Teaching unmanned intelligent car and virtual simulation method thereof

Abstract

This application relates to the fields of autonomous driving and computer simulation technology, and discloses a teaching unmanned intelligent vehicle and its virtual simulation method. The system includes a physical vehicle and a host computer. The host computer interacts with the physical vehicle to establish a global system timestamp. The main control module synchronously collects the back electromotive force signal of the motor and the three-axis acceleration signal, eliminates the time error of software polling through hardware-level binding, and integrates and generates a time-aligned data sequence. The host computer calculates the cross-power spectrum and time-domain electromechanical cross-correlation sequence of the data sequence, and constructs a hysteresis and flexible coupling matrix based on this. The host computer further uses this matrix to reconstruct ideal control commands and generate dynamic degradation commands, driving the virtual physics engine to output virtual sensor observation data. This invention improves the data synchronization accuracy of the underlying sensors and enhances the ability of the virtual simulation environment to reproduce the nonlinear dynamic characteristics of the actual physical vehicle.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving and computer simulation technology, specifically to a teaching unmanned intelligent vehicle and its virtual simulation method. Background Technology

[0002] In teaching and research related to autonomous driving, the design and verification of control algorithms typically involve combining physical intelligent vehicles with virtual simulation systems. Existing virtual simulation systems are mostly based on idealized rigid body dynamics models, rarely considering changes in stiffness and damping coefficients caused by wear and tear of mechanical components over long-term operation. This purely theoretical simulation environment struggles to reflect the dynamic degradation process at the physical level in the virtual space, resulting in discrepancies between the simulated vehicle state and the actual physical performance of the vehicle.

[0003] To reduce the discrepancy between the simulation environment and the real physical environment, it is necessary to collect the vehicle chassis's operating data to correct the simulation model. Currently, commonly used low-level data acquisition schemes generally employ a software polling mechanism to read the back electromotive force state of the motors and inertial sensor data one by one. Due to the influence of microcontroller internal task scheduling and communication bus occupancy, there are time delays and uncertainties in the sampling actions of different sensors. This results in a misalignment between the motor signals reflecting electrical states and the vibration signals reflecting mechanical motion in terms of timing. When this multi-source low-level data with time misalignment is input to the host computer for dynamic parameter evaluation or model updates, it directly increases the calculation error of system parameter identification, causing the virtual simulation system to fail to accurately reproduce the nonlinear dynamic characteristics of the real physical chassis. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a teaching unmanned intelligent vehicle and its virtual simulation method. It solves the problem that existing intelligent vehicle virtual simulation systems suffer from asynchronous data acquisition from multiple underlying sensors, which makes it impossible to accurately extract mechanical degradation parameters to update the virtual model, resulting in the simulation environment failing to reproduce the nonlinear dynamic characteristics of the real chassis.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a teaching unmanned intelligent vehicle and its virtual simulation method, comprising a physical vehicle and a host computer; The host computer is used to issue diagnostic control command sequences and establish a global system timestamp with the main control module of the vehicle entity; The main control module is used to trigger a hardware interrupt when the transient derivative of the back electromotive force signal of the acquired motor module exceeds a preset threshold. It integrates the global system timestamp, back electromotive force signal and triaxial acceleration signal synchronously acquired by the inertial sensor module into a data sequence and sends it to the host computer. The host computer is used to calculate the cross power spectrum and time-domain electromechanical cross-correlation sequence based on the data sequence to construct the hysteresis and flexible coupling matrix, and to use the matrix to reconstruct the ideal control command to generate the dynamic degradation command; The host computer is also used to execute dynamic degradation instructions, output virtual sensor observation data, and generate feedforward compensation instructions based on this data, which are then sent to the vehicle entity. The main control module then overlays these instructions onto the underlying driver of the vehicle entity for degradation compensation.

[0006] Furthermore, the host computer sends a synchronization request data packet carrying the system time. Upon receiving the synchronization request data packet, the main control module records the local microsecond-level count value of its internal hardware timer. The host computer and the main control module calculate the clock offset through multiple interactions involving the synchronization request data packets. The main control module adds the current local microsecond-level count value of the hardware timer to the clock offset to generate a global system timestamp. Through the interactive calculation of the hardware timer and the clock offset, a timing alignment mechanism for distributed nodes is established.

[0007] Furthermore, the diagnostic control command sequence includes step acceleration and step steering commands to trigger the dynamic response of the vehicle. The main control module includes an analog-to-digital converter (ADC) module, which sets an internal hardware timer to output a center-aligned pulse width modulation (PWM) signal. Within the dead-time window, the ADC module is activated to sample the back EMF signal. The main control module subtracts the historical back EMF signal value from the previous basic sampling period from the current back EMF signal, and divides the calculated potential difference by a fixed constant time interval to obtain the transient derivative.

[0008] Furthermore, the main control module includes a buffer module and a memory access module, extracting the global system timestamp at the moment of hardware interruption and writing it into the starting address space of the buffer module. The main control module hardware-level binds the hardware read interface of the inertial sensing module with the conversion completion event of the analog-to-digital conversion module. Within a preset time window, the memory access module moves the discrete data points of the synchronously acquired back EMF signal and triaxial acceleration signal to the buffer module. The buffer module is divided into a first buffer block and a second buffer block in memory, and the memory access module alternately writes data to the first and second buffer blocks. This hardware-level binding and alternating memory write mechanism avoids the time uncertainty caused by software polling, ensuring data synchronization of multiple sensors at the microsecond level.

[0009] Furthermore, the data sequence includes a cross-domain sensor sequence, composed of back EMF signals and triaxial acceleration signals interleaved within the same sampling period. The background communication thread within the main control module encapsulates the first or second buffer block filled with data into a data packet containing a fixed sequence frame header, a global system timestamp, the cross-domain sensor sequence, and a cyclic redundancy check (CRC) code. The first communication module of the main control module pushes the data packet to the second communication module of the host computer via an asynchronous non-blocking communication mechanism.

[0010] Furthermore, the frequency domain computation module of the host computer performs deinterleaving on the cross-domain sensor sequence, separating it into a one-dimensional back EMF time series and a three-dimensional triaxial acceleration time series. The frequency domain computation module calculates the square root of the sum of squares of the triaxial acceleration components at each sampling point in the three-dimensional triaxial acceleration time series to obtain a one-dimensional scalar acceleration sequence. The one-dimensional back EMF time series and the one-dimensional scalar acceleration sequence after mean removal and smoothing are independently subjected to Fast Fourier Transform (FFT) to obtain the frequency domain back EMF sequence and the frequency domain acceleration sequence, respectively. The conjugate complex numbers of the frequency domain back EMF sequence and the frequency domain acceleration sequence are multiplied to obtain the cross-power spectrum. A moving average filter is used to weighted smooth the cross-power spectrum, and an inverse discrete Fourier transform is performed on the smoothed cross-power spectrum to generate a time-domain electromechanical cross-correlation sequence. The cross-power spectrum and time-domain cross-correlation operation reflect the energy transfer relationship between electrical state and mechanical vibration.

[0011] Furthermore, the host computer's matrix generation module uses standard deviation normalization logic to process the time-domain electromechanical cross-correlation sequence to generate a normalized tensor. It then calls a one-dimensional convolutional neural network model to infer the normalized tensor, outputting the mechanical stiffness degradation compensation and damping coefficient degradation compensation amounts representing the parameters of the hysteresis and flexible coupling matrix construction. The one-dimensional convolutional neural network model internally includes an input layer, two sets of continuously cascaded feature extraction blocks, and a fully connected regression layer at the end. It uses a mean squared error loss function with regularization penalty for backpropagation training.

[0012] Furthermore, the vehicle entity includes a drive module, and the host computer includes a control strategy generation module and a virtual simulation module. The virtual simulation module calculates the virtual state trajectory by substituting dynamic parameters, including compensation for mechanical stiffness degradation and damping coefficient degradation. The control strategy generation module calculates a linear quadratic regulator algorithm for the virtual state trajectory, constructs a cost function encompassing state deviation loss and control input loss, calculates the feedback gain matrix, and generates feedforward compensation commands. The second communication module sends the feedforward compensation commands to the vehicle entity, where the main control module parses them into feedforward current compensation values ​​and superimposes them onto the underlying current closed-loop reference input of the drive module.

