Low-altitude robot motor multi-objective optimization method based on model predictive control

Abstract

The application relates to the technical field of motor drive control, and specifically discloses a low-altitude robot motor multi-objective optimization method based on model predictive control, which uses an onboard microphone to collect propeller aerodynamic noise in real time and analyzes the hearing masking band characteristics of the noise, simultaneously quantitatively evaluates the acoustic disturbance of the predicted current trajectory of each candidate voltage vector based on an offline calibrated electromechanical structure resonance cost matrix, and fuses the resonance penalty and the masking reward into a unified acoustic disturbance index, which is taken as an independent evaluation item to participate in rolling optimization together with torque tracking and switching loss in a multi-objective cost function, so that the inverter is guided to actively avoid the dangerous switching sequence of triggering structure resonance in a microsecond-level control cycle, and the inevitable electromagnetic harmonic energy is moved to the masking frequency area of the aerodynamic noise, thereby realizing the acoustic stealth of the low-altitude robot in all working conditions without sacrificing the dynamic thrust response.

Description

Technical Field

[0001] This application relates to the field of motor drive control technology, and more specifically, to a multi-objective optimization method for low-altitude robot motors based on model predictive control. Background Technology

[0002] With the rapid development of the low-altitude economy, low-altitude robots, represented by multi-rotor UAVs, are increasingly widely used in urban logistics, aerial inspection, and low-altitude commuting. The control performance of their rotor motor drive systems directly determines the dynamic response capability and mission execution quality of the flight platform. When operating in urban low-altitude airspace, low-altitude robots not only need to cope with the high dynamic torque response requirements under complex airflow disturbances, but also face increasingly stringent urban noise regulations. This requires the motor drive control system to ensure thrust tracking accuracy while also taking into account acoustic quality control. Therefore, the construction of a multi-objective optimization control scheme for low-altitude robot motors has urgent engineering needs and industrial value.

[0003] Model predictive control (MDC) has been widely applied in the field of motor drives due to its inherent multivariate handling capabilities and fast dynamic response characteristics. However, existing motor drive schemes based on finite control set MDC typically only consider electromagnetic torque tracking error and inverter switching losses as core optimization objectives in constructing multi-objective cost functions, completely ignoring the impact of electromagnetic noise caused by voltage vector switching on the urban acoustic environment. Due to the inherent characteristics of discrete voltage vectors, the switching sequence selected in finite control set MDC generates frequency converter current harmonics in each control cycle. When the frequencies of these harmonics happen to fall near the mechanical natural frequencies of the motor stator core and the arm connection structure, severe structural resonance will be induced, producing a piercing high-frequency electromagnetic howl. The cost functions of existing schemes are completely blind to this electro-acoustic coupling effect; the algorithm cannot predict the spectral characteristics of the disturbance generated by the currently selected voltage vector sequence in the frequency domain, nor can it assess the degree to which these spectral components are perceptible to the human ear. The root cause of this deficiency lies in the fact that the existing model predictive control framework lacks the technical means to incorporate the acoustic resonance characteristics of electromechanical structures and the psychoacoustic perception model of the human ear into the real-time evaluation system, resulting in a structural deficiency in the evaluation dimensions of multi-objective optimization.

[0004] Therefore, an optimized multi-objective optimization scheme for the motors of low-altitude robots is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a multi-objective optimization method for the motors of low-altitude robots based on model predictive control.

[0006] According to one aspect of this application, a multi-objective optimization method for the motor of a low-altitude robot based on model predictive control is provided, comprising: Step 1: Perform feature extraction and masking frequency band analysis on the collected operating status data to obtain the stator current feedback quantity in the rotating coordinate system and the feature vector of the dynamic auditory masking frequency band; Step 2: Based on the DC bus voltage and the real-time speed of the rotor motor, construct a finite control set and predict the forward state of the stator current feedback in the rotating coordinate system to obtain the inverter finite voltage vector set and the predicted stator current trajectory set. Step 3: Based on the pre-set electromechanical structure resonance cost matrix, perform electromechanical structure acoustic excitation feature targeted quantization on the predicted stator current trajectory set and the dynamic auditory masking frequency band feature vector to obtain the candidate vector acoustic excitation index array. Step 4: Based on the desired electromagnetic torque command and the desired stator flux linkage command, perform multi-objective rolling optimization for active noise shaping on the predicted stator current trajectory set, the inverter finite voltage vector set, and the candidate vector acoustic disturbance index array to obtain the optimal switching vector. Step 5: The optimal switching vector is subjected to anti-shoot-through dead-time compensation through dead-time compensation to obtain a three-phase pulse modulation drive signal with dead time. The three-phase pulse modulation drive signal with dead time is output to the inverter gate drive circuit.

[0007] Compared with existing technologies, this application provides a multi-objective optimization method for low-altitude robot motors based on model predictive control. It utilizes an onboard microphone to collect propeller aerodynamic noise in real time and analyze its auditory masking frequency band characteristics. Simultaneously, based on the offline calibrated electromechanical structure resonance cost matrix, it performs acoustic excitation quantification evaluation on the predicted current trajectory of each candidate voltage vector. The resonance penalty and masking reward are integrated into a unified acoustic excitation index, which is included as an independent evaluation item in the multi-objective cost function, participating in rolling optimization alongside torque tracking and switching losses. This guides the inverter to actively avoid dangerous switching sequences that trigger structural resonance within a microsecond-level control cycle, and shifts unavoidable electromagnetic harmonic energy to the masking frequency band of aerodynamic noise, achieving full-condition acoustic stealth for the low-altitude robot without sacrificing dynamic thrust response. Attached Figure Description

[0008] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1This is a flowchart of a multi-objective optimization method for low-altitude robot motors based on model predictive control, according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow of a multi-objective optimization method for low-altitude robot motors based on model predictive control according to an embodiment of this application; Figure 3 This is a flowchart of step three of the multi-objective optimization method for low-altitude robot motors based on model predictive control according to an embodiment of this application; Figure 4 This is a flowchart illustrating the process of performing acoustic perturbation feature-targeted fusion of the candidate vector fundamental resonance penalty set and the auditory masking reward weight set to obtain a candidate vector acoustic perturbation index array in the multi-objective optimization method for low-altitude robot motors based on model predictive control according to embodiments of this application. Figure 5 This is a flowchart of step four of the multi-objective optimization method for low-altitude robot motors based on model predictive control according to an embodiment of this application. Detailed Implementation

[0010] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0011] Existing low-altitude robot motor drive schemes based on finite control set model predictive control only focus on electromagnetic torque tracking and switching losses in the cost function, failing to incorporate the acoustic disturbance of frequency converter current harmonics caused by voltage vector switching to the electromechanical structure into the evaluation system. This results in the algorithm being unable to detect whether candidate vectors will excite the mechanical natural frequency resonance of the stator and arm during the optimization process, thus generating a piercing high-frequency electromagnetic whistling sound during low-altitude flight in urban areas. To address this problem, this application proposes a multi-objective optimization method for low-altitude robot motors based on model predictive control. In each control cycle, this method first performs coordinate transformation on the three-phase current of the rotor motor to obtain the stator current feedback. Simultaneously, it collects propeller aerodynamic noise through an airborne microphone and analyzes the dynamic auditory masking frequency band characteristics using a psychoacoustic model. Then, based on the DC bus voltage and the real-time motor speed, it constructs a finite voltage vector set and deduces the predicted current trajectory corresponding to each candidate vector. Finally, it uses the offline-calibrated electromechanical structure resonance cost matrix to evaluate the predicted current trajectory of the rotor motor. The current trajectory is measured and a quadratic mapping is performed to quantify the fundamental resonance penalty. The critical band distance between the equivalent electromagnetic harmonic frequency of each vector excitation and the masking band is calculated to generate the masking reward weight. The two are then fused into a candidate vector acoustic disturbance index. Finally, this acoustic index, together with the torque flux error and switching action loss, constitutes a multi-objective cost function for rolling extremum optimization. The optimal switching vector with the minimum global cost is selected and output to the inverter gate drive circuit after dead-zone compensation. This achieves active avoidance and spectrum shaping of electromechanical structure resonance noise while maintaining high dynamic thrust response.

[0012] Figure 1 This is a flowchart of a multi-objective optimization method for low-altitude robot motors based on model predictive control, according to an embodiment of this application. Figure 2 The data flow diagram for the multi-objective optimization method of low-altitude robot motors based on model predictive control according to the embodiments of this application is as follows: Figure 1 and Figure 2As shown, the multi-objective optimization method for low-altitude robot motors based on model predictive control according to an embodiment of this application includes: Step 1, performing feature extraction and masking frequency band analysis on the collected operating state data to obtain the stator current feedback quantity in the rotating coordinate system and the dynamic auditory masking frequency band feature vector; Step 2, based on the DC bus voltage and the real-time rotational speed of the rotor motor, constructing a finite control set and predicting the forward state of the stator current feedback quantity in the rotating coordinate system to obtain the inverter finite voltage vector set and the predicted stator current trajectory set; Step 3, based on the preset electromechanical structure resonance cost matrix, performing analysis on the predicted stator current trajectory set and... Step 4: Based on the desired electromagnetic torque command and desired stator flux linkage command, perform multi-objective rolling optimization for active noise shaping on the predicted stator current trajectory set, the inverter finite voltage vector set, and the candidate vector acoustic excitation index array to obtain the optimal switching vector; Step 5: Perform anti-shoot-through dead-time compensation on the optimal switching vector through dead-time compensation to obtain the three-phase dead-time pulse modulation drive signal, which is output to the inverter gate drive circuit.

[0013] Specifically, in step one, feature extraction and masking frequency band analysis are performed on the collected operational status data to obtain the stator current feedback quantity in the rotating coordinate system and the dynamic auditory masking frequency band feature vector. It is worth noting that the operational status data includes the three-phase current sampling values ​​of the rotor motor, the rotor mechanical angle of the rotor motor, and the audio time series of the onboard microphone. It should be understood that when the low-altitude robot performs logistics delivery or inspection tasks in urban low-altitude airspace, its rotor motor's model predictive control system needs to perform multi-objective evaluation of the inverter candidate voltage vector within each microsecond-level control cycle. The prerequisite for this evaluation is obtaining the precise electromagnetic state of the current motor and the auditory masking conditions formed by the propeller aerodynamic noise under the current flight environment. The former determines whether the initial iteration base point of the predictive model is accurate, and the latter determines whether the acoustic disturbance evaluation module can correctly determine which frequency bands of electromagnetic noise can be ignored by the human ear. Therefore, in the technical solution of this application, feature extraction and masking frequency band analysis are performed on the collected operating state data to obtain the stator current feedback quantity in the rotating coordinate system and the dynamic auditory masking frequency band feature vector. This provides an accurate initial current state for subsequent finite control set construction and forward state prediction, and provides a real-time updated masking frequency band criterion for targeted quantification of acoustic disturbance characteristics of electromechanical structures. In this way, it can ensure that the model predictive control system has a high-precision electromagnetic state starting point in each control cycle to ensure the reliability of torque prediction, and also has an acoustic masking reference benchmark that dynamically changes with flight conditions to support the physical validity of subsequent noise shaping evaluation.

