Electrical transmission system parameter self-tuning system based on neural network

Abstract

The invention relates to the technical field of motor control, and discloses an electrical transmission system parameter self-tuning system based on a neural network, and the system comprises a state monitoring and gating module which is used for activating the system when the high-frequency energy characteristic of a torque instruction exceeds a threshold value; the homomorphic mapping and reference reconstruction module is used for constructing a homomorphic digital filter to convert a torque instruction into a homomorphic instruction aligned with a feedback speed time sequence; the momentum residual calculation module is used for calculating the difference value between the theoretical impulse and the actual momentum increment based on the momentum theorem; the morphological decoupling and reasoning module is used for calculating morphological similarity between the residual vector and the reference vector by using the neuron structure so as to output confidence; and the parameter updating module is used for adjusting the adaptive updating gain according to the confidence coefficient and iteratively correcting the inertia parameter. According to the method, through signal homomorphic reconstruction and waveform form reasoning, the influence of loop delay and load disturbance is effectively inhibited, and the accuracy and robustness of parameter setting are improved.

Description

Technical Field

[0001] This invention relates to the field of motor control technology, specifically to a self-tuning system for electrical drive system parameters based on neural networks. Background Technology

[0002] In high-performance AC servo drive systems, the gain tuning of the speed loop controller is highly dependent on the accurate identification of the load's rotational inertia. Existing inertia identification techniques typically employ recursive least squares, model reference adaptive control, or Landau adaptive law. Their core logic is based on establishing a linear equation between electromagnetic torque and angular acceleration according to Newton's second law, and iteratively estimating the system inertia by minimizing the prediction error.

[0003] However, the above methods have limitations in practical engineering applications. First, angular acceleration signals typically require differential operations on discrete position or velocity feedback to obtain them. This process significantly amplifies the quantization and sampling noise of the position sensor, reducing the signal-to-noise ratio. If a low-pass filter is introduced to suppress noise, phase lag will inevitably be introduced, compromising the real-time performance of the dynamic signal. Second, traditional identification models often ignore the physical delay and filter response time within the servo drive, from the issuance of the torque command to the actual generation of electromagnetic torque. This results in a misalignment between the input command and the output response used for calculation, leading to a systematic bias. Furthermore, most existing parameter update algorithms only perform unidirectional correction based on the magnitude of the error, lacking the ability to analyze the waveform characteristics of the error. It is difficult to effectively distinguish whether the error is caused by internal parameter mismatch or by external load disturbances or nonlinear friction. When the system encounters sudden load changes or operates under complex conditions, the algorithm may misjudge external disturbances as parameter errors, causing significant fluctuations or divergence in the inertia estimate, affecting the system's operational stability.

[0004] Therefore, this invention proposes a neural network-based self-tuning system for electrical drive system parameters to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a neural network-based self-tuning system for electrical drive system parameters. This system solves the problems of low identification accuracy due to differential noise amplification and command response delay in existing servo inertia identification technologies, as well as the divergence in parameter updates caused by the inability to distinguish between load disturbances and parameter mismatch.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a neural network-based self-tuning system for electrical drive system parameters, comprising: The status monitoring and gating module is connected to the servo driver to synchronously acquire torque current commands and feedback mechanical angular velocity, and generates an activation signal when the high-frequency energy characteristics of the torque current command exceed a preset threshold. The homomorphic mapping and reference reconstruction module is used to construct a homomorphic digital filter based on the calibrated loop delay parameters, convert the torque current command into a homomorphic torque command that is aligned with the feedback mechanical angular velocity in time, and output the inertia dominance index. The momentum residual calculation module is used to calculate the theoretical electromagnetic impulse and the actual momentum increment within a preset sliding window based on the momentum theorem, and generate the difference between the theoretical electromagnetic impulse and the actual momentum increment as the momentum residual signal. The morphological decoupling and reasoning module is used to construct normalized residual vectors and normalized reference vectors, and to calculate the morphological similarity between the normalized residual vectors and the normalized reference vectors through the internally constructed neuron structure to output confidence scores. The parameter update module is used to adjust the adaptive update gain according to the confidence level value, iteratively update the inertia parameter of the system using the momentum residual signal, and map the inertia parameter to the velocity loop control gain.

[0007] Preferably, the state monitoring and gating module is configured to execute cross-correlation analysis logic: the state monitoring and gating module injects a high-frequency pseudo-random binary sequence into the current loop and records the response of the feedback mechanical angular velocity; the state monitoring and gating module calculates the discrete cross-correlation function between the torque current command sequence and the feedback mechanical angular velocity sequence, and searches for the global peak time of the discrete cross-correlation function to lock the loop delay parameter.

[0008] Preferably, the state monitoring and gating module calculates the high-frequency energy characteristic by: calculating the sum of squares of the difference between the torque current command within the current sliding window, and using the sum of squares as the high-frequency energy characteristic; the preset threshold is a dynamic variable set based on the motor's rated current; when the high-frequency energy characteristic is less than the preset threshold, the state monitoring and gating module outputs a suspension signal.

[0009] Preferably, the homomorphic mapping and reference reconstruction module is configured with a parameter monitoring mechanism for reading the filter configuration of the servo driver; the homomorphic digital filter includes a cascaded low-pass filter, a notch filter, and a pure time delay based on a circular queue; the homomorphic mapping and reference reconstruction module uses the homomorphic digital filter to generate the homomorphic torque command.

[0010] Preferably, the homomorphic mapping and benchmark reconstruction module is further used to calculate the inertia dominance index: the homomorphic mapping and benchmark reconstruction module calculates and estimates the friction torque based on the nominal friction parameters; the homomorphic mapping and benchmark reconstruction module calculates the proportion of the difference between the total electromagnetic torque and the estimated friction torque in the total electromagnetic torque to obtain the normalized inertia dominance index.

[0011] Preferably, the momentum residual calculation module adopts an integral domain observation mechanism: the homomorphic torque command is discretely summed within the sliding window to obtain the theoretical electromagnetic impulse; the difference between the feedback mechanical angular velocity at both ends of the sliding window is calculated, and combined with the inertia parameter of the previous moment to obtain the actual momentum increment.

[0012] Preferably, the neuron structure execution logic in the morphological decoupling and inference module is as follows: mapping the normalized reference vector to the dynamic synaptic weights of the neuron structure, and mapping the normalized residual vector to the input feature vector of the neuron structure; performing a vector inner product operation on the dynamic synaptic weights and the input feature vector to obtain the confidence value; the confidence value is used to quantify the probability that the momentum residual signal originates from internal parameter mismatch of the system.

