A noise reduction system and method for charging piles

Abstract

This invention discloses a noise reduction method and system for charging piles, aiming to solve the problem of noise reduction of complex noise from charging piles in open spaces. The method includes: acquiring complex signals such as broadband noise from cooling system fans and transformer vibration noise through a multi-channel microphone array, while simultaneously acquiring environmental data such as temperature, airflow, and obstacle status; preprocessing the noise signals and fusing them with quantized acoustic path features to form an input matrix; inputting an optimized RNN-GRU model to generate an inverted sound wave, the model employing a dual-mode architecture, combining transfer learning, domain adaptation, and acoustic path prediction optimization; outputting the inverted sound wave through a multi-band speaker, and dynamically adjusting parameters based on the frequency-domain weighted mean square error of residual noise. This invention can accurately capture complex noise, adapt to environmental changes, balance generalization and real-time performance, and meet the low-noise operation requirements of charging piles.

Description

Technical Field

[0001] This invention relates to the field of noise reduction technology, specifically to a noise reduction system and method for charging piles. Background Technology

[0002] With the rapid popularization of new energy vehicles, charging piles, as key supporting facilities, have been widely distributed in residential areas, commercial areas, and other areas with high population activity. However, charging piles generate various complex noises during operation: wide-frequency aerodynamic noise from the cooling system fan (covering 20Hz-10kHz), low-frequency noise from transformer core vibration (50-200Hz), and high-frequency switching noise from the power module (1-5kHz). When these noises propagate in open spaces, they can easily interfere with the surrounding environment, affecting user experience and environmental comfort.

[0003] Existing charging pile noise reduction technologies have the following shortcomings:

[0004] Passive noise reduction has significant limitations: traditional sound insulation materials (such as sound insulation cotton and soundproof covers) have poor blocking effect on high-frequency noise (>2kHz) and will increase the heat dissipation burden of charging piles, affecting the stability of equipment operation;

[0005] Traditional active noise cancellation has insufficient adaptability: Active noise cancellation technology based on adaptive filtering algorithms such as FxLMS has limited ability to handle the complex noise (multi-source, wideband, time-varying) of charging piles, and it is particularly difficult to capture the non-steady-state characteristics of instantaneous switching noise.

[0006] Weak environmental adaptability: In open spaces, noise propagation is significantly affected by environmental factors such as temperature, airflow, and obstacle movement (e.g., airflow can change the speed of sound wave propagation, and obstacle reflection can cause changes in the acoustic path). Traditional systems lack dynamic prediction of the acoustic path and are prone to cancellation failure due to the asynchronous nature of the antiphase sound wave and the original noise.

[0007] The contradiction between model generalization and real-time performance: Although deep learning models (such as RNNs) can improve the ability to capture noise features, the computational cost of complex models is large, making it difficult to achieve real-time inference on embedded devices in charging piles; at the same time, models trained on general noise datasets have poor adaptability to specific noises of charging piles (such as the unique vibration spectrum of transformers), requiring a large number of labeled samples for optimization, which is costly.

[0008] Therefore, there is an urgent need for a charging pile noise reduction solution that can comprehensively capture complex noise, dynamically adapt to environmental changes, and take into account both generalizability and real-time performance. Summary of the Invention

[0009] The purpose of this invention is to provide a charging pile noise reduction system and method to solve at least one of the above-mentioned problems.

[0010] The first aspect of this invention provides a method for noise reduction in charging piles, characterized by comprising the following steps:

[0011] (1) The composite noise signal during the operation of the charging pile is collected by a multi-channel microphone array. The composite noise signal includes: broadband noise of the cooling system fan, transformer vibration noise, and power module switching noise. At the same time, environmental data is collected by an environmental sensor. The environmental data includes temperature, humidity and airflow speed parameters. The positioning module monitors the movement status of obstacles within a preset range in real time.

[0012] (2) Perform time-domain normalization, frame-segmentation and data enhancement on the composite noise signal. The data enhancement includes additive white Gaussian noise injection and time-domain random offset, and converts it into time series data. Quantize the environmental data into acoustic path feature vectors, and merge them with the time series data through channel splicing to form an input matrix.

[0013] (3) Input the input matrix into the optimized RNN-GRU model to generate an antiphase acoustic signal;

[0014] (4) The anti-phase acoustic wave signal is output through a multi-band directional loudspeaker array;

[0015] (5) Collect residual noise through error microphone, calculate frequency domain weighted mean square error, and update acoustic path model in combination with real-time environmental data to dynamically adjust the output timing, amplitude and phase of antiphase sound wave to ensure that the generation delay of antiphase sound wave is within the preset range.