[0013] Furthermore, the host computer extracts the real feedback acceleration sequence and the virtual predicted acceleration sequence of the closed-loop feedback, calculates the difference sequence of corresponding sampling points of the two sets of sequences within the same time window to generate the root mean square residual value; using the recursive least squares algorithm and after adding a stable scalar to the main diagonal of the matrix to be inverted, iteratively updates the dynamic parameters inside the virtual simulation module; the mechanical stiffness degradation compensation amount, damping coefficient degradation compensation amount, and root mean square residual value are uniformly mapped and calculated through range standardization to obtain the comprehensive chassis health degradation index.

[0014] A second aspect of this invention provides a virtual simulation method for teaching unmanned intelligent vehicles, comprising the following steps: The host computer issues a sequence of diagnostic control commands and establishes a global system timestamp with the main control module of the vehicle entity; The main control module triggers a hardware interrupt when the transient derivative of the back electromotive force signal of the acquired motor module exceeds a preset threshold. The main control module integrates the global system timestamp, back electromotive force signal, and triaxial acceleration signal synchronously collected by the inertial sensing module into a data sequence and sends it to the host computer. The host computer calculates the cross-power spectrum and time-domain electromechanical cross-correlation sequence based on the data sequence to construct the hysteresis and flexible coupling matrix; The host computer uses matrix reconstruction of ideal control instructions to generate dynamic degradation instructions; The host computer executes the dynamic degradation command and outputs virtual sensor observation data, and generates a feedforward compensation command based on this data and sends it to the vehicle entity. The main control module then overlays the command onto the underlying driver of the vehicle entity for degradation compensation.

[0015] This invention provides a teaching unmanned intelligent vehicle and its virtual simulation method. It has the following beneficial effects: 1. This invention achieves hardware-level binding between the hardware reading interface of the inertial sensing module and the conversion completion event of the analog-to-digital conversion module. Combined with the mechanism of the memory access module alternately writing data to the first and second buffer blocks, it eliminates the time uncertainty caused by traditional software polling. This hardware-level timing control realizes the microsecond-level time alignment between the back electromotive force signal and the triaxial acceleration signal, improving the synchronization accuracy of the underlying multi-source sensor data acquisition.

[0016] 2. This invention utilizes a frequency domain computation module to calculate the cross-power spectrum of the back electromotive force time series and the scalar acceleration series and generate a time-domain electromechanical cross-correlation sequence. Subsequently, it uses a one-dimensional convolutional neural network model for inference. By calculating and establishing a correlation model between electrical state and mechanical vibration, it can directly extract the degradation compensation amount of mechanical stiffness and damping coefficient from discrete sensor sequences, thereby improving the accuracy of chassis mechanical state assessment.

[0017] 3. This invention generates dynamic degradation instructions by reconstructing ideal control instructions through hysteresis and flexible coupling matrices, inputs them into the virtual simulation module to drive the physics engine, and calculates feedforward current compensation values ​​in conjunction with the control strategy generation module and uses a recursive least squares algorithm to iteratively update dynamic parameters. This mechanism reduces the difference in dynamic performance between the pure theoretical simulation environment and the actual physical entity of the vehicle, and improves the ability of the virtual simulation system to reproduce the nonlinear dynamic characteristics of the actual vehicle. Attached Figure Description

[0018] Figure 1This is a perspective view of the present invention; Figure 2 This is a partial structural diagram of the present invention; Figure 3 This is a schematic diagram of the system architecture of the present invention; Figure 4 This is a flowchart of the method of the present invention; Figure 5 The diagram shows a comparison of the transient vibration response of the teaching vehicle chassis during a step start, as presented in this invention.

[0019] Among them, 100 is the data acquisition module for the vehicle; 110 is the main control module; 111 is the analog-to-digital conversion module; 112 is the memory access module; 113 is the buffer module; 114 is the first communication module; 120 is the drive module; 130 is the motor control module; 140 is the inertial sensing module; 200 is the host computer; 210 is the second communication module; 220 is the frequency domain operation module; 230 is the matrix generation module; 240 is the virtual simulation module; 241 is the instruction interception module; and 242 is the dynamics calculation module. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] See attached document Figure 1 - Appendix Figure 3 The present invention provides a teaching unmanned intelligent vehicle and its virtual simulation system, which may include: vehicle entity 100 and host computer 200.

[0022] The vehicle entity 100 includes a main control module 110, a drive module 120, a motor module 130, and an inertial sensing module 140. The main control module 110 integrates an analog-to-digital conversion module 111, a memory access module 112, a buffer module 113, and a first communication module 114.

[0023] The main control module 110 is electrically connected to the drive module 120 and controls the motor module 130 to perform actions. The analog-to-digital converter module 111 is connected to the output terminal of the drive module 120 and collects the electrical signals from both ends of the motor module 130. The inertial sensor module 140 is installed at the center of the chassis of the vehicle entity 100, electrically connected to the main control module 110, and collects the acceleration signals of the vehicle entity 100.

[0024] The host computer 200 includes a second communication module 210, a frequency domain operation module 220, a matrix generation module 230, and a virtual simulation module 240. The second communication module 210 establishes a data connection with the first communication module 114 via a data bus or wireless network. The virtual simulation module 240 internally includes an instruction interception module 241 and a dynamics calculation module 242.

[0025] Please see the appendix Figure 4 This invention provides a virtual simulation method for teaching unmanned intelligent vehicles, comprising the following steps: S101, the host computer 200 generates a diagnostic control instruction sequence and sends it to the vehicle entity 100. The host computer 200 and the main control module 110 perform clock synchronization operation to establish a global system timestamp. The vehicle entity 100 receives and executes the diagnostic control instruction sequence. S102, during the process of the vehicle entity 100 performing actions, the analog-to-digital conversion module 111 collects the back electromotive force signal of the motor module 130, the main control module 110 calculates the transient derivative of the back electromotive force signal, and triggers a hardware interrupt when it is determined that the transient derivative exceeds a preset threshold. S103, after triggering a hardware interrupt, the memory access module 112 records the corresponding global system timestamp. Within the preset time window, the memory access module 112 synchronously transfers the back electromotive force sequence collected by the analog-to-digital conversion module 111 and the acceleration sequence collected by the inertial sensing module 140 to the buffer module 113 for storage. S104, the main control module 110 sends the data sequence stored in the buffer module 113 and the global system timestamp to the host computer 200 through the first communication module 114; S105, the second communication module 210 receives the data sequence, and the frequency domain operation module 220 performs frequency domain transformation operation on the back electromotive force sequence and the acceleration sequence to calculate the cross power spectrum and time-domain electromechanical cross-correlation sequence of the two sets of sequences. S106, the frequency domain operation module 220 extracts the amplitude attenuation coefficient and delay time variable from the time domain electromechanical cross-correlation sequence, and the matrix generation module 230 uses the amplitude attenuation coefficient and delay time variable to construct the hysteresis and flexible coupling matrix in the discrete time state space. S107, During the operation of the virtual simulation module 240, the instruction interception module 241 intercepts the ideal control instruction output by the control algorithm and performs a timing comparison with the global system timestamp corresponding to the data sequence; S108, the instruction interception module 241 uses the hysteresis and flexible coupling matrix to reconstruct and map the ideal control instruction to generate a dynamic degradation instruction. The dynamic solution module 242 receives the dynamic degradation instruction, drives the virtual physics engine to run, and outputs virtual sensor observation data.

[0026] Please see the appendix Figure 1and attached Figure 2 In step S101, the host computer 200 generates a diagnostic control command sequence and sends it to the vehicle entity 100. The host computer 200 and the main control module 110 perform clock synchronization to establish a global system timestamp. The vehicle entity 100 receives and executes the diagnostic control command sequence, which specifically includes the following sub-steps: In step S1011, the host computer 200 generates a diagnostic control command sequence during the virtual simulation program initialization phase and sends this sequence to the vehicle entity 100 via the second communication module 210. The diagnostic control command sequence is set as a discrete control command set that varies with time, specifically including a step acceleration command and a step steering command. The step acceleration command is used to control the duty cycle of the pulse width modulation signal of the drive module 120 to undergo the maximum amplitude jump within a single control cycle. In this embodiment, the threshold for the duty cycle jump is typically set to 80% to 100% of the maximum drive range of the motor, based on ensuring that the transient electromagnetic torque of the motor is sufficient to overcome static friction and effectively excite the elastic deformation of the chassis material. The step steering command is used to control the steering servo of the vehicle entity 100 to reach the limit deflection angle position in a very short time. By introducing such a step excitation signal with high-frequency components, the vehicle entity 100 can generate the maximum transient physical response during stationary start-up or high-speed steering. This physical response manifests as sudden load changes within the motor module 130 and mechanical deformation of the non-rigid structure of the chassis of the vehicle entity 100, providing a physical prerequisite for subsequently acquiring feature data for degradation mapping.