[0013] Preferably, the morphological decoupling and inference module performs a reliability masking process before constructing the vector: it generates an effectiveness mask by combining the absolute value of the torque current command with the inertia dominance index calculated by the homomorphic mapping and benchmark reconstruction module; only when the torque current command avoids the current dead zone and the inertia dominance index is higher than a preset confidence threshold, the current data is allowed to be used to construct the normalized residual vector and the normalized benchmark vector.

[0014] Preferably, the parameter update module is configured to perform variable step-size adaptive update: the parameter update module obtains the inertia dominance index calculated by the homomorphic mapping and benchmark reconstruction module, and obtains the confidence value output by the morphological decoupling and inference module; the parameter update module uses the inertia dominance index and the confidence value to perform weighted modulation on the base learning rate to generate the adaptive update gain; the parameter update module uses the adaptive update gain, momentum residual signal, and mechanical angular velocity change to calculate the inertia correction.

[0015] Preferably, the parameter update module performs safety boundary projection and gain mapping: the updated inertia parameters are forcibly constrained within the physically feasible domain set based on the motor rotor inertia; the constrained inertia parameters are converted into a target proportional gain based on the preset speed loop bandwidth, and the final execution gain is generated through the ramp limiter.

[0016] This invention provides a self-tuning system for electrical drive system parameters based on neural networks. It has the following advantages: 1. This invention solves the signal timing asynchrony problem caused by communication delay and filtering in the servo control loop through a homomorphic mapping and reference reconstruction module. A homomorphic digital filter is constructed using calibrated loop delay parameters, ensuring that the torque current command is strictly aligned with the feedback mechanical angular velocity in the time domain response characteristics, eliminating spurious errors caused by signal phase differences. Combined with the integral domain observation mechanism used in the momentum residual calculation module, the momentum theorem is used to replace traditional differential calculations, effectively suppressing quantization noise and high-frequency interference from the position encoder, and improving the signal-to-noise ratio and accuracy of inertia parameter identification.

[0017] 2. This invention utilizes a morphological decoupling and inference module to achieve accurate classification and discrimination of momentum residual sources. By mapping the normalized residual vector and the normalized reference vector to the input features and dynamic weights of neurons, and calculating their vector inner product to obtain morphological similarity, the system can effectively distinguish whether the momentum residual originates from internal inertia parameter mismatch or external load disturbance. This waveform-morphology-based inference mechanism avoids erroneous parameter corrections when the algorithm encounters sudden load changes, impacts, or nonlinear frictional disturbances, ensuring the robustness of the self-tuning process under complex working conditions.

[0018] 3. This invention establishes a variable step-size adaptive update mechanism based on a multi-dimensional confidence index, balancing parameter convergence speed and system security. The state monitoring and gating module activates the system only when the command signal has sufficient high-frequency energy, while the parameter update module dynamically adjusts the update gain based on the inertia dominance and confidence values. This strategy ensures that the system performs significant corrections only when the effective excitation is sufficient and the proportion of frictional interference is low, while automatically limiting the update amplitude during current dead zones, low-speed creep, or low confidence, effectively preventing parameter divergence or oscillation. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the system structure of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention; Figure 3 This is a schematic diagram comparing the parameter identification performance of the embodiments of the present invention with that of the traditional FF-RLS method under all operating conditions; Figure 4 This is a schematic diagram illustrating the anti-interference performance analysis of the system of the present invention when it encounters a strong cutting force disturbance for 1.5 seconds; wherein, Figure (a) is the waveform of the applied cutting torque; and Figure (b) is a comparison curve of the inertia identification response of the method of the present invention and the conventional method under the disturbance. Figure 5This is a schematic diagram illustrating the temporal changes of key internal variables generated by the homomorphic mapping and benchmark reconstruction module and the morphological decoupling and inference module of the present invention.

[0020] Among them, 100 is the state monitoring and gating module; 200 is the homomorphic mapping and benchmark reconstruction module; 300 is the momentum residual calculation module; 400 is the morphological decoupling and inference module; and 500 is the parameter update module. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] See attached document Figure 1 , Figure 1 This is a schematic diagram of a neural network-based self-tuning system for electric drive system parameters according to an embodiment of the present invention. The present invention provides a neural network-based self-tuning system for electric drive system parameters, which is applied in an electric drive architecture including a servo driver and a motor. The system includes: The condition monitoring and gating module 100 communicates with the servo driver to acquire torque current commands, feedback speed, and feedback current data. The condition monitoring and gating module 100 is equipped with cross-correlation analysis logic to calibrate the total loop delay parameters of the system during the initialization phase. The condition monitoring and gating module 100 calculates the high-frequency energy characteristics of the torque current command and generates an activation signal when the energy characteristics exceed a preset threshold.

[0023] The homomorphic mapping and reference reconstruction module 200 is used to construct a reference signal that is timing-aligned with the physical system. The homomorphic mapping and reference reconstruction module 200 reads the filter configuration parameters of the servo driver and constructs a homomorphic digital filter by combining it with the total loop delay parameters. The homomorphic mapping and reference reconstruction module 200 inputs the torque current command into the homomorphic digital filter to generate a homomorphic torque command. The homomorphic mapping and reference reconstruction module 200 calculates the proportion of inertial torque in the total driving torque based on the nominal friction model and outputs the inertia dominance index.

[0024] The momentum residual calculation module 300 is used to calculate the observation error in the integral domain. It performs time integration on the homomorphic torque command to obtain the theoretical momentum increment and differential calculation on the feedback speed to obtain the actual momentum increment. The module then calculates the difference between the theoretical and actual momentum increments to generate a momentum residual signal.

[0025] The morphological decoupling and inference module 400 is used to determine the source of error. It generates a reliability mask by combining the feedback current amplitude and the inertia dominance index. The module then uses the reliability mask to weight the momentum residual signal and velocity change signal, and performs normalization on the weighted vectors. Finally, it calculates the morphological similarity between the normalized vectors and outputs a confidence score.

[0026] The parameter update module 500 is used to perform parameter correction. The parameter update module 500 calculates the inertia correction amount based on the confidence level and the amplitude of the momentum residual signal. The parameter update module 500 limits the corrected inertia parameters within a preset feasible range and sends the updated parameters to the servo driver 20.