[0014] More specifically, in the embodiments of this application, step one includes: performing Clark and Park cascaded coordinate mapping on the sampled values ​​of the three-phase current of the rotor motor based on the rotor mechanical angle and the number of pole pairs to obtain the stator current feedback in the rotating coordinate system; performing aerodynamic noise fundamental frequency optimization on the audio time series of the airborne microphone to obtain the aerodynamic noise fundamental frequency component; and performing critical band scale conversion and masking threshold estimation on the aerodynamic noise fundamental frequency component based on a nonlinear psychoacoustic mapping model to obtain the dynamic auditory masking frequency band feature vector.

[0015] Specifically, in a particular example of this application, firstly, the rotor electrical angle is calculated based on the rotor mechanical angle and the inherent pole pair number of the rotor motor. Using this electrical angle, a Clark and Park cascaded coordinate transformation matrix is ​​constructed. The three-phase current sampling values ​​of the rotor motor are then subjected to a Clark transformation from a three-phase stationary coordinate system to a two-phase stationary coordinate system, and a Park transformation from a two-phase stationary coordinate system to a two-phase rotating coordinate system. This maps the three-phase time-varying AC current signal into direct-axis and quadrature-axis components that rotate synchronously with the rotor magnetic field, obtaining the stator current feedback in the rotating coordinate system. Next, a window function is applied to the airborne microphone audio time series to suppress spectral leakage, and a fast Fourier transform is performed. Within a preset propeller aerodynamic noise characteristic frequency band, a global extremum search is conducted on the obtained frequency domain power spectral density. The physical frequency point corresponding to the peak power spectral density is extracted as the fundamental frequency component of the aerodynamic noise. This fundamental frequency component characterizes the frequency anchor point of the main aerodynamic sound source determined by the current rotor speed and the number of blades. Then, based on a nonlinear psychoacoustic mapping model, the fundamental frequency component of aerodynamic noise is mapped and transformed from the Hertzian physical frequency scale to the Bark critical band scale. Based on the center position of this critical band, the simultaneous masking threshold curve of aerodynamic noise on adjacent frequencies is calculated. The upper and lower limits of the effective masking band, imperceptible to the human ear under this sound pressure level, and the corresponding masking threshold amplitudes are determined. These masking band lower and upper limits, along with the masking threshold amplitudes, are concatenated to generate a dynamic auditory masking band feature vector. Furthermore, the stator current feedback quantity in the rotating coordinate system, as a precise digital representation of the motor's current true electromagnetic state, is passed to the subsequent step two as the initial iteration base point for the forward state prediction of the finite control set. The dynamic auditory masking band feature vector, as a quantitative description of the real-time acoustic environment, is passed to the subsequent step three to determine whether the electromagnetic harmonics excited by each candidate voltage vector fall into a safe frequency range that can be masked by aerodynamic noise. Together, these two constitute the dual input basis of the multi-objective evaluation system in the electromagnetic and acoustic domains of this scheme.

[0016] Specifically, in step two, based on the DC bus voltage and the real-time rotational speed of the rotor motor, a finite control set is constructed and forward state prediction is performed on the stator current feedback in the rotating coordinate system to obtain a finite voltage vector set and a predicted stator current trajectory set for the inverter. It should be understood that, due to the complex and rapidly changing airflow disturbances faced by low-altitude robots flying in urban canyons or between building clusters, the flight controller requires the rotor motor to accurately track the desired torque command within each microsecond-level control cycle. The core mechanism of model predictive control lies in forward prediction of all applicable control actions and selection of the best option. This requires clearly defining the decision space composed of all discrete voltage vectors that the current inverter topology can output, and deriving the stator current response trajectory in the next control cycle after the application of each candidate vector based on the motor electromagnetic dynamic model. This provides a complete set of candidate schemes and their corresponding electromagnetic state prediction results for the subsequent evaluation of the multi-objective cost function. Therefore, in the technical solution of this application, based on the DC bus voltage and the real-time rotational speed of the rotor motor, a finite control set is constructed and forward state prediction is performed on the stator current feedback quantity in the rotating coordinate system to obtain a finite voltage vector set and a predicted stator current trajectory set for the inverter. This exhaustively enumerates all physically feasible voltage vectors that the inverter can output under the current bus power supply conditions and deduces their electromagnetic consequences one by one. In this way, complete candidate control action domains and their corresponding current response prediction information can be provided for the subsequent quantitative evaluation of acoustic disturbance characteristics in step three and the multi-objective rolling optimization in step four, ensuring that the multi-objective evaluation is carried out without omission in the entire decision space.

[0017] More specifically, in the embodiments of this application, step two includes: performing space vector enumeration on the two-level three-phase inverter topology based on the amplitude of the DC bus voltage to obtain a finite voltage vector set of the inverter; calculating the electric angular velocity based on the real-time speed and number of pole pairs of the rotor motor, and combining the stator resistance and quadrature-direct axis inductance parameters, performing forward Euler discretization derivation on the continuous-time state-space equation to obtain the discrete parameter matrix of the motor prediction model; and based on the discrete parameter matrix of the motor prediction model, using the stator current feedback in the rotating coordinate system as the initial iteration base point, performing forward Euler stator current trajectory multiple derivation on each candidate vector in the finite voltage vector set of the inverter to obtain a predicted stator current trajectory set.

[0018] Accordingly, based on the amplitude of the DC bus voltage, spatial vector enumeration is performed on the two-level three-phase inverter topology to obtain a finite set of inverter voltage vectors. It should be understood that since the rotor motor drive system of the low-altitude robot uses a two-level three-phase voltage-source inverter as its power conversion topology, the three-phase bridge arms of this inverter can only output a finite number of discrete voltage vector combinations at any given time. The optimization mechanism of model predictive control requires an exhaustive evaluation of all physically feasible control actions in each control cycle. Therefore, it is necessary to first determine all discrete voltage vectors that the inverter can output under the current DC bus power supply conditions, along with their amplitudes and spatial orientations, to constitute a complete decision domain for candidate control actions. Therefore, in the technical solution of this application, spatial vector enumeration is further performed on the two-level three-phase inverter topology based on the amplitude of the DC bus voltage to obtain a finite set of inverter voltage vectors, thereby establishing a complete candidate action space for model predictive control under the current power supply state. This ensures that subsequent forward state prediction and multi-objective optimization are evaluated without omission on all physically feasible switch combinations, avoiding the omission of potential optimal control actions due to an incomplete candidate set.

[0019] In a specific example of this application, the construction process of the inverter's finite voltage vector set is as follows: First, based on the topology of a two-level three-phase inverter, the upper and lower transistors of each phase arm are complementary in conduction, and each phase has two independent switching states: on and off. The three-phase combination generates eight different combinations of arm switching states. Six of these combinations correspond to six non-zero effective voltage vectors with different spatial orientations, while the remaining two combinations, where all three upper transistors are either on or off, correspond to two zero vectors with zero amplitude. Next, the magnitude of each non-zero voltage vector is determined based on the amplitude of the acquired DC bus voltage. The six non-zero vectors are uniformly distributed in the complex plane space, and the spatial phase interval between adjacent vectors is fixed at sixty degrees. The amplitude of each vector is proportional to the DC bus voltage. Then, the six non-zero voltage vectors and the two zero vectors are indexed together to form a finite voltage vector set for the inverter containing eight discrete candidate elements. This set completely covers all voltage vector schemes that the current inverter topology can output under physical constraints, providing a deterministic input decision domain for subsequent forward Euler stator current trajectory deduction of each candidate vector based on the discrete parameter matrix of the motor prediction model.

[0020] Accordingly, based on the real-time rotational speed and pole pair number of the rotor motor, the electric angular velocity is calculated. Combined with the stator resistance and quadrature-direct axis inductance parameters, a forward Euler discretization derivation of the continuous-time state-space equations is performed to obtain the discrete parameter matrix of the motor predictive model. It should be understood that since the forward state prediction of model predictive control relies on the mathematical model of the motor's electromagnetic dynamic behavior, and the voltage-current relationship of the permanent magnet synchronous motor in the rotating coordinate system is described by continuous-time differential equations, the microcontroller cannot directly solve the continuous differential equations online. It must be transformed into a discrete difference form suitable for cycle-by-cycle recursive calculation by a digital processor. Furthermore, the parameters in the discrete model need to be dynamically updated according to the real-time rotational speed of the rotor motor to reflect the changes in the electromagnetic coupling strength between the quadrature and direct axes under the current flight conditions. Therefore, in the technical solution of this application, the electric angular velocity is further calculated based on the real-time speed and pole pair number of the rotor motor. Combined with the stator resistance and quadrature-axis inductance parameters, the continuous-time state-space equations are derived using forward Euler discretization to obtain the discrete parameter matrix of the motor prediction model. This transforms the continuous electromagnetic dynamic equations of the motor into a discrete recursive format that can be executed online by the microcontroller and adapts to the current speed condition in real time. This ensures that the prediction model used for current trajectory extrapolation of each candidate vector in the finite voltage vector set of the inverter accurately matches the current physical operating state of the motor, avoiding deviations in the predicted current from the true response due to model parameters lagging behind actual speed changes.

[0021] In a specific example of this application, the derivation process of the discrete parameter matrix of the motor prediction model is as follows: First, the rotor electric angular velocity is calculated by multiplying the real-time speed of the rotor motor by the inherent number of pole pairs of the motor. This electric angular velocity characterizes the rotational speed of the rotor magnetic field in the electrical space and determines the electromagnetic cross-coupling strength between the direct axis and the quadrature axis in the rotating coordinate system. Next, based on the calculated electric angular velocity, combined with the stator resistance, direct-axis inductance, and quadrature-axis inductance parameters specified on the motor nameplate, as well as the rotor permanent magnet flux linkage constant, a continuous-time state-space equation for the permanent magnet synchronous motor in the rotating coordinate system is established. This equation uses the direct-axis current and quadrature-axis current as state variables and the direct-axis voltage and quadrature-axis voltage as input variables, describing the transient evolution law of the stator current under the coupling effect of applied voltage excitation and back electromotive force. Then, using the forward Euler discretization method, a first-order time-domain truncation approximation is performed on the continuous-time state-space equation with the control period as the discrete step length. This yields the discrete parameter matrix of the motor prediction model, which consists of three parts: the system state transition matrix, the input control matrix, and the back EMF compensation vector. Its discrete expression is as follows:

[0022] in, The stator current feedback quantity in the rotating coordinate system for the current control cycle includes the direct-axis current component and the quadrature-axis current component; This is the predicted stator current state vector for the next control cycle; The input vector is the voltage in the rotating coordinate system currently applied; The resistance of each phase winding of the stator; It is a direct-axis synchronous inductor; It is a quadrature axis synchronous inductor; The rotor electric angular velocity is calculated based on the real-time rotational speed and number of pole pairs of the rotor motor; For a single control cycle in which the microcontroller performs model predictive control; The permanent magnet flux linkage constant generated by the rotor permanent magnet; This is the system state transition matrix; For input control matrix; This is the back electromotive force compensation vector.