[0016] Preferably, in step (2), quantizing the environmental data into an acoustic path feature vector includes:

[0017] The airflow velocity parameter is converted into an acoustic refraction correction term: δ air =k v · v air · cosθ

[0018] Where, k v The airflow-acoustic coupling coefficient is 0.12 s / m, v air The real-time airflow velocity is given by θ, which is the angle between the airflow direction and the main acoustic axis. Environmental data is quantized into an acoustic path feature vector P. τ =[v, d, H,δ air ] T Where τ is the current time, d is the distance to the obstacle, H is the humidity, T is the matrix dimension, the speed of sound v = 331.4 + 0.6t, and t is the temperature.

[0019] Preferably, the direction of the main sound axis is defined as the reference axis of the speaker array center pointing to the error microphone; the method for calculating the airflow direction angle θ is as follows: three airflow sensors distributed at 120° are arranged at the heat dissipation outlet of the charging pile for periodic detection, and the angle between the airflow vector direction and the main sound axis is calculated based on the triangulation method.

[0020] Preferably, the optimization of the RNN-GRU model includes:

[0021] (a) It adopts a dual-mode configurable architecture, which includes a basic mode and a high-performance mode. The basic mode includes two GRU layers, two Dropout layers and residual connections, while the high-performance mode includes three GRU layers, two Dropout layers and residual connections. The mode is dynamically switched according to the proportion of high-frequency noise energy.

[0022] (b) Optimize by transfer learning, pre-train the underlying GRU parameters on the basic dataset; freeze the underlying parameters and fine-tune the top-level network using specific samples from charging piles; use semi-supervised learning with a confidence level greater than 0.95 for unlabeled samples;

[0023] (c) During the pre-training phase, a domain adaptation mechanism is implemented through a gradient inversion layer to minimize the distribution difference between general noise and charging pile noise;

[0024] (d) The model input layer embeds acoustic path features, and the GRU layer synchronously learns the correlation between noise temporal features and acoustic path features. The time delay estimation network predicts the time delay and phase change of noise propagation and pre-compensates the generated antiphase sound waves.

[0025] Preferably, the RNN-GRU model is trained with mixed precision, quantized with INT8 and calibrated with dynamic range; after pruning, 63%-67% of the core parameters are retained, sparsity is maintained by the L2 regularization term of the AdamW optimizer, and the signal-to-noise ratio loss of the model after pruning is ≤1dB.

[0026] Preferably, when calculating the frequency domain weighted MSE, 24 critical frequency bands are divided, and the weight of each frequency band is positively correlated with the A-weighting curve; when instantaneous impulse noise is detected, the weight of the 2-4kHz frequency band is temporarily increased to 1.5 times the base weight.

[0027] Preferably, when a change in acoustic path propagation delay is detected due to obstacle movement, the output lead Δt is calculated based on the distance change rate Δd / Δt. adv Δt adv = (d new - d old ) / v where d new d represents the distance to the obstacle after the movement. oldLet be the distance to the obstacle before it moves, and v = 331.4 + 0.6t, where t is the temperature. Based on Δt... adv Adjust the output timing of the phase-reversing acoustic wave.

[0028] A second aspect of the present invention provides a charging pile noise reduction system, comprising the following modules:

[0029] The multi-source signal acquisition module includes: a multi-channel microphone array configured to acquire composite noise signals during the operation of the charging pile, wherein the composite noise signals include broadband noise from the cooling system fan, transformer vibration noise, and power module switching noise;

[0030] An environmental sensor array is configured to collect environmental data, including temperature, humidity, and airflow speed parameters; a positioning module is configured to monitor the movement of obstacles within a preset range in real time.

[0031] The data processing module is configured to perform time-domain normalization, frame-segmentation, and data enhancement on the composite noise signal, converting it into time-series data; quantize the environmental data into acoustic path feature vectors, and fuse them with the time-series data into an input matrix through channel splicing;

[0032] An adaptive noise reduction engine includes: an optimized RNN-GRU model that receives the input matrix and generates an antiphase acoustic signal; and a multi-band directional loudspeaker array that outputs the antiphase acoustic signal.

[0033] The closed-loop control module includes: an error microphone to collect residual noise; a dynamic adjustment unit to calculate the frequency domain weighted mean square error based on the residual noise; an acoustic path model updated in conjunction with real-time environmental data; and dynamic adjustment of the output timing, amplitude, and phase of the anti-phase sound wave to ensure that the generation delay of the anti-phase sound wave is within a preset range.

[0034] A third aspect of the present invention provides a charging pile noise reduction device, characterized in that the device includes a memory and a processor, the memory storing a computer program, the computer program being executable by the processor to implement the above-described method.