[0027] S1012, based on the aforementioned physical excitation requirements, the timing of data acquisition must be rigorous. Considering the inherent frequency drift between the operating system crystal oscillator of the host computer 200 and the internal hardware timer of the main control module 110, and the random transmission delay in the communication network, directly comparing local time would lead to misalignment of the simulation mapping timeline. Therefore, while issuing the diagnostic control command sequence, a bidirectional clock handshake connection is established between the host computer 200 and the main control module 110 to perform clock synchronization. To establish a unified reference time, the host computer 200 sends a synchronization request data packet carrying its local system time. Upon receiving the data packet, the main control module 110 records the local microsecond-level count value of its internal hardware timer. To avoid biased errors caused by single network jitter, the system typically performs multiple data packet exchanges and uses a weighted average logic after removing extreme values ​​to calculate the final clock offset between the two.

[0028] For the specific message exchange format of the clock synchronization protocol between multi-node devices and the calculation method for network round-trip delay compensation, those skilled in the art can refer to the relevant specifications of the Precision Time Protocol or Network Time Protocol. The principle of network delay compensation is a well-known technology in this field and will not be elaborated here.

[0029] After acquiring the clock offset, the main control module 110 defines a global system timestamp. To expand the protection scope of the technical solution, the specific logic for acquiring the global system timestamp is set as follows: the local count value of the hardware timer inside the main control module 110 at the current moment is directly added to the clock offset compensation amount obtained by calibration between the host computer 200 and the main control module 110. The physical purpose of this calculation process is to eliminate the relative time barrier between independent systems, thereby ensuring that the transient sensor data collected by the vehicle entity 100 in the physical environment can be in the same time dimension as the logical running time of the virtual simulation module 240 in the host computer 200, preventing simulation degradation and mapping failure due to timing misalignment.

[0030] S1013, after completing the aforementioned clock calibration, the vehicle entity 100 receives the diagnostic control command sequence through the first communication module 114. The control program inside the main control module 110 unpacks and parses the commands. The main control module 110 generates corresponding drive signals through its internal timer according to the timing specified in the diagnostic control command sequence. The main control module 110 sends the drive signals to the drive module 120, which amplifies the power and applies it to the motor module 130, driving the vehicle entity 100 into the set test motion state. During this process, due to the closing of the mechanical transmission gap and the flexible stress on the chassis material, the vehicle entity 100 transitions from static to dynamic operation. The chassis mechanical structure and the motor electrical state both exhibit transient characteristics that change drastically with the control commands. The system then prepares to enter the high-frequency cross-domain signal triggering and acquisition stage.

[0031] Please see the appendix Figure 1 and attached Figure 2 In step S102, during the operation of the vehicle entity 100, the analog-to-digital conversion module 111 acquires the back electromotive force signal of the motor module 130, the main control module 110 calculates the transient derivative of the back electromotive force signal, and triggers a hardware interrupt when it determines that the transient derivative exceeds a preset threshold. Specifically, this includes the following sub-steps: S1021, when the vehicle entity 100 responds to the control command and enters the dynamic running state, the analog-to-digital conversion module 111 acquires the electrical status of both ends of the motor module 130 in real time. The reason for choosing the back EMF signal as the input parameter is that the back EMF generated by the DC motor is strictly linearly positively correlated with the rotor angular velocity. The load fluctuations caused by the non-rigid deformation of the chassis will be directly coupled to the motor rotor through the mechanical transmission shaft, and thus reflected in the waveform fluctuations of the back EMF. To eliminate the power supply voltage interference applied by the motor drive module 120 during the conduction period, the main control module 110 sets the internal timer to output a center-aligned pulse width modulation signal, and generates a hardware trigger event during the off-time of this signal, i.e., the dead time window. This event starts the analog-to-digital conversion module 111 to perform signal sampling through the internal cascaded channel of the main control module 110. In this embodiment, to ensure the effectiveness of sampling, the duration of the dead time window is strictly set to be greater than the sum of the sampling setup time and conversion time of the analog-to-digital conversion module 111. The aforementioned underlying hardware timing linkage mechanism avoids voltage spikes when the switching transistor is turned on, ensuring that the system obtains the real back electromotive force signal generated by the motor rotor cutting magnetic field lines.

[0032] S1022, based on the discrete back EMF signal sequence obtained above, the main control module 110 uses its built-in arithmetic logic unit to calculate the transient derivative of the back EMF signal in real time. The general principle of transient derivative extraction is to use the backward difference of the discrete time series to approximate the derivative of continuous time. Its technical purpose is to filter out the low-frequency back EMF component during the steady-state operation of the vehicle entity 100, reflecting the high-frequency electrical fluctuation characteristics caused by sudden changes in mechanical load. The core calculation logic of the transient derivative is: subtract the historical value of the back EMF signal obtained in the previous basic sampling period from the back EMF signal collected at the current moment to obtain the potential difference, and then divide it by the time interval between two adjacent samplings. For the division operation logic here, in order to avoid the denominator approaching zero and causing a division error, the time interval between two adjacent samplings in this embodiment is constrained by the hardware clock divider of the main control module 110 and locked as a non-zero fixed constant. Thus, the above mathematical division operation is equivalently converted into basic subtraction and fixed shift operation at the microprocessor level, reducing the computational overhead of the real-time control loop.

[0033] S1023, after the main control module 110 obtains the transient derivative at the current moment, it compares it with the preset load mutation threshold in the system register. Flexible deformation of the chassis material of the vehicle entity 100 or instantaneous closure of mechanical transmission gaps will apply a transient impact torque to the motor rotor in the opposite direction through the transmission shaft system, causing a jump in rotor angular velocity and resulting in a spike in the transient derivative. The preset load mutation threshold is determined based on the pre-calibrated steady-state derivative extreme value of the vehicle entity 100 under no-load rigid support, and its value range is set to 3 to 5 times this steady-state derivative extreme value. This multiplier setting is used to distinguish between normal electrical fluctuations and actual mechanical deformation impacts. To prevent false triggering caused by sensor thermal noise or occasional electromagnetic interference, the output result is determined based on multi-condition logic in the time dimension, avoiding one-sided judgment relying on a single extreme value. When the absolute value of the transient derivative in two consecutive basic sampling periods is greater than the preset load change threshold, the main control module 110 determines that mechanical deformation has indeed occurred and immediately triggers the highest priority external hardware interrupt to start the subsequent high-frequency coherent data interception process.

[0034] Please see the appendix Figure 1 and attached Figure 2 In step S103, after triggering a hardware interrupt, the memory access module 112 records the corresponding global system timestamp. Within a preset time window, the memory access module 112 synchronously transfers the back electromotive force sequence collected by the analog-to-digital conversion module 111 and the acceleration sequence collected by the inertial sensing module 140 to the buffer module 113 for storage. This specifically includes the following sub-steps: S1031, when the core processor of the main control module 110 receives a hardware interrupt triggered by the above-mentioned judgment logic, the processor will suspend the current non-core background task. In order to anchor the initial moment of the physical deformation, the main control module 110 extracts the global system timestamp of the moment the hardware interrupt occurs and writes it into the first address space of the buffer module 113. This global system timestamp is established as the absolute time origin of the subsequent associated cross-domain sensor data. Its technical purpose is to ensure that the electrical mutation sequence and the mechanical deformation sequence obtained by the host computer have a deviation-free alignment reference on the simulation time axis.

[0035] S1032, based on the establishment of the aforementioned time reference, the system then opens a sampling time window of a fixed length. In this embodiment, the duration of this time window depends on the inherent mechanical resonant frequency of the chassis material, and is typically set to tens to hundreds of milliseconds to fully capture the low-frequency oscillation period from the occurrence of deformation to its decay. Considering that the back electromotive force sequence acquired by the analog-to-digital conversion module 111 and the acceleration sequence acquired by the inertial sensing module 140 are multi-source heterogeneous data, they have independent operating clocks and base sampling rates at the underlying hardware level. If an asynchronous reading method is directly adopted, accumulated time drift and operating condition misalignment will inevitably occur. To solve the problem of strict timing alignment of multi-source heterogeneous data, this embodiment adopts a lower-level hardware timer synchronization triggering mechanism. The main control module 110 establishes a hardware-level binding between the hardware reading interface of the inertial sensing module 140 and the conversion completion event of the analog-to-digital conversion module 111 through internal bus linkage. Specifically, it is set to use the lower of the maximum allowable sampling rates in the hardware manuals of both as the reference synchronization frequency, thereby forcing both to output discrete data points under the same hardware clock beat.