[0023] Furthermore, from the physical meaning of the above discrete expressions, the system state transition matrix... Diagonal elements and This reflects the attenuation effect of stator resistance on current; off-diagonal elements. and This reflects the electromagnetic cross-coupling strength between the direct and quadrature axes of the rotor motor at the current speed, and this coupling strength is updated in real time as the rotor motor's real-time speed changes; input control matrix This describes the gain effect of the voltage applied by the inverter on the stator current. Its elements are inversely proportional to the inductance parameter, indicating that the smaller the inductance, the more sensitive the current response to voltage; back EMF compensation vector. In The term characterizes the continuous disturbance effect of the back electromotive force generated by the rotation of the rotor permanent magnet on the quadrature-axis current. The absolute value of this term increases significantly during high-speed cruising of the low-altitude robot. The discrete parameter matrix of the motor prediction model encapsulates the above three components, providing a complete discrete recursive calculation framework for subsequent forward Euler stator current trajectory multiple extrapolation based on the stator current feedback in the rotating coordinate system as the initial iteration base point for each candidate vector in the finite voltage vector set of the inverter.

[0024] Accordingly, based on the discrete parameter matrix of the motor prediction model, and using the stator current feedback in the rotating coordinate system as the initial iteration base point, multiple forward Euler stator current trajectory extrapolations are performed on each candidate vector in the finite voltage vector set of the inverter to obtain the predicted stator current trajectory set. It should be understood that since the core decision-making logic of model predictive control lies in predicting and selecting the optimal system state after applying all candidate control actions, and the finite voltage vector set of the inverter contains eight discrete candidate voltage vectors, each of which, when applied, will cause the stator current to evolve along different directions and amplitudes, the system must calculate the current response results corresponding to each candidate vector one by one to provide a complete set of evaluation objects for subsequent acoustic disturbance assessment and torque flux linkage error calculation. Therefore, in the technical solution of this application, further based on the discrete parameter matrix of the motor prediction model, and using the stator current feedback in the rotating coordinate system as the initial iteration base point, multiple forward Euler stator current trajectory extrapolations are performed on each candidate vector in the finite voltage vector set of the inverter to obtain the predicted stator current trajectory set, thereby exhaustively calculating the complete response state of the motor stator current in the next control cycle after applying each candidate voltage vector. In this way, it can provide current prediction results of all candidate schemes for the subsequent step three, which is based on the pre-set electromechanical structure resonance cost matrix for targeted quantification of acoustic disturbance characteristics, and step four, which is based on electromagnetic torque prediction and stator flux error calculation, ensuring that the multi-objective evaluation covers the entire physical action space of the inverter.

[0025] In a specific example of this application, firstly, the rotating coordinate system stator current feedback quantity output in step one is used as a precise digital representation of the motor's electromagnetic state in the current control cycle, and established as the initial iteration base point for the forward Euler recursive calculation. This base point includes the direct-axis current component and the quadrature-axis current component at the current moment. Next, each candidate voltage vector is sequentially extracted from the finite voltage vector set of the inverter, and its projection transformation from the stationary coordinate system to the direct-axis-quadrature-axis coordinate system that rotates synchronously with the rotor magnetic field is performed to obtain the direct-axis voltage component and the quadrature-axis voltage component of the candidate vector in the rotating coordinate system. Then, the rotating coordinate system stator current feedback quantity and the rotating coordinate system voltage component of the candidate vector are substituted into the discrete state equation defined by the discrete parameter matrix of the motor prediction model to perform a forward Euler recursive operation, calculating the direct-axis predicted current and the quadrature-axis predicted current in the next control cycle after the application of the candidate vector, forming a set of current response vectors corresponding to the candidate vector. Furthermore, the above recursive calculation process is repeated for all eight candidate vectors in the finite voltage vector set of the inverter. The eight sets of current response vectors are aggregated and arranged according to the candidate vector index order to form a set of predicted stator current trajectories. This set completely describes the panoramic view of the current evolution caused by all possible switching actions of the inverter under the current electromagnetic state. It provides a unified and complete current prediction data foundation for the subsequent step three of targeted quantization of electromechanical acoustic disturbance characteristics of each candidate vector and the step four of electromagnetic torque prediction and stator flux linkage error calculation of each candidate vector.

[0026] Specifically, in step three, based on the pre-set electromechanical structure resonance cost matrix, the electromechanical structure acoustic excitation feature targeted quantization is performed on the predicted stator current trajectory set and the dynamic auditory masking frequency band feature vector to obtain a candidate vector acoustic excitation index array. It should be understood that when a low-altitude robot flies in urban low-altitude airspace, the discrete voltage vector selected by the finite control set model in each control cycle will generate frequency conversion current harmonics. When the frequencies of these harmonics fall near the mechanical natural frequencies of the motor stator core and the arm connection structure, they will excite structural resonance and produce a piercing electromagnetic howl. The cost function of existing schemes completely lacks the ability to evaluate this electromechanical acoustic coupling effect. Therefore, before multi-objective optimization, it is necessary to independently and quantitatively assess the degree of acoustic excitation that each candidate voltage vector may cause. Simultaneously, the auditory masking effect of propeller aerodynamic noise on adjacent frequency band electromagnetic noise is used to distinguish which harmonic excitations will be perceived by the human ear and which can be naturally masked without penalty. Therefore, in the technical solution of this application, based on the pre-set electromechanical structure resonance cost matrix, the electromechanical structure acoustic excitation feature targeted quantization is performed on the predicted stator current trajectory set and the dynamic auditory masking frequency band feature vector to obtain a candidate vector acoustic excitation index array. This generates a comprehensive acoustic evaluation index that integrates structural resonance risk and auditory masking advantage for each candidate voltage vector. In this way, the subsequent multi-objective rolling optimization module can have acoustic dimension discrimination capability when evaluating each candidate vector, actively avoid high noise vectors that trigger structural resonance, and prioritize low-perception vectors whose electromagnetic harmonics fall into the masking frequency region, thereby achieving active shaping of motor drive noise without sacrificing torque tracking performance.

[0027] Figure 3 This is a flowchart of step three of the multi-objective optimization method for low-altitude robot motors based on model predictive control according to an embodiment of this application. Figure 3 As shown, step three includes: S310, based on the preset electromechanical structure resonance cost matrix, performing a basic mechanical resonance penalty evaluation on each current response vector in the predicted stator current trajectory set to obtain a candidate vector basic resonance penalty set; S320, based on the masking frequency band boundary in the dynamic auditory masking frequency band feature vector, performing critical frequency band distance calculation and reward mapping on the equivalent electromagnetic harmonic frequencies derived from the predicted stator current trajectory set to obtain an auditory masking reward weight set; S330, performing acoustic excitation feature targeted fusion on the candidate vector basic resonance penalty set and the auditory masking reward weight set to obtain a candidate vector acoustic excitation index array.

[0028] Accordingly, in step S310, based on the preset electromechanical structure resonance cost matrix, a basic mechanical resonance penalty evaluation is performed on each current response vector in the predicted stator current trajectory set to obtain a candidate vector basic resonance penalty set. It should be understood that, since the stator core of the rotor motor of the low-altitude robot and its connection structure with the arm have specific natural mechanical frequencies, when the discrete voltage vector applied by the inverter causes the high-frequency harmonic components generated by the stator current to fall near these natural frequencies, severe structural resonance will be excited and high-sound-pressure level electromagnetic howling noise will radiate. Since the current response directions and amplitudes generated after applying different candidate voltage vectors vary, their excitation degrees to structural resonance are also drastically different. Therefore, it is necessary to independently and quantitatively assess the structural excitation risk of each candidate vector before multi-objective evaluation. Therefore, in the technical solution of this application, a basic mechanical resonance penalty evaluation is further performed on each current response vector in the predicted stator current trajectory set based on the preset electromechanical structure resonance cost matrix to obtain a candidate vector basic resonance penalty set, thereby quantifying the magnitude of the resonance excitation energy generated by each candidate voltage vector on the motor body structure. In this way, it can provide basic penalty data to characterize the structural resonance risk for subsequent targeted fusion of acoustic excitation features, enabling the system to have the quantitative discrimination ability to distinguish between high resonance risk vectors and low resonance risk vectors.

[0029] More specifically, in a specific example of this application, the generation process of the candidate vector fundamental resonance penalty set is as follows. First, a pre-set electromechanical structure resonance cost matrix is ​​loaded. This matrix is ​​a second-order symmetric positive definite matrix, generated offline through finite element modal analysis of the motor stator core and the arm connection structure before the product leaves the factory. Its eigenvalues ​​characterize the resonance sensitivity of each mechanical natural mode, and the eigenvectors characterize the current disturbance sensitive direction corresponding to each mode. A large penalty weight is assigned to the current distortion component falling into the direction of the structure's mechanical natural frequency. Next, the current response vector corresponding to each candidate voltage vector is extracted sequentially from the predicted stator current trajectory set. This vector contains the direct-axis predicted current component and the quadrature-axis predicted current component, characterizing the expected evolution state of the motor stator current in the rotating coordinate system after the candidate vector is applied. Then, the extracted current response vector is subjected to a quadratic form mapping operation with the pre-set electromechanical structure resonance cost matrix to calculate a scalar value as the fundamental mechanical resonance penalty value of the candidate vector. The calculation expression is as follows:

[0030] in, For the first The fundamental mechanical resonance penalty scalar value corresponding to each candidate voltage vector; To predict the first stator current trajectory in the set A current response vector, comprising direct-axis predicted current components and quadrature-axis predicted current components; This is the transpose of the current response vector; This is the pre-set electromechanical structure resonance cost matrix. From the physical meaning of this quadratic form operation, Partially, the projection transformation of the predicted current response vector onto the resonant sensitive ellipsoid defined by the pre-set electromechanical structure resonant cost matrix is ​​implemented, mapping the current vector from a uniform space to a non-uniform space weighted by structural modal sensitivity; then, it is right-multiplied... This completes the scalar aggregation of projected energy. When the direction of the predicted current response generated after applying a candidate vector is exactly aligned with the principal direction of the large eigenvalue of the pre-set electromechanical structure resonance cost matrix, the quadratic mapping result will be amplified sharply, indicating that the vector is very likely to excite mechanical resonance in the stator structure. Conversely, when the direction of the predicted current response is aligned with the direction of the small eigenvalue, the penalty value is smaller, indicating that the vector poses a lower risk of acoustic excitation to the structure. Furthermore, the above quadratic mapping calculation is repeated for all eight candidate vectors in the inverter's finite voltage vector set. The eight resonance penalty scalar values ​​are aggregated and arranged in index order to form the candidate vector basic resonance penalty set. This set is passed to the subsequent acoustic excitation feature targeted fusion stage and the auditory masking reward weight set for offset calculation.

