum as a broadband energy surge lasting 50-100ms; the effective impact event is located by energy threshold detection (more than 15dB higher than the background). (3) The phase difference information of the microphone array (3 microphones, spacing 8cm) is used; the sound source in the non-target direction (outside ±30° of the nozzle-vehicle line) is suppressed by the beamforming algorithm; the spatial filtering gain can reach 12-18dB.

[0126] This implementation effectively solves the integral drift problem of vehicle-side inertial measurement units (IMUs) during long-term operation. The method utilizes the audio signal generated by high-pressure water jets striking the vehicle body. Through analysis of echo delay and spectral attenuation, it achieves accurate measurement of the actual physical distance between the nozzle and the vehicle body. The acoustic ranging method is unaffected by optical obstruction from water mist and foam, enabling stable operation even in harsh car wash environments. By comparing the actual distance measured acoustically with the theoretical distance calculated inertial navigation, the cumulative error of the inertial navigation system can be detected promptly. When the distance deviation exceeds a threshold, an error correction process is triggered. The generation of the error correction vector and the dynamic adjustment of the static bounding box model achieve real-time compensation for IMU drift errors. A progressive correction strategy and a multi-point fusion algorithm ensure the smoothness and robustness of the correction process.

[0127] In one embodiment of this invention, the actual physical distance between the nozzle and the vehicle body is calculated based on the echo delay and spectral attenuation rate, including the following steps: S510: Extract the acoustic signature features generated by water flow hitting different materials of the car body from the audio signal; S520: Determine the vehicle body material type of the current cleaning area based on the voiceprint characteristics; S530: Load the sound wave propagation speed and energy attenuation coefficient corresponding to the vehicle body material type from the preset acoustic parameter database; S540. Calculate the actual physical distance between the nozzle and the vehicle body based on the echo delay, sound wave propagation speed, and energy attenuation coefficient.

[0128] In this embodiment, the materials may include glass, metal, and tires.

[0129] During car washing, the sounds produced when high-pressure water jets strike different parts of the car body exhibit significant material-related characteristics. These acoustic differences provide a physical basis for material identification. The car body surface is mainly composed of different materials such as glass, metal, and tires. Glass is primarily found in the front and rear windshields, side windows, and sunroof areas, and is characterized by high hardness, high density, and a smooth surface. Metal is mainly distributed in body panels such as doors, hood, roof, and fenders, and is typically made of steel or aluminum alloy with a paint coating. Tires are made of rubber composite materials, which have high elasticity and a rough surface texture.

[0130] When high-pressure water jets impact these different materials at speeds of 15 to 25 meters per second, the spectral characteristics of the resulting sound waves differ fundamentally. The extraction of acoustic signature features begins with preprocessing the audio signals acquired by acoustic sensors, including removing background noise and mechanical vibration interference. The preprocessing employs an adaptive filtering algorithm, which can identify and suppress continuous noise sources such as the operation of the car wash robot arm motor and the water pump. In the denoised audio signal, an energy threshold detection algorithm is used to locate the time window of the water jet impact event. This window is typically 50 to 100 milliseconds, encompassing the shock wave at the moment of impact and the subsequent vibration decay process.

[0131] A short-time Fourier transform was performed on the extracted time-window signal to obtain a time-frequency spectrum, which shows the distribution characteristics of sound energy in both time and frequency dimensions. Due to its high stiffness and low damping characteristics, glass produces a sound spectrum concentrated in the high-frequency range when water impacts it, with the main energy distributed in the 2kHz to 8kHz range. The time-frequency spectrum exhibits sharp peaks, slow decay, and a long duration. Metal materials have a relatively broad sound spectrum, with energy distributed in the 500Hz to 5kHz range. Due to the resonant characteristics of metal sheets, multiple resonance peaks appear in the spectrum, resulting in a multi-peak structure in the time-frequency spectrum, with a moderate decay rate. Tire materials, due to the high damping characteristics of rubber, produce a sound spectrum concentrated in the low-frequency range, with the main energy distributed in the 100Hz to 1kHz range. The time-frequency spectrum is broad and low-frequency, with a fast decay rate and short duration.

[0132] In addition to frequency domain features, time domain features are also extracted, including peak amplitude, rise time, decay time constant, and zero-crossing rate. Glass materials exhibit high peak amplitude, short rise time, and large decay time constant; metal materials have medium peak amplitude, and both rise and decay times are moderate; tire materials have low peak amplitude, long rise time, and small decay time constant. To construct a more robust voiceprint feature vector, the Mel frequency cepstral coefficient extraction method is employed. This method simulates the human ear's perception of sound, converting the linear spectrum to a Mel scale and then calculating the cepstral coefficients. Typically, 13-dimensional Mel frequency cepstral coefficients are extracted as the core part of the voiceprint feature, combined with time-domain statistical features, to form a comprehensive voiceprint feature vector of approximately 20 dimensions.

[0133] Specifically, the specific steps for extracting Mel frequency cepstral coefficients (MFCC) are as follows: (1) High-pass filtering is performed on the acquired audio signal s(n) to highlight the high-frequency components: The pre-emphasis factor of 0.97 was determined based on the spectral characteristics of the sound of water flow. (2) The signal was divided into 25ms frames, with a frame shift of 10ms (overlap rate of 60%); a Hamming window was applied to each frame to reduce spectral leakage. N is the number of sampling points corresponding to the frame length (N=1200 at a sampling rate of 48kHz); (3) Perform a 512-point fast Fourier transform on the windowed frame to obtain the spectrum X(k); calculate the power spectrum: (4) Design 26 triangular Mel filters, covering a frequency range of 20Hz-8kHz; the conversion relationship between Mel scale and frequency: The filter center frequencies are uniformly distributed on the Mel scale and non-linearly distributed on the linear frequency range; calculate the energy of each filter: , where H_m(k) is the frequency response of the m-th Mel filter. (5) Take the logarithm of the filter bank energy to simulate the logarithmic perception characteristics of the human ear. (6) Perform discrete cosine transform on the logarithmic energy to extract 13 cepstral coefficients: (7) Calculate the first-order and second-order differences; (8) Extract 7-dimensional time-domain statistical features: short-time energy. Zero crossing rate Peak amplitude, rise time, decay time constant, waveform factor, peak factor; (9) Final feature vector dimension = 13(MFCC) + 13(first-order difference) + 13(second-order difference) + 7(time domain) = 46 dimensions.

[0134] To reduce computational complexity, principal component analysis is used to reduce the dimensionality to 20 dimensions, while retaining more than 95% of the variance contribution rate.

[0135] The training process for a Support Vector Machine (SVM) classifier is as follows: Step 1: Collect 500 samples each of glass, metal and tire materials, for a total of 1500 samples; the samples cover different water pressures (8-15MPa), different angles (30-90°), and different ambient temperatures (5-35℃); divide the samples into training set (1050), validation set (300) and test set (150) in a 7:2:1 ratio.

[0136] Step 2: Use a radial basis function (RBF) kernel: The parameters γ and C are optimized through grid search. Search range: γ∈[0.001,0.1],C∈[1,100]; Optimal parameters: γ=0.01, C=10 (determined through 5-fold cross-validation).

[0137] Step 3: Using a one-to-one strategy, train three binary classifiers: glass-metal, glass-tire, and metal-tire. Voting determines the final category, and the category probability distribution is output.