[0027] See attached document Figure 2 , Figure 2 This is a schematic flowchart of a neural network-based self-tuning method for electrical drive system parameters according to an embodiment of the present invention. The present invention provides a neural network-based self-tuning method for electrical drive system parameters, comprising the following steps: S100 calibrates the total loop delay through cross-correlation analysis of commands and feedback during the system initialization phase, and monitors the high-frequency energy of torque commands during the operation phase. When the energy exceeds the threshold, the parameter tuning process is activated. S200 reads the driver filter parameters and constructs a homomorphic filter by combining the total loop delay, maps the torque command to a homomorphic torque command, and calculates the inertia dominance based on the friction model; S300 calculates the theoretical momentum increment based on the homomorphic torque command and the actual momentum increment based on the feedback speed, and calculates the difference between the two to generate the momentum residual. S400 generates a reliability mask based on the current amplitude and the dominance of inertia, performs weighted normalization on the momentum residual and velocity change, and calculates the morphological similarity between the two. The S500 determines the parameter correction direction and step size based on morphological similarity, updates the rotational inertia parameter under safety boundary constraints, and simultaneously adjusts the velocity loop control gain.

[0028] To further clarify the implementation of each technical aspect of the present invention, the following will provide a detailed description of the implementation of each functional module involved above and its internal processing flow.

[0029] See attached document Figure 1 The status monitoring and gating module 100 is configured to perform data interaction and logic control of the execution system. On the one hand, the status monitoring and gating module 100 is used to ensure strict timing synchronization of multi-source signals, and on the other hand, it is used to identify the valid operating conditions for parameter identification.

[0030] S110, Synchronous acquisition and buffering of multi-channel signals. In this embodiment, the status monitoring and gating module 100 establishes a high-frequency data link with the underlying control core of the servo driver via a bus or direct memory access. To ensure the mathematical accuracy of the homomorphic mapping, it is necessary to ensure that each frame of data is strictly aligned on the time axis. Therefore, the status monitoring and gating module 100 synchronously latches the current torque current command within each servo control cycle (typically 62.5 microseconds to 250 microseconds). Feedback mechanical angular velocity and feedback torque current .here The current index representing the discrete time series.

[0031] These real-time streaming data are not processed immediately, but are sequentially pushed into a fixed-depth, first-in-first-out (FIFO) circular buffer. The length of this buffer is designed to cover the longest observation window required by subsequent algorithms, thereby providing a unified time reference for multiple signals and eliminating random jitter caused by transmission links.

[0032] S120, System loop delay calibration based on cross-correlation. In electric drive systems, there is an inherent physical lag between command issuance and sensor response, including PWM dead time, inverter transmission delay, sample-and-hold delay, and filter group delay. To eliminate the interference of this lag on high-frequency identification, this embodiment does not use empirical values ​​for estimation, but instead introduces a cross-correlation self-calibration mechanism. It should be noted that this calibration is not performed in real time, but is strictly limited to the motor power-on initialization phase or system maintenance mode to avoid interfering with normal processing and production.

[0033] In practice, the state monitoring and gating module 100 injects a small-amplitude, high-frequency pseudo-random binary sequence into the current loop as an excitation. This strategy utilizes the inherent mechanical inertia of the motor rotor as a natural low-pass filter. Because the frequency of the PRBS sequence is set much higher than the system's mechanical cutoff frequency, and the amplitude is limited to 5% to 10% of the motor's rated torque, the motor rotor will only experience micro-level vibrations without macroscopic rotation, thus ensuring the safety of the testing process.

[0034] The system records the rotational speed response under this excitation and calculates the discrete cross-correlation function between the torque current command sequence and the feedback mechanical angular velocity sequence. Its operational logic is as follows: ; In the formula, For the injected test stimulus sequence, To ensure statistical significance, the sequence length (e.g., 1024 or 2048 points) should be selected. This represents the number of lag beats within the search range (the search range typically covers 0 to 20 control cycles). This is feedback speed data with a corresponding lag.

[0035] Physically, the peak of the cross-correlation function corresponds to the phase difference between two signals with the highest degree of overlap. Therefore, by searching... The global maximum value can be used to lock the total loop delay parameter of the system. : ; Once determined, This parameter is then permanently stored. During subsequent normal operation, the system will directly call this parameter to build a digital twin model, ensuring precise synchronization between the model and the physical object on the timeline.

[0036] S130, a global gating system driven by the potential energy of the command source. After solving the timing synchronization problem, another challenge is how to sift out valuable segments from massive amounts of runtime data. Traditional methods often rely on the rate of change of feedback speed, but this is easily misled by encoder quantization noise. This embodiment proposes a more robust strategy: high-frequency energy determination based on the "command source". Due to the torque command... It directly reflects the controller's operational intent and does not contain measurement noise. Monitoring its rate of change can more accurately capture the system's true dynamic intent.

[0037] The state monitoring and gating module 100 calculates the high-frequency energy of the torque command within the current sliding window in real time. The calculation method is as follows: ; In the formula, The value represents the sliding window length, typically covering one to two mechanical time constants (e.g., 64 to 128 sampling points) to balance sensitivity and stability. This formula essentially extracts the high-frequency dynamic components from the command stream through differential operations, filtering out the DC component corresponding to the steady-state load.

[0038] To ensure the gating mechanism is applicable to motors of different power ratings, this embodiment employs a relative threshold criterion. The calculated energy... With adaptive threshold A comparison was conducted, among which It is not a fixed constant, but rather linked to the rated capacity of the motor: ; In the formula, The effective value of the rated current read from the motor's electronic nameplate; This is a dimensionless sensitivity coefficient, and its value is preferably within the range of 10.-3 Up to 10 -2 Between. In harsh electromagnetic environments, the setting can be appropriately increased. The value can be adjusted to increase the immunity threshold; however, in a precise experimental environment, it can be reduced. It is valuable for capturing subtle dynamic stimuli.

[0039] When detected When the system determines that it is in a "strong stimulus" state, it immediately activates the subsequent parameter identification process; otherwise, if... This indicates that the system is in a steady state or weak excitation region, and the module will output a suspension signal to keep subsequent algorithms silent. This mechanism ensures that parameter updates only occur at the moment of highest signal-to-noise ratio, eliminating the risk of false updates from the source.