[0031] Accordingly, in step S320, based on the masking frequency band boundary in the dynamic auditory masking frequency band feature vector, the critical frequency band distance is calculated and reward mapping is performed on the equivalent electromagnetic harmonic frequencies derived from the predicted stator current trajectory set to obtain the auditory masking reward weight set. It should be understood that since the aerodynamic broadband noise generated by the propeller rotation during the flight of the low-altitude robot is objectively present and cannot be eliminated, and psychoacoustic research shows that when a strong sound source is present, the human ear's perception threshold for weaker sounds with similar frequencies will be significantly increased. This means that if the electromagnetic harmonic frequency excited by the inverter switching action happens to fall within the masking frequency band covered by the propeller aerodynamic noise, the electromagnetic noise will be naturally ignored by the human ear without additional penalty. Conversely, if the electromagnetic harmonic frequency is far from the masking frequency band, it will be clearly perceived by the human ear. Therefore, it is necessary to quantitatively evaluate the frequency proximity between the electromagnetic harmonics generated by each candidate vector and the masking frequency band and allocate differentiated masking rewards accordingly. Therefore, in the technical solution of this application, based on the masking frequency band boundary in the dynamic auditory masking frequency band feature vector, the critical frequency band distance is calculated and reward mapping is performed on the equivalent electromagnetic harmonic frequencies derived from the predicted stator current trajectory set to obtain the auditory masking reward weight set. This is used to quantify the effectiveness of the electromagnetic noise generated by each candidate voltage vector being masked by the propeller aerodynamic noise. In this way, reward data characterizing the masking protection capability can be provided for subsequent acoustic disturbance feature targeted fusion, enabling the system to distinguish between low-perceived noise that can be naturally masked and high-perceived noise that cannot be masked when evaluating the acoustic cost of each candidate vector.

[0032] More specifically, in a concrete example of this application, firstly, the current response vector corresponding to each candidate voltage vector is extracted sequentially from the predicted stator current trajectory set. The transient difference between this current response vector and the stator current feedback in the rotating coordinate system is calculated. Based on the ratio of the amplitude of this difference to the control period, the center frequency of the equivalent electromagnetic harmonic excited by the switching action of the candidate vector in the frequency domain is estimated. This frequency reflects the main energy concentration location of the current ripple caused by the inverter state switching in the frequency spectrum. Next, the estimated center frequency of the equivalent electromagnetic harmonic is transformed to the critical band scale through a nonlinear psychoacoustic mapping to obtain the position coordinates of the harmonic at the critical band scale. At the same time, the lower boundary and upper boundary of the masking band are extracted from the dynamic auditory masking band feature vector. Then, the unidirectional distance between the position coordinates of the equivalent electromagnetic harmonic at the critical band scale and the upper and lower boundaries of the masking band is calculated. The calculation expression is as follows:

[0033] in, For the first The equivalent electromagnetic harmonics excited by each candidate vector extend beyond the masking band boundary by a unidirectional distance at the critical frequency band scale. For the first The position coordinates of the equivalent electromagnetic harmonic frequencies excited by each candidate vector on the critical frequency band scale after nonlinear mapping; This represents the lower boundary of the masking frequency band in the dynamic auditory masking frequency band feature vector; This represents the upper boundary of the masking frequency band in the dynamic auditory masking frequency band feature vector. From the physical meaning of this distance calculation, It partially measures the amount by which the equivalent harmonic frequency escapes below the lower limit of the masking band. Partially, it measures the offset of the equivalent harmonic frequency escaping in a direction higher than the upper limit of the shielding frequency band; the sum of the two is the comprehensive escape distance for the harmonic to completely escape the shielding protection area. fall into and Both terms are zero within the defined closed interval, indicating that the electromagnetic harmonic is completely within the shielding and protection range of propeller aerodynamic noise.

[0034] Furthermore, based on the calculated one-way distance and combined with the masking threshold amplitude in the dynamic auditory masking frequency band feature vector, the distance is mapped to a normalized reward coefficient using a negative exponential decay function. The calculation expression is as follows:

[0035] in, For the first Auditory masking reward weight coefficients for candidate vectors; The masking threshold amplitude in the dynamic auditory masking frequency band feature vector represents the upper limit of the masking ability of aerodynamic noise on adjacent frequency bands under the current propeller speed and sound pressure level conditions. To penalize the steepness adjustment factor, which controls the rate of reward decay after exceeding the masking frequency band, the larger the value, the steeper the transition between the masked and unprotected areas. This refers to the one-way distance calculated above. From the physical meaning of this reward mapping, As a multiplicative factor, the absolute magnitude of the reward is anchored to the actual masking capability of the current aerodynamic noise, ensuring that the masking reward does not exceed the physically achievable masking limit; Partially acting as a decay factor, it achieves a non-linear mapping from distance to reward. When the exponential function outputs one when the value is zero, the reward reaches its maximum value. This indicates that the electromagnetic noise of this vector can be completely masked by aerodynamic noise; when As the exponential function output increases, it approaches zero according to a negative exponential law, indicating that the electromagnetic noise of the vector has escaped the shielded area and will be clearly perceived by the human ear. The above distance calculation and reward mapping process is repeated for all eight candidate vectors. The eight reward coefficients are aggregated and arranged in index order to form an auditory masking reward weight set. This set is passed to the subsequent acoustic disturbance feature targeted fusion stage for offset calculation with the candidate vector fundamental resonance penalty set.

[0036] Accordingly, in step S330, the candidate vector fundamental resonance penalty set and the auditory masking reward weight set are subjected to targeted fusion of acoustic excitation features to obtain a candidate vector acoustic excitation index array. It should be understood that since the candidate vector fundamental resonance penalty set characterizes the degree of resonance excitation risk of each candidate voltage vector to the motor structure, while the auditory masking reward weight set characterizes the effectiveness of the electromagnetic noise generated by each candidate vector being masked by propeller aerodynamic noise, both evaluate the acoustic characteristics of the same candidate vector from two independent dimensions: structural vibration hazard and human auditory visibility. The evaluation results of these two dimensions need to be fused into a unified comprehensive acoustic index before it can be included as an independent evaluation item in the subsequent multi-objective cost function for rolling optimization decision-making. Therefore, in the technical solution of this application, the candidate vector fundamental resonance penalty set and the auditory masking reward weight set are further subjected to targeted fusion of acoustic excitation features to obtain a candidate vector acoustic excitation index array, thereby integrating the structural resonance risk and auditory masking protection capability into a single acoustic evaluation dimension for multi-objective optimization. This allows the subsequent multi-objective rolling optimization module to directly read a comprehensive acoustic value that integrates physical disturbance risk and psychoacoustic perception characteristics when traversing each candidate vector, without having to process two independent sets separately. This reduces the computational complexity of the optimization process while ensuring the integrity of the evaluation.

[0037] More specifically, in a concrete example of this application, the generation process of the candidate vector acoustic perturbation index array is as follows. First, each resonance penalty scalar value in the candidate vector fundamental resonance penalty set is paired element-wise with the masking reward coefficient at the corresponding index position in the auditory masking reward weight set to establish a one-to-one fusion calculation relationship. Next, element-wise division is used as the fusion operator. Each penalty value in the candidate vector fundamental resonance penalty set is used as the numerator, and the corresponding masking reward coefficient in the auditory masking reward weight set is used as the denominator for division. A very small positive bias constant is added to the denominator to ensure that the microcontroller's floating-point operation does not experience division-by-zero overflow when the masking reward approaches zero. The calculation expression is as follows:

[0038] in, For the first Acoustic perturbation feature index of candidate vectors after targeted fusion; For the candidate vector fundamental resonance penalty set, the first A resonance penalty value characterizes the excitation energy of the vector on the motor structure; For auditory masking reward weighting concentration A masking reward coefficient characterizes the degree to which the electromagnetic noise of the vector is effectively masked by aerodynamic noise; This is a constant biased for extremely small positive numbers to prevent overflow during division by zero in floating-point operations. From the physical meaning of this fusion operation, the molecule... The denominator represents the degree of structural resonance hazard of the candidate vector. When the resonant penalty of a candidate vector is large, but its electromagnetic harmonics fall within the masking frequency band, resulting in a large masking reward, the division operation is effectively suppressed, and the acoustic excitation index remains at a low level. This indicates that although the vector excites a certain degree of structural vibration, the noise it generates can be naturally masked by the propeller's aerodynamic noise and is not perceived by city residents. When the resonant penalty of a candidate vector is large, and its electromagnetic harmonic frequency is far from the masking frequency band, causing the masking reward to approach zero, the division operation result is sharply amplified, and the acoustic excitation index reaches a maximum value. This indicates that the vector not only excites severe structural resonance, but its noise cannot be protected by aerodynamic masking and will be clearly perceived by the human ear. Then, the above element-by-element division fusion calculation is repeated for all eight candidate vectors, and the eight acoustic excitation characteristic indices are aggregated and arranged in index order to form a candidate vector acoustic excitation index array.

[0039] Specifically, when low-altitude robots perform urban patrol or logistics delivery tasks, their rotor motor model predictive control system needs to complete the targeted quantitative evaluation of acoustic excitation characteristics within a microsecond-level control cycle. In the first embodiment, sub-step S330 uses a simple element-wise division to fuse the candidate vector fundamental resonance penalty set with the auditory masking reward weight set. This implies a key physical assumption: the masking effectiveness is independent of the resonance severity, that is, no matter how large the resonance excitation energy is, as long as the masking reward weight is high enough, the penalty value can be suppressed proportionally.

[0040] However, from the physical perspective of electroacoustics, this assumption carries a serious risk of failure under extreme conditions. When a low-altitude robot encounters a sudden crosswind in an urban canyon, the flight controller requests extremely high torque compensation. At this point, the resonant energy of the stator structure excited by certain candidate voltage vectors will far exceed the normal cruise level. Once the resonant excitation energy exceeds the linear elastic vibration threshold of the connection structure between the motor stator core and the arm, the stator structure will enter a nonlinear vibration range. Within this range, the magnetostrictive saturation effect of the material and the collision effect of the mechanical connection gaps will generate a large number of high-order harmonic distortion components. The frequencies of these distortion components will far exceed the narrow band masking range covered by the original aerodynamic noise. Even if the masking conditions at the fundamental frequency are good, the overflowing high-order harmonics will still be clearly perceived by the human ear.

[0041] The first embodiment completely ignores the nonlinear saturation coupling relationship between the resonance amplitude and the masking effectiveness. Under extreme wind conditions, the algorithm still incorrectly assigns too low an acoustic cost score to candidate vectors that generate large structural disturbances, causing the multi-objective optimization module to select these vectors, resulting in the local failure of the acoustic stealth strategy and causing the low-altitude robot to produce a piercing high-frequency howl over urban residential areas.