[0036] S1033, within the entire time window of high-frequency interception of the aforementioned multi-source signals, if the core processor of the main control module 110 is relied upon to perform repeated data reading and transfer operations, it will significantly preempt the calculation cycle of the underlying motor real-time control loop, potentially causing the physical movement of the vehicle entity 100 to become uncontrollable. The system utilizes the memory access module 112 to take over the data transfer task. For the specific settings of the direct memory access technology, those skilled in the art can refer to the direct memory access specification in the underlying architecture of the microcontroller. Its principle of bypassing the central processing unit to independently control the data bus is a well-known technology in the field and will not be elaborated here. The memory access module 112 transports the synchronized back electromotive force sequence and acceleration sequence to the buffer module 113 in a predetermined sequence through an independent direct memory access channel.

[0037] S1034, this embodiment constructs a dual-buffered alternating architecture to cope with intensive operating conditions. The buffer module 113 is divided into a first buffer block and a second buffer block with independent addresses in the microcontroller's random access memory. The memory access module 112 continuously writes sequence data to the first buffer block by default. If the system encounters a secondary transient impact that meets the trigger threshold condition before the current time window has ended, the memory access module 112 automatically redirects and moves the subsequent data sequence to the spare second buffer block according to the underlying hardware pointer monitoring mechanism. This alternating storage logic prevents the collision of new and old data overwriting caused by sudden high-frequency deformation, ensuring the integrity of the feature record corresponding to each impact. The aligned data sequence stored in the buffer module 113 is directly set as a multi-dimensional discrete vector set in terms of data structure. Each set of discrete data points in this vector set is composed of the back electromotive force signal value and the inertial acceleration component obtained at the same synchronous sampling time. This vectorized data binding method, based on the same underlying hardware timing index, eliminates the spatial and temporal separation between electrical and mechanical domain data, providing an input source with a strict timing mapping relationship for cross-domain frequency domain transformation of the host computer.

[0038] Please see the appendix Figure 1 and attached Figure 2 In step S104, the main control module 110 sends the data sequence stored in the buffer module 113 and the global system timestamp to the host computer 200 through the first communication module 114, which specifically includes the following sub-steps: S1041, after the memory access module 112 completes the task of transferring multi-source heterogeneous data within a specific time window, it will trigger a status flag indicating that the buffer block is full at the hardware level. Due to the limited interrupt resources of the microcontroller, to avoid performing time-consuming string concatenation and memory copy operations within the highest priority control interrupt, the real-time operating system set up inside the main control module 110 will respond to this flag and wake up a low-priority background communication thread that is currently suspended. This background communication thread is responsible for taking over the buffer block that is currently full of data and performing data packet encapsulation according to the predefined communication protocol specifications. In this embodiment, the data packet organization structure in memory is sequentially set as follows: a fixed sequence frame header, a global system timestamp recording the initial trigger time of physical deformation, a cross-domain sensor sequence as the effective data payload, and a cyclic redundancy check code for integrity verification. To ensure time alignment of the multi-source heterogeneous data during unpacking by the host computer, the specific interleaving arrangement logic of the cross-domain sensor sequence is strictly limited to: a single hardware synchronous sampling period as the basic unit, sequentially pushing in the back electromotive force data value and the triaxial acceleration data value at that specific moment, thereby forming a compact structure array. The purpose of adding a checksum to the end of the data is to ensure that the host computer can identify and discard erroneous data frames caused by radio frequency interference or physical link jitter, so as to avoid introducing incorrect timing characteristics into subsequent virtual simulation mapping.

[0039] In step S1042, after the background communication thread completes the data packet encapsulation logic, the main control module 110 passes the pointer to the first address of the data packet to be sent in memory and the total length of the data packet to the sending controller associated with the first communication module 114. Given the large amount of data accumulated from high-frequency continuous sampling, directly using processor instructions for polling and sending could easily block the execution flow of the main program. Based on this real-time challenge, this embodiment adopts an asynchronous non-blocking communication mechanism, with the hardware controller inside the first communication module 114 taking over the byte stream shifting and output task. Considering that under intensive testing conditions, the actual transmission rate of the physical layer may temporarily be lower than the data sampling generation rate, a circular buffer and overflow protection mechanism are added inside the communication system. When the queue to be sent is detected to be full, the system will trigger a protection judgment, actively discarding the oldest data frame to ensure the absolute priority of the real-time control task. The first communication module 114 continuously pushes the encapsulated data stream to the designated receiving port of the host computer 200 through a wireless fidelity network or industrial serial bus according to the predetermined baud rate and underlying level protocol.

[0040] S1043, during the entire cycle of the aforementioned data sequence being stripped, encapsulated, and asynchronously sent to the host computer 200, the main control module 110 needs to maintain high-frequency computation of the underlying real-time motor control loop. Based on this requirement, this solution utilizes three techniques—double-buffered alternating storage, independent transport of memory access modules, and low-priority thread packaging—to achieve decoupling of computational power. The physical significance of this mechanism lies in isolating the system overhead caused by high-frequency control computational power and large data throughput, ensuring that the microprocessor can solve the closed-loop proportional-integral-derivative control algorithm without interference within extremely short control cycles. Through this decoupling, the vehicle entity 100 can robustly execute diagnostic control instructions to complete physical excitation, while ensuring that the data packet carrying complete timing information and mechanical force characteristics is smoothly transferred, thus providing a timing-rigorous raw input for subsequent cross-domain frequency domain reconstruction on the host computer 200.

[0041] Please see the appendix Figure 1 and attached Figure 2 In step S105, the second communication module 210 receives the data sequence, and the frequency domain operation module 220 performs a frequency domain transformation operation on the back electromotive force sequence and the acceleration sequence, calculating the cross power spectrum and time-domain electromechanical cross-correlation sequence of the two sets of sequences. Specifically, this includes the following sub-steps: S1051, the second communication module 210 of the host computer 200 continuously monitors the network port or serial bus. After receiving a data packet asynchronously sent by the vehicle entity 100, the low-level driver of the host computer 200 executes the checksum comparison logic. When the data packet is determined to be complete and error-free, the system parses out the cross-domain sensor sequence carrying the global system timestamp. To meet the stringent requirements of subsequent frequency domain analysis on data dimensionality, the frequency domain operation module 220 uses a deinterleaving algorithm to re-extract the cross-domain sensor sequence into an independent one-dimensional back EMF time series and a three-dimensional triaxial acceleration time series. The reason for choosing acceleration as the input parameter is that chassis deformation usually exhibits multi-dimensional mechanical vibration characteristics in space. To achieve equal-dimensional cross-power spectrum matching with the back EMF signal, the system performs a dimensionality reduction operation on the three-dimensional acceleration data, specifically by calculating the arithmetic square root of the sum of squares of the three-axis components at each discrete sampling point, and equivalently merging them into a one-dimensional scalar acceleration sequence. Based on the reference synchronization frequency of the aforementioned lower-level hardware timer, these two sets of one-dimensional sequences maintain a strict homogeneous and equal-length characteristic in memory addresses, avoiding abnormal conditions such as matrix operations exceeding the limit due to inconsistent array lengths.

[0042] S1052, after extracting the independent time series, the frequency domain operation module 220 performs frequency domain transformation operations on the back electromotive force time series and the one-dimensional scalar acceleration time series. The general principle of multi-dimensional signal time-frequency conversion is to decompose the transient impacts and periodic oscillations mixed in the time domain into combinations of orthogonal basis functions of different frequencies, thereby intuitively quantifying the energy distribution law of each frequency band in the complex space of the frequency domain. Before performing the numerical transformation, in order to prevent the static bias voltage and the gravitational acceleration reference surface from interfering with the extraction of high-frequency features, the system independently performs a mean-removal operation on the above two sets of sequences to filter out the background DC component. At the same time, in order to suppress the spectral leakage phenomenon generated at the truncation edge of the finite-length discrete sampling sequence, the system synchronously applies a Hanning window to the preprocessed data sequence for smooth attenuation. After completing the above preparations, the system uses the fast Fourier transform algorithm to complete the mapping, obtaining the frequency domain back electromotive force sequence and the frequency domain acceleration sequence, respectively.