[0040] See attached document Figure 1 The homomorphic mapping and reference reconstruction module 200, as a key link connecting the control command domain and the physical response domain, primarily functions to accurately reproduce the physical transmission characteristics of the electrical drive system in digital space, thereby providing a reference signal with strictly aligned timing and frequency domain characteristics for subsequent parameter identification. Its specific implementation method and logical details are as follows: S210, Dynamic monitoring and synchronization of control loop parameters. In actual operation scenarios of electric drive systems, the internal control parameters of servo drives are not static. Users may adjust the cutoff frequency of the low-pass filter according to operating conditions, or the system's adaptive vibration damping function may dynamically change the center frequency and bandwidth of the notch filter. If the model parameters inside the identification algorithm fail to update accordingly, a non-negligible phase deviation will occur between the mathematical model and the physical entity, leading to divergence in the identification results.

[0041] To address this issue, this embodiment configures a background parameter monitoring mechanism in the homomorphic mapping and reference reconstruction module 200. This mechanism does not occupy high-priority control interrupt resources; instead, it periodically (e.g., every 10ms to 100ms) polls the key registers inside the driver at a lower priority, monitoring system status including current loop sampling period and filter configuration. Once a change in register value is detected, the homomorphic mapping and reference reconstruction module 200 immediately reads the latest parameter set and triggers the subsequent coefficient recalculation logic of the homomorphic digital filter, thereby ensuring that the digital channel always maintains dynamic consistency with the physical channel.

[0042] S220, Construction and Signal Mapping of the Homomorphic Digital Filter. This step aims to construct a purely digital signal processing channel whose response characteristics to input commands, particularly amplitude attenuation and phase hysteresis, closely match the response characteristics of the actual physical controller and power loop. The homomorphic mapping and reference reconstruction module 200 constructs the homomorphic digital filter based on the real-time read control parameters and the total loop delay parameters calibrated in the previous steps. .

[0043] In practical implementation, the homomorphic mapping and reference reconstruction module 200 will convert the original torque current command... As an excitation source, a homomorphic torque command is generated through a cascaded digital filtering circuit. The mathematical transitivity of this process is described as follows: ; In the formula, For discrete Z-transform operators; and These are the Z-domain representations of the original instruction and the homomorphic instruction, respectively. To implement the above transfer function in a microprocessor, this embodiment employs a discretization approximation strategy: For the low-pass filter stage With notch filter Based on the cutoff frequency of reading Center frequency and bandwidth parameters (Usually defined as the bandwidth where the amplitude-frequency response drops by 3dB), using the bilinear transform method or the Tustin approximation method, the S-domain transfer function of the continuous domain is mapped to the Z-domain transfer function of the discrete domain, and then transformed into the coefficients of the difference equation that can be implemented programmably.

[0044] For pure time delay Considering its physical meaning is simply a time delay, this embodiment employs a fixed-length circular queue structure in its software implementation. The queue depth is set to... The system writes the latest instruction to the tail of the queue at each sampling time and reads the lagging instruction from the head of the queue. The data from each beat is used as the output. Through this signal mapping, the generated... In terms of mathematical properties, it becomes a digital mirror signal that actually acts on the electromagnetic torque of the motor rotor.

[0045] S230, Inertia Dominance Assessment Based on Nominal Model. In engineering practice of parameter identification, friction torque often manifests as a nonlinear disturbance term. When the motor is running at low speed or with low output torque, most of the electromagnetic torque is used to overcome frictional resistance rather than to drive the load to accelerate. At this time, the system is in a "low signal-to-noise ratio" range. If inertia identification is forced, it is very easy to introduce large errors.

[0046] To quantify the effectiveness of current data for inertia identification, the homomorphic mapping and baseline reconstruction module 200 introduces the "inertia dominance" metric. First, the system calculates the current estimated friction torque based on preset nominal friction parameters. It should be noted that the nominal friction parameters here (including the viscosity coefficient) Coulomb friction torque This is not required to be exactly equal to the true value; it is usually set to zero or a typical empirical value (such as 1% to 3% of the rated torque) during system initialization, and is only used for order-of-magnitude estimation. The estimation uses a Coulomb-viscous composite model: ; In the formula, To provide feedback on mechanical angular velocity, The sign function is used. Based on this estimate, the inertia dominance index is calculated. Its core logic lies in assessing the proportion of "net acceleration torque" in "total electromagnetic torque": ; In the formula, This is the torque constant of the motor; The numerator represents the amplitude of the total electromagnetic torque; the physical meaning of the numerator is the effective torque component theoretically used to drive the acceleration of the inertial load after removing frictional losses. This is a numerical stability constant, and its value is recommended to be set to 10% of the motor's rated torque. -6 Up to 10 -4 The multiple is mainly used to prevent calculation overflow when the denominator is zero due to the motor being unloaded and stationary.

[0047] Calculated It is a normalized value: when When the value approaches 1, it indicates that the impact of friction is negligible, and the data is highly reliable; conversely, when... When the value approaches 0, it indicates that the torque is mainly used to counteract friction, resulting in low data confidence. This index will subsequently be used as a dynamic weight to adjust the step size of the parameter update algorithm, thereby achieving a robust strategy for high-gain identification only in the "inertia-dominated region".

[0048] See attached document Figure 1 The momentum residual calculation module 300 outputs a homomorphic torque command that is strictly aligned with the feedback signal in timing to the homomorphic mapping and reference reconstruction module 200. Subsequently, the main function of the momentum residual calculation module 300 is to construct an error observation equation with a high signal-to-noise ratio, thereby driving the subsequent parameter identification process. Its specific implementation method and logical details are as follows: S310, the construction principle of the integral domain observation mechanism. In existing inertia identification techniques, the instantaneous acceleration observation model based on Newton's second law (i.e., However, considering the characteristics of digital sampling control systems, the feedback mechanical angular velocity... The velocity signal is typically obtained by differential calculation from pulse counting of an optical encoder. This process inevitably introduces high-frequency quantization noise into the signal. If the velocity signal is directly differentiated twice to obtain the acceleration, according to the basic principles of signal processing, the noise component will be amplified by a square multiple of the frequency, easily causing the useful acceleration signal to be submerged in the noise background. Although introducing a low-pass filter can alleviate the noise problem to some extent, the inherent phase lag characteristic of the filter will disrupt the timing synchronization of both sides of the equation, thus affecting the identification accuracy.

[0049] Therefore, this embodiment chooses to use the momentum theorem to construct the observation mechanism in the integration domain. By utilizing the inherent low-pass filtering characteristic of integration operations in the frequency domain, i.e., the amplitude response decays as the frequency increases, high-frequency quantization noise tends to cancel each other out during the time integration process. This method preserves the low-frequency dynamic characteristics of the signal and improves the signal-to-noise ratio of the observed data without the need for additional filtering.

