A transformer noise reduction method and system based on deep learning

Abstract

This application provides a transformer noise reduction method and system based on deep learning. The method includes: determining several real-time noise sources currently existing in the transformer and corresponding noise source identification information; applying corresponding marker signals to each of the real-time noise sources according to the noise source identification information and a preset signal modulation parameter set to obtain a mixed noise signal; inputting the mixed noise signal into a preset noise separation model to perform noise identification and noise separation on the mixed noise signal through the noise separation model to obtain noise identification signals corresponding to each of the real-time noise sources; performing feature quantization analysis on each of the noise identification signals to determine the current dominant noise mode of the transformer; generating a corresponding noise reduction strategy according to the dominant noise mode and a preset noise reduction target, and performing noise reduction on the transformer according to the noise reduction strategy to improve the noise reduction effect of the transformer.

Description

Technical Field

[0001] This application relates to the field of transformer noise reduction technology, and in particular to a transformer noise reduction method and system based on deep learning. Background Technology

[0002] Transformers, as critical equipment in power systems, generate noise that not only pollutes the environment but can also interfere with equipment condition monitoring and fault diagnosis. Traditional noise reduction methods primarily rely on passive sound insulation, sound-absorbing materials, or active noise control technologies based on fixed filters. These methods typically treat the overall noise of the transformer as a single object, lacking the ability to finely analyze the internal components of the noise. In actual transformer operation, noise is generated by multiple physical mechanisms, including core magnetostriction, winding electromagnetic forces, and cooling systems. Its spectral characteristics dynamically change with load, temperature, and equipment status. Traditional methods struggle to adaptively track and distinguish these complex sound sources, resulting in noise reduction measures that are not targeted enough and have limited effectiveness. Furthermore, the fixed processing frequency band may introduce new noise problems or affect the normal operation of the equipment.

[0003] Furthermore, transformer noise exhibits typical characteristics of complexity and time-varying nature, with various noise sources overlapping in both the time and frequency domains to form complex, non-stationary, and non-Gaussian signals. Most existing technologies directly filter or cancel the mixed noise signal as a whole, failing to achieve physical separation and independent analysis of the noise sources. Therefore, existing technologies have limitations in precisely separating and specifically suppressing transformer composite noise, resulting in overall noise reduction accuracy and adaptability that cannot meet the requirements for quiet operation and intelligent maintenance of high-performance transformers. Summary of the Invention

[0004] To address the aforementioned technical issues, this application provides a transformer noise reduction method and system based on deep learning, which effectively separates and identifies the dominant noise sources, thereby improving the noise reduction effect of the transformer.

[0005] In a first aspect, embodiments of this application provide a transformer noise reduction method based on deep learning, comprising: The real-time noise signal of the transformer during a preset time period is acquired, and the real-time noise signal is input into a preset noise identification model to identify and determine several real-time noise sources currently existing in the transformer and the corresponding noise source identification information. The noise identification model is constructed based on several historical noise signals of the transformer. Each of the real-time noise sources is given a corresponding marker signal according to the noise source identification information and the preset signal modulation parameter set to obtain a mixed noise signal. The signal modulation parameter set is set according to the noise source identification information of all historical noise sources of the transformer. The mixed noise signal is input into a preset noise separation model to identify and separate noise based on each marker signal in the mixed noise signal, thereby obtaining the noise identification signal corresponding to each of the real-time noise sources. The noise separation model is constructed based on a deep learning model and trained based on several historical mixed noise signals. The noise identification signals are subjected to feature quantization analysis, and the current dominant noise mode of the transformer is determined based on the feature quantization analysis results and preset classification rules. Based on the dominant noise mode and the preset noise reduction target, a corresponding noise reduction strategy is generated, and the transformer is noise-reduced according to the noise reduction strategy.