[0035] A fourth aspect of the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program that can be executed by a processor to implement the above-described method.

[0036] This invention, through multi-dimensional technological innovation, achieves a significant improvement in noise reduction effect, specifically addressing the characteristics of charging pile composite noise and open space environments. Specifically, it includes:

[0037] (1) Comprehensive capture and accurate cancellation of composite noise: Multi-channel microphone array is used to synchronously collect broadband noise from cooling system fan, transformer vibration noise and power module switching noise. Combined with data augmentation methods such as time-domain normalization, frame segmentation and additive white Gaussian noise injection, the model can learn the multi-source characteristics of composite noise. The optimized RNN-GRU model dynamically adapts to the noise energy distribution through a dual-mode architecture (basic mode / high-performance mode), which significantly improves the capture accuracy of broadband (20Hz-10kHz) and instantaneous impulse noise, and solves the problem of incomplete processing of composite noise by traditional algorithms.

[0038] (2) Environmental Adaptability and Dynamic Optimization of Acoustic Path: Temperature, humidity, and airflow velocity are collected by environmental sensors, and the movement status of obstacles is monitored by the positioning module. The environmental data is quantified into acoustic path feature vectors (including airflow refraction correction terms, sound velocity correction values, etc.) and embedded into the model input layer. The GRU layer synchronously learns the correlation between noise temporal features and acoustic path, and predicts propagation delay and phase change through a time delay estimation network to compensate for the output timing of the anti-phase sound wave in advance. This mechanism effectively reduces the error between the anti-phase sound wave and the original noise. In scenarios with obstacle movement or airflow changes, the noise reduction effect (SNR improvement) is significantly improved compared with traditional systems.

[0039] (3) Improved model generalization and sample efficiency: Based on the transfer learning framework, the underlying parameters of the model are pre-trained using a general noise dataset, and the top-level network is fine-tuned using a small number of specific samples from charging piles. The data is expanded using unlabeled samples through semi-supervised learning. A gradient inversion layer is introduced during the pre-training stage to achieve domain adaptation, which significantly reduces the distribution difference between general noise and charging pile noise. This approach improves the noise reduction SNR of the model for new fault noises of untrained charging piles (such as abnormal switching sound of power modules) and reduces the sample requirement.

[0040] (4) The model balances real-time performance and lightweight design. It adopts mixed precision training (INT8 quantization) and dynamic range calibration. After pruning, 63%-67% of the core parameters are retained. The inference latency on embedded devices (such as NVIDIA Jetson Nano) is ≤10ms, which meets the real-time requirements. At the same time, sparsity is maintained by L2 regularization of AdamW optimizer to ensure that the signal-to-noise ratio loss after pruning is ≤1dB, thus achieving a balance between lightweight design and high performance.

[0041] (5) In the frequency domain weighted MSE calculation for targeted optimization of high frequency and transient noise, the weights of the 24 critical frequency bands are positively correlated with the A-weighting curve, and the weights of the 2-4kHz frequency band are temporarily increased to 1.5 times when transient impulse noise occurs, which specifically enhances the cancellation effect of noise in the frequency band sensitive to human ears and improves the noise suppression effect of the frequency (2-5kHz).

[0042] In summary, this invention significantly improves the noise reduction effect, environmental adaptability, and engineering practicality of charging piles in open spaces through a full-process innovation of multi-source acquisition, environmental fusion, intelligent modeling, and closed-loop optimization, providing an effective solution for low-noise operation of charging piles. Attached Figure Description

[0043] Figure 1 This is a deployment diagram of the charging pile noise reduction system provided in Embodiment 1 of the present invention.

[0044] Figure 2 Flowchart of the charging pile noise reduction method provided in Embodiment 1 of the present invention

[0045] Figure 3 This is a block diagram of the charging pile noise reduction system provided in Embodiment 2 of the present invention. Detailed Implementation

[0046] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. 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. Various modifications or adjustments made to the present invention by those skilled in the art after reading this invention, as long as they do not depart from the essential spirit of the invention, are all within the scope defined by the claims of the present invention.

[0047] Example 1

[0048] This invention provides a noise reduction method for charging piles, which achieves efficient suppression of operating noise through multi-sensor fusion and adaptive algorithms. Specifically, it includes the following steps:

[0049] The charging pile's operation is monitored by a multi-channel microphone array, which collects composite noise signals including broadband noise from the cooling system fan, transformer vibration noise, and power module switching noise. Simultaneously, environmental data, including temperature, humidity, and airflow speed parameters, is collected by an environmental sensor, and the movement of obstacles within a preset range is monitored in real time by a positioning module.

