A fan noise reduction method, system, and medium

Abstract

A fan noise reduction method, system and medium, comprising collecting operation data of a fan motor; preprocessing the operation data, and respectively performing time domain feature extraction and frequency domain feature extraction on the preprocessed operation data to generate a feature vector set; standardizing the feature vector set, and according to a pre-set sliding window, intercepting the standardized feature vector set to obtain a model input matrix; inputting the model input matrix into a pre-trained load prediction model, and through model forward propagation calculation, outputting a load change trend, a mutation amplitude and a mutation time; based on the load change trend, the mutation amplitude, the mutation time and real-time electrical data, through an optimization algorithm, obtaining optimized magnetic field oriented control parameters and auxiliary control parameters and issuing them to the fan motor to adjust the motor operating state. The present application realizes the advance prediction of the fan load change trend, mutation amplitude and mutation time, and effectively improves the fan noise reduction effect.

Description

Technical Field

[0001] This invention relates to the field of high-speed motor technology, and in particular to a method, system and medium for reducing noise in fans. Background Technology

[0002] The noise generated during the operation of high-speed motors is becoming increasingly prominent, not only affecting the physical and mental health of operators but also accelerating equipment wear. To address the noise problem of high-speed motors, existing technologies have developed various noise reduction solutions, mainly divided into two categories: passive noise reduction and active noise reduction.

[0003] Traditional passive noise reduction technologies, such as soundproof enclosures and damping materials, employ fixed structural designs, and their noise reduction performance is optimized for specific operating conditions. When fluctuations in the fan load cause changes in the motor's operating state, resulting in changes in noise frequency and amplitude, passive noise reduction structures cannot adjust in real time, and their effectiveness in suppressing instantaneous noise peaks caused by sudden load changes is poor.

[0004] Traditional active noise cancellation technologies often employ a closed-loop "detect first, control later" approach, using sensors to collect noise or vibration signals and then generating control commands to adjust motor parameters. However, in scenarios with rapidly fluctuating wind turbine loads, there is an unavoidable time delay in the acquisition, transmission, and processing of noise signals. This causes the control command output to lag behind noise changes, making it impossible to suppress instantaneous noise caused by sudden load changes in a timely manner.

[0005] Furthermore, existing technologies do not consider the advance prediction of wind turbine load changes, and can only respond passively after load changes cause noise. They fail to take advantage of the correlation and trend of load changes to adjust motor parameters in advance, resulting in poor overall noise reduction effect. Summary of the Invention

[0006] To address the problem of inadequate noise reduction performance in existing technologies, this invention provides a method for reducing fan noise, comprising the following steps:

[0007] Collect operating data of the wind turbine motor; the operating data includes electrical data, mechanical vibration data, and environmental data; The running data is preprocessed, and time-domain and frequency-domain features are extracted from the preprocessed running data to generate a feature vector set. The feature vector set is standardized, and the standardized feature vector set is truncated according to a preset sliding window to obtain the model input matrix; The model input matrix is ​​input into the pre-trained load prediction model, and the load change trend, mutation magnitude and mutation time are output through forward propagation calculation. Based on the load change trend, abrupt change amplitude, abrupt change time, and real-time electrical data, optimized field-oriented control parameters and auxiliary control parameters are calculated through an optimization algorithm and sent to the fan motor to adjust the motor's operating status.

[0008] Optionally, the electrical data includes three-phase phase current, bus voltage, estimated rotational speed, and rotor angle; The mechanical vibration data includes X-axis vibration acceleration, Y-axis vibration acceleration, Z-axis vibration acceleration, and vibration frequency spectrum; The environmental data includes winding temperature and ambient temperature.

[0009] Optionally, the preprocessing includes outlier removal, multi-source data synchronization and alignment, and low-pass filtering for noise reduction.

[0010] Optionally, temporal feature extraction is performed on the preprocessed runtime data, including: The mean, variance, peak value, peak-to-peak value, and kurtosis of the preprocessed running data are calculated to obtain statistical characteristics; The dynamic characteristics are obtained by calculating the ripple coefficient of the current and the root mean square value of the vibration in the preprocessed running data. Based on the statistical and dynamic features, a time-domain feature vector is obtained.

[0011] Optionally, frequency domain feature extraction is performed on the preprocessed running data, including: Perform a Fast Fourier Transform or Wavelet Transform on the preprocessed running data to convert it into frequency domain data; Identify spectral peaks and their corresponding frequencies from the frequency domain data, and calculate the energy percentage of a preset frequency band; When using Fast Fourier Transform, a frequency domain feature vector is obtained based on the energy proportion, spectral peak value and its corresponding frequency; When wavelet transform is used, the energy value and entropy value of each scale are extracted from the wavelet transform coefficients, and the frequency domain feature vector is obtained based on the energy ratio, energy value, entropy value, spectral peak value and its corresponding frequency.

[0012] Optionally, the training steps of the load prediction model include: Obtain the historical dataset; the historical dataset includes historical feature vectors and corresponding historical payload values; The historical dataset is truncated using the sliding window method to generate training samples with temporal correlation. The training samples are divided into a training set, a validation set, and a test set; Construct an initial prediction model; the initial prediction model is a long short-term memory network, a gated recurrent unit, or a temporal convolutional network. The initial prediction model is iteratively trained using the training set, and the model hyperparameters are tuned using the validation set to obtain the load prediction model.