[0138] After extracting the voiceprint feature vector, a pattern recognition algorithm is needed to determine the vehicle body material type of the current cleaning area. Material classification employs a supervised learning method, requiring offline training before system deployment. The training process collects a large number of water flow impact sound samples from different vehicle models and different material parts, with at least 500 samples collected for each material type, covering sound variations under different water pressures, impact angles, and environmental conditions. Voiceprint feature vectors are extracted from each sample and labeled with the corresponding material type to construct the training dataset.

[0139] The classifier selection considered a balance between real-time performance and accuracy, employing Support Vector Machines (SVM) as the core classification algorithm. SVM, by finding the optimal classifying hyperplane in a high-dimensional feature space, effectively handles nonlinearly separable problems and exhibits good generalization ability for small sample data. The training process uses a radial basis function kernel, and the kernel and penalty parameters are optimized through cross-validation. The trained classification model is then deployed to the edge computing unit of the car wash machine to achieve real-time material recognition.

[0140] In actual operation, after extracting new voiceprint feature vectors, they are input into the classifier for inference calculations. The classifier outputs probability distributions for three categories, corresponding to the confidence levels of glass, metal, and tires, respectively. The category with the highest probability is selected as the recognition result, and this probability value is recorded as the recognition confidence level. To improve the robustness of recognition, a time-series smoothing strategy is adopted, that is, voting or probability fusion is performed on the recognition results of multiple consecutive frames. Only when a certain material type is recognized in 5 consecutive frames or the cumulative probability exceeds a threshold is the material type switch confirmed.

[0141] During the cleaning operation, the robotic arm moves along a predetermined path. When the nozzle moves from the metal area of ​​the car door to the glass area of ​​the car window, the acoustic signature changes significantly. The classifier can detect the material switching event in a timely manner and update the current material type label. The material recognition results are not only used for subsequent acoustic ranging correction, but also for optimizing cleaning parameters. For example, in the glass area, the water pressure and approach distance can be appropriately increased; in the metal area, the water pressure needs to be controlled to avoid damaging the paint; and in the tire area, the scrubbing intensity can be increased.

[0142] After determining the type of vehicle body material in the current cleaning area, it is necessary to load the corresponding sound wave propagation speed and energy attenuation coefficient from a pre-set acoustic parameter database. These parameters are key inputs for accurately calculating the actual physical distance between the nozzle and the vehicle body. The acoustic parameter database was established during the system development phase through extensive experimental measurements. The experiments were conducted in standard laboratory environments and actual car wash environments to measure the propagation and reflection characteristics of sound waves on different materials.

[0143] For glass, the reflection coefficient of sound waves on the glass surface is relatively high, approximately 0.9 to 0.95, indicating that most of the sound energy is reflected back, with only a small amount being absorbed or transmitted. During the process of sound waves propagating through water-laden air to the glass surface and then reflecting back to the sensor, the energy attenuation coefficient is approximately 0.15 dB per meter. For metal, due to the paint coating on the surface, the sound wave reflection characteristics are between rigid reflection and partial absorption, with a reflection coefficient of approximately 0.7 to 0.8. The vibration and resonance that may exist in the metal sheet will consume some sound energy, making the reflected sound energy lower than that of glass. The energy attenuation coefficient of sound waves during propagation is approximately 0.20 dB per meter, slightly higher than that of glass.

[0144] For tire materials, the porous structure and high damping properties of rubber cause a large amount of sound waves to be absorbed, with a reflection coefficient of only 0.3 to 0.4. Most of the sound energy is converted into heat energy or attenuated within the material. The energy attenuation coefficient of sound wave propagation is approximately 0.35 dB per meter, significantly higher than that of glass and metal. The speed of sound propagation is approximately 343 meters per second under standard atmospheric conditions, but in a car wash environment, due to the presence of water mist, the effective speed of sound will be slightly reduced to approximately 340 meters per second.

[0145] The microscopic roughness of different material surfaces also affects the scattering characteristics of sound waves. Glass surfaces are smooth, with specular reflection being dominant; metal surfaces are relatively smooth, with specular reflection being dominant but with some diffuse reflection; tire surfaces are rough, with diffuse reflection being the predominant factor. The parameters stored in the database are organized in the form of lookup tables, indexed by material type. Each material corresponds to a set of acoustic parameters, including reflection coefficient, attenuation coefficient, and sound velocity correction value. When the material identification module outputs the current material type, the distance calculation module immediately reads the corresponding parameters from the database and loads them into the calculation formula. To adapt to different environmental conditions, the database also includes temperature and humidity correction tables, dynamically adjusting the acoustic parameters based on real-time measured environmental parameters.

[0146] After obtaining the echo delay, sound wave propagation speed, and energy attenuation coefficient, the actual physical distance between the nozzle and the vehicle body can be calculated comprehensively. The distance calculation uses a modified acoustic ranging model, which considers the influence of material-related acoustic properties on the measurement results. The basic distance calculation formula is: distance equals sound wave propagation speed multiplied by echo delay, then divided by 2. The division by 2 is because the sound wave undergoes a round trip path from the nozzle to the vehicle body and back.

[0147] However, this basic formula assumes that sound waves propagate in a homogeneous medium with no energy loss, requiring several corrections in a real car wash environment. First, energy attenuation correction is needed. The energy of a sound wave attenuates exponentially during propagation, and the intensity of the attenuated sound wave affects the accuracy of echo detection. By measuring the actual intensity of the echo signal and comparing it with the theoretical intensity, the actual propagation distance of the sound wave can be calculated. In the energy attenuation correction formula, the actual intensity equals the initial intensity multiplied by an exponential attenuation factor, where the exponent of the attenuation factor is a negative attenuation coefficient multiplied by the propagation distance.

[0148] Secondly, material reflection correction is performed. Different materials have different reflection coefficients, resulting in differences in the phase and amplitude characteristics of the echo signal. The high reflectivity of glass makes the echo signal clear and the time delay measurement accuracy high; the low reflectivity of tires makes the echo signal weak and the time delay measurement has greater uncertainty. By introducing material-related confidence weights, the distance measurement results are weighted and averaged or Bayesian fusion is performed to improve the reliability of distance measurement.

[0149] Further multipath propagation correction is performed. In a car wash environment, sound waves may reach the sensor after multiple reflections from various surfaces of the car body, creating a multipath effect. Multipath signals can interfere with the detection of direct echoes, leading to time delay measurement errors. By analyzing the time-domain waveform of the echo signal, the first arriving echo peak is identified, and subsequent multipath interference peaks are suppressed to ensure that the time delay measurement corresponds to the shortest propagation path.

[0150] Taking all the above correction factors into account, the final distance calculation formula is: corrected distance equals the basic distance multiplied by the energy attenuation correction factor, then multiplied by the material reflection correction factor, and finally subtracted from the deviation introduced by multipath propagation. This correction model significantly improves the accuracy of the calculated actual physical distance between the nozzle and the vehicle body.