[0006] This application provides a deep learning-based transformer noise reduction method. By actively adding identification signals to each current noise source, it achieves accurate identification and separation of mixed noise signals, and then targets the dominant noise mode for targeted noise reduction, effectively improving the noise reduction effect of the transformer. Traditional noise reduction methods often treat the overall noise of the transformer as a single object of processing, lacking fine differentiation of different internal noise sources, resulting in insufficient targeting of noise reduction measures and poor noise reduction effect. In contrast, this embodiment first analyzes the real-time noise signal through a preset noise identification model, thereby accurately identifying each current noise source and its characteristics, laying the foundation for subsequent targeted processing. Then, using the identified noise source identification information and a preset signal modulation parameter set, a labeling signal is applied to each real-time noise source to generate a mixed noise signal, and a deep learning-based noise separation model is used to separate the mixed signal, realizing the physical decomposition of composite noise. This step overcomes the limitation of traditional methods in handling noise aliasing, enabling various noise sources to be analyzed and processed independently. Finally, by performing feature quantization analysis on the separated noise identification signals, the dominant noise mode is determined, and a noise reduction strategy is generated and executed accordingly, achieving precision and adaptability in the noise reduction process. Overall, this embodiment demonstrates the ability to dynamically adapt to noise changes under different operating conditions, significantly improving the accuracy and efficiency of transformer noise reduction, thereby effectively reducing environmental noise pollution from transformers.

[0007] Furthermore, the step of constructing the noise identification model based on several historical noise signals of the transformer includes: Several historical noise signals of the transformer are acquired through several sensors; Based on the preset positions of each sensor, source component analysis is performed on each historical noise signal to obtain several noise source type samples and corresponding historical noise source identification information. The historical noise source identification information includes the noise generation location, noise propagation path, and noise dominant frequency distribution range. The noise recognition model is constructed based on the correspondence between each of the historical noise signals, each of the noise source type samples, and each of the historical noise source identification information.

[0008] This application details the construction process of a noise identification model. Historical noise signals are acquired through multiple sensors, and source component analysis is performed based on the preset locations of the sensors. This yields noise source type samples and corresponding historical noise source identification information, fully utilizing multi-source sensor data and spatial location information to more comprehensively capture the spatiotemporal distribution characteristics of transformer noise. Through in-depth analysis of historical noise signals, this embodiment can learn the typical acoustic characteristics of different noise sources (such as core magnetostriction, winding electromagnetic force, cooling systems, etc.) and their correlation with noise generation locations, propagation paths, and dominant frequency distributions. This model, built based on historical data, not only improves the accuracy of noise identification but also enhances its adaptability to different transformer types and operating conditions. Furthermore, by establishing the correspondence between noise signals, noise source types, and identification information, the model can more effectively generalize to new real-time noise signal processing processes. This provides a more reliable basis for subsequent noise separation and denoising strategy generation, ensuring the accuracy of subsequent noise identification and separation, and improving the noise reduction effect of the transformer.

[0009] Furthermore, the step of applying corresponding marker signals to each of the real-time noise sources according to the noise source identification information and the preset signal modulation parameter set to obtain a mixed noise signal includes: Based on the noise source identification information and the signal modulation parameter set, determine the modulation mode type, modulation frequency, modulation amplitude and modulation input corresponding to each real-time noise source; A modulation signal corresponding to each of the real-time noise sources is generated based on each of the modulation scheme types, each of the modulation frequencies, and each of the modulation amplitudes. According to the power requirements of each modulation input, each modulation signal is amplified to generate a corresponding marker signal; According to each of the modulation inputs, a corresponding marker signal is applied to each of the real-time noise sources to obtain the mixed noise signal.

[0010] This application provides a method for applying a marker signal. First, based on the identification information, the modulation method, frequency, amplitude, and modulation input of each noise source are determined, ensuring the targeting and specificity of the marker signal. After generating the modulation signal, it is amplified according to the power requirements of each modulation input, ensuring that the marker signal can be effectively embedded into the original noise without being submerged. Finally, the amplified marker signal is precisely applied to the corresponding position on the transformer, forming a mixed noise signal. This embodiment actively introduces an identifiable marker signal, providing clear "clues" or "guidance" for subsequent noise separation. Compared to traditional blind source separation methods that rely entirely on signal statistical characteristics, this embodiment utilizes prior modulation information, greatly reducing the difficulty of noise separation and improving separation accuracy under strong background noise and dynamic operating conditions. Simultaneously, this controllable marker signal application method enhances the ability to independently analyze different noise sources, creating favorable conditions for subsequent feature extraction and noise reduction strategy formulation, and improving the noise reduction effect of the transformer.