[0050] like Figure 1As shown, to comprehensively capture the complex noise of the charging pile, the system employs a circular array of eight high-precision MEMS microphones 1, evenly distributed within a 1.5-meter radius around the charging pile. These microphones have a frequency response covering 20Hz to 20kHz, enabling them to fully capture wideband fan noise (20Hz-10kHz), low-frequency transformer vibration (50-200Hz), and high-frequency switching noise from the power module (1-5kHz). Simultaneously, to adapt to changes in the open environment, the system is equipped with an environmental sensor suite, including a temperature and humidity sensor 2 installed at the top ventilation opening of the charging pile, which can monitor ambient temperature and humidity in real time, with a temperature measurement error of ±0.3℃ and a humidity measurement error of ±2%; an ultrasonic anemometer 3 positioned near the heat dissipation vent to measure airflow speed and correct the propagation path of sound waves in the flowing air; and a millimeter-wave radar 4 integrated into the charging pile's casing, which scans obstacles within a 5-meter radius at a 20Hz refresh rate to dynamically adjust the noise reduction strategy.

[0051] The composite noise signal is subjected to time-domain normalization, frame-segmentation, and data enhancement. The data enhancement includes additive white Gaussian noise injection and time-domain random offset, which converts it into time-series data. The environmental data is quantized into acoustic path feature vectors, which are then fused with the time-series data through channel concatenation to form an input matrix.

[0052] The core computing module of the system uses an NVIDIA Jetson Xavier NX as the main processor, paired with a Xilinx Zynq UltraScale+ FPGA for real-time signal processing. This combination ensures low-latency response in complex environments. The microphone array records synchronously at a sampling rate of 48kHz, generating a frame of 512 audio data every 10ms. The FPGA performs time-domain normalization (normalized to [-1,1]), framing (25ms frame length, 10ms frame shift) on the original signal, and injects slight Gaussian white noise (SNR=15dB) to enhance model robustness. Temperature, humidity, wind speed, and obstacle distance data are updated every 100ms and converted into acoustic path feature vectors, which are input into the noise reduction model along with the audio data.

[0053] In open spaces, sound wave propagation is influenced by multiple physical factors. The system achieves accurate modeling by constructing acoustic path feature vectors. Quantifying environmental data into acoustic path feature vectors includes:

[0054] First, a dynamic correction is made to the speed of sound. The speed of sound in air is significantly affected by temperature. Based on the physical property that the speed of sound in air increases with temperature (0.6 m / s per 1°C increase), the following formula for the speed of sound is given:

[0055] v = 331.4 + 0.6t, where t is the real-time data from the temperature sensor (in °C).

[0056] For example, when the temperature sensor measures the ambient temperature to be 32°C, the speed of sound is calculated as follows:

[0057] v = 331.4 + 0.6 × 32 = 350.6 m / s

[0058] This correction ensures precise synchronization between the phase-reversed acoustic wave and the original noise, guaranteeing the timing control accuracy of the phase-reversed acoustic wave.

[0059] In open environments, airflow alters the propagation path and speed of sound waves. The system quantifies this effect, converting the airflow velocity parameter into an acoustic refraction correction term: δ air =k v · v air · cosθ

[0060] Where, k v θ is the airflow-acoustic coupling coefficient, with a value of 0.12 s / m. vair is the real-time airflow velocity measured by the ultrasonic anemometer. θ is the angle between the airflow direction and the main acoustic axis direction, which is defined as the reference axis of the speaker array center pointing to the error microphone.

[0061] Three airflow sensors are deployed at a 120° angle at the heat dissipation vent. The angle θ between the airflow direction and the main acoustic axis is calculated using triangulation. For example, when the measured wind speed is 1.8 m / s and the angle with the main acoustic axis is 45°, the system applies an acoustic-airflow coupling model.

[0062] δair=0.12×1.8×cos45°≈0.153 s

[0063] The coupling coefficient of 0.12 s / m was calibrated through wind tunnel experiments, characterizing the delay effect of unit wind speed on sound wave propagation. This correction term is used to adjust the timing of out-of-phase sound waves to counteract the sound wave offset caused by airflow.

[0064] Finally, the system quantizes the environmental data into an acoustic path feature vector P. τ =[v, d, H,δ air ] T Where τ is the current time, d is the distance to the obstacle measured by the millimeter-wave radar, H is the normalized humidity value, and T is the matrix dimension.

[0065] Example: Under the conditions of temperature 32℃, obstacle distance 2.3m, humidity 65%, and wind speed 1.8m / s, the feature vector is:

[0066] [350.6,2.3,0.65,0.153] T

[0067] This vector, concatenated with audio data, is input into the model, enabling the deep learning network to synchronously perceive changes in the physical environment.