[0042] To address the aforementioned shortcomings, a second embodiment is proposed. Specifically, in another specific example of this application, Figure 4 This is a flowchart illustrating the process of performing acoustic perturbation feature-targeted fusion of the candidate vector fundamental resonance penalty set and the auditory masking reward weight set to obtain a candidate vector acoustic perturbation index array, according to the multi-objective optimization method for low-altitude robot motors based on model predictive control according to embodiments of this application. Figure 4As shown, step S330 includes: S331, based on a preset stator structure linear elastic vibration threshold, quantifying the resonance overflow degree of each element in the candidate vector fundamental resonance penalty set to obtain a candidate vector resonance overflow saturation factor array; S332, dynamically attenuating the masking trust degree of each reward coefficient in the auditory masking reward weight set and the corresponding factor in the candidate vector resonance overflow saturation factor array to obtain a masking effective weight set after saturation correction; S333, performing saturation-sensory acoustic excitation targeted fusion of the candidate vector fundamental resonance penalty set and the masking effective weight set after saturation correction to obtain a candidate vector acoustic excitation index array.

[0043] More specifically, in step S331, based on a preset stator structure linear elastic vibration threshold, the resonance overflow degree of each element in the candidate vector fundamental resonance penalty set is quantified to obtain a candidate vector resonance overflow saturation factor array. It is understandable that, due to the clear physical boundary point of the rotor motor stator structure of the low-altitude robot, when the resonant excitation energy is below this boundary point, the structure is in the linear elastic vibration range, and the electromagnetic noise spectrum generated is concentrated near the fundamental frequency, which can be effectively masked by the propeller aerodynamic broadband noise. When the resonant excitation energy exceeds this boundary point, the structure enters the nonlinear distortion range. The magnetostriction saturation effect of the material and the collision effect of the mechanical connection gap will generate a large number of high-order harmonic distortion components. The frequency of these distortion components will far exceed the narrow band masking range covered by the original aerodynamic noise. Even if the masking conditions at the fundamental frequency are good, the overflowing high-order harmonics can still be clearly perceived by the human ear. However, when the first embodiment mentioned above uses a simple element-wise division to fuse the candidate vector fundamental resonance penalty set with the auditory masking reward weight set, it completely ignores this nonlinear saturation coupling relationship between the resonance amplitude and the masking effectiveness. Under extreme wind conditions, the algorithm still incorrectly gives those candidate vectors that generate large structural excitations too low an acoustic cost score. Therefore, in the technical solution of this application, based on a preset linear elastic vibration threshold of the stator structure, the resonance overflow degree of each element in the candidate vector fundamental resonance penalty set is quantified to obtain a candidate vector resonance overflow saturation factor array. This allows for real-time online determination of whether each penalty scalar value in the candidate vector fundamental resonance penalty set is about to break through the linear limit of the structure. This provides a continuous quantization factor characterizing whether each candidate vector resonance excitation has approached or broken through the linear elastic boundary of the stator structure for subsequent dynamic decay correction of masking confidence. This enables the system to adaptively sense the nonlinear distortion region of the structure under extreme conditions, fundamentally eliminating the problem of local failure of the acoustic stealth strategy caused by blindly trusting masking.

[0044] In a specific example of this application, the generation process of the candidate vector resonance overflow saturation factor array is as follows. First, a preset stator structure linear elastic vibration threshold is loaded. This threshold, calibrated through offline finite element modal analysis, represents the upper limit of the maximum resonant excitation energy that the motor stator can withstand without generating significant high-order harmonic distortion. When a low-altitude robot encounters a sudden crosswind in an urban canyon, the flight controller requests a large torque compensation. At this time, the stator structure resonance energy excited by some candidate voltage vectors will far exceed the normal cruise level. Once the resonance excitation energy exceeds this threshold, the stator structure will enter the nonlinear vibration range, generating a large number of high-order harmonic distortion components escaping the masking frequency band. Next, the resonance penalty scalar value corresponding to each candidate vector is extracted sequentially from the candidate vector basic resonance penalty set and compared with the preset stator structure linear elastic vibration threshold to calculate the normalized exceedance degree of the resonance penalty value relative to the threshold. Then, using the soft saturation mapping characteristics of the hyperbolic tangent function, the exceedance degree of the resonance penalty value of each candidate vector relative to the threshold is transformed into a continuous saturation factor between zero and one, the calculation expression of which is:

[0045] in, For the first The resonant overflow saturation sensing factor corresponding to each candidate voltage vector takes values ​​in a closed interval from zero to one. For the candidate vector fundamental resonance penalty set, the first Each element represents the structural resonance energy excited by the vector. The preset linear elastic vibration threshold of the stator structure is calibrated through offline finite element modal analysis; This is a saturation transition steepness adjustment factor, which controls the width of the transition band from the linear region to the nonlinear region. The larger the value, the steeper the transition. This is an array of candidate vector resonant overflow saturation factors, containing eight saturation factors corresponding one-to-one with the eight candidate voltage vectors. From the physical meaning of this formula, Partial calculation of the candidate vector fundamental resonance penalty set. The normalized offset of each resonance penalty value relative to the preset linear elastic vibration threshold of the stator structure is zero when the resonance penalty value is equal to the threshold, negative when it is below the threshold, and positive when it exceeds the threshold. Part of the normalized offset is scaled by a saturation transition steepness adjustment factor to control the response sensitivity of the hyperbolic tangent function near the threshold. Part of the scaled offset is mapped to the interval from negative one to positive one using the hyperbolic tangent function, achieving a smooth and continuous transition from the linear region to the nonlinear region; Part of the linear transformation maps the output of the hyperbolic tangent function from the interval of negative one to positive one to the interval of one to zero, so that the final saturation factor approaches one when the structure is safe and approaches zero when the structure is dangerous.

[0046] It should be noted that when the resonance penalty value is much lower than the preset linear elastic vibration threshold of the stator structure, the output of the hyperbolic tangent function approaches negative one, and the saturation factor approaches one after linear transformation, indicating that the structure is in a safe linear vibration region and the masking mechanism is completely reliable. When the resonance penalty value significantly exceeds the preset linear elastic vibration threshold of the stator structure, the output of the hyperbolic tangent function approaches positive one, and the saturation factor approaches zero, indicating that the structure has entered the nonlinear distortion region and the masking mechanism tends to fail. The smooth and continuous characteristics of the hyperbolic tangent function ensure that the saturation factor does not produce a step change in the transition range, avoiding numerical chattering interference in subsequent multi-objective optimization.

[0047] Furthermore, under normal cruise conditions of the low-altitude robot, the resonant excitation of each candidate vector is within the linear range of the structure, and each element in the candidate vector resonant overflow saturation factor array is close to one. The subsequent dynamic decay correction of the masking confidence does almost not change the original value of the auditory masking reward weight set, and the system behavior is almost the same as that of the first embodiment, without introducing additional performance loss. However, under extreme wind resistance or rapid maneuver conditions, when the resonant excitation of some candidate vectors approaches or exceeds the preset linear elastic vibration threshold of the stator structure, the corresponding elements in the candidate vector resonant overflow saturation factor array rapidly decay to near zero, providing a physical basis for the forced compression of the subsequent masking weights. The candidate vector resonant overflow saturation factor array is passed to the next stage and coupled with the auditory masking reward weight set for element-wise multiplication and correction.

[0048] More specifically, in step S332, the masking confidence is dynamically attenuated and corrected for each reward coefficient in the auditory masking reward weight set and the corresponding factor in the candidate vector resonance overflow saturation factor array to obtain the masking effective weight set after saturation correction. It should be understood that after obtaining the candidate vector resonance overflow saturation factor array, it needs to be applied to the auditory masking reward weight set to achieve a physically adaptive contraction of the masking confidence. In situations where low-altitude robots encounter extreme gusts and require high torque output, some candidate vectors, although their fundamental frequency noise falls within the propeller aerodynamic masking zone (i.e., their auditory masking reward weight is large), may experience structural nonlinear overflow due to excessive excitation. The system needs to automatically revoke the masking exemption for these vectors; otherwise, subsequent fusion calculations will still incorrectly suppress the acoustic excitation index of these vectors due to their large masking reward weight, causing the multi-objective optimization module to select these dangerous vectors that have actually generated high-order harmonic escape. Therefore, in the technical solution of this application, the masking confidence level is further dynamically attenuated and corrected for each reward coefficient in the auditory masking reward weight set and the corresponding factor in the candidate vector resonance overflow saturation factor array to obtain the masking effective weight set after saturation correction. This allows the confidence level of the masking protection capability to be adaptively reduced according to the degree of structural resonance overflow of each candidate vector. In this way, it can be ensured that when the resonance excitation of a candidate vector approaches or exceeds the linear elastic boundary of the stator structure, the system actively revokes the masking exemption right originally enjoyed by the vector, so that the subsequent fusion calculation applies a penalty intensity that matches the actual acoustic risk.

[0049] In a specific example of this application, the process of generating the effective masking weight set after saturation correction is as follows: First, the original masking reward coefficients corresponding to each candidate vector are extracted sequentially from the auditory masking reward weight set, and the saturation factors at the corresponding index positions are extracted from the candidate vector resonance overflow saturation factor array, establishing a one-to-one correction calculation relationship. Next, each masking reward coefficient in the auditory masking reward weight set is multiplied element-wise with the saturation factor at the corresponding index in the candidate vector resonance overflow saturation factor array to generate the effective masking weight set after saturation correction, the calculation expression of which is:

[0050] in, For the first The actual effective masking weights of the candidate vectors after saturation correction; For auditory masking reward weighting concentration One original cover reward coefficient; The first in the candidate vector resonance overflow saturation factor array One saturation factor; This is the set of effective masking weights after saturation correction, containing eight effective masking weights that have undergone dynamic decay correction. From the physical meaning of this multiplicative correction operation... Partially characterizes the first The frequency domain proximity of the electromagnetic harmonic frequency of a candidate vector to the propeller aerodynamic noise masking frequency band is the maximum masking and protection capability that the vector can obtain under ideal linear vibration conditions. The multiplicative attenuation factor characterizes whether the resonant excitation of the candidate vector is still within the linear elastic range of the structure, i.e., whether the protective commitment of the masking mechanism for this vector is still physically reliable; the product of the two... It achieves physically adaptive shrinking of the masking trust level, scaling the original masking reward proportionally according to the structural safety margin.

[0051] When the saturation factor is close to one, it indicates that the resonant excitation of the candidate vector is within the linear elastic range of the structure, the electromagnetic noise spectrum generated by the stator vibration is concentrated near the fundamental frequency without high-order harmonic distortion, and the protection commitment of the masking mechanism is completely reliable. At this time, the masking weight is hardly attenuated, and the full effectiveness of the original masking reward is retained. When the saturation factor approaches zero, it indicates that the resonant excitation of the candidate vector has significantly exceeded the preset linear elastic vibration threshold of the stator structure, the structure has entered the nonlinear distortion range and generated a large number of high-order harmonic components that escape from the masking frequency band, and the protection commitment of the masking mechanism has physically failed. At this time, the masking weight is greatly compressed to a value close to zero, which means that the system no longer trusts the protection capability of the masking mechanism for this vector, and subsequent fusion calculations will impose a great acoustic penalty on it.

[0052] Furthermore, the element-wise multiplication correction process is repeated for all eight candidate vectors. The eight corrected effective masking weights are aggregated and arranged in index order to form a set of effective masking weights after saturation correction. This set replaces the uncorrected auditory masking reward weight set in the first embodiment and is passed to the next stage. It is then used to perform saturation sensing acoustic excitation targeted fusion calculation with the candidate vector basic resonance penalty set, so that the final generated candidate vector acoustic excitation index array has the ability to adaptively sense the nonlinear distortion region of the structure.