[0011] In one possible implementation, obtaining the signal modulation parameter set based on the noise source identification information of all historical noise sources of the transformer includes: The generation location and propagation path of each historical noise source are determined based on the noise source identification information, and then the modulation input corresponding to each historical noise source is determined based on the generation location and propagation path. Based on the identification information of each noise source, a candidate modulation parameter set corresponding to each historical noise source is determined. The candidate modulation parameter set includes several candidate modulation methods, several candidate modulation frequencies, and several candidate modulation amplitudes. For each historical noise source, several candidate parameter combinations are generated in a random combination manner based on the corresponding candidate modulation parameter set, and modulation separation evaluation is performed on each candidate parameter combination to obtain each corresponding evaluation result. Then, the optimal parameter combination is determined based on each evaluation result, and the signal modulation parameter combination of the noise source is generated based on the optimal parameter combination and the modulation input of the historical noise source. The signal modulation parameter set is constructed by combining the signal modulation parameters of each of the historical noise sources.

[0012] This application provides a method for constructing a signal modulation parameter set. Based on the identification information of all historical noise sources, a systematic parameter optimization process is used to determine the optimal modulation parameter combination for each noise source. First, the modulation entry point is determined according to the noise source's generation location and propagation path, ensuring that the marker signal can be effectively coupled to the target noise. Second, a candidate modulation parameter set is defined for each historical noise source, and multiple candidate schemes are generated through random combinations. Then, modulation separability is evaluated for each candidate parameter combination, and the optimal combination is selected. This evaluation and optimization process is crucial because it ensures sufficient distinguishability between marker signals designed for different noise sources, enabling accurate identification and separation in subsequent mixed signal separation. By integrating the optimal parameter combinations of all historical noise sources, a systematic signal modulation parameter set is finally constructed, laying an important data foundation for subsequent marker signal application and noise identification, and improving the performance and reliability of the entire noise reduction system.

[0013] Furthermore, for any of the candidate parameter combinations, the modulation separation evaluation of each candidate parameter combination to obtain the corresponding evaluation results includes: Random candidate parameter combinations for each of the other historical noise sources are obtained by random sampling. A first marker signal is generated in the simulation environment based on the candidate parameter combination, and a first index of the candidate parameter combination is determined based on the signal-to-noise ratio of the marker signal. Based on each of the random candidate parameter combinations, generate corresponding second marker signals; The correlation or orthogonality between the first marker signal and each of the second marker signals is calculated respectively, and then the second index of the candidate parameter combination is determined based on the correlation or orthogonality. The first indicator and the second indicator are weighted and summed according to preset weight parameters to obtain the evaluation result of the candidate parameter combination.

[0014] This application further refines the specific steps of modulation separability evaluation, providing a quantitative evaluation system for selecting the optimal modulation parameter combination. The evaluation process first generates a first labeled signal corresponding to the candidate parameter combination through simulation, and determines a first index based on its signal-to-noise ratio (SNR), reflecting the clarity and anti-interference capability of the labeled signal itself. Simultaneously, candidate parameter combinations from other noise sources are randomly selected to generate corresponding second labeled signals, and the correlation or orthogonality between the first labeled signal and each second labeled signal is calculated, thereby determining a second index. The second index directly measures the separability between labeled signals from different noise sources and is crucial for ensuring successful subsequent noise separation. Finally, the two indices are weighted and summed using preset weights to obtain a comprehensive evaluation result. This dual-index evaluation mechanism takes into account both the individual quality of the labeled signals and the differences between groups, avoiding the bias that may arise from considering only a single factor. Through the modulation separability evaluation provided in this embodiment, modulation parameters that perform well in both SNR and separation can be selected, thereby maximizing the effectiveness of the labeled signals and providing a high-quality input data foundation for subsequent deep learning models to achieve high-precision noise source separation, thus improving the noise reduction effect of the transformer.

[0015] In one possible implementation, the noise separation model includes a first feature encoding sub-network, a second feature encoding sub-network, a feature joint network, an association attention network, and a separation decoding network. The step of inputting the mixed noise signal into a preset noise separation model to perform noise identification and separation on the mixed noise signal based on each labeled signal in the mixed noise signal, and obtaining noise identification signals corresponding to each of the real-time noise sources, includes: The mixed noise signal is input into the noise separation model; The first feature encoding subnetwork extracts mixed noise features related to acoustic properties from the mixed noise signal; Based on the signal modulation parameter set, the second feature coding subnetwork encodes several signal modulation parameters corresponding to the mixed noise signal into mixed noise signal modulation parameter features. The feature combination network is used to combine the mixed noise features and the mixed noise signal modulation parameter features to obtain joint feature samples. Based on the signal modulation parameter set, the noise source weight distribution of the joint feature samples is calculated through the associated attention network; The joint feature samples are decoded by the separation decoding network according to the noise source weight distribution to obtain each noise identification signal.