[0050] S320, impulse and momentum calculation based on a sliding window. To capture the dynamic changes of the system in real time, the momentum residual calculation module 300 does not integrate the data over the entire time period, but instead uses a fixed-length sliding window to perform local definite integral operations on the flowing data. This means that during the system initialization startup phase, the momentum residual calculation module 300 is internally configured with a waiting counter, which waits for the accumulated number of sampling points to reach a certain threshold. Exceeding window length Only when the data buffer is insufficient will subsequent calculation logic be triggered, in order to prevent calculation errors caused by insufficient data buffer filling.

[0051] The calculation process involves two parallel paths. On one hand, the momentum residual calculation module 300 calculates the momentum residual based on the homomorphic torque command. Calculate the length of the sliding window. The electromagnetic impulse applied to the system is denoted as the theoretical electromagnetic impulse. To adapt to the discrete computing environment of microprocessors, this embodiment uses the rectangular summation method for discretization, as shown in the following formula: ; In the formula, Represents the current sampling time; This is the motor's torque constant (N·m / A), which needs to be obtained in advance by reading the motor nameplate or through offline calibration; The sampling period of the servo controller (in seconds); This indicates the number of sampling points contained in the sliding window. Typically, the window length... The selection covers a time span from 5ms to 20ms (e.g., when...). At μs, The value ranges from 50 to 200, and the length is chosen to balance noise suppression effectiveness with response speed to dynamic changes.

[0052] On the other hand, the momentum residual calculation module 300 calculates the momentum residual based on the feedback mechanical angular velocity. and the estimated total inertia of the system at the previous moment Calculate the actual change in momentum of the system within the same time window, denoted as the estimated momentum increment. : ; In the formula, This is the estimated total moment of inertia updated in the previous control cycle; and These are the endpoint (current time) and the starting point (of the sliding window), respectively. The feedback mechanical angular velocity (one beat before) is used. This formula uses the difference in velocity at both ends of the time window to accurately express the change in momentum, thus mathematically avoiding the differential calculations that may be involved in the intermediate process.

[0053] S330, Generation and physical characterization of momentum residual signal. Based on the two physical quantities calculated above, the momentum residual calculation module 300 generates the current momentum residual signal. : ; The residual signal Physically, this represents the imbalance between the "product of input forces" and the "change in output momentum." In an ideal, frictionless environment with accurate inertia parameters, according to the momentum theorem, this difference should theoretically always be zero. If... A significant deviation from zero means that, based on the current parameters... The predicted system behavior deviates from actual observations. This deviation mainly consists of two parts: first, the error between the estimated inertia and the true value; and second, external load disturbances (such as friction and cutting forces) not included in the model. Therefore, The core error-driven signal is sent to the subsequent morphological decoupling and inference module 400 to further separate the error sources and correct the parameters.

[0054] See attached document Figure 1 The morphological decoupling and inference module 400 generates a momentum residual signal containing system error information in the momentum residual calculation module 300. Subsequently, the main function of the morphological decoupling and inference module 400 is to accurately distinguish whether the residual originates from a mismatch in the system's internal inertia parameters or from disturbances caused by external unmodeled loads by analyzing the time-domain waveform morphological characteristics of the signal, thereby providing a high-confidence decision basis for subsequent parameter correction. Its specific implementation method and logical details are as follows: S410, Generation of the composite reliability mask. To ensure the robustness of the parameter identification process under complex operating conditions, the system needs to pre-remove invalid or low signal-to-noise ratio data points before performing core inference calculations. Considering the physical characteristics of the inverter and the nonlinear effects of low-speed friction, this embodiment constructs a composite reliability mask to shield data in the current zero-crossing region and the strong friction region.

[0055] The morphological decoupling and inference module 400 calculates the current effective mask value in real time. This value is defined as a binary logic variable, and its generation logic is as follows: ; In the formula, This is the Heaviside step function, which outputs 1 when the input variable is greater than or equal to 0, and 0 otherwise. This is the absolute value of the torque current command; This is the preset current dead zone threshold. Regarding this threshold... The determination of this value mainly depends on the power device characteristics of the servo drive inverter and the PWM dead time. It is usually set to 1% to 2% of the motor's rated current. The purpose of setting this threshold is to avoid the nonlinear distortion region caused by the dead time effect when the current crosses zero. The inertia dominance index is calculated and output by the homomorphic mapping and baseline reconstruction module 200. The preset confidence threshold is preferably in the range of 0.6 to 0.8. This means that the current data is considered valid and allowed to proceed to the next stage only when more than 60% to 80% of the electromagnetic torque is clearly used to drive the acceleration of the inertial load.

[0056] S420, Vector Construction and Normalization Processing with Noise Floor Suppression. To extract statistically significant morphological features from time-domain waveforms, this embodiment abandons traditional single-point scalar calculations and instead employs a vectorized analysis method based on a short-time window. The module internally configures two lengths... shift register ( (This is a vector dimension, typically ranging from 5 to 10), and data is stored following a first-in, first-out (FIFO) principle. Specifically, the system constructs separate vectors containing the most recently accessed vectors. The residual vector of each sampling point With reference vector : ; ; Because signal amplitude can change significantly under different operating conditions, vector direction normalization is necessary to standardize the measurement. However, considering that the signal amplitude may be lower than the sensor's physical noise floor when the system is stationary or undergoing minor movements, direct division could lead to numerical overflow or randomization of the direction. Therefore, this embodiment introduces "noise floor suppression" logic.

[0057] Normalized reference vector The calculation method is as follows: ; In the formula, The L2 norm of a vector is used to represent the Euclidean norm. This is a preset noise floor threshold, the value of which needs to be rigorously determined based on the quantization accuracy of the position sensor. It is typically set to 3 to 5 times the momentum increment corresponding to the encoder's minimum resolution within a single sampling period. For the residual vector... The same logic is used to generate the normalized residual vector. .

[0058] S430, morphological similarity inference based on neuronal structure. After obtaining the normalized vector, the morphological decoupling and inference module 400 constructs a single-layer dynamic neuronal structure (or dynamic perceptron) to evaluate the morphological similarity of the waveform. The physical principle of this step is based on the following fact: if the generated momentum residual is caused by inertia parameter error, according to the error propagation model, the residual waveform... The trend of change should be related to the excitation source that caused the error (i.e., the theoretical impulse waveform). The residuals are highly correlated (in phase or out of phase); conversely, if the residuals are caused by external random disturbances (such as cutting chatter or frictional abrupt changes), since the disturbance signals are usually not causally related to the control commands, the two tend to be orthogonal or weakly correlated in terms of waveform morphology.