[0053] More specifically, in step S333, the candidate vector fundamental resonance penalty set and the saturated-corrected masking effective weight set are fused using saturated sensing acoustic excitation to obtain a candidate vector acoustic excitation index array. It should be understood that after completing the saturation correction of the masking weights, the candidate vector fundamental resonance penalty set and the saturated-corrected masking effective weight set need to be finally fused to generate the acoustic evaluation basis flowing to the multi-objective optimization module. Compared to the first embodiment, which directly uses the uncorrected auditory masking reward weights as the denominator, the second embodiment uses the effective weights after saturation factor decay as the denominator. When the resonance excitation of a certain vector exceeds the structural linear threshold, the effective weights are compressed to a minimum, the division result is drastically amplified, and the acoustic excitation index of that vector is non-linearly penalized and amplified, which is physically equivalent to establishing an insurmountable acoustic safety boundary based on structural limits. Therefore, in the technical solution of this application, the candidate vector fundamental resonance penalty set and the masking effective weight set after saturation correction are further fused with saturation-sensing acoustic excitation to obtain a candidate vector acoustic excitation index array. This is used to finally hedge and quantify the structural resonance risk and the masking protection capability after saturation-sensing correction, generating a comprehensive acoustic evaluation index with adaptive sensing capability for structural nonlinear distortion. In this way, it can be ensured that when the resonance excitation of some candidate vectors approaches or exceeds the linear limit of the stator structure under extreme conditions, their acoustic excitation index is nonlinearly amplified, thereby forcing the subsequent multi-objective optimization module to actively avoid these high acoustic risk vectors. This fundamentally eliminates the problem of local failure of the acoustic stealth strategy caused by blindly trusting the masking under extreme conditions in the original scheme.

[0054] In a specific example of this application, the generation process of the candidate vector acoustic perturbation index array is as follows. First, each penalty element in the candidate vector fundamental resonance penalty set is paired element-wise with the corrected effective weight at the corresponding index position in the saturated corrected masking effective weight set to establish a one-to-one fusion calculation relationship. Next, each penalty element in the candidate vector fundamental resonance penalty set is fused by element-wise division with the corresponding corrected weight in the saturated corrected masking effective weight set, and a very small positive bias constant is added to the denominator to ensure the numerical safety of the microcontroller's floating-point operations, generating the candidate vector acoustic perturbation index array, the calculation expression of which is:

[0055] in, For the first The final acoustic perturbation feature index after saturated sensing fusion of candidate vectors; For the candidate vector fundamental resonance penalty set, the first One resonance penalty value; For the masking effective weight set after saturation correction, the first The corrected effective masking weights; It is a constant bias for extremely small positive numbers to prevent overflow during division by zero in floating-point operations; The candidate vector acoustic perturbation index array contains eight fusion index terms, which serve as the final output of this macroscopic step and flow to the multi-objective cost function weighted optimization module. From the physical meaning of this fusion operation, molecules... The denominator characterizes the resonant excitation energy of the candidate vector on the motor stator structure; a larger value indicates a higher risk of the vector inducing structural resonance. This characterizes the actual shielding strength that the electromagnetic noise of the vector can achieve after dynamic attenuation correction by the saturation factor. This value has been adaptively contracted based on whether the structure is within the linear elastic range. When the resonant excitation of a candidate vector is within the linear range of the structure, the saturation factor is close to one. After saturation correction, the effective weights corresponding to the effective weight set of the shielding remain at a large value, the division result is effectively suppressed, and the acoustic excitation index is maintained at a reasonable level. This indicates that although the vector generates some structural vibration, its noise can be effectively masked by aerodynamic noise. When the resonant excitation of a candidate vector exceeds the linear threshold of the structure, the saturation factor rapidly decays to near zero, the effective weights are compressed to a minimum value, the division result is sharply amplified, and the acoustic excitation index of the vector is nonlinearly penalized and amplified. Physically, this is equivalent to establishing an insurmountable acoustic safety boundary based on the structural limit, forcing the subsequent multi-objective optimization module to actively avoid the vector in the cost function evaluation.

[0056] Then, the above element-wise division fusion calculation is repeated for all eight candidate vectors. The eight acoustic excitation characteristic indices are aggregated and arranged in index order to form a candidate vector acoustic excitation index array. This array is passed to step four as an independent acoustic evaluation dimension and participates in multi-objective cost function fusion and rolling extreme value optimization in parallel with the candidate vector torque and flux linkage error evaluation array and the candidate vector switching action loss cost array.

[0057] Furthermore, under normal cruise conditions of the low-altitude robot, the resonant excitation of each candidate vector is within the linear range of the structure, and the saturation factor is close to one. The behavior of the second embodiment is almost the same as that of the first embodiment, without introducing additional performance loss. However, under extreme wind resistance or rapid maneuver conditions, when the resonant excitation of some candidate vectors approaches or exceeds the linear limit of the structure, the saturation factor rapidly decays to near zero, the masking weight is forcibly compressed, and the fused acoustic excitation index is nonlinearly amplified, thereby forcing the subsequent multi-objective optimization module to actively avoid these high acoustic risk vectors.

[0058] By introducing a resonant overflow saturation sensing factor based on the linear elastic vibration threshold of the stator structure, and then multiplying and coupling it with the auditory masking reward weight before performing fusion calculation, the second embodiment enables the quantitative evaluation of acoustic disturbance characteristics to have adaptive sensing capabilities for the nonlinear distortion region of the structure. This mechanism fundamentally eliminates the problem of local failure of the acoustic stealth strategy caused by blindly trusting masking under extreme conditions in the original scheme, ensuring that the low-altitude robot can maintain reliable acoustic stealth performance in the entire operating range above urban residential areas. At the same time, the smooth and continuous characteristics of the hyperbolic tangent function ensure that the saturation factor does not produce a step change in the transition range, avoiding unnecessary chattering interference to the motor torque output.

[0059] Specifically, in step four, based on the desired electromagnetic torque command and the desired stator flux linkage command, a multi-objective rolling optimization oriented to active noise shaping is performed on the predicted stator current trajectory set, the inverter finite voltage vector set, and the candidate vector acoustic disturbance index array to obtain the optimal switching vector. It should be understood that since the flight controller of the low-altitude robot generates the desired electromagnetic torque command and the desired stator flux linkage command in each control cycle based on the current attitude error and trajectory planning requirements, the model predictive control system needs to select an optimal voltage vector from the inverter finite voltage vector set that can simultaneously consider torque tracking accuracy, switching heat loss constraints, and acoustic stealth performance. The aforementioned steps have already generated the predicted stator current trajectory set for calculating the electromagnetic performance prediction of each candidate vector, the inverter finite voltage vector set for evaluating the switching cost, and the candidate vector acoustic disturbance index array for quantifying acoustic risk. The evaluation results of these three dimensions need to be uniformly incorporated into a global cost function for comprehensive optimization decision-making. Therefore, in the technical solution of this application, based on the desired electromagnetic torque command and the desired stator flux linkage command, a multi-objective rolling optimization oriented to active noise shaping is performed on the predicted stator current trajectory set, the inverter finite voltage vector set, and the candidate vector acoustic disturbance index array to obtain the optimal switching vector. This allows for the selection of a unique control action from all candidate voltage vectors that comprehensively optimizes the three objectives of electromagnetic torque tracking, switching loss suppression, and acoustic stealth. This ensures that the voltage vector output by the low-altitude robot in each microsecond-level control cycle not only meets the power requirements for maintaining flight attitude but also actively avoids vector schemes with high switching losses and high acoustic disturbance risks, achieving a globally optimal balance between electrical performance, thermal safety, and quiet urban flight.

[0060] Figure 5 This is a flowchart of step four of the multi-objective optimization method for low-altitude robot motors based on model predictive control according to an embodiment of this application. Figure 5As shown, step four includes: S410, obtaining the desired electromagnetic torque command and desired stator flux command issued by the flight controller; S420, based on the quadrature and direct axis components, permanent magnet flux, and inductance parameters of each candidate sequence in the predicted stator current trajectory set, performing electromagnetic torque prediction and stator flux error calculation on the desired electromagnetic torque command and desired stator flux command respectively to obtain candidate vector torque and flux error evaluation arrays; S430, based on the bridge arm switching state of each candidate vector in the inverter finite voltage vector set, performing inverter dynamic switching action loss evaluation on its physical switching state with the previous control cycle to obtain candidate vector switching action loss cost arrays; S441, performing multi-objective cost function fusion and rolling extremum optimization on the candidate vector torque and flux error evaluation arrays, candidate vector switching action loss cost arrays, and candidate vector acoustic excitation index arrays to obtain the optimal switching vector.

[0061] Accordingly, in steps S410 and S420, the desired electromagnetic torque command and desired stator flux linkage command issued by the flight controller are obtained. Based on the quadrature and direct axis components, permanent magnet flux linkage, and inductance parameters of each candidate sequence in the predicted stator current trajectory set, electromagnetic torque prediction and stator flux linkage error calculation are performed on the desired electromagnetic torque command and desired stator flux linkage command respectively to obtain candidate vector torque and flux linkage error evaluation arrays. It should be understood that since the flight controller of the low-altitude robot calculates and generates the desired electromagnetic torque command and desired stator flux linkage command in real time according to the current attitude error and trajectory tracking requirements in each control cycle, the core task of the model predictive control system is to select the optimal solution from multiple candidate voltage vectors that makes the actual output torque and flux linkage of the motor closest to the above-mentioned desired command. The premise for judging the electromagnetic tracking performance of each candidate vector is to calculate the degree of deviation between the predicted torque and predicted flux linkage generated by the motor after the application of each candidate vector and the expected value. Therefore, in the technical solution of this application, the desired electromagnetic torque command and desired stator flux command issued by the flight controller are obtained. Based on the quadrature and direct axis components, permanent magnet flux, and inductance parameters of each candidate sequence in the predicted stator current trajectory set, electromagnetic torque prediction and stator flux error calculation are performed on the desired electromagnetic torque command and desired stator flux command respectively to obtain candidate vector torque and flux error evaluation arrays. This is used to quantify the degree to which each candidate voltage vector meets the power requirements of the flight controller in the electromagnetic drive performance dimension. In this way, basic evaluation data characterizing the electrical tracking accuracy of each candidate vector can be provided for subsequent multi-objective cost function fusion and rolling extremum optimization, ensuring that the finally selected optimal switching vector prioritizes the torque output accuracy required for flight attitude maintenance while satisfying acoustic stealth and switching loss constraints.