[0043] S1053, based on the acquired frequency domain sequences, the frequency domain operation module 220 calculates the cross-power spectrum of the two sequences. The cross-power spectrum can effectively quantify the coherence of the non-rigid deformation (mechanical domain) of the chassis of the vehicle entity 100 and the motor load fluctuation (electrical domain) at a specific frequency point. The core calculation formula for the cross-power spectrum is set as follows: ; In the formula, Indicates the first Cross-power spectrum values ​​at discrete frequency points; The frequency domain back electromotive force sequence represents the sequence at the 1st... Complex values ​​at discrete frequency points; The frequency domain acceleration sequence represents the first... Complex conjugate values ​​at discrete frequency points; The discrete frequency index is defined as ranging from zero to half the length of the Fast Fourier Transform window minus one, constrained by the Nyquist sampling theorem. The technical purpose of this calculation step is to amplify the resonant frequency band characteristics of tightly coupled mechanical and electrical systems while suppressing uncorrelated background random noise. Considering noise interference in actual physical systems, the calculated cross-power spectrum is usually further processed by a weighted average using a smoothing window function. In this embodiment, the system employs a fixed-bandwidth moving average filter to perform multi-point weighted smoothing of the cross-power spectrum across adjacent frequency bands. This operation aims to reduce the random error in spectral variance estimation, ensuring that the coherence determination of the output results is based on weighted logic along the frequency band dimension, avoiding biased judgments caused by relying solely on a single discrete frequency extremum.

[0044] In step S1054, after obtaining the cross-power spectrum, the frequency domain computation module 220 maps it back to the time dimension to extract the specific physical delay time. The system uses inverse discrete Fourier transform to perform reverse derivation on the smoothed cross-power spectrum to generate a time-domain electromechanical cross-correlation sequence. This sequence eliminates high-frequency incoherent clutter in the original time-domain signal, and its peak position and envelope shape reflect the time hysteresis response characteristics of mechanical impact transmitted to the electrical end of the motor. Through this closed-loop transformation logic from the time domain to the frequency domain and back to the time domain, the system completes the deep fusion and feature purification of cross-domain multi-source signals, providing a data foundation for constructing the hysteresis and flexible coupling relationship of the underlying dynamics.

[0045] Please see the appendix Figure 1 and attached Figure 2 In step S106, the feature mapping module 230 extracts degradation features based on the above-mentioned time-domain electromechanical cross-correlation sequence, and combines them with a preset neural network model to convert them into the underlying dynamic parameters required by the virtual simulation module 240. Finally, the time-series mapping and state update of the non-rigid degradation state of the physical chassis are completed in the digital space, which specifically includes the following sub-steps: S1061, the feature mapping module 230 of the host computer 200 takes over the time-domain electromechanical cross-correlation sequence from the frequency domain operation module 220. In actual physical operation, due to the thermal drift effect of sensor devices and environmental electromagnetic background radiation, the original sequence is often accompanied by baseline shift and random impulse noise. To meet the strict requirements of subsequent high-dimensional feature extraction algorithms on the stability of input data distribution, the feature mapping module 230 uses standard deviation normalization logic to preprocess the one-dimensional sequence. This preprocessing logic calculates the local arithmetic mean and standard deviation of the input sequence and performs translation and scaling, thereby forcing the mean of the data distribution to tend to zero and the variance to normalize. In this process, to prevent the variance from approaching zero due to the extreme flatness of the sequence and thus causing a division-by-zero anomaly, the system forcibly introduces a very small positive floating-point constant (e.g., 10) into the denominator of the standard deviation. -7 This serves as a lower bound for protection. After preprocessing, considering the potential slight fluctuations in the length of the extracted time series under actual working conditions, the feature mapping module 230 performs zero-padding or equidistant resampling operations on the sequence based on the preset maximum sampling length. This low-level alignment operation ensures that non-uniformly long dynamic features are forcibly reorganized into a strictly fixed-dimensional tensor format, ready for injection into the feature evaluation network.

[0046] S1062, based on the standardized tensor generated by the above preprocessing, the feature mapping module 230 calls the internally deployed one-dimensional convolutional neural network model to perform mapping inference. The reason for choosing a one-dimensional convolutional architecture to address the local feature coupling problem of cross-domain sensor time series is that its sliding convolution kernel can efficiently capture the local peak hysteresis features of mechanical impact within its coherent decay period, without relying on expert experience for manual feature selection. In this embodiment, the internal hierarchical structure of the neural network model is rigorously configured as follows: a data input layer, two consecutively cascaded sets of feature extraction blocks, and a fully connected regression layer at the end. Each feature extraction block sequentially includes a one-dimensional convolutional layer, a batch normalization layer, a linear rectified activation function, and a max-pooling layer. The data flow follows the forward propagation law; the standardized one-dimensional sequence tensor passes through each level of network nodes sequentially. Its spatial dimension is gradually compressed under pooling operations, while the channel depth characterizing electromechanical hysteresis characteristics increases layer by layer. Finally, the fully connected regression layer maps it to the physical state output under the specific business scenario. The output of the model is defined as a two-dimensional parameter vector, which respectively represents the mechanical stiffness degradation compensation and damping coefficient degradation compensation in the chassis structure of the virtual simulation module 240.

[0047] S1063, regarding the construction and parameter establishment of the one-dimensional convolutional neural network, the system pre-executed supervised training steps in an offline environment. The training sample dataset comes from a large number of measured electromechanical cross-correlation sequences obtained from multiple calibrated vehicle entities 100 at different wear stages throughout their entire life cycle. The corresponding ideal labels are obtained by back-calculation of the true values ​​of physical mechanical deformation synchronously recorded by a high-precision laser vibrometer. The network training process is based on the error backpropagation mechanism, aiming to find the weight combination that minimizes the overall prediction bias. To extend the comprehensiveness of equivalent feature protection, the core loss function on which this training iteration depends is logically expressed as a mean squared error criterion with regularization penalty. The specific calculation steps are as follows: First, calculate the square mean of the difference between the true physical degradation label vector corresponding to all samples in the current batch and the predicted degradation vector output by the current forward inference of the network model. On the basis of this mean squared error, forcibly superimpose a penalty term obtained by multiplying the manually set regularization penalty coefficient by the square of the L2 norm of the weight matrix of the fully connected layer of the network. The regularization penalty coefficient, which is manually set, is typically between 0.001 and 0.01 to constrain network complexity. The technical purpose of this loss function is not only to rigorously quantify the mean squared error between the model's predicted values ​​and the actual physical state, but also to suppress abrupt changes in the high-dimensional weight matrix by introducing a regularization penalty term, thus preventing the model from overfitting when faced with noisy experimental sequences.

[0048] S1064, after obtaining the mechanical stiffness degradation compensation and damping coefficient degradation compensation from the network output, the feature mapping module 230 does not directly perform a rigid numerical overwrite. Considering that a single extreme mechanical impact may cause an instantaneous extreme value jump in the model output, to prevent numerical divergence in the virtual simulation system, the validity determination of the output result is based on a time-dimensional weighted smoothing logic. The feature mapping module 230 uses an exponential moving average algorithm to smoothly fuse the current predicted component with the historical state baseline to obtain the final simulation update parameters. Subsequently, the virtual simulation module 240 uses these update parameters to reconstruct its internal multibody dynamics matrix. In the process of reconstructing the chassis stiffness matrix, the system forcibly executes matrix singularity check logic. By solving the determinant of the updated stiffness matrix, if the absolute value of the determinant is determined to be lower than the preset non-singular safety threshold, the system will automatically use a very small unit diagonal matrix for perturbation compensation. In this embodiment, the determination of the non-singular safety threshold depends on the machine precision of the host computer's floating-point arithmetic unit, and is usually set to 10. -6 Up to 10 -5 This verification and compensation mechanism ensures at the physical level that the chassis virtual model always possesses the most basic anti-deformation recovery capability, and prevents the simulation solver from crashing due to matrix inversion failure from the bottom layer of the mathematical algorithm, thus robustly realizing the mapping of deteriorated state from physical entity to digital space.