[0013] Optionally, based on the load change trend, abrupt change amplitude, abrupt change time, and real-time electrical data, optimized field-oriented control parameters and auxiliary control parameters are calculated using an optimization algorithm, including: Based on the load change trend, abrupt change amplitude, abrupt change time, and real-time electrical data, optimized field-oriented control parameters and auxiliary control parameters are dynamically calculated using reinforcement learning algorithms or model predictive control algorithms. Verify the magnetic field orientation control parameters and auxiliary control parameters to ensure that they meet the preset operating constraints.

[0014] Optionally, the field-oriented control parameters include excitation current command, torque current command, proportional coefficient of proportional-integral controller, integral coefficient of proportional-integral controller, and pulse width modulation frequency; the auxiliary control parameters include notch filter frequency and field weakening control coefficient.

[0015] Corresponding to the aforementioned fan noise reduction method, the present invention provides a fan noise reduction system, comprising: The data acquisition module is used to collect operating data of the wind turbine motor; the operating data includes electrical data, mechanical vibration data, and environmental data. The preprocessing module is used to preprocess the running data and extract time-domain features and frequency-domain features from the preprocessed running data to generate a feature vector set. The standardization processing module is used to standardize the feature vector set, and to extract the standardized feature vector set according to a preset sliding window to obtain the model input matrix and input it to the load prediction module. The load prediction module is used to calculate and output the load change trend, mutation magnitude, and mutation time through model forward propagation. The noise reduction processing module is used to calculate optimized magnetic field orientation control parameters and auxiliary control parameters based on the load change trend, abrupt change amplitude, abrupt change time and real-time electrical data, and send them to the fan motor to adjust the motor's operating status.

[0016] In addition, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a fan noise reduction program, which, when executed by a processor, implements the steps of the fan noise reduction method as described above.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) By collecting multi-dimensional operating data such as electrical data, mechanical vibration data and environmental data, and after time-domain and frequency-domain feature extraction and standardization, a model input matrix with time-series correlation is constructed and input into the load prediction model to realize the prediction of the wind turbine load change trend, sudden change amplitude and sudden change time, which effectively improves the wind turbine noise reduction effect; Based on the prediction results and real-time electrical data, the optimized magnetic field orientation control parameters and auxiliary control parameters are calculated by the optimization algorithm and sent to the motor, which can actively adjust the motor operating state before the load sudden change causes noise, forming a feedforward control mechanism.

[0018] (2) This invention constructs a multi-physics coupled data acquisition system: electrical data directly reflects the electromagnetic state of the motor and its correlation with the load, mechanical vibration data captures the mechanical root causes of noise, and environmental data characterizes the impact of thermal effects on motor characteristics. The three types of data work together to achieve full-dimensional perception of the motor's operating state, providing a high signal-to-noise ratio and high correlation data foundation for subsequent feature extraction and load prediction, and significantly improving the input quality and prediction accuracy of the load prediction model.

[0019] (3) Outlier removal eliminates outliers caused by sensor malfunctions or electromagnetic interference; multi-source data synchronization and alignment resolves timestamp misalignment caused by differences in sampling rates between different sensors; and low-pass filtering removes high-frequency electromagnetic noise and mechanical harmonic interference. These three preprocessing operations work synergistically to improve the quality and ensure temporal consistency of multi-source heterogeneous data. This preprocessing process ensures the accuracy of subsequent feature extraction, avoids prediction bias caused by outlier data or temporal misalignment, and provides data quality assurance for the stable operation and reliable output of the load prediction model.

[0020] (4) By calculating statistical characteristics such as mean, variance, peak value, peak-to-peak value, and kurtosis, the static distribution characteristics and abnormal fluctuation tendency of the operating data are captured; by calculating dynamic characteristics such as current ripple coefficient and root mean square value of vibration, the degree of current harmonic distortion and vibration energy level are quantified. The statistical and dynamic characteristics work together to construct a time-domain feature vector that has both steady-state description capability and dynamic response sensitivity. It can effectively characterize the amplitude, rate and degree of abnormality of load changes, provide a low-dimensional representation of the essential law of load state for the load prediction model, and improve the model's ability to identify the precursory features of load changes.

[0021] (5) By identifying the peak frequency and its corresponding frequency and calculating the energy ratio of the preset frequency band, the accurate location of the noise main frequency and energy distribution can be achieved; when using wavelet transform, the time-frequency localization characteristics of non-stationary signals are captured by extracting the energy values ​​and entropy values ​​at each scale. The frequency domain features and time domain features work together to construct a joint feature representation system, which comprehensively reflects the distribution law of load changes in the frequency domain and the energy transfer characteristics, and significantly enhances the load prediction model's ability to identify the load evolution trend under complex working conditions.

[0022] (6) By using the sliding window method to extract historical datasets, training samples with temporal correlation are generated, enabling the model to learn the time dependence and evolution of load changes; by dividing the training set, validation set and test set, the process of model training, hyperparameter tuning and generalization ability evaluation is decoupled; Long short-term memory network, gated recurrent unit or temporal convolutional network are selected as the initial prediction model, and their gating mechanism or causal convolutional structure is used to model the long-term temporal dependence to achieve high-precision prediction of load change trend, mutation amplitude and mutation time.

[0023] (7) Based on load change trends, abrupt change amplitudes, abrupt change times, and real-time electrical data, the adaptive dynamic optimization of control parameters is achieved through reinforcement learning algorithms, utilizing their trial-and-error learning and strategy optimization capabilities; or the optimal control sequence is solved under multi-step prediction through model predictive control algorithms, utilizing their rolling optimization and constraint handling capabilities. By verifying the preset operating constraints, it is ensured that the optimized field-oriented control parameters and auxiliary control parameters are within the physical limits and safe operating boundaries of the motor. The predictive information and optimization algorithm work together to achieve closed-loop feedforward control from "load prediction" to "parameter optimization" and then to "control execution," maximizing the technical advantages of predictive noise reduction while ensuring the safe operation of the motor.