[0068] The input matrix is ​​input into an optimized RNN-GRU model to generate an antiphase acoustic signal.

[0069] The optimization of the RNN-GRU model includes:

[0070] The system employs a dual-mode configurable architecture, comprising a basic mode and a high-performance mode. The basic mode includes two GRU layers, two Dropout layers, and residual connections, while the high-performance mode includes three GRU layers, two Dropout layers, and residual connections. The mode dynamically switches based on the proportion of high-frequency noise energy. The model includes two operating modes to adapt to different noise scenarios.

[0071] Basic model: A 2-layer GRU network with 128 neurons per layer. Suitable for steady-state noise (such as low-frequency vibrations of transformers). Low computational latency (~3ms), suitable for normal operation.

[0072] High-performance mode: 3-layer GRU network, 256 neurons per layer. Optimized for transient impulse noise (such as power module switching). Slightly higher computation latency (~6ms), but 30% improved noise reduction depth.

[0073] Switching logic: The system analyzes the noise spectrum in real time. If the energy in the frequency band above 2kHz exceeds 30% of the total energy for three consecutive frames, it switches to high-performance mode to ensure that high-frequency noise is effectively suppressed.

[0074] The innovation of the RNN-GRU model lies in its environment-aware dual-mode architecture. Based on this architecture, the model can dynamically balance computational efficiency and noise reduction accuracy.

[0075] Under normal operation, the system operates in basic mode: a two-layer GRU network with 128 neurons each, coupled with a Dropout layer with a dropout rate of 0.3, enabling low-power operation. When frequent switching of the power module causes the energy proportion of high-frequency noise (>2kHz) to exceed the 30% threshold, the model automatically switches to high-performance mode: a three-layer deep network with 256 neurons is immediately activated. Although power consumption increases, the ability to suppress transient impulse noise is significantly improved.

[0076] The underlying GRU parameters are pre-trained using a base dataset to optimize the network through transfer learning; the underlying parameters are frozen, and the top-level network is fine-tuned using specific samples from charging piles; semi-supervised learning with a confidence level greater than 0.95 is used for unlabeled samples.

[0077] To address the scarcity of noise samples from charging piles, the system employs a three-stage progressive training strategy. First, it utilizes a large-scale, general-purpose noise dataset for underlying pre-training. During this pre-training phase, the model learns the generation patterns of inverse waveforms from common noises such as traffic and industrial equipment noise, thus fully optimizing the parameters of the underlying GRU network. Next, in the fine-tuning phase, the underlying parameters are frozen, and a charging pile-specific dataset is introduced, containing noise samples from different models of charging piles under various operating conditions. Targeted training allows the top-level network to adapt to the unique acoustic characteristics of charging piles. To address the insufficient labeled data, the system employs a semi-supervised learning mechanism: the massive amount of unlabeled noise collected on-site undergoes confidence screening, and pseudo-labels are automatically generated and added to the training set only when the model's prediction confidence exceeds 0.95. This strategy both expands the amount of training data and improves noise reduction accuracy.

[0078] During the pre-training phase, a gradient inversion layer is used to implement a domain adaptation mechanism, minimizing the distribution difference between general noise and charging pile noise.

[0079] To bridge the distribution differences between general noise and charging pile noise, a gradient inversion layer is embedded in a shared feature extraction layer during the pre-training phase. Its working mechanism is as follows: feature information is transmitted normally during forward propagation, while the gradient is inverted and scaled during backward propagation. This design forces the feature extractor to be unable to distinguish the type of input noise, thus effectively reducing the distribution difference. During training, the system simultaneously inputs general noise samples and charging pile noise samples, quantifies the feature distribution distance between the two types of noise using the maximum mean difference algorithm, and uses this distance as an auxiliary loss function to optimize the model. This mechanism significantly improves the model's generalization ability; when the system is applied to new types of charging piles that were not included in the training, the noise reduction depth is still effectively maintained.

[0080] (d) The model input layer embeds acoustic path features, and the GRU layer synchronously learns the correlation between noise temporal features and acoustic path features. The time delay estimation network predicts the time delay and phase change of noise propagation and pre-compensates the generated antiphase sound waves.

[0081] The system incorporates an acoustic path feature vector into the model input layer. This four-dimensional vector includes key environmental parameters such as sound speed, obstacle distance, humidity, and airflow refraction correction terms. Within the GRU network, these physical features are deeply coupled with the noise time-series data through a gating mechanism, enabling the model to synchronously perceive the impact of environmental changes on sound wave propagation. When the millimeter-wave radar detects a change in obstacle distance, the time delay estimation network calculates the change in sound wave propagation delay in real time. For example, if an obstacle moves 1.2 meters, resulting in a 3.47-millisecond delay, the system performs time-domain offsetting and phase compensation on the out-of-phase sound wave. This technique ensures the stability of noise reduction performance in dynamic environments.