[0049] Please see the appendix Figure 1 and attached Figure 2 In step S107, the virtual simulation module 240 performs forward deduction based on the updated underlying dynamic parameters, and the control strategy generation module 250 generates feedforward compensation commands based on the deduction results and sends them to the vehicle entity 100 to achieve virtual-real closed-loop control. This specifically includes the following sub-steps: S1071, after the feature mapping module 230 completes parameter reconstruction, the virtual simulation module 240 initiates a forward dynamic simulation within a leading time window. The general principle of multibody system dynamics derivation lies in establishing a system of second-order differential equations containing the mass matrix, damping matrix, and stiffness matrix based on the Lagrange equations, thereby solving for the system's state evolution under a given input excitation. In this embodiment, the system uses the current actual electromagnetic torque of the motor as a known external input, substituting it into a digital model that has incorporated degraded stiffness and damping coefficients, to calculate a virtual state trajectory including the chassis's flexible deformation displacement. Since this model has absorbed near-real-time physical wear characteristics, the generated virtual state trajectory can faithfully reproduce the potential oscillation trend and phase shift of the vehicle's chassis 100 over several future control cycles.

[0050] S1072, the control strategy generation module 250 takes over the aforementioned virtual state trajectory and begins to calculate the feedforward compensation command used to suppress mechanical oscillations. For this type of flexible vibration suppression problem involving multivariable coupling, this embodiment employs a linear quadratic regulator algorithm to construct the optimal control law. The core concept of this algorithm is to transform the control objective into minimizing an infinite time cost function encompassing state deviation and control energy. This infinite time cost function is implemented at the underlying logic level as: continuously integrating the sum of the state deviation loss and the control input loss over the entire integral interval. Here, the state deviation loss is quantified as a time variable. The state deviation vector at any given time (i.e., the difference between the virtual simulated flexible displacement and the ideal rigid displacement) is the continuous product of its transpose and the positive definite state weighting matrix. The control input loss is quantized as the control input vector to be solved (i.e., the physical feedforward compensation torque acting on the motor) and its transpose and the positive definite control weighting matrix. The determination of these two weighting matrices depends on the weighted trade-off between the system state convergence speed and the upper limit of motor output energy consumption in the actual control requirements. They are typically statically calibrated by technicians during offline commissioning based on the rated power of the specific motor. To prevent singularity collapse or floating-point division-by-zero overflow due to the inversion operation of the control weighting matrix when solving the feedback gain matrix, as a preferred method, the system sets an absolute value of not less than 10 for the main diagonal elements of the positive definite control weighting matrix at the underlying level. -4The lower bound of the positive constraint. The technical purpose of this function calculation step is to find an optimal control path that allows the system to decay the non-rigid mechanical deformation of the chassis as quickly as possible while consuming the least electrical energy.

[0051] The control strategy generation module 250 calculates the feedback gain matrix by solving the associated algebraic Riccati equation, and then multiplies it with the state deviation vector to obtain the current control input vector.

[0052] S1073, after acquiring the control input vector, the control strategy generation module 250 converts it into an equivalent feedforward current given component required for the chassis drive control of the vehicle entity 100. Considering the instruction lag that may be caused by industrial Ethernet network congestion or scheduling delay of the host computer operating system, in order to avoid the delayed instruction being out of phase with the current physical phase and thus causing positive feedback divergence, the output result is determined based on a multi-dimensional timeliness weighted logic of communication delay and calculation time. Specifically, the system checks the difference between the generation timestamp corresponding to the feedforward current given component and the current absolute clock. If the difference is strictly less than one physical chassis control cycle, the system determines that the timing is fully aligned and retains the instruction with full weight (i.e., gain coefficient of 1.0); if the difference exceeds one cycle but is within a preset tolerance buffer window, the system multiplies the current given component by a linear attenuation coefficient between 0 and 1 according to the proportion of time deviation, so as to avoid secondary sudden changes in motor torque caused by abrupt truncation; if it completely exceeds the tolerance buffer window, the component is actively zeroed out. The current given component after the above weighted determination will be encapsulated into a downlink control frame containing a cyclic redundancy check code, and then asynchronously sent to the vehicle entity 100 by the second communication module 210 via an industrial Ethernet bus or wireless link.

[0053] In S1074, the main control module 110 of the vehicle entity 100 receives and verifies the downlink control frame. After confirming that the data link is intact and error-free, it parses it into a specific feedforward current compensation value. In the next low-level hardware timer interrupt cycle, the main control module 110 uses direct memory access or register overwrite operations to directly superimpose this compensation value onto the low-level current closed-loop reference input of the drive module 120. This signal superposition operation electrically alters the motor output torque command originally based on the assumption of an absolute rigid body. The updated electromagnetic torque acts on the drive wheels, and the resulting mechanical reaction force can accurately counteract the low-frequency resonance induced by chassis structural degradation in both the time and frequency domains. Through this closed-loop compensation mechanism, the system robustly achieves complete closed-loop control from physical entity perception and digital space degradation assessment to physical entity compensation drive.

[0054] Please see the appendix Figure 1 and attached Figure 2In step S108, the host computer 200, based on the operational feedback data after virtual-real closed-loop control and the updated underlying dynamic parameters, performs adaptive evolution of the digital model and multi-dimensional assessment and early warning of the chassis health status of the vehicle entity 100. This specifically includes the following sub-steps: S1081, the second communication module 210 of the host computer 200 continuously acquires the closed-loop feedback acceleration sequence of the vehicle entity 100 after performing feedforward compensation. The virtual simulation module 240 synchronously extracts the virtual predicted acceleration sequence within the corresponding absolute time window. The general technical principle of multi-dimensional system parameter evolution lies in adjusting the implicit deviation coefficient inside the model in reverse by quantifying the system residual between the predicted value in digital space and the measured value in physical space, so that the virtual model continuously approximates the real physical state. To ensure the tight alignment of the two sets of multi-source data in time and operating conditions, the system performs sequence truncation based on the pre-embedded global system timestamp to ensure that the starting beats of the two are consistent. Subsequently, the system calculates the difference sequence between the feedback acceleration and the virtual acceleration at the corresponding sampling time. Considering the high-frequency glitches that may be caused by sensor noise floor and mechanical transmission clearance, the system uses root mean square error logic for time-domain smoothing to prevent noise abrupt changes from causing artificially high residuals. The core formula of this evaluation index is set as follows: ; In the formula, This represents the root mean square residual value within the current time window; This indicates the total number of discrete sampling points contained within the time window; Indicates the first The actual feedback acceleration value at each discrete point; Indicates the first Virtual predicted acceleration values ​​at discrete points; This represents the discrete point index. The technical purpose of this calculation step is to filter out single-point extreme value jumps and objectively and stably quantify the approximation accuracy of the virtual and real models on a macroscopic time scale. When performing division operations, given... The total number of discrete sampling points within this time window is always a positive integer under normal communication conditions, naturally avoiding the risk of arithmetic overflow where the denominator approaches zero. To address the extreme anomaly of no valid sampling points within a specific time window due to a physical interruption of the communication link, the system forcibly presets a low-level verification judgment: if a violation is detected... If the value is zero, the residual calculation for the current cycle is skipped and the evaluation value of the previous effective cycle is inherited, thus achieving absolute completeness of the algorithm logic under all working conditions.

[0055] S1082, based on the generated root mean square residual value, the system uses a recursive least squares algorithm with a forgetting factor to iteratively update the unmodeled perturbation matrix in the virtual simulation module 240. In the update stage of inverting the high-dimensional covariance matrix, to address the matrix singularity problem caused by insufficient frequency band of the physical input excitation signal, the system forcibly implements a diagonal loading regularization strategy. A very small positively stable scalar is appended to the main diagonal of the matrix to be inverted. In this embodiment, the value range of this positively stable scalar is limited to 10. -5 Up to 10 -4 The specific determination of this value depends on the addressing bit width and rounding tolerance limit of the floating-point arithmetic unit of the host computer processor. This underlying protection logic mathematically avoids parameter update divergence caused by local singularities, ensuring the algorithmic completeness of the adaptive evolution process of the digital model.