[0024] (8) Magnetic field orientation control parameters include excitation current command, torque current command, proportional coefficient of proportional-integral controller, integral coefficient of proportional-integral controller, and pulse width modulation frequency. These parameters directly affect the decoupling control core of motor flux linkage and torque, achieving precise control of the root cause of motor electromagnetic noise. Auxiliary control parameters include notch frequency and field weakening control coefficient. The two types of parameters work together to form a joint optimization mechanism of main control loop parameters and auxiliary control loop parameters: the adjustment of excitation current command and torque current command changes the electromagnetic excitation characteristics of the motor, suppressing electromagnetic noise at the source; the adjustment of PI controller parameters optimizes current tracking performance and reduces additional noise caused by control error; the adjustment of PWM frequency balances switching losses and current harmonics, suppressing high-frequency carrier noise; the setting of notch frequency effectively filters out resonant noise in specific frequency bands, while the optimization of field weakening control coefficient reduces the iron loss and noise radiation of the motor in the high-speed operating region. Through the joint optimization of the above multi-dimensional parameters, the fine adjustment of the operating state of the wind turbine motor is achieved, and the generation and propagation of noise are suppressed from multiple levels such as electromagnetic, mechanical and control. Attached Figure Description

[0025] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a simplified flowchart of an embodiment of the fan noise reduction method of the present invention; Figure 2 This is a technical schematic diagram of an embodiment of the fan noise reduction method of the present invention; Figure 3 This is a framework diagram of an embodiment of the fan noise reduction system of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of 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, not all embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. 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.

[0027] like Figure 1 As shown, a fan noise reduction method of the present invention includes the following steps: Collect operating data of the wind turbine motor; the operating data includes electrical data, mechanical vibration data, and environmental data; The running data is preprocessed, and time-domain and frequency-domain features are extracted from the preprocessed running data to generate feature vector sets. The feature vector set is standardized, and the standardized feature vector set is truncated according to a preset sliding window to obtain the model input matrix; The model input matrix is ​​fed into a pre-trained load prediction model. Through forward propagation, the model outputs the load change trend, abrupt change magnitude, and abrupt change time. Preferably, the load change trend is represented by a 0-1 value (0 = stable load, 0.3-0.7 = slow load fluctuation, 0.7-1 = load about to change abruptly). The abrupt change magnitude is used to predict the magnitude of the load change (unit: percentage of the load baseline, such as +20%, -15%). The abrupt change time is used to predict the specific moment when the load change occurs (unit: ms, advance prediction time = 50-200ms, adapting to parameter adjustment response time). Based on load change trends, abrupt change amplitudes, abrupt change times, and real-time electrical data, optimized field-oriented control parameters and auxiliary control parameters are calculated using optimization algorithms and sent to the fan motor to adjust the motor's operating status.

[0028] In this embodiment, the standardization of the feature vector set is specifically Z-score standardization (mean=0, variance=1) to eliminate the dimensional differences of different physical quantities (such as current, vibration, and temperature); the synchronized data is segmented by a preset sliding window (such as 1s / segment), with each segment containing 1000-5000 data points to facilitate subsequent feature calculation.

[0029] This invention collects multi-dimensional operational data, including electrical data, mechanical vibration data, and environmental data. After time-domain and frequency-domain feature extraction and standardization, it constructs a time-series-related model input matrix and inputs it into a load prediction model. This enables early prediction of the wind turbine load change trend, abrupt change amplitude, and abrupt change time, effectively improving the wind turbine noise reduction effect. Based on the prediction results and real-time electrical data, an optimized magnetic field-oriented control parameter and auxiliary control parameter are calculated through an optimization algorithm and sent to the motor. This allows the motor operating state to be proactively adjusted before load abrupt changes cause noise, forming a feedforward control mechanism.

[0030] In this embodiment, the electrical data includes three-phase current (current harmonics can cause electromagnetic noise), bus voltage, estimated speed, and rotor angle; wherein, the three-phase current and bus voltage are both collected as instantaneous values, and the sampling frequency is synchronized with the PWM frequency; the estimated speed refers to the real-time speed (unit: rpm) and rotor angle (unit: rad) estimated by the motor control chip, and the timing data is output at a frequency of 10ms / time; Mechanical vibration data includes X-axis, Y-axis, and Z-axis vibration accelerations and vibration frequency spectra from a MEMS accelerometer (Micro-Electro-Mechanical System Accelerometer); the sampling frequency is set to 1kHz-5kHz to obtain continuous time series vibration amplitude data. Environmental data includes winding temperature (temperature changes affect motor structural noise) and ambient temperature.

[0031] This invention constructs a multi-physics coupled data acquisition system: electrical data directly reflects the electromagnetic state of the motor and its correlation with the load; mechanical vibration data captures the mechanical root causes of noise generation; and environmental data characterizes the impact of thermal effects on motor characteristics. These three types of data work synergistically to achieve a comprehensive perception of the motor's operating state, providing a high signal-to-noise ratio and high correlation data foundation for subsequent feature extraction and load prediction, significantly improving the input quality and prediction accuracy of the load prediction model.

[0032] In this embodiment, the preprocessing includes outlier removal, multi-source data synchronization and alignment, and low-pass filtering for noise reduction.