[0082] Furthermore, given the limited computing resources of charging piles, the RNN-GRU model can be further optimized for embedded platforms. The system employs a triple lightweighting technique: First, mixed-precision training is implemented, converting the original floating-point weights to 8-bit integer format (INT8). Dynamic range calibration ensures minimal quantization error—specifically, a statistical weight distribution histogram is used to linearly map the range [-2.56, 2.54] to the integer interval 0-255, significantly reducing storage requirements by 75%. Next, structured pruning is performed, guided by L2 regularization of the AdamW optimizer, gradually removing redundant connections with an impact of less than 0.01 on the output, ultimately retaining 65% of the core parameters (error ±2%).

[0083] The anti-phase acoustic signal is output through a multi-band directional loudspeaker array.

[0084] The multi-band directional speaker array comprises four wideband directional speakers mounted at a 30° tilt angle to ensure precise coverage of the noise radiation area by anti-phase sound waves. The system employs a multi-band collaborative beamforming architecture, dividing the speaker array into three dedicated frequency band units: a low-frequency unit (40-800Hz, transformer noise), a mid-frequency unit (200-5kHz, fan noise), and a high-frequency unit (3-12kHz, switching noise). The signal delay of each speaker is calculated based on the target location, and each unit achieves directional sound beam emission through a precise time-delay control algorithm. Based on the speaker spacing and current sound velocity, the system calculates the phase difference between units at different positions and dynamically adjusts the output timing. This design concentrates sound energy onto the noise source area, forming a 1.5-meter diameter noise reduction core zone 3 meters in front of the charging pile, while simultaneously reducing the overall system power consumption.

[0085] Residual noise is collected by an error microphone, the frequency domain weighted mean square error is calculated, and the acoustic path model is updated in combination with real-time environmental data to dynamically adjust the output timing, amplitude and phase of the antiphase sound wave, ensuring that the generation delay of the antiphase sound wave is within the preset range.

[0086] Error microphone 6 is positioned in the noise reduction target area (e.g., 1.5 meters in front of a charging pile) to monitor residual noise in real time. The residual noise collected by the error microphone is first converted into a frequency domain signal and divided into 24 critical frequency bands according to the characteristics of human hearing. The system introduces a dynamic weight adjustment mechanism: under normal conditions, the weights of each frequency band conform to the A-weighting curve, but when impulse noise with a duration of less than 5 milliseconds is detected, the weight of the 2-4kHz sensitive frequency band is automatically increased to 1.5 times the base value. The environmental adaptive module continuously optimizes the acoustic path model parameters through real-time data streams—when the temperature change exceeds 2°C, the wind speed fluctuation is greater than 1 m / s, or the obstacle displacement reaches 0.3 meters, the system smoothly updates by fusing historical parameters with real-time measurements at a ratio of 0.8:0.2. This strategy has an extremely fast response speed to typical interferences such as passing vehicles and sudden strong winds, and the system restores noise reduction performance in a very short time through amplitude and phase adjustments.

[0087] In response to sudden changes in the acoustic path, when the millimeter-wave radar detects movement of an obstacle (such as a vehicle entering a charging station), the system immediately initiates a four-step response process:

[0088] 1. The distance change monitoring radar tracks the target at a refresh rate of 20Hz. Compensation is triggered when a distance change Δd > 0.2m is detected for three consecutive frames. For example, if a vehicle moves from 3.2m to 4.1m, the distance difference Δd = 0.9m.

[0089] 2. Acoustic parameter updates synchronously utilize temperature and humidity sensor data:

[0090] Current temperature t=26℃

[0091] The speed of sound v = 331.4 + 0.6 × 26 = 346.84 m / s

[0092] 3. Accurate calculation of lead time: Compensation value derived based on the principle of sound wave propagation.

[0093] Δt adv =Δd*v=0.9 *346.84≈2.60ms, which means that the phase-reversed sound wave needs to be output 2.60ms earlier to compensate for the path delay.

[0094] 4. Multi-dimensional compensation implementation

[0095] Timing adjustment: The audio buffer is moved forward by 125 samples (48kHz sampling rate).