[0059] Based on this principle, the morphological decoupling and inference module 400 internally constructs a single-layer dynamic neuron structure (or dynamic perceptron). In this structure, the morphological decoupling and inference module 400 normalizes the reference vector... Defined as the "dynamic synaptic weights" of neurons, the normalized residual vector The confidence factor is calculated by defining the "input feature vector" of a neuron and performing vector inner product operations and linear / nonlinear activation processing. : ; In the formula, This represents the dot product operation of vectors. Geometrically, this formula is equivalent to calculating the cosine of the angle between two high-dimensional vectors.

[0060] Calculation results Reliability for quantization parameter updates: when When the value approaches 1, it indicates that the residual waveform is highly linearly correlated with the excitation waveform. The system determines that the current error is mainly due to the mismatch of the inertia parameters, and thus fully accepts the residual for parameter correction. when When the value approaches 0, it indicates that the residual waveform is independent of the excitation waveform shape. The system determines that the current error mainly originates from unmodeled external disturbances. At this point, it utilizes... As an attenuation coefficient, it automatically reduces or blocks parameter updates, thereby achieving the function of anti-disturbance.

[0061] See attached document Figure 1 The parameter update module 500 outputs a confidence factor that characterizes the morphological similarity of the signal to the morphological decoupling and inference module 400. Subsequently, the core task of the parameter update module 500 is to perform the final inertia parameter iteration and map these identification results into the gain parameters of the servo control loop in a safe and controllable manner. This embodiment aims to solve engineering challenges such as the tendency of traditional fixed-step-size algorithms to diverge under transient disturbances and the potential for mechanical shocks caused by parameter mutations by introducing a confidence-based variable-step-size mechanism and multiple safety boundary constraint strategies. Its specific implementation and logical details are as follows: S510, Confidence-Based Variable Step Size Adaptive Update. To find the optimal trade-off between parameter convergence speed and the overall system's robustness against disturbances, this embodiment constructs a dynamically adjusted variable step size update law. Its underlying physical logic lies in fully utilizing the reliability of the data to modulate the learning rate in real time: when the system is in the inertia-dominated region and the residual waveform closely matches the theoretical waveform, the algorithm increases the step size to quickly approximate the true value; conversely, if external disturbances are detected or signal characteristics are unclear, the system automatically decreases the step size to maintain parameter stability.

[0062] Specifically, the parameter update module 500 updates the inertia dominance based on the previously calculated inertia dominance. Morphological confidence factor Real-time synthesis of current adaptive update gain The calculation logic is as follows: ; In the formula, This represents the base learning rate, which is typically set to 10. -3 Up to 10 -2The magnitude and specific value need to be matched with the system's sampling frequency and the expected convergence time constant (usually designed to be from several hundred milliseconds to several seconds); and These are weighted indices for inertia dominance and morphological similarity, typically ranging from 1.5 to 2.0. This design utilizes the non-linear characteristics of the exponential function to significantly enhance the weight of high-confidence data while rapidly attenuating the influence of low-confidence data.

[0063] After determining the synthesis gain Then, the system immediately performs the inertia estimation. Iterative updates: ; In the formula, This is the estimated value of the inertia at the previous moment; This is the momentum residual signal; This represents the change in mechanical angular velocity within the current sliding window; It is a symbolic function; To prevent the use of tiny constants that divide by zero, the physical meaning of this formula is based on the principle of linearization of momentum deviation within a short-time window: given the momentum residual within a small observation window... Correction item Essentially, it represents the instantaneous observation of inertia error. By introducing... and dynamic gain This not only avoids the situation when the motor speed is constant (i.e. The system calculates the risk of divergence and ensures that parameters are updated only at valid moments when the signal-to-noise ratio is high and the shape is matched.

[0064] S520, Safety Boundary Projection and Integral Saturation Suppression. Considering that in actual physical systems, the load inertia of a motor is limited by its physical structure and inevitably has a definite numerical range, this embodiment introduces a projection algorithm to strictly constrain the estimation results to prevent algorithm divergence caused by the accumulation of numerical calculation errors or extreme operating conditions (such as instantaneous loss of encoder signals).

[0065] The system has a preset lower limit for inertia. With the upper limit of inertia .in, It is generally set to the rotor inertia of the motor itself, and this parameter can be obtained directly from the motor datasheet; This value is typically set based on the maximum allowable load of the mechanical structure, generally taking 30 to 50 times the motor rotor inertia. The parameter update module 500 executes the following decision logic after each iteration: ; By forcing the estimated values ​​to be within the physical feasible region This step effectively prevents the integral saturation phenomenon common in adaptive algorithms, thus ensuring that the identification results are always kept within a reasonable range of physical meaning.

[0066] S530, gain mapping and rate limiting protection. Identified inertia parameters. Ultimately, this needs to be converted into control parameters for the servo driver speed loop, namely the proportional gain. With integral gain However, directly applying the identification results of real-time fluctuations to the control loop could potentially cause a sudden change in gain, leading to a step jump in the motor's output torque and causing impacts or even squealing in the mechanical structure. Therefore, this embodiment incorporates a rate limiting protection mechanism in the mapping stage.

[0067] The parameter update module 500 first updates the parameter based on the preset speed loop bandwidth. (Unit: rad / s) Calculate the target proportional gain : ; Regarding speed loop bandwidth The selection of ...

[0068] The target gain is then used to generate the final execution gain via a ramp limiter. Its logical description is as follows: ; In the formula, It is a saturation function, that is, when the input... The absolute value exceeds Time output The sign value is not specified; otherwise, output the value. itself; This is the maximum allowable gain change within a single control cycle. This threshold must be set to ensure that torque fluctuations caused by gain changes are less than 0.5% to 1% of the motor's rated torque, thus ensuring both timely parameter updates and smooth transition of control parameters. Integral gain The update follows a similar smooth mapping logic, based on ( Synchronous adjustment is performed for the integral time constant.

[0069] To enable those skilled in the art to more fully understand the technical details and practical application effects of the present invention, the embodiments of the present invention will be further described below in conjunction with a specific industrial application scenario, namely, the Z-axis servo system of a PCB CNC drilling machine. It should be noted that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0070] I. Application Scenario Configuration and System Initialization; In this embodiment, the parameter self-tuning system provided by the present invention is integrated into the Z-axis vertical motion controller of a high-performance PCB drilling machine. This working condition is highly representative, characterized by the need for the Z-axis to perform high-frequency reciprocating motion (pecking action) at frequencies ranging from several hertz to tens of hertz, and the sudden change in cutting force during the processing, from unloaded acceleration to the moment the drill bit contacts the board, poses an extremely high challenge to the anti-disturbance capability of parameter identification.