[0033] Preferably, outlier removal processing uses 3 The criteria (mean ± 3 standard deviations) identify extreme outliers in the X-axis / Y-axis / Z-axis vibration acceleration and three-phase current (such as sudden data changes caused by sensor failures), and replace outliers with linear interpolation of adjacent data.

[0034] In this embodiment, the multi-source data synchronization and alignment processing uses the rotor angle signal as the time reference to unify the timestamps of the mechanical vibration data (error ≤ 10μs) in order to solve the problem of multi-sensor acquisition delay; the low-pass filtering noise reduction processing uses a Butterworth low-pass filter (cutoff frequency set to 500Hz) to filter out high-frequency electromagnetic interference and sensor inherent noise, while retaining low-frequency vibration and current ripple components related to motor operation.

[0035] This invention eliminates outliers caused by sensor malfunctions or electromagnetic interference through outlier removal, resolves timestamp misalignment issues due to differences in sampling rates between different sensors through multi-source data synchronization and alignment, and filters out high-frequency electromagnetic noise and mechanical harmonic interference through low-pass filtering. These three preprocessing operations work synergistically to improve the quality and ensure temporal consistency of multi-source heterogeneous data. This preprocessing workflow ensures the accuracy of subsequent feature extraction, avoids prediction biases caused by outliers or temporal misalignments, and provides data quality assurance for the stable operation and reliable output of the load prediction model.

[0036] In this embodiment, temporal feature extraction is performed on the preprocessed running data, including: The mean (reflecting the average intensity of vibration / current), variance (reflecting the degree of data fluctuation), peak value (maximum vibration acceleration / maximum current value, reflecting instantaneous impact intensity), peak-to-peak value (the difference between the maximum and minimum values, reflecting the range of vibration amplitude), and kurtosis (to determine whether there is an impact component in the vibration; preferably, when the kurtosis is >3, it indicates the presence of abrupt impact) of the preprocessed running data are calculated to obtain statistical characteristics. The dynamic characteristics are obtained by calculating the ripple coefficient of the current (the ratio of the peak-to-peak value of the current to the effective value, reflecting the stability of the current) and the root mean square value of the vibration (RMS, reflecting the magnitude of the vibration energy) in the preprocessed running data. Based on statistical and dynamic characteristics, time-domain feature vectors are obtained; each data segment corresponds to one set of time-domain feature vectors, such as "X-axis vibration mean: 0.8 m / s², X-axis vibration variance: 0.05 m² / s²". 4 "Phase A current peak: 12.5A, current ripple coefficient: 8%", used for subsequent model input or preliminary judgment of noise sources (e.g., high kurtosis value indicates possible mechanical shock noise).

[0037] This invention captures the static distribution characteristics and abnormal fluctuation trends of operating data by calculating statistical features such as mean, variance, peak value, peak-to-peak value, and kurtosis; and quantifies the degree of current harmonic distortion and vibration energy level by calculating dynamic features such as current ripple coefficient and root mean square value of vibration. The synergistic effect of statistical and dynamic features constructs a time-domain feature vector that combines steady-state description capability with dynamic response sensitivity. This effectively characterizes the amplitude, rate, and degree of anomaly of load changes, providing a low-dimensional representation of the essential laws of load state for load prediction models and improving the model's ability to identify precursory features of load abrupt changes.

[0038] In this embodiment, frequency domain feature extraction is performed on the preprocessed running data, including: Perform a Fast Fourier Transform or Wavelet Transform on the preprocessed running data to convert it into frequency domain data; Identify the spectral peaks and their corresponding frequencies (e.g., 200Hz, 350Hz) from the frequency domain data, corresponding to the motor bearing resonance frequency and current harmonic frequencies (e.g., 5th and 7th harmonics), and calculate the energy proportion of preset frequency bands (e.g., low frequency band 10-100Hz, mid frequency band 100-300Hz, high frequency band 300-500Hz) to determine the main source of noise (e.g., a high energy proportion in the low frequency band indicates mechanical noise, and a high proportion in the mid frequency band indicates electromagnetic noise). When using the Fast Fourier Transform, the frequency domain feature vector is obtained based on the energy ratio, spectral peak and its corresponding frequency; When wavelet transform is used, the energy and entropy values ​​at each scale are extracted from the wavelet transform coefficients to reflect the frequency change characteristics of non-stationary signals (such as a significant increase in wavelet entropy values ​​in the high-frequency band when the load changes). Based on the energy ratio, energy value, entropy value, spectral peak and its corresponding frequency, the frequency domain feature vector is obtained. Each data segment corresponds to a set of frequency domain feature vectors, such as "FFT peak frequency: 220Hz (amplitude: 1.2m / s²), 5th harmonic frequency: 300Hz (current amplitude: 2.1A), low frequency energy ratio: 65%, wavelet entropy: 1.8", which are used to accurately locate noise sources (e.g., the 220Hz peak corresponds to bearing resonance noise, and the 300Hz peak corresponds to electromagnetic harmonic noise).

[0039] Preferably, the preprocessing further includes data augmentation processing, specifically including: Zero-padding extension: Zero-padding is applied to each segment of time-domain data to ensure the data length meets the requirement of 2. n (For example, padding 1000 data points with zeros to make 1024), to ensure the computational efficiency and frequency resolution of the FFT transform; Windowing to suppress spectral leakage: Applying a Hanning window to the zero-padded data reduces spectral leakage caused by aperiodic truncation of the time-domain signal (especially for the accurate identification of current harmonics and vibration resonant frequencies). DC component removal: Calculate the DC component (mean) of each data segment and remove (subtract) it to avoid the DC component occupying the energy peak in the frequency domain and affecting the identification of high-frequency features.