[0056] S1083, after obtaining the latest chassis dynamic parameters after evolution, the host computer 200 enters the high-level evaluation stage of health status. In complex industrial environments, drastic fluctuations in a single parameter often stem from accidental external impacts rather than structural degradation. Therefore, the judgment of the output results must be based on multi-dimensional weighted logic, and it is strictly forbidden to rely solely on the one-sided judgment of a single extreme value. The system simultaneously extracts the mechanical stiffness degradation compensation amount, damping coefficient degradation compensation amount, and the root mean square residual value calculated above after reconstruction. These three heterogeneous dimensional data with strong physical causal relationships are input into the preset evaluator, and the range standardization algorithm is used to uniformly map them to a dimensionless interval. The underlying derivation logic of range standardization is to subtract the sample minimum value from the original variable and then divide by the difference between the sample maximum and minimum value. To strictly prevent the division-by-zero crash anomaly caused by the maximum and minimum values ​​being completely equal under extremely stable physical conditions or data bus jamming conditions, the system forcibly superimposes a very small positive floating-point constant (e.g., 10) when calculating the range denominator. -6 The system uses these three sets of normalized data as a baseline for numerical protection. Subsequently, it assigns different static weight coefficients to these three sets of normalized data, calculating a comprehensive chassis health degradation index that characterizes the current physical state. In this embodiment, the values ​​of the aforementioned static weight coefficients are strictly limited to between 0 and 1, and the sum of the three is forcibly constrained to 1.0. The specific numerical ratios are not blindly specified, but rather determined in advance by technicians using the analytic hierarchy process (AHP) combined with sensitivity calibration from a large number of offline historical failure cases. The reason for choosing these three parameters as input dimensions is that mechanical stiffness directly reflects the absolute decrease in chassis load-bearing capacity, damping coefficient characterizes the increased energy dissipation caused by microcrack propagation, and root mean square residuals macroscopically confirm the actual interference of physical degradation on control accuracy. These three parameters are interdependent, jointly constructing a complete health assessment coordinate system.

[0057] S1084, Regarding the aforementioned comprehensive chassis health degradation index, the system maintains a historical data queue with a fixed time span in memory. When assessing the service life of the current physical entity, the system does not use the highest index in a single frame as the judgment criterion, but continuously calculates the exponential moving average of all index elements in the historical queue. The system compares this smoothed moving average with a preset service life termination threshold in real time. The service life termination threshold is determined based on the lower limit of the extreme value recorded at the moment when a micro-crack macroscopic occurs during destructive fatigue bench testing of the same batch of standard vehicle entities. In this embodiment, this threshold is typically set between 0.85 and 0.92 in the normalized interval. Once it is detected that the moving average exceeds the set service life termination threshold for multiple consecutive check cycles, the system ultimately determines that the physical chassis of vehicle entity 100 has suffered irreversible plastic damage or severe fatigue. In this state, the host computer 200 will trigger a highlighted alarm log on the visual interface and generate a digital maintenance ledger carrying precise degradation coordinates and component replacement suggestions according to the protocol specifications, thereby completing the entire business process from bottom-level signal perception to top-level status intervention in a closed loop.

[0058] Specific application examples: Application Background: In a university's autonomous vehicle laboratory, an autonomous vehicle used as a teaching tool has been in service for two years. Recently, when students were verifying the path tracking algorithm, they discovered that the vehicle frequently experienced chassis vibration and control delays during static starts or high-speed turns. This was suspected to be due to non-rigid degradation of the chassis mechanical structure caused by long-term collisions and wear (such as loose screws and microcracks in the carbon fiber chassis). To eliminate the interference of this mechanical degradation on the upper-level control algorithm, the laboratory implemented the virtual simulation and compensation system described in this invention.

[0059] Specific implementation steps: Command triggering and data synchronization: During the simulation initialization phase, the host computer generates diagnostic commands to control the duty cycle of the vehicle's pulse width modulation (PWM) signal to jump directly from 0 to 90% within a single control cycle (step acceleration). Simultaneously, the host computer and the vehicle's main control module synchronize their clocks and establish a global system timestamp.

[0060] Low-level high-frequency data extraction: The vehicle starts instantaneously upon receiving a step command, and the instantaneous conversion of static friction into kinetic friction triggers microscopic deformation of the chassis. The chassis deformation torque is transmitted back to the motor rotor, causing a local spike in the motor's back electromotive force (EMF). The main control module calculates the transient derivative and finds it exceeds 4.5 times the steady-state extreme value, immediately triggering the highest-level hardware interrupt. Within the subsequently set 150-millisecond time window, the memory access module (DMA) automatically and synchronously moves the back EMF sequence and the triaxial acceleration sequence to a dual buffer block and sends them back to the host computer via Ethernet with a timestamp.

[0061] Cross-domain feature frequency domain extraction: After receiving the data, the host computer's frequency domain computation module removes the mean from the aforementioned time series, adds a Hanning window, performs a Fast Fourier Transform, and calculates the cross-power spectrum of back electromotive force and acceleration (the core calculation basis is...). The sequence is then inversely transformed back to a time-domain electromechanical cross-correlation sequence. A one-dimensional convolutional neural network (CNN) infers from this sequence that the current mechanical stiffness degradation compensation of the chassis is -18%, and the damping coefficient degradation compensation is -12%.

[0062] Digital Reconstruction and Feedforward Compensation: The virtual simulation module substitutes the aforementioned attenuation parameters into the dynamic equations and performs forward derivation to accurately predict the oscillation amplitude and phase of the vehicle in the next few control cycles. Based on this, the control strategy generation module uses the LQR algorithm (minimizing an infinite time cost function covering state deviation and control energy) to calculate a feedforward current compensation command to suppress oscillations and sends it to the vehicle entity. The vehicle motor superimposes this reverse compensation current on top of the original drive command to actively cancel mechanical resonance from the physical electrical end.

[0063] Health assessment and early warning: After the control loop is closed, the system is based on the formula The system continuously compares the virtual and real accelerations and calculates the root mean square residual. At this point, the comprehensive chassis health degradation index, derived by mapping the residuals with the degradation parameters, is 0.81. Although it has not reached the absolute scrap threshold (0.88) for the end of its service life, the system has already written the relevant parameters into the maintenance log, indicating that the connection between the right front suspension and the chassis needs to be tightened again.

[0064] Experimental verification, effect comparison, and analysis of figures: Please see the appendix Figure 5 To verify the actual effectiveness of the technical solution of the present invention in a physical entity, the laboratory conducted two sets of control experiments on the aforementioned deteriorated vehicle in the same testing site. Figure 5 The results of the comparison of the transient response of chassis vibration during the step start of the teaching vehicle were visually recorded.

[0065] Test conditions: On a smooth epoxy resin surface, step drive commands are sent to change the duty cycle from 10% to 85%.

[0066] Control group: Traditional PID control is used (without the feedforward compensation mechanism of this invention).

[0067] Experimental group: Start the virtual-real closed-loop feedforward compensation system of this invention.

[0068] Combined with appendix Figure 5The coordinate system and the physical meaning of each curve in the figure, along with the explanation of the core comparative data, are as follows: Basic meaning of chart coordinate elements: The horizontal axis (X-axis) - time (s): represents the absolute time after the vehicle receives the step start command and begins to execute the action, in seconds. As can be observed in the graph, the curve begins to fluctuate sharply around 0.02 seconds, indicating the exact moment the physical action actually begins.

[0069] Vertical axis (Y-axis) - Chassis vibration acceleration (m / s²) 2 The value represents the longitudinal / comprehensive vibration acceleration collected by the inertial sensing module installed on the chassis of the vehicle. The larger the absolute value, the more severe the mechanical deformation and vibration of the chassis. The 0 mark indicates that the vehicle is in a stable running or stationary state.

[0070] Key comparative data and curve trend analysis: Acceleration transient extreme value suppression: As shown by the traditional control (without feedforward compensation) represented by the black dashed line in the figure, the peak acceleration at the moment of start-up is extremely large (indicating a large starting impact force), with the peak longitudinal acceleration reaching 2.85 m / s². 2 Such drastic deformation can easily lead to severe distortion of sensor data; in contrast, the system of this invention (feedforward degradation compensation), represented by the black solid line, shows that the highest extreme value at the moment of start-up is significantly reduced to 1.21 m / s. 2 The impact extreme value decreased by approximately 57.5%.

[0071] The duration of mechanical oscillations was shortened: Observing to the right along the time axis, the black dashed line representing the experimental group without the system of this invention experienced a relatively long period of up-and-down oscillations, lasting approximately 420ms before gradually converging to 0. This reflects the prolonged mechanical resonance phenomenon caused by the lack of active electrical suppression after non-rigid degradation of the chassis. Conversely, the black solid line representing the closed-loop system of this invention showed that subsequent oscillations were rapidly absorbed and completely suppressed within 140ms, with the convergence time shortened by 66.6%. This strongly demonstrates that this invention, by accurately calculating the feedforward reverse current, cleverly utilizes the electromagnetic torque of the motor to forcibly counteract the mechanical deformation torque at the moment when the physical chassis is about to oscillate.

[0072] Root mean square residual ( Verification: During closed-loop operation, the residual calculation logic of the multi-dimensional system was used to evaluate that the residual of the virtual and real motion model of the control group (dashed line condition) was as high as 0.48; while after the evolution compensation of this system was enabled (real line condition), the residual stabilized below 0.11, which greatly improved the high fidelity and physical following characteristics of the virtual simulation.