[0062] More specifically, in a concrete example of this application, firstly, the desired electromagnetic torque command and desired stator flux linkage command issued by the flight controller are obtained. The desired electromagnetic torque command is generated in real time by the attitude loop and velocity loop of the flight controller based on the current flight state deviation, representing the rigid power requirement required to maintain the current flight attitude of the low-altitude robot. The desired stator flux linkage command is generated online based on the maximum torque-to-current ratio strategy, representing the optimal magnetic field operating point of the motor at the current speed. Next, the current response vector corresponding to each candidate voltage vector is extracted sequentially from the set of predicted stator current trajectories, and its direct-axis predicted current component and quadrature-axis predicted current component are obtained. Combined with the permanent magnet flux linkage constant and quadrature-axis and direct-axis inductance parameters of the motor body, the predicted electromagnetic torque after the application of each candidate vector is calculated according to the torque equation of the permanent magnet synchronous motor, and the predicted stator flux linkage amplitude after the application of each candidate vector is calculated according to the flux linkage equation. Then, the predicted electromagnetic torque of each candidate vector is subtracted from the desired electromagnetic torque command, and the predicted stator flux linkage amplitude is subtracted from the desired stator flux linkage command. The torque error and flux linkage error are squared respectively, and their respective weighting coefficients are added before summing to obtain the composite error evaluation index for each candidate vector. Its calculation expression is as follows:

[0063] in, To apply the first The predicted electromagnetic torque is calculated based on the predicted current after considering several candidate voltage vectors. This represents the number of pole pairs of the motor. The rotor permanent magnet flux linkage constant; To predict the first stator current trajectory in the set The cross-axis components of a current response vector; To predict the first stator current trajectory in the set The direct-axis components of the current response vector; It is a direct-axis inductor; It is a quadrature axis inductor; To apply the first The predicted stator flux linkage amplitude is calculated after considering several candidate voltage vectors; The desired electromagnetic torque command issued by the flight controller; The desired stator flux linkage command issued by the flight controller; This is the weighting coefficient for electromagnetic torque tracking error; This is the stator flux linkage tracking error weighting coefficient; For the first The torque and flux linkage combined error evaluation index for each candidate vector. From the physical meaning of the above formula, in the torque equation... This partially characterizes the main electromagnetic torque component generated by the interaction between the permanent magnet flux and the quadrature axis current, which is absolutely dominant in surface-mount permanent magnet motors. This partially characterizes the reluctance torque component arising from the asymmetry of the quadrature and direct-axis inductances, which provides an additional torque contribution in embedded permanent magnet motors. In the flux linkage equations... This partially characterizes the vector superposition of permanent magnet flux linkage and current-induced flux linkage along the direct axis. This partially characterizes the current-induced flux linkage component in the quadrature axis direction; the vector sum of these two components represents the total stator flux linkage. In the error evaluation formula... This partially quantifies the degree to which the actual output torque of the motor deviates from the power demand of the flight controller after the candidate vector is applied. The smaller this value is, the more accurately the vector can meet the thrust required to maintain the current flight attitude. It partially quantifies the degree to which the stator flux deviates from the optimal operating point; the smaller this value, the more efficient the motor is operating in a magnetic field.

[0064] Furthermore, the torque prediction and flux linkage error calculation process is repeated for all eight candidate vectors. The eight composite error evaluation indices are aggregated and arranged in index order to form a candidate vector torque and flux linkage error evaluation array. This array represents the ranking of the candidate vectors in terms of electromagnetic drive performance. It is then passed to the subsequent multi-objective cost function fusion and rolling extreme value optimization stage, and participates in the global optimal decision-making process alongside the candidate vector switching action loss cost array and the candidate vector acoustic disturbance index array.

[0065] Accordingly, in step S430, based on the bridge arm switching states of each candidate vector in the inverter's finite voltage vector set, a dynamic switching action loss assessment of the inverter is performed between these states and the physical switching states of the previous control cycle to obtain a candidate vector switching action loss cost array. It should be understood that since the inverter power switches of the low-altitude robot generate switching losses and electromagnetic interference each time they switch on or off, the more times the switches flip, the more severe the heat accumulation in the power devices and the greater the high-frequency electromagnetic noise generated. Furthermore, the finite control set model predicts that the control may select voltage vectors with spatially distant locations between adjacent control cycles, causing multiple phase bridge arms to flip simultaneously. If the switching cost is not constrained, the system may face the risk of inverter overheating or electromagnetic compatibility degradation under high dynamic conditions. Therefore, in the technical solution of this application, a dynamic switching action loss assessment of the inverter is further performed based on the bridge arm switching states of each candidate vector in the inverter's finite voltage vector set, compared with the physical switching states of the previous control cycle, to obtain a candidate vector switching action loss cost array. This quantifies the total amount of bridge arm flipping action required to switch from the current physical switching state to each candidate vector. This provides evaluation data characterizing the switching thermal cost of each candidate vector for subsequent multi-objective cost function fusion and rolling extremum optimization, enabling the system to take into account the thermal safety constraints of inverter power devices when selecting the optimal switching vector, and avoid unnecessary high-frequency switching chattering.

[0066] More specifically, in a specific example of this application, the generation process of the candidate vector switching action loss cost array is as follows: First, each candidate voltage vector is sequentially extracted from the finite voltage vector set of the inverter. Based on its topological definition in the space vector enumeration, the expected switching level state of the three-phase bridge arm corresponding to the candidate vector is decoded. The switching level of each bridge arm is one of two logic states: on or off. Next, the physical switching state that was actually issued and executed in the previous control cycle is read from the microcontroller's internal register. This state records the actual on / off configuration of the current three-phase bridge arm of the inverter. Then, the expected switching level state of the three-phase bridge arm of each candidate vector is compared phase by phase with the physical switching state of the previous control cycle. The absolute difference between the expected state and the actual state is calculated for each phase. If the expected state and the actual state of a phase are inconsistent, the difference is one, indicating that the bridge arm of that phase needs to perform a switching action. If they are consistent, the difference is zero, indicating that the phase does not need to be switched. The sum of the absolute differences of the three phases is used to obtain the total number of bridge arm switching operations required for the candidate vector to switch from the current state, which is its switching action loss cost. Furthermore, the above phase-by-phase comparison and summation process is repeated for all eight candidate vectors in the finite voltage vector set of the inverter. The eight switching action loss costs are aggregated and arranged in index order to form a candidate vector switching action loss cost array. The elements with a value of zero in this array correspond to candidate vectors that are completely consistent with the current physical state, i.e., no switching action is required. The elements with a value of three correspond to candidate vectors that all three phase bridge arms need to be flipped simultaneously, i.e., the switching cost is the maximum. The candidate vector switching action loss cost array is passed to the subsequent multi-objective cost function fusion and rolling extreme value optimization stage, and participates in the global optimal decision-making in parallel with the candidate vector torque and flux error evaluation array and the candidate vector acoustic disturbance index array.

[0067] Accordingly, in step S440, the candidate vector torque and flux linkage error evaluation array, the candidate vector switching action loss cost array, and the candidate vector acoustic disturbance index array are subjected to multi-objective cost function fusion and rolling extremum optimization to obtain the optimal switching vector. It should be understood that, since the aforementioned steps have quantitatively evaluated all candidate vectors in the inverter's finite voltage vector set from three independent physical dimensions—electromagnetic drive performance, switching heat loss, and acoustic stealth—and the model predictive control system can only issue one voltage vector to the inverter in each control cycle, it is necessary to unify and fuse the evaluation results of the three dimensions into the same scalar cost space and select the unique optimal solution that minimizes the overall cost through exhaustive comparison, thereby achieving a globally optimal balance decision between electrical performance, thermal safety, and quiet urban flight. Therefore, in the technical solution of this application, multi-objective cost function fusion and rolling extremum optimization are further performed on the candidate vector torque and flux linkage error evaluation array, the candidate vector switching action loss cost array, and the candidate vector acoustic disturbance index array to obtain the optimal switching vector. This allows for the selection of a unique control action from all candidate voltage vectors that comprehensively optimizes the three objectives of electromagnetic torque tracking, switching loss suppression, and acoustic stealth. This ensures that the voltage vector output by the low-altitude robot in each microsecond-level control cycle meets the power requirements for maintaining flight attitude while actively avoiding high switching losses and high acoustic disturbance risks, achieving a globally optimal rolling decision for active noise shaping.

[0068] More specifically, in a concrete example of this application, the process of generating the optimal switching vector is as follows. First, for each candidate vector index, the torque-flux composite error evaluation index corresponding to that index is extracted from the candidate vector torque-flux error evaluation array; the switch flipping cost corresponding to that index is extracted from the candidate vector switch action loss cost array; and the acoustic excitation characteristic index corresponding to that index is extracted from the candidate vector acoustic excitation index array. This establishes an alignment relationship between the three evaluation dimensions on the same candidate vector. Next, the values ​​of the above three evaluation dimensions are linearly weighted and summed after being appended with their respective weighting coefficients to construct the global multi-objective cost function value of the candidate vector. Its calculation expression is as follows:

[0069] in, For the first The global multi-objective cost function value of candidate voltage vectors; The candidate vector torque and flux linkage error evaluation array is the first... Each element represents the degree of electromagnetic drive tracking deviation of the vector; The first in the candidate vector switch action loss cost array Each element represents the total amount of bridge arm flipping motion required to switch from the current physical state to this vector; The i-th element in the candidate vector acoustic perturbation index array represents the comprehensive acoustic perturbation risk of the vector after saturated sensing fusion. This is the switching loss constraint weighting coefficient, used to adjust the system's emphasis on inverter thermal safety; This represents the acoustic stealth constraint weighting coefficient, used to adjust the system's emphasis on silent flight in urban areas. From the physical meaning of this cost function, This ensures that the motor torque and flux linkage accurately track the flight controller commands to maintain the flight attitude stability of the low-altitude robot. The terms constrain the switching frequency of the inverter bridge arms to control the heat loss and electromagnetic interference of the power devices. The project incorporates acoustic disturbance evaluation into the decision-making process to achieve active noise shaping. The linear weighted summation of the three factors unifies the optimization objectives of the three physical domains—electrical performance, thermal safety, and acoustic stealth—into the same scalar cost space.

[0070]

[0071] in, The optimal switching vector; For the inverter, it is a finite set of voltage vectors; To find the value of the global multi-objective cost function The independent variable indexing operation reaches the minimum value. From the physical meaning of this optimization operation, the candidate vector with the minimum cost function value is the globally optimal control action that can simultaneously take into account torque tracking accuracy, switching thermal safety and acoustic stealth performance in the current control cycle. This vector is closest to the power requirements of the flight controller in terms of electromagnetic performance, generates the least bridge arm flipping heat loss in switching, and has the lowest risk of resonance excitation to the motor structure in terms of acoustic characteristics, and its electromagnetic noise is masked by the propeller aerodynamic noise to the greatest extent.

[0072] Furthermore, based on the locked minimum index, the corresponding discrete voltage vector is extracted from the finite voltage vector set of the inverter and output as the optimal switching vector. This vector is the optimal physical control action that takes into account both high dynamic disturbance rejection thrust response and full-spectrum audible stealth under the current control cycle, and is passed to step five for anti-shoot-through dead zone compensation and inverter hardware drive signal generation.