[0040] Subsequently, the running data, after outlier removal, multi-source data synchronization and alignment, low-pass filtering and noise reduction, and data augmentation, is converted into frequency domain data by Fast Fourier Transform or Wavelet Transform.

[0041] In this embodiment, the preprocessed running data is subjected to Fast Fourier Transform or Wavelet Transform to convert it into frequency domain data; specifically, this includes: For conventional frequency analysis: Fast Fourier Transform (FFT) is used to convert the preprocessed time-domain sequence into frequency-domain amplitude and power spectra. The frequency resolution Δf = sampling frequency / data length (e.g., when the sampling frequency is 1kHz and the data length is 1024, Δf ≈ 0.98Hz). For non-stationary signal analysis: Wavelet transform (such as db4 wavelet basis) is used to perform multi-scale decomposition on non-stationary vibration / current data during load abrupt changes to obtain wavelet coefficients in different frequency bands (such as 10-50Hz, 50-200Hz, 200-500Hz) and capture the frequency variation over time.

[0042] This invention achieves precise localization of noise dominance frequency and energy distribution by identifying spectral peaks and their corresponding frequencies and calculating the energy proportion of preset frequency bands. When using wavelet transform, it captures the time-frequency localization characteristics of non-stationary signals by extracting energy and entropy values ​​at various scales. The synergistic effect of frequency domain and time domain features constructs a joint feature representation system that comprehensively reflects the distribution law and energy transfer characteristics of load changes in the frequency domain, significantly enhancing the load prediction model's ability to identify load evolution trends under complex operating conditions.

[0043] In this embodiment, the training steps of the load prediction model include: Obtain historical datasets; historical datasets include historical feature vectors and corresponding historical load values; preferably, the feature vector set includes time-domain feature vectors and frequency-domain feature vectors, and the historical load values ​​are obtained by conversion from wind pressure / air volume sensors and estimation of motor output torque; Historical datasets are truncated using a sliding window method to generate training samples with temporal correlations; preferably, the window length is set to 50-100 sampling points (corresponding to 500ms-1s, matching the time scale of wind turbine load changes), and the window step size is 10 sampling points to ensure data continuity; The training samples are divided into a training set, a validation set, and a test set; preferably, the training set (historical normal operation data + load fluctuation scenario data), the validation set (used for hyperparameter tuning), and the test set (simulating real load change scenario data) are split in a 7:2:1 ratio. Construct an initial prediction model; the initial prediction model can be a long short-term memory network, a gated recurrent unit, or a temporal convolutional network. The initial prediction model is iteratively trained using the training set, and the model hyperparameters are tuned using the validation set to obtain the load prediction model.

[0044] In this embodiment, to address the problem of scarce load mutation samples, time stretching, additive Gaussian noise (signal-to-noise ratio ≥30dB), data flipping, and other methods can be used to expand the mutation scenario samples and improve the model's generalization ability.

[0045] This invention extracts historical datasets using a sliding window method to generate training samples with temporal correlations, enabling the model to learn the time dependence and evolution of load changes. By dividing the dataset into training, validation, and test sets, the process of model training, hyperparameter tuning, and generalization ability evaluation is decoupled. Long Short-Term Memory (LSTM) networks, gated recurrent units (ROUs), or temporal convolutional networks (TRNs) are selected as the initial prediction model. Leveraging their gating mechanisms or causal convolutional structures to model long-range temporal dependencies, this invention achieves high-precision prediction of load change trends, mutation magnitudes, and mutation times.

[0046] In this embodiment, the initial prediction model is preferably selected as a Long Short-Term Memory (LSTM) network / Gated Recurrent Unit (GRU): Input layer (feature dimension = number of preprocessed features, such as 20 dimensions) → Hidden layer (2-3 layers, 64-128 neurons per layer, activation function is ReLU) → Dropout layer (dropout rate = 0.2, to prevent overfitting) → Fully connected layer (output dimension = 3, corresponding to load change trend, mutation magnitude, mutation time). Temporal Convolutional Networks (TCNs) for Complex Scenes: Input layer → Causal convolutional layer (3-4 layers, kernel size = 3, dilation coefficients are set to 2) n (Incremental) → Residual Connection Layer → Global Average Pooling Layer → Output Layer; Loss function design: A hybrid loss function (MSE+MAE) is adopted, that is, Loss=0.7×MSE (mean squared error, to optimize the overall prediction accuracy)+0.3×MAE (mean absolute error, to suppress prediction bias in abrupt scenarios).

[0047] In this embodiment, hyperparameter initialization includes: Optimizer selection: Adam optimizer (initial learning rate = 0.001, using a learning rate decay strategy, decaying to 95% of the original rate every 100 epochs); Training parameters: batch size = 32-64, number of epochs = 200-300, early stopping strategy (training stops if the validation set loss does not decrease for 15 consecutive epochs).

[0048] In this embodiment, the model iterative training process includes: Phase 1 (Basic Training): Full training is performed using the training set, and the model's performance on the validation set is validated every 10 epochs (evaluation metrics: accuracy of load change trend, error in prediction of mutation magnitude, and deviation in prediction of mutation time). The second stage (focused optimization): For the prediction error of the load mutation scenario in the validation set, a hard example mining strategy (screening samples with prediction deviation > 10%) is adopted to carry out reinforcement training separately, and the learning rate is adjusted to 1 / 5 of the initial value to enhance the model's ability to capture mutation features.