[0151] This implementation method extracts the acoustic fingerprint features generated by water flowing onto the vehicle body and identifies the material type, enabling real-time perception of the vehicle body material during the washing process and providing material-related correction parameters for acoustic ranging. This method utilizes the acoustic fingerprint differences of different materials and employs a machine learning classification algorithm to accurately identify three main materials: glass, metal, and tires, significantly improving the recognition accuracy. By loading the sound wave propagation speed and energy attenuation coefficient corresponding to each material from an acoustic parameter database, material characteristics are incorporated into the distance calculation model, effectively eliminating the impact of material differences on ranging accuracy. The corrected acoustic ranging model comprehensively considers multiple physical factors such as energy attenuation, material reflection, and multipath propagation, effectively improving ranging accuracy. The material-adaptive acoustic ranging method is unaffected by the optical obstruction of water mist and foam, and can still work stably under the harsh conditions of a car wash environment.

[0152] In one embodiment of this invention, driving the intelligent car wash machine to perform a vehicle washing operation based on a dynamically locked coordinate system includes the following steps: S610, real-time monitoring of the real-time latency value of the 5G heterogeneous data synchronization link; S620: When the real-time delay value is lower than the first preset threshold, drive the robotic arm to maintain a high-rigidity mode and perform high-pressure rinsing or close-fitting brushing. S630. When the real-time delay value is higher than the second preset threshold, the proportional-integral-derivative parameters of the robotic arm joint are determined based on the real-time delay value, and the stiffness coefficient and damping coefficient are reduced based on the proportional-integral-derivative parameters, so that the robotic arm enters the flexible following mode. S640, in flexible following mode, drives the robotic arm to perform a passive retreat mode based on the vehicle's current contact physical data to prevent scratches on the paint.

[0153] Throughout the cleaning operation, the communication quality of the 5G heterogeneous data synchronization link directly affects the timeliness of vehicle motion information transmission, thus determining the synchronization accuracy of the robotic arm's compensating motion. The real-time latency value refers to the time interval between the moment the vehicle's inertial measurement unit collects motion data and the moment the car wash machine receives and processes that data. This latency includes the cumulative time of multiple stages such as data encapsulation, network transmission, data parsing, and processing.

[0154] To achieve accurate monitoring of communication latency, a timestamp mechanism is embedded in the data transmission protocol. The vehicle-side writes a high-precision timestamp of the transmission time into each data packet, provided by the vehicle's GPS timing system or a 5G network synchronized clock. Upon receiving a data packet, the car wash machine immediately reads its local clock, compares the received time with the transmission time in the data packet, and calculates the one-way transmission delay. Considering the potential clock discrepancy between the vehicle and car wash machines, a round-trip time (RTD) measurement method is used for calibration. The car wash machine periodically sends a time synchronization request to the vehicle, which responds immediately. The one-way delay is estimated using half of the RTD, and the clock discrepancy is corrected accordingly.

[0155] The real-time latency monitoring frequency is set to 100Hz, consistent with the inertial data update frequency; the latency value is calculated once for each received frame of data. To eliminate the impact of transient network jitter, a sliding window averaging filter is applied to the latency value, with a window length of 10 frames, averaging the latency values ​​within the most recent 100 milliseconds to obtain a smooth latency trend. The monitoring system also calculates the standard deviation of the latency value to assess network stability; a large standard deviation indicates severe network jitter, requiring a reduction in the system's reliance on real-time performance.

[0156] Under ideal 5G network conditions, end-to-end latency is typically in the range of 3 to 8 milliseconds. At this latency, vehicle movement information can be transmitted to the car wash machine almost in real time, and the robotic arm's compensating movements can remain highly synchronized with the vehicle's movement. However, under high network load, degraded signal quality, or during cell handover, latency can rise to 20 to 50 milliseconds or even higher. Increased latency means that the car wash machine receives vehicle location information with a time lag. If the robotic arm still compensates based on this delayed information, the actual relative position of the spray nozzle and the vehicle body will deviate from the expected position, increasing the risk of collision or inadequate cleaning. The monitoring system compares the latency value with a preset threshold and triggers a corresponding control mode switch.

[0157] When the real-time latency is below the first preset threshold, it indicates that the 5G heterogeneous data synchronization link is in good communication condition, and vehicle motion information can be transmitted to the car wash machine in a timely and accurate manner. At this time, the performance advantages of the dynamic virtual coupling model can be fully utilized. The first preset threshold is usually set to 10 milliseconds. This threshold is selected based on the response time of the robotic arm control system and the speed characteristics of the vehicle's micro-movements. With a latency of 10 milliseconds, even if the vehicle moves at a speed of 5 centimeters per second, the position lag is only 0.5 millimeters, which is far less than the positioning accuracy of the robotic arm and the tolerance range of the cleaning operation.

[0158] Under low-latency conditions, the robotic arm maintains a high-stiffness mode. This mode is characterized by high stiffness and high damping coefficients at the joints, enabling the robotic arm to respond quickly and accurately to control commands and remain stable without oscillation after reaching the target position. In high-stiffness mode, the joint stiffness coefficient is set to 100% of the nominal value, typically 10,000 Nm per radian, and the damping coefficient is set to 80% of the critical damping, ensuring rapid system convergence without over-damping.

[0159] In this mode, the robotic arm can perform cleaning operations requiring precise position control, such as high-pressure rinsing or close-fitting brushing. High-pressure rinsing involves aiming the nozzle at the vehicle surface from a distance of 8 to 12 centimeters, spraying a high-pressure water stream with a pressure of 8 to 15 MPa, using the impact force of the water stream to remove stubborn stains and dirt from the vehicle surface. This operating mode requires the nozzle to maintain a stable distance from the vehicle body; too close a distance will result in excessive water impact force, potentially damaging the paint or seals; too far a distance will result in insufficient impact force and reduced cleaning effectiveness. The high-rigidity mode ensures that the robotic arm can accurately maintain the preset distance, and even if the vehicle undergoes slight movements, the robotic arm can quickly follow and compensate, maintaining a constant relative position.

[0160] Close-fitting brushing refers to a rotating brush head that gently presses against the vehicle's surface, removing adhering dirt and insect remains through the mechanical friction of the bristles. This operating mode requires the brush head to maintain a constant contact pressure with the vehicle body; excessive pressure will scratch the paint, while insufficient pressure will result in poor cleaning. In high-rigidity mode, the robotic arm's force control is highly precise, able to accurately adjust the brush head pressure through feedback from the end effector force sensor, maintaining it within a safe range of 0.5 to 2 Newtons. Simultaneously, the high rigidity ensures that the robotic arm does not experience significant positional shift when subjected to brushing reaction forces, guaranteeing the accuracy of the brushing trajectory.

[0161] When the real-time latency exceeds the second preset threshold, it indicates a decline in the communication quality of the 5G heterogeneous data synchronization link, resulting in a significant time lag in vehicle motion information. Continuing to maintain the high-rigidity mode may cause the robotic arm's movement to become asynchronous with the vehicle's actual movement, increasing the risk of collision. The second preset threshold is typically set to 25 milliseconds, a threshold determined comprehensively based on system safety margins and operational quality requirements. With a 25-millisecond latency, if the vehicle moves at a speed of 5 centimeters per second, the positional lag reaches 1.25 millimeters, an error approaching the upper tolerance limit for cleaning operations.

[0162] To address high latency, the control parameters of the robotic arm need to be dynamically adjusted to reduce the system's dependence on real-time performance and enhance its robustness to latency. The core of control parameter adjustment is determining the proportional, integral, and derivative parameters of the robotic arm joints based on real-time latency values. This set of parameters is the key parameter of the control system, determining its response speed, steady-state accuracy, and stability. The proportional gain parameter controls the instantaneous response strength to errors; a larger gain results in a faster response but is more prone to oscillations. The integral parameter is used to eliminate steady-state errors; a smaller integral time constant indicates a stronger integral effect. The derivative parameter is used to predict error trends and provide damping; a larger derivative gain indicates stronger damping.