[0073] Experimental conclusion: This invention can effectively identify the cross-domain hysteresis characteristics caused by mechanical structure aging, and through micro current feedforward compensation in the electrical domain, it enables an old car with non-rigid degradation to recover to a control response steady state close to that of a brand-new rigid car in terms of chassis vibration transient response.

Claims

1. A teaching unmanned intelligent car, characterized in that, Includes a vehicle body (100) and a host computer (200); The host computer (200) is used to issue diagnostic control command sequences and establish a global system timestamp with the main control module (110) of the vehicle entity (100); The main control module (110) is used to trigger a hardware interrupt when the transient derivative of the back electromotive force signal of the collected motor module (130) exceeds a preset threshold, and integrate the global system timestamp, the back electromotive force signal and the triaxial acceleration signal collected synchronously by the inertial sensing module (140) into a data sequence and send it to the host computer (200). The host computer (200) is used to calculate the cross power spectrum and time-domain electromechanical cross-correlation sequence based on the data sequence to construct the hysteresis and flexible coupling matrix, and to use the matrix to reconstruct the ideal control command to generate the dynamic degradation command; The host computer (200) is also used to execute the dynamic degradation instruction to output virtual sensor observation data, and generate a feedforward compensation instruction based on it and send it to the vehicle entity (100), which is then superimposed by the main control module (110) to the bottom driving end of the vehicle entity (100) for degradation compensation.

2. The unmanned intelligent smart car for teaching according to claim 1, characterized in that, The host computer (200) is used to send a synchronization request data packet carrying the host computer system time; The main control module (110) is used to record the local microsecond-level count value of the internal hardware timer of the main control module (110) at the instant of receiving the synchronization request data packet; The host computer (200) and the main control module (110) are used to calculate the clock offset by interacting with the synchronization request data packet multiple times; The main control module (110) is used to add the local microsecond-level count value of the hardware timer at the current moment to the clock offset to generate the global system timestamp.

3. The unmanned intelligent smart car for teaching according to claim 2, characterized in that, The diagnostic control command sequence includes a step acceleration command and a step steering command for triggering the dynamic response of the vehicle entity (100); The main control module (110) includes an analog-to-digital conversion module (111). The main control module (110) is used to set the internal hardware timer to output a center-aligned pulse width modulation signal, and to start the analog-to-digital conversion module (111) to perform signal sampling on the back electromotive force signal within the dead time window; The main control module (110) is used to subtract the historical value of the back EMF signal from the previous basic sampling period from the back EMF signal collected at the current moment, and divide the calculated potential difference by a fixed constant time interval to obtain the transient derivative.

4. The unmanned intelligent smart car for teaching according to claim 3, characterized in that, The main control module (110) includes a buffer module (113) and a memory access module (112). The main control module (110) is used to extract the global system timestamp at the moment when the hardware interrupt occurs and write it into the first address space of the buffer module (113); The main control module (110) is used to bind the hardware reading interface of the inertial sensing module (140) to the conversion completion event of the analog-to-digital conversion module (111) at the hardware level. Within a preset time window, the memory access module (112) is used to transfer the discrete data points of the synchronously acquired back electromotive force signal and the triaxial acceleration signal to the buffer module (113). The buffer module (113) is divided into a first buffer block and a second buffer block in memory, and the memory access module (112) is used to alternately write data to the first buffer block and the second buffer block.

5. The teaching unmanned intelligent trolley according to claim 4, characterized in that, The data sequence includes a cross-domain sensor sequence, which is composed of the back electromotive force signal and the triaxial acceleration signal interleaved and arranged in the same sampling period; The background communication thread inside the main control module (110) is used to encapsulate the first buffer block or the second buffer block filled with data into a data packet. The data packet includes a fixed sequence frame header, the global system timestamp, the cross-domain sensor sequence, and a cyclic redundancy check code. The main control module (110) includes a first communication module (114), and the host computer (200) includes a second communication module (210). The first communication module (114) is used to push the data packet to the second communication module (210) through an asynchronous non-blocking communication mechanism.

6. The teaching unmanned intelligent trolley according to claim 5, characterized in that, The host computer (200) includes a frequency domain operation module (220). The frequency domain operation module (220) is used to perform a deinterleaving operation on the cross-domain sensor sequence to separate it into a one-dimensional back electromotive force time series and a three-dimensional triaxial acceleration time series; The arithmetic square root of the sum of squares of the triaxial acceleration components at each sampling point in the three-dimensional triaxial acceleration time series is used to obtain the one-dimensional scalar acceleration series. The one-dimensional back EMF time series and the one-dimensional scalar acceleration series after mean removal and smoothing window processing are independently subjected to fast Fourier transform to obtain frequency domain back EMF series and frequency domain acceleration series respectively. The cross power spectrum is obtained by multiplying the conjugate complex numbers of the frequency domain back EMF series and the frequency domain acceleration series. The cross-power spectrum is weighted and smoothed using a moving average filter, and an inverse discrete Fourier transform is performed on the smoothed cross-power spectrum to generate a time-domain electromechanical cross-correlation sequence.

7. The teaching unmanned intelligent trolley according to claim 6, characterized in that, The host computer (200) includes a matrix generation module (230); The matrix generation module (230) is used to process the time-domain electromechanical cross-correlation sequence using standard deviation normalization logic to generate a normalized tensor; A one-dimensional convolutional neural network model is invoked to infer the standardized tensor, and the mechanical stiffness degradation compensation amount and damping coefficient degradation compensation amount representing the parameters of the hysteresis and flexible coupling matrix construction are output. The one-dimensional convolutional neural network model includes an input layer, two sets of continuously cascaded feature extraction blocks, and a fully connected regression layer at the end. It is trained by backpropagation using a mean squared error loss function with regularization penalty.

8. The teaching unmanned intelligent trolley according to claim 7, characterized in that, The host computer (200) includes a control strategy generation module (250) and a virtual simulation module (240), and the vehicle entity (100) includes a drive module (120). The virtual simulation module (240) is used to calculate the virtual state trajectory by substituting the dynamic parameters, which include the mechanical stiffness degradation compensation amount and the damping coefficient degradation compensation amount. The control strategy generation module (250) is used to construct a cost function covering state deviation loss and control input loss for the linear quadratic regulator algorithm for the virtual state trajectory, calculate the feedback gain matrix and generate feedforward compensation instructions; The second communication module (210) is used to send the feedforward compensation command to the vehicle entity (100), which is then parsed by the main control module (110) into a feedforward current compensation value and superimposed on the bottom current closed-loop reference input terminal of the drive module (120).

9. The teaching unmanned intelligent vehicle according to claim 1, characterized in that, The host computer (200) is used to extract the real feedback acceleration sequence and the virtual predicted acceleration sequence of the closed-loop feedback, and to calculate the difference sequence of the sampling points of the two sets of sequences within the same time window to generate the root mean square residual value. The host computer (200) is used to iteratively update the dynamic parameters inside the virtual simulation module (240) by using the recursive least squares algorithm and adding a stable scalar to the main diagonal of the matrix to be inverted; The mechanical stiffness degradation compensation amount, damping coefficient degradation compensation amount, and the root mean square residual value are uniformly mapped and calculated through range standardization to obtain the comprehensive chassis health degradation index.

10. A virtual simulation method for teaching an unmanned intelligent smart car, characterized in that, The application of a teaching unmanned intelligent vehicle as described in any one of claims 1-9 includes the following steps: The host computer (200) issues a sequence of diagnostic control instructions and establishes a global system timestamp with the main control module (110) of the vehicle entity (100); The main control module (110) triggers a hardware interrupt when the transient derivative of the back electromotive force signal of the collected motor module (130) exceeds a preset threshold. The main control module (110) integrates the global system timestamp, the back electromotive force signal and the triaxial acceleration signal synchronously collected by the inertial sensing module (140) into a data sequence and sends it to the host computer (200). The host computer (200) calculates the cross power spectrum and time-domain electromechanical cross-correlation sequence based on the data sequence to construct the hysteresis and flexible coupling matrix; The host computer (200) uses the matrix to reconstruct the ideal control instruction and generate a dynamic degradation instruction; The host computer (200) executes the dynamic degradation instruction to output virtual sensor observation data, and generates a feedforward compensation instruction based on this and sends it to the vehicle entity (100), which is then superimposed by the main control module (110) to the bottom driving end of the vehicle entity (100) for degradation compensation.

Images