[0049] In this embodiment, model pruning and quantization include: Pruning optimization: Remove redundant connections with absolute weight values ​​< 0.001 in the model (pruning rate ≤ 30%) to reduce model inference latency and adapt to the hardware computing power of the wind turbine controller; Quantization optimization: The model parameters are quantized from 32-bit floating-point (FP32) to 16-bit floating-point (FP16) to improve the model running speed while ensuring that the prediction accuracy loss is ≤3%.

[0050] In this embodiment, based on the load change trend, abrupt change amplitude, abrupt change time, and real-time electrical data, optimized field-oriented control parameters and auxiliary control parameters are calculated using an optimization algorithm, including: Based on load change trends, abrupt change amplitudes, abrupt change times, and real-time electrical data, optimized field-oriented control parameters and auxiliary control parameters are dynamically calculated using reinforcement learning algorithms or model predictive control algorithms. The field-oriented control parameters and auxiliary control parameters are verified to ensure that they meet the preset operating constraints. Preferably, the preset operating constraints include hard constraints and soft constraints. Hard constraints include upper limit of current and speed range; soft constraints include noise reduction efficiency ≥85% and torque ripple ≤5%.

[0051] In this embodiment, based on load change trends, abrupt change amplitudes, abrupt change times, and real-time electrical data, optimized field-oriented control parameters and auxiliary control parameters are dynamically calculated using reinforcement learning algorithms or model predictive control algorithms, including: Predictive data normalization: Independent normalization is performed on three types of predictive data: load change trend, abrupt change amplitude, and abrupt change time, to eliminate dimensional differences and ensure that data of different dimensions (such as current, time, and frequency) can participate in optimization calculations on the same scale, avoiding algorithm deviations caused by numerical differences; linear mapping or min-max normalization is preferred to linearly scale the original data to the [0,1] interval, preserving the relative relationships between data. Define the parameter optimization range: Based on rated parameters (such as motor rated current) and hardware limits (such as maximum allowable speed, PWM frequency hardware support range, etc.), ensure that the adjustment of field-oriented control parameters and auxiliary control parameters does not exceed physical limits; combine the fan load fluctuation law (such as load change trend, sudden change amplitude) to dynamically adjust the range of field-oriented control parameters and auxiliary control parameters. For example, in response to the dynamic response requirements when the load changes suddenly, allow the current to briefly exceed the motor rated current within a safe range. When the load increases, increase the torque current command to quickly respond to the load change, while reducing the PWM frequency to reduce high-frequency noise energy. Optimization model solution: Based on the normalized load change trend, abrupt change amplitude, abrupt change time and real-time electrical data, the optimized field-oriented control parameters and auxiliary control parameters are solved by reinforcement learning or model predictive control algorithms.

[0052] In this embodiment, the load change trend (range 0-1) determines the control sensitivity. The severity of load fluctuation is judged by the load change trend, and the control response speed and noise suppression accuracy are dynamically adjusted. The higher the value (e.g., 0.7-1), the more likely the load is to change abruptly. Kp (proportional coefficient) and Ki (integral coefficient) need to be increased to improve the control response speed, so that the motor can quickly track the load change and reduce transient noise caused by control lag. When it is predicted that a sudden load change may cause a new resonant frequency, the center frequency of the notch filter can be aligned with that frequency in advance to directly suppress the resonant noise that may be generated after the change.

[0053] In this embodiment, the amplitude of the sudden change determines the electromagnetic excitation intensity. Based on the amplitude of the load sudden change, the electromagnetic excitation and switching frequency are adjusted to balance the torque demand and noise level. When the load increases, the torque current command increases, the excitation current command decreases slightly, and the PWM frequency increases. When the load decreases, the torque current command decreases, the excitation current command increases slightly, and the PWM frequency decreases moderately.

[0054] In this embodiment, the mutation time (unit: ms) determines the optimization strategy. The optimization strategy of the control algorithm is dynamically adjusted according to the mutation time to balance real-time performance and multi-objective optimization. The shorter the mutation time, the smaller the control time domain and the fewer steps the algorithm needs to solve. An aggressive adjustment strategy is adopted to prioritize the real-time performance of parameter adjustment and avoid noise peaks caused by delay. The longer the mutation time, the smoother the parameter adjustment, taking into account multi-objective optimization. While reducing noise, motor efficiency (such as field weakening control coefficient optimization) and operational stability (such as speed fluctuation suppression) are also considered.

[0055] This invention, based on load change trends, abrupt change amplitudes, abrupt change times, and real-time electrical data, utilizes reinforcement learning algorithms, leveraging their trial-and-error learning and strategy optimization capabilities, to achieve adaptive dynamic optimization of control parameters; or it employs model predictive control algorithms, utilizing their rolling optimization and constraint handling capabilities, to solve for the optimal control sequence under multi-step prediction. Verification through preset operating constraints ensures that the optimized field-oriented control parameters and auxiliary control parameters remain within the motor's physical limits and safe operating boundaries. The synergistic effect of predictive information and optimization algorithms achieves closed-loop feedforward control from "load prediction" to "parameter optimization" and then to "control execution," maximizing the technical advantages of predictive noise reduction while ensuring safe motor operation.

[0056] In this embodiment, the field-oriented control parameters include the excitation current command, the torque current command, the proportional coefficient of the proportional-integral controller, the integral coefficient of the proportional-integral controller, and the pulse width modulation frequency. The auxiliary control parameters include the notch filter frequency and the field weakening control coefficient.