[0096] Phase correction: Apply a linear phase shift ΔΦ = -2πf*Δt to the out-of-phase acoustic wave. adv (f is the dominant noise frequency)

[0097] Amplitude fine-tuning: According to the inverse square law of distance, the gain is increased (d new / d old)2≈1.28

[0098] This embodiment collects composite noise, embeds acoustic path features into the model input layer, and employs a dual-mode RNN-GRU model architecture. It combines transfer learning, domain adaptation, acoustic path prediction optimization, and sample expansion to output anti-phase sound waves through multi-band speakers. Parameters are dynamically adjusted based on the frequency-domain weighted mean square error of residual noise. This invention can accurately capture composite noise, adapt to environmental changes, balance generalization and real-time performance, and meet the low-noise operation requirements of charging piles.

[0099] Example 2

[0100] This embodiment provides a charging pile noise reduction system, which includes the following modules:

[0101] The multi-source signal acquisition module includes: a multi-channel microphone array configured to acquire composite noise signals during the operation of the charging pile, wherein the composite noise signals include broadband noise from the cooling system fan, transformer vibration noise, and power module switching noise;

[0102] An environmental sensor array is configured to collect environmental data, including temperature, humidity, and airflow speed parameters; a positioning module is configured to monitor the movement of obstacles within a preset range in real time.

[0103] The data processing module is configured to perform time-domain normalization, frame-segmentation, and data enhancement on the composite noise signal, converting it into time-series data; quantize the environmental data into acoustic path feature vectors, and fuse them with the time-series data into an input matrix through channel splicing;

[0104] An adaptive noise reduction engine includes: an optimized RNN-GRU model that receives the input matrix and generates an antiphase acoustic signal; and a multi-band directional loudspeaker array that outputs the antiphase acoustic signal.

[0105] The closed-loop control module includes: an error microphone to collect residual noise; a dynamic adjustment unit to calculate the frequency domain weighted mean square error based on the residual noise; an acoustic path model updated in conjunction with real-time environmental data; and dynamic adjustment of the output timing, amplitude, and phase of the anti-phase sound wave to ensure that the generation delay of the anti-phase sound wave is within a preset range.

[0106] Example 3

[0107] This embodiment provides a charging pile noise reduction device, which includes a memory and a processor. The memory stores a computer program, which can be executed by the processor to implement the aforementioned charging pile noise reduction method. For example, an NVIDIA Jetson Nano is used as the processor, equipped with 4GB of memory. The written noise reduction algorithm program (such as using Python combined with TensorFlow or PyTorch deep learning framework) is stored in the memory. The processor reads and executes the program to complete a series of noise reduction operations, including noise signal acquisition from a multi-channel microphone array, environmental sensor data acquisition, data preprocessing, RNN-GRU model inference to generate inverted sound waves, outputting inverted sound waves from a multi-band directional speaker array, and error microphone feedback and dynamic adjustment, thereby achieving low-noise operation of the charging pile.

[0108] Example 4

[0109] In this embodiment, the computer-readable storage medium is a 32GB SD card. The computer program corresponding to the above-described charging pile noise reduction method (e.g., written in C++ and compiled into an executable file that can run on an embedded Linux system) is stored on the SD card. When the SD card is inserted into the card reader of the charging pile noise reduction device, the device's processor can read and execute the program, thereby realizing various functions of charging pile noise reduction, including signal acquisition, data processing, inverted sound wave generation and output, closed-loop control, etc., effectively reducing noise interference from the charging pile during operation and improving the comfort of the surrounding environment.

[0110] The above embodiments are for reference only and can be adjusted and optimized according to actual application scenarios to meet the noise reduction needs of different charging piles.

Claims

1. A method for reducing noise in charging piles, characterized in that, Includes the following steps: (1) The composite noise signal during the operation of the charging pile is collected by a multi-channel microphone array. The composite noise signal includes: broadband noise of the cooling system fan, transformer vibration noise, and power module switching noise. At the same time, environmental data is collected by an environmental sensor. The environmental data includes temperature, humidity and airflow speed parameters. The positioning module monitors the movement status of obstacles within a preset range in real time. (2) Perform time-domain normalization, frame-segmentation and data enhancement on the composite noise signal. The data enhancement includes additive white Gaussian noise injection and time-domain random offset, and converts it into time series data. Quantize the environmental data into acoustic path feature vectors, and merge them with the time series data through channel splicing to form an input matrix. (3) Input the input matrix into the optimized RNN-GRU model to generate an antiphase acoustic signal; the optimization of the RNN-GRU model includes: embedding acoustic path features into the model input layer, synchronously learning the correlation between noise temporal features and acoustic path features in the GRU layer, predicting the time delay and phase change of noise propagation through the time delay estimation network, and pre-compensating the generated antiphase acoustic signal. (4) The anti-phase acoustic wave signal is output through a multi-band directional loudspeaker array; (5) Collect residual noise through error microphone, calculate frequency domain weighted mean square error, and update acoustic path model in combination with real-time environmental data to dynamically adjust the output timing, amplitude and phase of antiphase sound wave to ensure that the generation delay of antiphase sound wave is within the preset range. In step (2), quantizing the environmental data into an acoustic path feature vector includes: Convert the airflow velocity parameter into an acoustic refraction correction term: δ air = k v · v air · cosθ Where, k v The airflow-acoustic coupling coefficient is 0.12 s / m, v air θ represents the real-time airflow velocity, and θ is the angle between the airflow direction and the main acoustic axis direction. Environmental data is quantized into an acoustic path feature vector P. τ =[v, d, H,δ air ] T Where τ is the current time, d is the distance to the obstacle, H is the humidity, T is the matrix dimension, the speed of sound v = 331.4 + 0.6t, and t is the temperature.