[0163] Under high latency conditions, maintaining the original high proportional gain will cause the system to overreact to the lagging position information, resulting in the robotic arm's movement lagging behind the vehicle's movement and causing oscillations. Therefore, it is necessary to reduce the proportional gain to slow down the system's response speed and make the robotic arm's movement smoother. Simultaneously, extending the integral time constant weakens the integral action and avoids excessive accumulation of lag errors. The derivative gain also needs to be reduced because, under latency conditions, the rate of change of the error is also lagging, and excessively strong derivative action will introduce unnecessary disturbances.

[0164] The specific parameter adjustment strategy employs a delay-adaptive algorithm. This algorithm calculates the parameter adjustment factor based on the amount by which the delay exceeds a second preset threshold. The larger the delay exceeds the threshold, the greater the parameter adjustment. The adjustment formula for the proportional gain is: dynamic proportional gain equals the baseline proportional gain multiplied by the attenuation factor. The attenuation factor decreases as the delay exceeds the threshold, typically using an exponential or linear attenuation function. The adjustment formula for the integral time constant is: dynamic integral time constant equals the baseline integral time constant multiplied by the amplification factor. The amplification factor increases as the delay exceeds the threshold. The adjustment formula for the derivative gain is: dynamic derivative gain equals the baseline derivative gain multiplied by the suppression factor. The suppression factor decreases as the delay exceeds the threshold.

[0165] By adjusting the proportional-integral-derivative (PID) parameters, the stiffness and damping coefficients of the robotic arm joints are further calculated. The stiffness coefficient is positively correlated with the proportional gain, and the damping coefficient is positively correlated with the derivative gain. Reducing the proportional and derivative gains results in a corresponding decrease in both the stiffness and damping coefficients, allowing the robotic arm to transition from a high-stiffness mode to a flexible following mode. The flexible following mode is characterized by lower stiffness and damping at the robotic arm joints, resulting in a slower response speed but enhanced adaptability to external disturbances. In this mode, even with a delay in vehicle motion information, the robotic arm can absorb positional deviations through flexible deformation, avoiding rigid collisions. In flexible following mode, the stiffness coefficient is reduced to 30% to 50% of its nominal value, and the damping coefficient is reduced to 40% to 60% of its critical damping.

[0166] In flexible following mode, the reduced stiffness of the robotic arm makes it more sensitive to external contact forces. Therefore, a passive yielding mechanism is needed to prevent accidental contact between the robotic arm and the vehicle body during delays, which could scratch the paint. The passive yielding mode is a force feedback-based safety protection strategy. By monitoring the physical data of the contact between the robotic arm and the vehicle in real time, when the contact force exceeds a safety threshold, the robotic arm is actively controlled to yield in the opposite direction of the contact force until the contact force is reduced to a safe range.

[0167] The current physical data primarily includes contact force and torque information measured by a six-dimensional force sensor at the end effector of the robotic arm. This sensor, installed between the end flange of the robotic arm and the nozzle or brush head, measures force components in three directions and torque components in three directions, with a measurement accuracy of 0.1 Newtons and 0.01 Newton-meters. During normal cleaning operations, the nozzle should not come into contact with the vehicle body, and the force sensor reading should be close to zero, with only the reaction force of the water flow and the small torque generated by the rotation of the brush head. When the nozzle or brush head comes into contact with the vehicle body due to delay or control error of the robotic arm, the force sensor will detect a sudden increase in the contact force signal.

[0168] The safe threshold for contact force is set based on the paint's tolerance and the characteristics of the cleaning tools. For soft brush heads, the safe threshold is set to 5 Newtons; for spray nozzles, since they should not come into contact with the car body, the threshold is set to 2 Newtons. The passive yielding mode execution process first reads force sensor data in real time, sampling at a frequency of 1kHz to ensure timely detection of contact events. The sampled data is compared with the safe threshold, and once the contact force exceeds the threshold, the yielding procedure is immediately triggered.

[0169] The yielding procedure first calculates the direction vector of the contact force, which is obtained by normalizing the three-dimensional force components measured by a force sensor and points in the direction of the contact force. Then, it calculates the target direction of the yielding motion, which is the opposite direction of the contact force direction vector, pointing away from the vehicle body. The speed of the yielding motion is adaptively adjusted according to the magnitude of the contact force; the greater the contact force, the faster the yielding speed. It is typically set to multiply the portion of the contact force exceeding a threshold by a speed gain coefficient, with the gain coefficient being 10 mm / s per Newton.

[0170] Upon receiving a retraction command, the robotic arm controller pauses the current cleaning trajectory and prioritizes the retraction motion, driving coordinated movement of all joints to move the end effector along the retraction direction. During the retraction process, the contact force is continuously monitored. When the contact force drops below a safety threshold, the retraction motion stops, and the robotic arm remains in the retracted position, awaiting delayed recovery or a replanning of the cleaning path. This passive retraction mode serves as a last line of defense, enabling timely detection and response to contact events even in extreme situations such as communication delays, sensor errors, or control failures through a force feedback mechanism.

[0171] This implementation achieves adaptive optimization of the cleaning operation under different communication quality conditions by dynamically adjusting the control mode of the robotic arm in real time based on the latency value of the 5G heterogeneous data synchronization link. Under low latency conditions, the high stiffness mode fully utilizes the real-time tracking performance of the dynamic virtual coupling model to achieve high-precision operations such as high-pressure rinsing and close-fitting brushing. Under high latency conditions, by adaptively adjusting the proportional-integral-derivative parameters based on the latency value, the stiffness coefficient and damping coefficient of the robotic arm are reduced, enabling the system to enter a flexible following mode. This effectively enhances the robustness to communication latency and avoids control oscillations and collision risks caused by latency. The passive yielding mechanism in the flexible following mode detects contact events through real-time force feedback and actively yields, providing a final safety guarantee for the system. This adaptive control strategy reduces the system's dependence on communication quality from an absolute requirement to a relative requirement, enabling the car wash machine to continue operating safely under non-ideal conditions such as 5G network load fluctuations and signal fading.

[0172] In one embodiment of this invention, the proportional-integral-derivative (PID) parameters of the robotic arm joints are determined based on the real-time delay value, and the stiffness coefficient and damping coefficient are reduced based on the PID parameters to enable the robotic arm to enter a flexible following mode. This includes the following steps: S701. Calculate the amount by which the real-time delay value exceeds the second preset threshold; S702. Calculate the proportional gain attenuation factor based on the delay excess, where the larger the delay excess, the smaller the proportional gain attenuation factor. S703. Multiply the preset reference proportional gain of the robotic arm joint by the proportional gain attenuation factor to obtain the dynamically adjusted proportional gain parameter. S704. Calculate the integral time constant amplification factor based on the delay excess, where the larger the delay excess, the larger the integral time constant amplification factor. S705. Multiply the reference integral time constant of the robotic arm joint by the integral time constant amplification factor to obtain the dynamically adjusted integral parameters. S706. Calculate the differential gain suppression factor based on the delay excess, where the larger the delay excess, the smaller the differential gain suppression factor. S707. Multiply the reference differential gain of the robotic arm joint by the differential gain suppression factor to obtain the dynamically adjusted differential parameters. S708. Calculate the reduction ratio of stiffness coefficient and damping coefficient based on the dynamically adjusted proportional gain parameter, integral parameter and differential parameter. S709. Multiply the reference stiffness coefficient of the robotic arm joint by the stiffness coefficient reduction ratio to obtain the actual stiffness coefficient in the flexible following mode. S710. Multiply the reference damping coefficient of the robotic arm joint by the damping coefficient reduction ratio to obtain the actual damping coefficient in the flexible following mode.