[0071] In terms of system hardware, a rated thrust is selected. The N-type permanent magnet synchronous linear motor serves as the actuator, with a mover base mass of approximately 5.0 kg. Position feedback utilizes a high-precision optical grating ruler with a resolution of 0.1 μm. The control algorithm runs in an FPGA-based servo drive unit, with a current loop sampling period of... Set to 100μs.

[0072] During the static initialization phase of the system upon power-up, the state monitoring and gating module 100 first injects a high-frequency pseudo-random binary sequence with an amplitude of 5% of the rated thrust into the current loop. By monitoring the response of the feedback velocity and performing cross-correlation analysis, the system automatically locks the total loop delay parameter. Under the test conditions of this embodiment, the delay was calibrated to 3 control cycles, or 300 μs. This parameter was then read by the homomorphic mapping and reference reconstruction module 200 to construct a digital filter that is strictly synchronized with the physical system. This process lays the timing reference for subsequent high-precision identification.

[0073] II. Parameter self-tuning logic in high-frequency machining process; As the machine tool enters the automated machining process, the Z-axis begins to perform a drilling motion downwards with an acceleration of 2g. At this time, the internal operating logic of the system is as follows: When the Z-axis is in the high-speed downward acceleration phase, the state monitoring and gating module 100 calculates the high-frequency energy characteristics of the torque command in real time. Because the acceleration is extremely high at this point, the calculated energy far exceeds the preset activation threshold. (Set as) This triggers the parameter tuning process by increasing the rated energy by multiples of the original thrust command. Simultaneously, the homomorphic mapping module uses the previously calibrated delay parameters and low-pass filter coefficients to reconstruct the original thrust command into a homomorphic torque command. This ensures that it is perfectly aligned with the speed signal fed back by the grating ruler on the time axis.

[0074] In the momentum residual calculation stage, this embodiment does not employ a noise-sensitive differential algorithm. Instead, the momentum residual calculation module 300 selects a sliding window of length L=100 (i.e., 10ms) to calculate the theoretical electromagnetic impulse based on the command. Compared with actual momentum increments based on feedback The difference between the two generates the momentum residual signal. It is worth noting that in the initial stage of processing, the preset inertia parameters may deviate from the actual load (including the mass of the drill bit chuck). It usually exhibits a significant non-zero value.

[0075] The most critical disturbance rejection judgment occurs at the moment the drill bit contacts the PCB board.

[0076] At this point, the significant cutting resistance intervenes in the system, leading to momentum residuals. Dramatic fluctuations occurred. Using the traditional recursive least squares (RLS) method, this drag could easily be misinterpreted as a sudden increase in inertia. However, in this embodiment, the morphological decoupling and inference module 400 plays a crucial role. This module extracts the residual vectors from the most recent eight sampling points. With reference vector And perform normalized morphological analysis.

[0077] Since the cutting force is a random high-frequency disturbance, its waveform characteristics have no causal relationship with the smooth electromagnetic thrust command, resulting in the two being nearly orthogonal in the vector space. The calculated morphological similarity... It dropped sharply to around 0.15. Based on this "low confidence" judgment, the parameter update module 500 automatically adjusted the adaptive gain. The parameters decay to near zero, thus "freezing" the parameter updates. This mechanism ensures that the system learns only in the "inertia-dominated region" where the morphology is highly consistent, while maintaining parameter stability in the cutting interference region.

[0078] III. Experimental verification and data comparison; To quantitatively evaluate the performance of this invention, a full-condition comparison was conducted with the traditional recursive least squares method with forgetting factor (FF-RLS).

[0079] Experimental results show that this invention demonstrates advantages in handling grating ruler quantization noise and resisting cutting force interference. Specific performance indicators are compared in Table 1 below: Table 1: Comparison of parameter identification performance under PCB drilling conditions

[0080] As can be seen from the table above, thanks to the integral domain observation mechanism, the parameter fluctuations in this embodiment are minimal in steady state; and when faced with cutting disturbances, compared with the severe parameter drift that occurs in traditional schemes, this scheme controls the deviation to within 3.8%, basically achieving immunity to non-modeled disturbances.

[0081] Based on the above application embodiments, the following describes the appendix. Figures 3 to 5 A detailed explanation will be provided to visually demonstrate the technical effects of this invention.

[0082] See attached document Figure 3 This figure visually illustrates the difference in response between the embodiment of the present invention and the traditional method during the dynamic process of parameter identification. The horizontal axis represents time (seconds), and the vertical axis represents the estimated moment of inertia (kg·m). 2 The black dashed line represents the actual physical inertia. Observing the system's static interval from 0s to 0.5s, it is evident that the traditional FF-RLS method, which involves the second differentiation of the position signal, greatly amplifies the quantization noise of the grating ruler, resulting in a sawtooth-like high-frequency oscillation in the estimation curve. In contrast, the proposed solution utilizes the integral smoothing characteristic of the momentum theorem to effectively filter out high-frequency background noise, resulting in a smooth curve that closely matches the true value. When the time advances to 1.0s, the simulation addresses the situation where the operator changes the heavy drill bit chuck, causing a step jump in the actual inertia. At this point, the proposed solution, benefiting from the confidence-based variable step size mechanism, responds rapidly after detecting a high-confidence residual, achieving convergence in only 45ms with no significant overshoot. In contrast, the traditional method, due to the hysteresis effect of the forgetting factor, exhibits a sluggish convergence process, taking approximately 120ms to stabilize.

[0083] See attached document Figure 4 This figure focuses on analyzing the robustness of the system under strong external disturbances. Figure (a) depicts the cutting resistance waveform with an amplitude of approximately 1.5 Nm and random high-frequency vibration characteristics generated at t=1.5 when the drill bit contacts the PCB board. Figure (b) records the drastically different responses of the two algorithms under this disturbance. For traditional methods, because the algorithm cannot distinguish between the "force driving inertia" and the "resistance to overcome cutting" from a physical mechanism perspective, it incorrectly interprets the huge cutting force as a dramatic increase in system inertia, causing the estimated value to spike to more than 3.5 times the true value instantly. This divergence is very likely to cause oscillations in practical engineering. However, in this invention, the morphological decoupling module keenly captures the morphological mismatch between the residual waveform and the command waveform at the instant the disturbance occurs, and quickly cuts off the parameter update path. Therefore, throughout the entire cutting process, the curve of this invention remains as stable as a horizontal line, confirming the shielding capability of the "morphological decoupling" mechanism against non-modeled disturbances.