2. The method according to claim 1, characterized in that, The direction of the main sound axis is defined as the reference axis of the center of the speaker array pointing to the error microphone; the method for calculating the airflow direction angle θ is as follows: three airflow sensors distributed at 120° are arranged at the heat dissipation outlet of the charging pile for periodic detection, and the angle between the airflow vector direction and the main sound axis is calculated based on the triangulation method.

3. The method according to claim 1, characterized in that, The optimization of the RNN-GRU model includes: (a) It adopts a dual-mode configurable architecture, which includes a basic mode and a high-performance mode. The basic mode includes two GRU layers, two Dropout layers and residual connections, while the high-performance mode includes three GRU layers, two Dropout layers and residual connections. The mode is dynamically switched according to the proportion of high-frequency noise energy. (b) Optimize by transfer learning, pre-train the underlying GRU parameters on the basic dataset; freeze the underlying parameters and fine-tune the top-level network using specific samples from charging piles; use semi-supervised learning with a confidence level greater than 0.95 for unlabeled samples; (c) During the pre-training phase, a domain adaptation mechanism is implemented through a gradient inversion layer to minimize the distribution difference between general noise and charging pile noise.

4. The method according to claim 1, characterized in that, The RNN-GRU model is trained with mixed precision, quantized with INT8 and calibrated for dynamic range. After pruning, 63%-67% of the core parameters are retained, and sparsity is maintained by the L2 regularization term of the AdamW optimizer. The signal-to-noise ratio loss of the model after pruning is ≤1dB.

5. The method according to claim 1, characterized in that, When calculating the frequency domain weighted MSE, 24 critical frequency bands are divided, and the weight of each frequency band is positively correlated with the A-weighting curve. When instantaneous impulse noise is detected, the weight of the 2-4kHz frequency band is temporarily increased to 1.5 times the base weight.

6. The method according to claim 1, characterized in that, When an obstacle movement is detected that causes a change in the acoustic path propagation delay, the output lead Δt is calculated based on the distance change rate Δd / Δt. adv Δt adv = (d new - d old ) / v where d new d represents the distance to the obstacle after the movement. old Let be the distance to the obstacle before it moves, and v = 331.4 + 0.6t, where t is the temperature. Based on Δt... adv Adjust the output timing of the phase-reversing acoustic wave.

7. A charging pile noise reduction system, used to implement the method according to any one of claims 1-6, characterized in that, Includes the following modules: The multi-source signal acquisition module includes: a multi-channel microphone array configured to acquire composite noise signals during the operation of the charging pile, wherein the composite noise signals include broadband noise from the cooling system fan, transformer vibration noise, and power module switching noise; An environmental sensor array is configured to collect environmental data, including temperature, humidity, and airflow speed parameters; a positioning module is configured to monitor the movement of obstacles within a preset range in real time. The data processing module is configured to perform time-domain normalization, frame-segmentation, and data enhancement on the composite noise signal, converting it into time-series data; quantize the environmental data into acoustic path feature vectors, and fuse them with the time-series data into an input matrix through channel splicing; An adaptive noise reduction engine includes: an optimized RNN-GRU model that receives the input matrix and generates an antiphase acoustic signal; and a multi-band directional loudspeaker array that outputs the antiphase acoustic signal. The closed-loop control module includes: an error microphone to collect residual noise; a dynamic adjustment unit to calculate the frequency domain weighted mean square error based on the residual noise; an acoustic path model updated in conjunction with real-time environmental data; and dynamic adjustment of the output timing, amplitude, and phase of the anti-phase sound wave to ensure that the generation delay of the anti-phase sound wave is within a preset range.

8. A noise reduction device for charging piles, characterized in that, The apparatus includes a memory and a processor, the memory storing a computer program that can be executed by the processor to implement the method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that can be executed by a processor to implement the method as described in any one of claims 1-6.

Images