[0173] After detecting that the real-time latency of the 5G heterogeneous data synchronization link exceeds the second preset threshold, the first step is to accurately calculate the latency excess. This parameter is the basis for adjusting all subsequent control parameters. The latency excess is defined as the difference between the real-time latency value and the second preset threshold, reflecting the degree to which the current communication latency exceeds the system's acceptable latency limit. The calculation is achieved through a simple subtraction operation: the latency excess equals the real-time latency value minus the second preset threshold. For example, when the second preset threshold is set to 25 milliseconds and the real-time monitored latency value is 35 milliseconds, the latency excess is 10 milliseconds.

[0174] The magnitude of this excess directly reflects the degree of communication quality degradation; a larger excess indicates a more severe latency problem and a more significant impact on system real-time performance. To avoid frequent parameter adjustments due to instantaneous fluctuations in latency values, a time-window smoothing process is applied to the real-time latency value before calculating the latency excess. A moving average filter is used to average the latency values ​​of the most recent 10 frames to obtain a smoothed latency trend value, which is then compared with a threshold. This smoothing process filters out short-term network jitter, ensuring that only a sustained increase in latency triggers parameter adjustments, thus preventing control instability caused by frequent switching between normal and flexible modes.

[0175] Excess latency also needs to be subject to saturation limiting, with a maximum excess limit set, typically 50 milliseconds. Even if the actual latency exceeds 75 milliseconds, the excess is still calculated as 50 milliseconds. The purpose of this saturation limit is to prevent over-adjustment of parameters under extreme latency conditions, which could cause the system to completely lose its responsiveness. After the excess latency is determined, it is normalized to the range of 0 to 1. The normalized value is equal to the excess latency divided by the maximum excess limit. The resulting normalized excess latency serves as a unified input parameter for the calculation of subsequent adjustment factors.

[0176] Proportional gain is the most critical parameter in a control system, determining the system's immediate response to position errors. Under high latency conditions, position information lags behind. If a high proportional gain is maintained, the system will overrespond to the delayed error signal, causing the robotic arm's movement to lag behind the vehicle's actual movement and resulting in tracking oscillations. Therefore, a proportional gain attenuation factor needs to be calculated based on the excess latency. This factor reduces the proportional gain and slows down the system's response.

[0177] The proportional gain attenuation factor is inversely related to the delay excess; the larger the delay excess, the smaller the attenuation factor, and the greater the reduction in proportional gain. The attenuation factor is calculated using an exponential attenuation function, specifically, the proportional gain attenuation factor equals an exponential function, where the exponent is a negative attenuation coefficient multiplied by the normalized delay excess. The attenuation coefficient is a parameter that adjusts the attenuation rate, typically ranging from 2 to 5; the larger the coefficient, the faster the attenuation rate. For example, when the attenuation coefficient is 3 and the normalized delay excess is 0.2, the proportional gain attenuation factor is approximately 0.55, indicating that the proportional gain will decrease to 55% of its original value. When the normalized delay excess reaches 1, the attenuation factor decreases to approximately 0.05, and the proportional gain decreases to 5% of its original value.

[0178] The choice of the exponential decay function is based on stability analysis in control theory. This function maintains a high gain to sustain response performance when the delay is small, and rapidly reduces the gain to ensure stability when the delay is large. After calculating the proportional gain decay factor, it is multiplied by the preset reference proportional gain of the robotic arm joints to obtain the dynamically adjusted proportional gain parameters. The reference proportional gain is the optimal gain value obtained through system identification and optimization tuning under no-delay or low-delay conditions. For a typical six-degree-of-freedom car wash robotic arm, the reference proportional gain of each joint varies depending on the load inertia and stiffness requirements, typically ranging from 100 to 500.

[0179] Integral parameters are used in control systems to eliminate steady-state errors. By integrating the error signal over time, the control quantity is gradually accumulated until the error returns to zero. The integral time constant defines the strength of the integral action; a smaller time constant results in a stronger integral action and faster error elimination, but it is more prone to overshoot and oscillation. In high-delay situations, the position error signal itself is lagging. Integrating this lag error can cause the direction of control quantity accumulation to deviate from the actual demand, leading to overshoot or oscillation. Therefore, it is necessary to increase the integral time constant to weaken the integral action and avoid excessive accumulation of lag errors.

[0180] The integral time constant amplification factor is positively correlated with the delay excess; the larger the delay excess, the larger the amplification factor, and the greater the increase in the integral time constant. The amplification factor is calculated using a linear amplification function, specifically as the integral time constant amplification factor equals 1 plus the amplification coefficient multiplied by the normalized delay excess. The amplification coefficient is a parameter that adjusts the amplification speed, typically ranging from 3 to 8. The larger the coefficient, the more significant the reduction in integral action. For example, when the amplification coefficient is 5 and the normalized delay excess is 0.2, the integral time constant amplification factor is 2, meaning the integral time constant will increase to twice its original value, and the integral action will decrease to 50% of its original value. When the normalized delay excess reaches 1, the amplification factor reaches 6, and the integral time constant increases to six times its original value.

[0181] The selection of the linear amplification function is based on the sensitivity analysis of the integral element to delay. This function form can smoothly reduce the integral effect as delay increases, avoiding integral saturation or control divergence. After calculating the integral time constant amplification factor, it is multiplied by the reference integral time constant of the robotic arm joint to obtain the dynamically adjusted integral parameters. The reference integral time constant is the optimal value tuned under standard operating conditions. For a car wash robotic arm, the reference integral time constant of each joint is typically in the range of 0.1 to 0.5 seconds, and this value is determined based on the load characteristics and the desired steady-state accuracy.

[0182] In control systems, differential parameters are used to predict the trend of error changes. By differentiating the error signal over time, control inputs are generated in advance to suppress error growth, providing a damping effect. The differential gain determines the strength of the differential action; a larger gain results in stronger damping and smaller system overshoot, but also greater sensitivity to noise. In cases of high delay, the rate of change of the error signal is also lagging. Differential actions based on this lag information may produce incorrect predictions and introduce unnecessary control disturbances. Therefore, it is necessary to reduce the differential gain to weaken the differential action and avoid over-responding to the lagging rate of change.

[0183] The differential gain suppression factor is inversely related to the delay excess; the larger the delay excess, the smaller the suppression factor, and the greater the reduction in differential gain. The suppression factor is calculated using an exponential suppression function, specifically as follows: the differential gain suppression factor equals an exponential function, where the exponent is a negative suppression coefficient multiplied by the normalized delay excess. The suppression coefficient is a parameter that adjusts the suppression rate, typically ranging from 1.5 to 4. This coefficient is smaller than the attenuation coefficient of the proportional gain because the differential action is relatively less sensitive to delay. For example, when the suppression coefficient is 2.5 and the normalized delay excess is 0.2, the differential gain suppression factor is approximately 0.61, indicating that the differential gain will be reduced to 61% of its original value.