[0073] Specifically, in step five, dead-time compensation is used to perform anti-shoot-through dead-time compensation on the optimal switching vector to obtain a three-phase pulse modulation drive signal with dead time. This three-phase pulse modulation drive signal with dead time is output to the inverter gate drive circuit. It should be understood that since the inverter of the low-altitude robot adopts a two-level three-phase full-bridge topology, the upper and lower transistors of each phase arm share the same DC bus. If the upper and lower transistors overlap and conduct simultaneously during the switching instant, a short circuit will occur on the DC bus, causing the power devices to burn out instantly. However, actual power switches have a non-zero turn-off tail time; that is, the lower transistor receives the turn-on command before the upper transistor is completely turned off. Therefore, a dead-time delay must be forcibly inserted based on the ideal complementary logic instruction to ensure that the upper and lower transistors of the same bridge arm are not simultaneously in a conducting state at any time. Therefore, in the technical solution of this application, dead-time compensation is further applied to the optimal switching vector to obtain a three-phase pulse modulation drive signal with dead time. This three-phase pulse modulation drive signal with dead time is output to the inverter gate drive circuit, thereby safely converting the optimal control decision obtained from multi-objective optimization into a physical drive waveform that can be directly executed by the inverter power devices. This ensures that the electromagnetic torque tracking and acoustic stealth optimization intentions carried by the optimal switching vector are transmitted to the motor physical level for execution without loss, while absolutely eliminating the risk of power device damage caused by bridge arm shoot-through short circuit, thus guaranteeing the hardware safety of the low-altitude robot drive system under all operating conditions.

[0074] More specifically, in the embodiments of this application, step five includes: performing optimal space vector decoding on the optimal switching vector based on the space vector modulation mapping rule to obtain the original logic instruction of the three-phase upper bridge arm; performing complementary flipping and channel splicing on the original logic instruction of the three-phase upper bridge arm to obtain the complementary logic instruction of the six-way upper and lower bridge arms; and performing forced delay and Boolean AND logic truncation on the rising edge of the complementary logic instruction of the six-way upper and lower bridge arms based on the preset dead time delay operator to obtain the three-phase pulse modulation drive signal with dead time.

[0075] Accordingly, in a specific example of this application, step five includes the following processing steps: First, based on the space vector modulation mapping rule, the optimal switching vector is decoded using the optimal space vector. According to the topological index definition of the optimal switching vector in the inverter's finite voltage vector set, the index is converted into a three-bit binary number through bit operations. The expected turn-on or turn-off logic levels that the upper arms of the inverter's A, B, and C phases should execute are extracted bit by bit to form the original logic instructions for the three-phase upper arms. Next, the original logic instructions for the three-phase upper arms are subjected to complementary flipping and channel splicing. Boolean logic NOT operations are performed on the upper arm logic level of each phase to generate the corresponding lower arm logic level, ensuring that the lower arm is turned off when the upper arm is turned on and turned on when the upper arm is turned off. The logic levels of the three-phase upper arms and the logic levels of the three-phase lower arms are alternately spliced ​​in phase sequence to form six complementary logic instructions for the upper and lower arms. Then, based on the preset dead-time delay operator, the rising edges of the complementary logic instructions of the six upper and lower bridge arms are forcibly delayed and truncated using Boolean AND logic. The current logic level of each signal is then subjected to a Boolean AND operation with the historical logic level after the dead-time delay. This mechanism forces the rising edge of the signal to be delayed by a dead-time length while the falling edge remains instantaneous, thus inserting a safe interval between the upper transistor's turn-off and the lower transistor's turn-on, where both transistors are in a turned-off state, generating a three-phase pulse modulation drive signal with dead time. Furthermore, the three-phase pulse modulation drive signal with dead time is written into the pulse width modulation peripheral register of the microcontroller. After level conversion and power amplification by the inverter gate drive circuit, it is applied to the gate terminals of the power switching transistors, driving the low-altitude robot rotor motor to perform physical commutation according to the optimal switching vector determined by multi-objective optimization, completing the full-link closed loop from algorithm decision to physical execution in the current control cycle.

[0076] In summary, the multi-objective optimization method for low-altitude robot motors based on model predictive control according to the embodiments of this application is explained. It utilizes an airborne microphone to collect propeller aerodynamic noise in real time and analyze its auditory masking frequency band characteristics. At the same time, based on the offline calibrated electromechanical structure resonance cost matrix, it performs acoustic excitation quantification evaluation on the predicted current trajectory of each candidate voltage vector. The resonance penalty and masking reward are integrated into a unified acoustic excitation index, which is included as an independent evaluation item in the multi-objective cost function, participating in rolling optimization alongside torque tracking and switching losses. Thus, within a microsecond-level control cycle, the inverter is guided to actively avoid dangerous switching sequences that trigger structural resonance, and the unavoidable electromagnetic harmonic energy is shifted to the masking frequency band of aerodynamic noise. This achieves full-condition acoustic stealth of the low-altitude robot without sacrificing dynamic thrust response.

[0077] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A multi-objective optimization method for motors of low-altitude robots based on model predictive control, characterized in that, include: Step 1: Perform feature extraction and masking frequency band analysis on the collected operating status data to obtain the stator current feedback quantity in the rotating coordinate system and the feature vector of the dynamic auditory masking frequency band; Step 2: Based on the DC bus voltage and the real-time speed of the rotor motor, construct a finite control set and predict the forward state of the stator current feedback in the rotating coordinate system to obtain the inverter finite voltage vector set and the predicted stator current trajectory set. Step 3: Based on the pre-set electromechanical structure resonance cost matrix, perform electromechanical structure acoustic excitation feature targeted quantization on the predicted stator current trajectory set and the dynamic auditory masking frequency band feature vector to obtain the candidate vector acoustic excitation index array. Step 4: Based on the desired electromagnetic torque command and the desired stator flux linkage command, perform multi-objective rolling optimization for active noise shaping on the predicted stator current trajectory set, the inverter finite voltage vector set, and the candidate vector acoustic disturbance index array to obtain the optimal switching vector. Step 5: The optimal switching vector is subjected to anti-shoot-through dead-time compensation through dead-time compensation to obtain a three-phase pulse modulation drive signal with dead time. The three-phase pulse modulation drive signal with dead time is output to the inverter gate drive circuit.

2. The multi-objective optimization method for low-altitude robot motors based on model predictive control according to claim 1, characterized in that, The operational status data includes the three-phase current sampling values ​​of the rotor motor, the rotor mechanical angle of the rotor motor, and the audio time series of the airborne microphone.

3. The multi-objective optimization method for low-altitude robot motors based on model predictive control according to claim 2, characterized in that, Step one includes: Based on the rotor mechanical angle and pole pair number of the rotor motor, the sampled values ​​of the three-phase current of the rotor motor are subjected to Clark and Park cascaded coordinate mapping to obtain the stator current feedback in the rotating coordinate system. Aerodynamic noise fundamental frequency optimization is performed on the airborne microphone audio time series to obtain the aerodynamic noise fundamental frequency component; Based on a nonlinear psychoacoustic mapping model, a critical band scale conversion and masking threshold estimation are performed on the fundamental frequency component of aerodynamic noise to obtain the dynamic auditory masking band feature vector.

4. The multi-objective optimization method for low-altitude robot motors based on model predictive control according to claim 1, characterized in that, Step two includes: Based on the magnitude of the DC bus voltage, space vector enumeration is performed on the two-level three-phase inverter topology to obtain a finite set of inverter voltage vectors. The electric angular velocity is calculated based on the real-time speed and number of pole pairs of the rotor motor. Combined with the stator resistance and quadrature-direct axis inductance parameters, the continuous-time state-space equation is derived by forward Euler discretization to obtain the discrete parameter matrix of the motor prediction model. Based on the discrete parameter matrix of the motor prediction model, and taking the stator current feedback in the rotating coordinate system as the initial iteration base point, the forward Euler stator current trajectory multiple derivation is performed on each candidate vector in the finite voltage vector set of the inverter to obtain the predicted stator current trajectory set.

5. The multi-objective optimization method for low-altitude robot motors based on model predictive control according to claim 1, characterized in that, Step three includes: Based on the pre-set electromechanical structure resonance cost matrix, a basic mechanical resonance penalty evaluation is performed on each current response vector in the predicted stator current trajectory set to obtain a candidate vector basic resonance penalty set. Based on the masking frequency band boundary in the dynamic auditory masking frequency band feature vector, the critical frequency band distance is calculated and reward mapping is performed on the equivalent electromagnetic harmonic frequency derived from the predicted stator current trajectory set to obtain the auditory masking reward weight set. Acoustic perturbation feature-targeted fusion is performed on the candidate vector fundamental resonance penalty set and the auditory masking reward weight set to obtain the candidate vector acoustic perturbation index array.

6. The multi-objective optimization method for low-altitude robot motors based on model predictive control according to claim 1, characterized in that, Step four includes: Obtain the desired electromagnetic torque command and desired stator flux command issued by the flight controller; Based on the quadrature and direct axis components, permanent magnet flux linkage and inductance parameters of each candidate sequence in the predicted stator current trajectory set, electromagnetic torque prediction and stator flux linkage error calculation are performed on the desired electromagnetic torque command and desired stator flux linkage command respectively to obtain candidate vector torque and flux linkage error evaluation arrays. Based on the bridge arm switching state of each candidate vector in the inverter's finite voltage vector set, the inverter's dynamic switching action loss is evaluated by comparing it with the physical switching state of the previous control cycle to obtain the candidate vector switching action loss cost array. The optimal switching vector is obtained by fusing multi-objective cost functions and rolling extremum optimization on the candidate vector torque and flux linkage error evaluation array, the candidate vector switching action loss cost array, and the candidate vector acoustic disturbance index array.

7. The multi-objective optimization method for low-altitude robot motors based on model predictive control according to claim 1, characterized in that, Step five includes: Based on the space vector modulation mapping rule, the optimal switching vector is decoded by the optimal space vector to obtain the original logic instruction of the three-phase upper bridge arm. The original logic commands of the three-phase upper arm are complementary flipped and spliced ​​to obtain the complementary logic commands of the six upper and lower arms. Based on the preset dead-time delay operator, the rising edge of the complementary logic instruction of the six bridge arms is forcibly delayed and truncated by Boolean AND logic to obtain a three-phase pulse modulation drive signal with dead time.

8. The multi-objective optimization method for low-altitude robot motors based on model predictive control according to claim 5, characterized in that, The candidate vector fundamental resonance penalty set and the auditory masking reward weight set are subjected to acoustic perturbation feature-targeted fusion to obtain a candidate vector acoustic perturbation index array, including: Based on the preset linear elastic vibration threshold of the stator structure, the degree of resonance overflow of each element in the candidate vector fundamental resonance penalty set is quantified to obtain the candidate vector resonance overflow saturation factor array. Dynamic decay correction of masking confidence is performed on each reward coefficient in the auditory masking reward weight set and the corresponding factor in the candidate vector resonance overflow saturation factor array to obtain the masking effective weight set after saturation correction. Saturation-sensing acoustic perturbation-targeted fusion is performed on the candidate vector fundamental resonance penalty set and the masked effective weight set after saturation correction to obtain the candidate vector acoustic perturbation index array.

Images