[0084] See attached document Figure 5This figure further reveals the internal mechanism by which the present invention achieves the aforementioned intelligent decision-making. The solid line in the figure represents the inertia dominance index. The dashed line represents the shape similarity index. Observing the timing changes reveals that at the 1.0s parameter abrupt change, due to the residual originating from internal system parameter mismatch, the residual waveform changes in phase with the command waveform. The value rapidly climbed to a high level of 0.98, which the system identified as an "internal factor" and updated the parameters at full speed. However, at the 1.5s cutting disturbance point, although the momentum residual amplitude was large, the randomness of the disturbance waveform meant that... The value plummeted to below 0.15. This combination of "high residual and low similarity" was accurately identified by the system as external interference, thus forcibly blocking parameter correction. The dynamic interplay of these two curves vividly illustrates the logical reasoning process of this system, similar to that of a neuron.

[0085] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A self-tuning system for electrical drive system parameters based on neural networks, characterized in that, include: The status monitoring and gating module is connected to the servo driver to synchronously acquire torque current commands and feedback mechanical angular velocity, and generates an activation signal when the high-frequency energy characteristics of the torque current command exceed a preset threshold. The homomorphic mapping and reference reconstruction module is used to construct a homomorphic digital filter based on the calibrated loop delay parameters, convert the torque current command into a homomorphic torque command that is aligned with the feedback mechanical angular velocity in time, and output the inertia dominance index. The momentum residual calculation module is used to calculate the theoretical electromagnetic impulse and the actual momentum increment within a preset sliding window based on the momentum theorem, and generate the difference between the theoretical electromagnetic impulse and the actual momentum increment as the momentum residual signal. The morphological decoupling and reasoning module is used to construct normalized residual vectors and normalized reference vectors, and to calculate the morphological similarity between the normalized residual vectors and the normalized reference vectors through the internally constructed neuron structure to output confidence scores. The parameter update module is used to adjust the adaptive update gain according to the confidence level value, iteratively update the inertia parameter of the system using the momentum residual signal, and map the inertia parameter to the velocity loop control gain.

2. The self-tuning system for electrical drive system parameters based on neural networks according to claim 1, characterized in that, The state monitoring and gating module is configured to perform cross-correlation analysis logic: The state monitoring and gating module injects a high-frequency pseudo-random binary sequence into the current loop and records the response of the feedback mechanical angular velocity; The state monitoring and gating module calculates the discrete cross-correlation function between the torque current command sequence and the feedback mechanical angular velocity sequence, and searches for the global peak time of the discrete cross-correlation function to lock the loop delay parameter.

3. The self-tuning system for electrical drive system parameters based on neural networks according to claim 1, characterized in that, The state monitoring and gating module calculates the high-frequency energy characteristics in the following way: Calculate the sum of squared differences of the torque current command within the current sliding window, and use the sum of squared differences as the high-frequency energy characteristic; The preset threshold is a dynamic variable set based on the motor's rated current; when the high-frequency energy characteristic is less than the preset threshold, the status monitoring and gating module outputs a suspension signal.

4. The self-tuning system for electrical drive system parameters based on neural networks according to claim 1, characterized in that, The homomorphic mapping and benchmark reconstruction module is equipped with a parameter monitoring mechanism for reading the filter configuration of the servo driver; The homomorphic digital filter includes a cascaded low-pass filter, a notch filter, and a pure time delay based on a circular queue; the homomorphic mapping and reference reconstruction module uses the homomorphic digital filter to generate the homomorphic torque command.

5. The self-tuning system for electrical drive system parameters based on neural networks according to claim 1, characterized in that, The homomorphic mapping and benchmark reconstruction module is also used to calculate the inertia dominance index: The homomorphic mapping and benchmark reconstruction module calculates and estimates the friction torque based on the nominal friction parameters; The homomorphic mapping and benchmark reconstruction module calculates the proportion of the difference between the total electromagnetic torque and the estimated frictional torque in the total electromagnetic torque, and obtains the normalized inertia dominance index.

6. The self-tuning system for electrical drive system parameters based on neural networks according to claim 1, characterized in that, The momentum residual calculation module adopts an integral domain observation mechanism: The homomorphic torque command is discretely summed within the sliding window to obtain the theoretical electromagnetic impulse; The difference between the feedback mechanical angular velocity at both ends of the sliding window is calculated, and the actual momentum increment is obtained by combining it with the inertia parameter at the previous moment.

7. The self-tuning system for electrical drive system parameters based on neural networks according to claim 1, characterized in that, The execution logic of the neuron structure in the morphological decoupling and reasoning module is as follows: The normalized baseline vector is mapped to the dynamic synaptic weights of the neuron structure, and the normalized residual vector is mapped to the input feature vector of the neuron structure. The dynamic synaptic weights and the input feature vector are subjected to a vector inner product operation to obtain the confidence value; the confidence value is used to quantify the probability that the momentum residual signal originates from internal parameter mismatch of the system.

8. The neural network-based electric drive system parameter self-tuning system according to claim 7, characterized in that, The morphological decoupling and inference module performs a reliability masking process before constructing the vector: An effectiveness mask is generated by combining the absolute value of the torque current command with the inertia dominance index calculated by the homomorphic mapping and reference reconstruction module. The current data is allowed to be used to construct the normalized residual vector and the normalized reference vector only when the torque current command avoids the current dead zone and the inertia dominance index is higher than a preset confidence threshold.

9. The self-tuning system for electrical drive system parameters based on neural networks according to claim 1, characterized in that, The parameter update module is configured to perform variable step size adaptive updates: The parameter update module obtains the inertia dominance index calculated by the homomorphic mapping and benchmark reconstruction module, and obtains the confidence value output by the morphological decoupling and inference module; The parameter update module uses the inertia dominance index and the confidence value to perform weighted modulation on the base learning rate to generate the adaptive update gain. The parameter update module uses adaptive update gain, momentum residual signal, and mechanical angular velocity change to calculate inertia correction.

10. The neural network-based self-tuning system for electrical drive system parameters according to claim 1, characterized in that, The parameter update module performs safety boundary projection and gain mapping: The updated inertia parameters are forcibly constrained within the physically feasible region set based on the motor rotor inertia. Based on the preset velocity loop bandwidth, the constrained inertia parameters are converted into target proportional gain, and the final execution gain is generated through the ramp limiter.

Images