[0184] The exponential suppression function ensures that the differential action weakens smoothly with increasing delay, avoiding dynamic performance degradation caused by abrupt changes in system damping. After calculating the differential gain suppression factor, it is multiplied by the reference differential gain of the robotic arm joints to obtain the dynamically adjusted differential parameters. The reference differential gain is the optimal value tuned under standard operating conditions. For a car wash robotic arm, the reference differential gain of each joint is typically in the range of 10 to 50, and this value is determined based on the system's inertia and the desired damping ratio.

[0185] After dynamically adjusting the proportional gain, integral, and differential parameters, it is necessary to further calculate the reduction ratio of the stiffness coefficient and damping coefficient of the robotic arm joints. These two parameters directly determine the degree to which the robotic arm transitions from a high-stiffness mode to a flexible following mode. The stiffness coefficient describes the ability of the robotic arm joint to resist external disturbances; the higher the stiffness, the smaller the displacement of the joint under external force, and the stronger its position-keeping ability. The damping coefficient describes the ability of the robotic arm joint to dissipate kinetic energy; the higher the damping, the smaller the oscillation of the joint movement, and the better the stability.

[0186] In control theory, there is a direct positive correlation between the stiffness coefficient and the proportional gain parameter, and their quantitative relationship can be established through linear or nonlinear mapping functions. The calculation of the stiffness coefficient reduction ratio is based on the ratio of the dynamically adjusted proportional gain parameter to the baseline proportional gain. Specifically, the stiffness coefficient reduction ratio equals the dynamically adjusted proportional gain parameter divided by the baseline proportional gain. This ratio reflects the relative magnitude of the proportional gain change. Since the proportional gain is reduced after adjustment by the attenuation factor, this ratio is less than 1, indicating that the stiffness coefficient needs to be reduced accordingly.

[0187] To prevent excessive reduction in stiffness coefficient from causing the robotic arm to lose its basic position-keeping ability, a lower limit of 0.3 is set for the stiffness coefficient reduction ratio. Even under extreme delay conditions, the stiffness coefficient will not fall below 30% of the baseline value. The damping coefficient reduction ratio is calculated based on the ratio of the dynamically adjusted differential parameter to the baseline differential gain. The specific formula is: damping coefficient reduction ratio equals dynamically adjusted differential gain divided by baseline differential gain. This ratio reflects the relative change in differential gain. Since the differential gain is reduced after adjustment by the suppression factor, this ratio is less than 1, indicating that the damping coefficient needs to be reduced accordingly. A lower limit of 0.4 is also set for the damping coefficient reduction ratio to ensure the system maintains basic stability.

[0188] After determining the reduction ratios of stiffness and damping coefficients, the final step is to calculate the actual stiffness and damping coefficients of the robotic arm joints in flexible following mode. These parameters will be directly loaded into the joint's servo controller to achieve real-time adjustment of the robotic arm's dynamic characteristics. The reference stiffness coefficient of the robotic arm joint is a nominal value set in high-stiffness mode. This value is determined based on the joint's mechanical structure, transmission system, and load characteristics. For a typical car wash robotic arm, the reference stiffness coefficient of each joint is typically in the range of 8,000 to 12,000 Nm per radian.

[0189] The actual stiffness coefficient in flexible following mode is obtained by multiplying the reference stiffness coefficient by a reduction factor. For example, if the reference stiffness coefficient is 10,000 Nm / radian, a reduction factor of 0.55 results in an actual stiffness coefficient of 5,500 Nm / radian. This reduction in stiffness makes the joint more prone to angular displacement under external forces, weakening the overall rigidity of the robotic arm and exhibiting flexible characteristics. The actual stiffness coefficient is set by adjusting the position loop gain and torque limiting parameters of the servo controller. Reducing the position loop gain decreases the response torque of the joint to position errors, thus effectively reducing stiffness.

[0190] The reference damping coefficient of a robotic arm joint is a nominal value set in high-stiffness mode. This value is determined based on the desired damping ratio and the system's natural frequency, and is typically set to 70% to 90% of the critical damping to achieve rapid response without over-damping. For a typical car wash robotic arm, the reference damping coefficient for each joint is usually in the range of 80 to 150 Nm / s per radian. The actual damping coefficient in flexible following mode is obtained by multiplying the reference damping coefficient by a damping coefficient reduction factor. This reduction in damping value reduces energy dissipation during joint movement and enhances the smoothness of motion, but decreases oscillation suppression.

[0191] After the actual stiffness and damping coefficients are calculated, the control system sends these parameters to the servo drives of each joint via a real-time communication bus. Upon receiving the parameters, the drives immediately update the gain matrix of their internal control algorithm, enabling online reconstruction of the robotic arm's dynamic characteristics. The parameter update process employs a smooth transition strategy, meaning it does not switch to new parameters instantaneously, but rather uses linear interpolation over 10 to 20 control cycles to avoid control shocks caused by abrupt parameter changes.

[0192] This implementation achieves dynamic adaptation of the robotic arm control system to 5G communication latency by establishing an adaptive adjustment mechanism for control parameters based on latency excess. This method provides a quantitative basis for parameter adjustment by accurately calculating the latency excess, enabling the system to respond in stages according to the actual degree of latency. The calculation of the proportional gain attenuation factor, integral time constant amplification factor, and differential gain suppression factor optimizes the three core parameters of the control system, ensuring coordinated changes in the system's response speed, steady-state performance, and damping characteristics as latency increases. By mapping the adjusted control parameters to the reduction ratio of stiffness and damping coefficients, a parameter conversion from the control domain to the physical domain is achieved, allowing the robotic arm's dynamic characteristics to adapt to changes in communication conditions in real time. The calculation and loading of the actual stiffness and damping coefficients in flexible following mode completes the smooth transition of the robotic arm from high-stiffness mode to flexible mode, enabling the system to maintain basic following capability and safety even under high latency conditions.

[0193] In one embodiment of this invention, the current physical contact data includes contact force data when the robotic arm contacts the vehicle body. In flexible following mode, based on the vehicle's current physical contact data, the robotic arm is driven to execute a passive retreat mode to prevent scratching the paint, including the following steps: S810. Determine whether the contact force data exceeds the safety threshold. S820. If the contact force data exceeds the safety threshold, calculate the direction vector of the contact force; S830: Control the robotic arm to perform a yielding motion in the opposite direction of the contact force vector until the contact force data is lower than the safety threshold.

[0194] In flexible following mode, the reduced stiffness of the robotic arm makes it more sensitive to external contact. Therefore, real-time monitoring of contact force data is necessary to determine if accidental contact with the vehicle body has occurred. The contact force data is collected by a six-dimensional force sensor mounted on the robotic arm's end effector. This sensor can simultaneously measure force components in three orthogonal directions and torque components along three rotational axes, with a typical measurement range of ±50 Newtons and ±5 Nm, and measurement accuracy of 0.1 Newtons and 0.01 Nm, respectively. The force sensor continuously samples at a high frequency of 1 kHz, generating a set of six-dimensional force data every 1 millisecond to ensure timely detection of contact events.