[0057] The field-oriented control parameters of this invention include excitation current command, torque current command, proportional coefficient of the proportional-integral controller, integral coefficient of the proportional-integral controller, and pulse width modulation frequency. These parameters directly act on the core of decoupling control between the motor flux linkage and torque, achieving precise control over the root cause of motor electromagnetic noise. Auxiliary control parameters include notch frequency and field weakening control coefficient. These two types of parameters work synergistically to form a joint optimization mechanism between the main control loop parameters and the auxiliary control loop parameters: adjusting the excitation current command and torque current command alters the electromagnetic excitation characteristics of the motor, suppressing electromagnetic noise at its source; adjusting the PI controller parameters optimizes current tracking performance, reducing additional noise caused by control errors; adjusting the PWM frequency balances switching losses and current harmonics, suppressing high-frequency carrier noise; setting the notch frequency effectively filters out resonant noise in specific frequency bands; and optimizing the field weakening control coefficient reduces iron losses and noise radiation in the high-speed operating region. Through the joint optimization of these multi-dimensional parameters, refined adjustment of the wind turbine motor's operating state is achieved, synergistically suppressing noise generation and propagation from multiple levels, including electromagnetic, mechanical, and control aspects.

[0058] In this embodiment, the method further includes: acquiring noise reduction effect monitoring data, and iteratively updating the load prediction model based on the noise reduction effect monitoring data. Preferably, noise is collected through a microphone, and speed fluctuations are monitored through MEMS to evaluate the current noise reduction level. Furthermore, this invention avoids the one-sidedness of a single indicator by establishing multi-dimensional quantifiable indicators (see the examples in Table 1 below).

[0059] Table 1 Examples of Multidimensional Quantifiable Indicators

[0060] For a better understanding of this invention, please refer to the following references. Figure 2 The technical solution of this invention is realized through a five-layer closed-loop architecture of "perception-analysis-decision-execution-feedback". The perception layer collects electrical data, mechanical vibration data and environmental data through multi-dimensional sensors; the analysis layer preprocesses the collected data and extracts features, and outputs the load change trend, abrupt change amplitude and abrupt change time through the load prediction model; the decision layer calculates the optimized control parameters based on the prediction results through optimization algorithms; the execution layer sends the optimized parameters to the wind turbine motor and adjusts the motor's operating state through field-oriented control; the feedback layer collects noise reduction effect data and iteratively updates the load prediction model.

[0061] In this embodiment, after the optimized field-oriented control parameters and auxiliary control parameters are sent to the wind turbine motor, the noise reduction parameters are implemented through the following hardware and control coordination process: Parameter reception and parsing: The motor controller receives the standardized parameter package output by the decision layer. The parameter package includes field-oriented control parameters (excitation current command Id, torque current command Iq, proportional coefficient Kp, integral coefficient Ki, pulse width modulation frequency f_PWM) and auxiliary control parameters (notch frequency f_notch, field weakening control coefficient k_weak).

[0062] The parameters are integrated into the core logic of field-oriented control: the excitation current command and torque current command are used as the target values ​​of the Park inverse transformation; based on the three-phase phase currents collected in real time by the sensing layer, the three-phase stationary coordinate system currents (Ia, Ib, Ic) are converted into two-phase stationary coordinate system currents (Iα, Iβ) through Clarke transformation, and then converted into two-phase rotating coordinate system currents (Id, Iq) through Park transformation; combined with the optimized proportional coefficient Kp and integral coefficient Ki, the deviations between Id and Id, and between Iq and Iq are adjusted by PI, and the voltage commands (Vd, Vq) in the two-phase rotating coordinate system are output. Based on the current rotational speed and the optimized Id and Iq, adjust the gain parameters of the sliding mode observer or extended Kalman filter to improve the accuracy of rotor position and rotational speed estimation; incorporate the notch frequency f_notch in the auxiliary control parameters into the observer's filtering logic to suppress high-frequency noise generated during the observation process. The pulse width modulation frequency f_PWM is written into the frequency register of the space vector pulse width modulation module to set the switching frequency of the switching transistor; based on Vd and Vq output by the Park inverse transform, the three-phase voltage vector is synthesized through the space vector pulse width modulation algorithm to control the switching transistors of the gallium nitride or silicon carbide drive module to turn on and off. Harmonic suppression and noise source control: Adjusting the excitation current command and torque current command alters the electromagnetic excitation characteristics of the motor, suppressing electromagnetic noise at its source; adjusting the proportional and integral coefficients optimizes current tracking performance, reducing additional noise caused by control errors; adjusting the pulse width modulation frequency balances switching losses and current harmonics, suppressing high-frequency carrier noise; setting the notch filter frequency suppresses high-frequency noise from the observer; adjusting the field weakening control coefficient optimizes operational stability in the high-speed range. Through the synergistic effect of these multi-dimensional parameters, comprehensive intervention in the motor noise generation mechanism is achieved.

[0063] like Figure 3 As shown, the present invention also provides a fan noise reduction system, which includes: The data acquisition module 10 is used to collect the operating data of the wind turbine motor; the operating data includes electrical data, mechanical vibration data and environmental data; The preprocessing module 20 is used to preprocess the running data and extract time-domain features and frequency-domain features from the preprocessed running data to generate a feature vector set. The standardization processing module 30 is used to standardize the feature vector set and extract the standardized feature vector set according to the preset sliding window to obtain the model input matrix and input it to the load prediction module 40. The load prediction module 40 is used to calculate and output the load change trend, mutation magnitude and mutation time through model forward propagation; The noise reduction processing module 50 is used to calculate optimized magnetic field orientation control parameters and auxiliary control parameters based on load change trends, abrupt change amplitude and abrupt change time, and send them to the fan motor to adjust the motor's operating status.