[0195] During normal cleaning operations, when the nozzles spray high-pressure water, a reaction force is generated, typically ranging from 0.5 to 2 Newtons, in the opposite direction to the spray direction. When the rotating brush head operates, a rotational inertial torque is generated, typically ranging from 0.1 to 0.5 Newton-meters. These normal operating forces are expected and should not trigger the yielding mechanism. However, when the robotic arm's nozzles or brush heads accidentally come into contact with the vehicle body, a significantly increased contact force is generated. This force is directed in the direction of the vehicle body's reaction force on the robotic arm, and its value can reach 5 to 20 Newtons depending on the contact speed and contact stiffness.

[0196] The setting of safety thresholds needs to take into account the paint's tolerance and the characteristics of the cleaning tools. For soft sponge brush heads, due to their soft material, they are not likely to cause damage even if they come into contact with the car body, so the safety threshold can be set to 5 Newtons. For hard nylon brush heads, the contact force needs to be strictly controlled, so the safety threshold is set to 3 Newtons. For high-pressure nozzles, since they should not have physical contact with the car body, the safety threshold is set to 2 Newtons.

[0197] The judgment process is achieved by calculating the resultant force amplitude of the six-dimensional force data. The resultant force amplitude is equal to the square root of the sum of the squares of the three force components, and this value represents the total intensity of the contact force. The calculated resultant force amplitude is compared with the safety threshold of the corresponding cleaning tool in real time. When the resultant force amplitude exceeds the safety threshold, an over-limit contact event is determined to have occurred, and a passive retreat procedure is immediately triggered. To avoid false triggering caused by sensor noise or instantaneous impact, a continuous judgment strategy is adopted, that is, a contact event is confirmed only when the resultant force amplitude exceeds the threshold for three consecutive sampling periods.

[0198] Once it is confirmed that the contact force data exceeds the safety threshold, the direction vector of the contact force needs to be calculated immediately. This vector indicates the spatial direction of the force exerted by the vehicle body on the robotic arm and is a key basis for determining the direction of the retraction movement. The direction vector of the contact force is calculated using three force components measured by a six-dimensional force sensor. Let the force components in the three orthogonal directions be... , and These three components are defined in the sensor coordinate system. Typically, the Z-axis of the sensor coordinate system is aligned with the axial direction of the end flange of the robotic arm, while the X-axis and Y-axis correspond to two radial directions, respectively.

[0199] The direction vector of the contact force is a three-dimensional vector composed of the three force components. The direction of this vector points in the direction of the contact force, that is, from the vehicle body towards the end effector of the robotic arm. To obtain the unit direction vector, it needs to be normalized. Normalization is achieved by dividing each component of the vector by its magnitude, which is equal to the square root of the sum of the squares of the three components. The normalized unit direction vector is denoted as [vector name missing]. The magnitude of this vector is 1, which retains only the direction information and eliminates the influence of the force magnitude.

[0200] The unit direction vector is expressed in the sensor coordinate system. To facilitate the motion control of the robotic arm, it needs to be transformed to the robotic arm's base coordinate system or world coordinate system. Coordinate transformation is achieved through a rotation matrix, which describes the attitude relationship between the sensor coordinate system and the base coordinate system. This rotation matrix can be obtained through the forward kinematics of the robotic arm. Specifically, based on the current angles of each joint of the robotic arm, the pose matrix of the end effector is calculated using forward kinematics. This matrix includes a rotation submatrix and a translation vector; the rotation submatrix is ​​the required coordinate transformation matrix. Multiplying the unit direction vector in the sensor coordinate system by the rotation matrix on the left yields the contact force direction vector expressed in the base coordinate system.

[0201] This direction vector clearly indicates the direction of the contact force in space, providing accurate geometric information for subsequent calculation of the retraction direction. In practical implementation, considering that contact may occur at different locations at the end of the robotic arm, such as the side of the nozzle or the edge of the brush head, the direction of the contact force may not be entirely along a single principal axis, but rather has a complex spatial orientation. Through complete measurement and precise coordinate transformation by a six-dimensional force sensor, contact forces in any direction can be accurately captured.

[0202] After calculating the direction vector of the contact force, the robotic arm is immediately controlled to retract in the opposite direction of this vector to quickly disengage from the contact state and avoid continuous pressure or friction damage to the car paint. The retraction direction is defined as the opposite direction of the contact force direction vector, that is, taking the negative value of each component of the direction vector to obtain the retraction direction vector. This direction points away from the vehicle body. Moving in this direction increases the distance between the end effector of the robotic arm and the surface of the vehicle body, gradually reducing the contact force until it disappears.

[0203] The speed of the yielding motion needs to be adaptively adjusted according to the magnitude of the contact force. A larger contact force indicates a tighter contact or higher contact stiffness, requiring a faster yielding speed to disengage quickly. When the contact force is smaller, a slower yielding speed can be used to avoid overreaction and resulting uneven motion. The formula for calculating the yielding speed is: yielding speed equals the baseline yielding speed plus the portion of the contact force exceeding the safety threshold multiplied by a speed gain coefficient. The speed gain coefficient is typically taken as 10 to 20 mm / s per Newton. For example, when the baseline yielding speed is set to 50 mm / s, the contact force is 8 Newtons, the safety threshold is 3 Newtons, and the speed gain coefficient is 15 mm / s per Newton, the yielding speed is 50 + 5 multiplied by 15 equals 125 mm / s.

[0204] The retraction motion is executed through the robot arm's speed control mode. The controller calculates the velocity vector of the end effector in Cartesian space based on the retraction direction vector and the retraction velocity. Then, it maps the Cartesian velocity to the angular velocity of each joint using an inverse Jacobian matrix, driving the servo motors to coordinate the joint movements. During the retraction motion, force sensors continuously monitor contact force data, updating the magnitude and direction of the contact force in real time. When the direction of the contact force changes, the retraction direction is adjusted accordingly to ensure that the movement always follows the optimal direction for disengagement.

[0205] The termination condition for the yielding motion is that the contact force data falls below a safety threshold. When the amplitude of the resultant force measured by the force sensor decreases below the safety threshold, it is determined that the contact state has been released, and the yielding motion immediately stops, with the robotic arm remaining in its current position. To avoid oscillations caused by repeated triggering and stopping near the threshold, the yielding termination threshold is set slightly lower than the trigger threshold, typically 80% of the trigger threshold, forming a hysteresis judgment logic. For example, if the trigger threshold is 3 Newtons, the termination threshold is set to 2.4 Newtons; the yielding motion only stops when the contact force decreases below 2.4 Newtons.

[0206] After the retreating motion is completed, the control system records detailed information about the retreating event, including the time and location of the contact, the peak contact force, and the retreat distance, for subsequent trajectory optimization and fault analysis. Simultaneously, the system reassesses the current cleaning path to determine whether path parameters need adjustment or whether to switch to a more conservative operating mode to prevent similar contact events from recurring.

[0207] This implementation method ensures safety while preventing false triggering by setting material-related safety thresholds and a continuous judgment strategy. Precise calculation and coordinate transformation of the contact force direction vector ensure the accuracy of the retreat movement direction, enabling the robotic arm to quickly disengage along the optimal path. An adaptive retreat speed adjustment strategy dynamically adjusts the movement speed based on the magnitude of the contact force, ensuring rapid disengagement in emergencies while avoiding motion shocks caused by overreaction. The introduction of hysteresis judgment logic eliminates oscillations near the threshold, resulting in a smooth and stable retreat process. This passive retreat mechanism, as a supplement to active control, still provides reliable safety protection through force feedback when communication delays cause position tracking failures or sensor drift leads to coordinate system deviations.

[0208] Reference Figure 2 This application also provides an intelligent car wash machine, including: Camera; The memory is configured to store instructions; and The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the aforementioned 5G-based intelligent car wash machine and vehicle collaborative perception and communication method.

[0209] Reference Figure 3 This application also provides a car wash system, including: vehicle; Intelligent car wash machines connect to vehicles.

[0210] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0211] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0212] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0213] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0214] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0215] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0216] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0217] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0218] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A 5G communication-based intelligent car washing machine and vehicle cooperative perception communication method, characterized in that, Applied to a car wash system, which includes a smart car wash machine and a vehicle, the smart car wash machine includes a camera, the method includes: Construct a 5G heterogeneous data synchronization link between the smart car wash machine and the vehicle, and based on the 5G heterogeneous data synchronization link, acquire visual data through the camera and acquire data from the vehicle's inertial measurement unit. Extracting visual features from visual data; Based on visual data and vehicle-mounted inertial measurement unit data, the initial attitude of the vehicle entering the car wash area is determined; A dynamic virtual coupling model is constructed based on data from the vehicle-mounted inertial measurement unit. In response to receiving real-time acceleration and angular velocity data uploaded by the vehicle through the 5G heterogeneous data synchronization link, the vehicle in the dynamic virtual coupling model is driven to move, and the coordinate system of the smart car wash machine is dynamically locked to the vehicle coordinate system based on visual features. A dynamically locked coordinate system is determined, and the intelligent car wash machine is driven to perform cleaning operations on the vehicle according to the dynamically locked coordinate system.

2. The method of claim 1, wherein, The intelligent car wash machine also includes a robotic arm equipped with nozzles and servo motors. Based on data from the vehicle-side inertial measurement unit, a dynamic virtual coupling model is constructed, including: The vehicle's attitude change is calculated based on data from the vehicle-mounted inertial measurement unit. Calculate the displacement of each preset feature point on the vehicle surface based on the attitude change; The robotic arm drives a servo motor to perform compensating movements in the same direction and with equal amount of displacement to keep the relative position of the nozzle and the vehicle surface constant.

3. The method of claim 2, wherein, The vehicle-side inertial measurement unit (IMU) data includes roll angle, pitch angle, and yaw angle. Based on the attitude changes, it calculates the displacement of each preset feature point on the vehicle surface, including: The static bounding box model of the vehicle is determined based on visual features, and the vehicle size data in the static bounding box model is obtained. Based on the roll angle, pitch angle, yaw angle and vehicle size data, calculate the three-dimensional spatial displacement vector of each preset feature point on the vehicle surface; By mapping the three-dimensional spatial displacement vector to the coordinate system of the intelligent car wash machine, the displacement of each preset feature point on the vehicle surface is obtained.

4. The method of claim 3, wherein, After the robotic arm drives the servo motor to perform a compensating motion in the same direction and with the same amount based on the displacement, it also includes: In response to the audio signal reflected from the impact of high-pressure water flow on the vehicle body, the echo delay and spectral attenuation rate of the audio signal are determined. The actual physical distance between the nozzle and the vehicle body is calculated based on the echo delay and spectral attenuation rate. The actual physical distance is compared with the theoretical distance calculated based on data from the vehicle-mounted inertial measurement unit to obtain the distance deviation; An error correction vector is generated based on the distance deviation, and the spatial coordinates of the static bounding box model are adjusted based on the error correction vector to eliminate the integral drift error of the vehicle-side inertial measurement unit.

5. The method of claim 4, wherein, The actual physical distance between the nozzle and the vehicle body is calculated based on the echo delay and spectral attenuation rate, including: Extract the acoustic signature features generated by water flow hitting different materials on the car body from the audio signal; The type of vehicle body material in the current cleaning area is determined based on the voiceprint characteristics. Load the sound wave propagation speed and energy attenuation coefficient corresponding to the vehicle body material type from the preset acoustic parameter database; The actual physical distance between the nozzle and the vehicle body is calculated based on the echo delay, sound wave propagation speed, and energy attenuation coefficient.

6. The method of claim 2, wherein, The intelligent car wash machine operates by cleaning vehicles based on a dynamically locked coordinate system, including: Real-time monitoring of the real-time latency value of the 5G heterogeneous data synchronization link; When the real-time delay value is lower than the first preset threshold, the drive robotic arm maintains a high-rigidity mode and performs high-pressure rinsing or close-fitting brushing. When the real-time delay value is higher than the second preset threshold, the proportional-integral-derivative parameters of the robotic arm joint are determined based on the real-time delay value, and the stiffness coefficient and damping coefficient are reduced based on the proportional-integral-derivative parameters, so that the robotic arm enters the flexible following mode. In flexible following mode, the robotic arm is driven to perform a passive retreat mode based on the vehicle's current contact physical data to prevent scratches on the paint.

7. The method according to claim 6, characterized in that, The proportional-integral-derivative (PID) parameters of the robotic arm joints are determined based on real-time delay values. Then, the stiffness and damping coefficients are reduced based on these parameters, enabling the robotic arm to enter a flexible following mode. This includes: Calculate the amount by which the real-time delay value exceeds the second preset threshold; The proportional gain attenuation factor is calculated based on the delay excess, where the larger the delay excess, the smaller the proportional gain attenuation factor. Multiply the preset reference proportional gain of the robotic arm joint by the proportional gain attenuation factor to obtain the dynamically adjusted proportional gain parameter. Calculate the integral time constant amplification factor based on the delay excess; the larger the delay excess, the larger the integral time constant amplification factor. The dynamic-adjusted integral parameters are obtained by multiplying the reference integral time constant of the robotic arm joint by the integral time constant amplification factor. The differential gain suppression factor is calculated based on the delay excess, where the larger the delay excess, the smaller the differential gain suppression factor. The differential parameters are obtained by multiplying the baseline differential gain of the robotic arm joint by the differential gain suppression factor. Based on the dynamically adjusted proportional gain parameters, integral parameters, and differential parameters, calculate the reduction ratio of stiffness coefficient and damping coefficient. The actual stiffness coefficient in the flexible following mode is obtained by multiplying the reference stiffness coefficient of the robotic arm joint by the stiffness coefficient reduction ratio. The actual damping coefficient in the flexible following mode is obtained by multiplying the reference damping coefficient of the robotic arm joint by the damping coefficient reduction ratio.

8. The method according to claim 6, characterized in that, Current physical contact data includes contact force data when the robotic arm contacts the vehicle body. In flexible following mode, based on the vehicle's current physical contact data, the robotic arm is driven to execute a passive retreat mode to prevent scratches to the paint, including: Determine whether the contact force data exceeds the safety threshold; If the contact force data exceeds the safety threshold, calculate the direction vector of the contact force; The robotic arm is controlled to retract in the opposite direction of the contact force vector until the contact force data is below the safety threshold.

9. An intelligent car wash machine, characterized in that, include: Camera; The memory is configured to store instructions; as well as The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the intelligent car wash machine and vehicle collaborative perception communication method based on 5G communication according to any one of claims 1 to 8.

10. A car wash system, applied to the method described in any one of claims 1-8, characterized in that, include: vehicle; Intelligent car wash machines connect to vehicles.

Images