[0064] Furthermore, the system of the present invention may also include: a feedback module, used to acquire noise reduction effect monitoring data and iteratively update the load prediction model based on the noise reduction effect monitoring data.

[0065] This invention also provides a computer-readable storage medium, which may be a computer-readable storage medium included in the memory described in the above embodiments; or it may be a standalone computer-readable storage medium not assembled into a device. The computer-readable storage medium stores at least one instruction, which is loaded and executed by a processor to implement... Figure 1 The method for reducing noise in a wind turbine is shown. The computer-readable storage medium may be a read-only memory, a hard disk, or an optical disk, etc.

[0066] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For system embodiments and storage medium embodiments, since they are basically similar to method embodiments, the descriptions are relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0067] Furthermore, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0068] The foregoing description illustrates and describes preferred embodiments of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept by means of the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for reducing noise in a fan, characterized in that, Includes the following steps: Collect operating data of the wind turbine motor; the operating data includes electrical data, mechanical vibration data, and environmental data; The running data is preprocessed, and time-domain and frequency-domain features are extracted from the preprocessed running data to generate a feature vector set. The feature vector set is standardized, and the standardized feature vector set is truncated according to a preset sliding window to obtain the model input matrix; The model input matrix is ​​input into the pre-trained load prediction model, and the load change trend, mutation magnitude and mutation time are output through forward propagation calculation. Based on the load change trend, abrupt change amplitude, abrupt change time, and real-time electrical data, optimized field-oriented control parameters and auxiliary control parameters are calculated through an optimization algorithm and sent to the fan motor to adjust the motor's operating status.

2. The fan noise reduction method according to claim 1, characterized in that, The electrical data includes three-phase current, bus voltage, estimated speed and rotor angle; The mechanical vibration data includes X-axis vibration acceleration, Y-axis vibration acceleration, Z-axis vibration acceleration, and vibration frequency spectrum; The environmental data includes winding temperature and ambient temperature.

3. The fan noise reduction method according to claim 1, characterized in that, The preprocessing includes outlier removal, multi-source data synchronization and alignment, and low-pass filtering for noise reduction.

4. The fan noise reduction method according to claim 1, characterized in that, Temporal feature extraction is performed on the preprocessed runtime data, including: The mean, variance, peak value, peak-to-peak value, and kurtosis of the preprocessed running data are calculated to obtain statistical characteristics; The dynamic characteristics are obtained by calculating the ripple coefficient of the current and the root mean square value of the vibration in the preprocessed running data. Based on the statistical and dynamic features, a time-domain feature vector is obtained.

5. The fan noise reduction method according to claim 1, characterized in that, Frequency domain feature extraction is performed on the preprocessed running data, including: Perform a Fast Fourier Transform or Wavelet Transform on the preprocessed running data to convert it into frequency domain data; Identify spectral peaks and their corresponding frequencies from the frequency domain data, and calculate the energy percentage of a preset frequency band; When using Fast Fourier Transform, a frequency domain feature vector is obtained based on the energy proportion, spectral peak value and its corresponding frequency; When wavelet transform is used, the energy value and entropy value of each scale are extracted from the wavelet transform coefficients, and the frequency domain feature vector is obtained based on the energy ratio, energy value, entropy value, spectral peak value and its corresponding frequency.

6. The fan noise reduction method according to claim 1, characterized in that, The training steps for the load prediction model include: Obtain the historical dataset; the historical dataset includes historical feature vectors and corresponding historical payload values; The historical dataset is truncated using the sliding window method to generate training samples with temporal correlation. The training samples are divided into a training set, a validation set, and a test set; Construct an initial prediction model; the initial prediction model is a long short-term memory network, a gated recurrent unit, or a temporal convolutional network. The initial prediction model is iteratively trained using the training set, and the model hyperparameters are tuned using the validation set to obtain the load prediction model.

7. The fan noise reduction method according to claim 1, characterized in that, Based on the load change trend, abrupt change amplitude, abrupt change time, and real-time electrical data, optimized field-oriented control parameters and auxiliary control parameters are calculated using an optimization algorithm, including: Based on the load change trend, abrupt change amplitude, abrupt change time, and real-time electrical data, optimized field-oriented control parameters and auxiliary control parameters are dynamically calculated using reinforcement learning algorithms or model predictive control algorithms. Verify the magnetic field orientation control parameters and auxiliary control parameters to ensure that they meet the preset operating constraints.

8. The fan noise reduction method according to claim 7, characterized in that, The field-oriented control parameters include excitation current command, torque current command, proportional coefficient of proportional-integral controller, integral coefficient of proportional-integral controller, and pulse width modulation frequency. The auxiliary control parameters include notch filter frequency and field weakening control coefficient.

9. A fan noise reduction system, characterized in that, include: The data acquisition module is used to collect operating data of the wind turbine motor; the operating data includes electrical data, mechanical vibration data, and environmental data. The preprocessing module is used to preprocess the running data and extract time-domain features and frequency-domain features from the preprocessed running data to generate a feature vector set. The standardization processing module is used to standardize the feature vector set, and to extract the standardized feature vector set according to a preset sliding window to obtain the model input matrix and input it to the load prediction module. The load prediction module is used to calculate and output the load change trend, mutation magnitude, and mutation time through model forward propagation. The noise reduction processing module is used to calculate optimized magnetic field orientation control parameters and auxiliary control parameters based on the load change trend, abrupt change amplitude, abrupt change time and real-time electrical data, and send them to the fan motor to adjust the motor's operating status.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a fan noise reduction program, which, when executed by a processor, implements the steps of the fan noise reduction method as described in any one of claims 1 to 8.

Images