[0016] This application clarifies the data processing flow within the noise separation model. A first feature encoding sub-network extracts deep features related to acoustic characteristics from the original mixed noise signal, capturing the essential time-frequency domain information of the noise. Simultaneously, a second feature encoding sub-network encodes known signal modulation parameters as features, providing the model with prior knowledge about the "identity" of the noise sources. Then, a feature joint network fuses these two types of features to form joint feature samples containing original signal characteristics and manually labeled information, greatly enriching the information dimensions available to the model. Based on this, an association attention network calculates the weight distribution of different noise sources in the joint features. Its core function is to simulate and strengthen the model's differentiated allocation of attention to each noise source, thereby highlighting the target source in the aliased signal. Finally, a separation decoding network decodes the joint features based on this weight distribution, reconstructing each independent noise identification signal. This end-to-end network architecture incorporating an attention mechanism significantly improves the model's separation performance and generalization ability in complex aliasing scenarios, providing a technical guarantee for achieving accurate physical separation of noise sources and improving the noise reduction effect of transformers.

[0017] Furthermore, the step of performing feature quantization analysis on each of the noise identification signals, and determining the current dominant noise mode of the transformer based on the feature quantization analysis results and preset classification rules, includes: For each of the noise identification signals, the noise energy characteristics of the noise identification signal are determined by analyzing the sound power or vibration energy of the noise identification signal; By analyzing the main frequency amplitude and the amplitude of each harmonic of the noise identification signal, the main frequency and harmonic distribution characteristics of the noise identification signal are determined. The temporal stability characteristics of the noise identification signal are determined by calculating the variance of the energy or main frequency amplitude fluctuation of the noise identification signal over several preset time periods. By analyzing the peak-to-peak value and fluctuation rate of the instantaneous amplitude of the noise identification signal, the signal amplitude fluctuation characteristics of the noise identification signal are determined; Based on the noise energy characteristics, harmonic distribution characteristics, time stability characteristics, signal amplitude fluctuation characteristics, and preset classification rules of each noise identification signal, the current dominant noise mode of the transformer is determined.

[0018] This application provides a method for multi-dimensional feature quantization analysis of separated noise identification signals. The feature quantization analysis covers four core dimensions: determining noise energy characteristics through sound power or vibration energy, directly reflecting the total intensity of the noise source and its contribution to the overall noise; determining the distribution characteristics of the dominant frequency and harmonics by analyzing the dominant frequency amplitude and subharmonic amplitude, revealing the spectral composition of the noise and helping to identify its physical generation mechanism (such as specific mechanical vibration frequencies or electromagnetic harmonics); determining time stability characteristics by calculating the fluctuation variance of energy or dominant frequency amplitude over multiple time periods, characterizing the time-varying characteristics of the noise, and serving as an important basis for distinguishing steady-state noise from transient and impulsive noise; and determining signal amplitude fluctuation characteristics by analyzing the peak-to-peak value and volatility of the instantaneous amplitude, describing the instantaneous dynamic range of the noise signal. By integrating these four types of characteristics and comprehensively evaluating each noise source from multiple perspectives such as energy, spectrum, and time-domain stability according to preset classification rules, the dominant noise mode that needs to be prioritized can be determined more scientifically and robustly. This provides accurate data support for subsequently developing the most targeted noise reduction strategy and improves the noise reduction effect of transformers.

[0019] Furthermore, the step of generating a corresponding noise reduction strategy based on the dominant noise mode and a preset noise reduction target, and then performing noise reduction on the transformer according to the noise reduction strategy, includes: The corresponding noise reduction strategy is determined from a preset strategy matching library based on the dominant noise pattern. Based on the feature quantization analysis results of the target noise identification signal corresponding to the dominant noise mode and the noise reduction target, the specific execution parameters of the noise reduction strategy are determined. The transformer is subjected to noise reduction according to the specific execution parameters of the noise reduction strategy.

[0020] This application illustrates how to generate and execute noise reduction strategies based on a determined dominant noise mode and preset noise reduction targets, achieving a closed loop from "analysis and diagnosis" to "precise intervention." First, a corresponding basic noise reduction strategy is retrieved from a preset strategy matching library based on the dominant noise mode. This strategy library integrates expert knowledge and historical successful experiences for different noise modes, such as adjusting operating parameters, applying reverse sound waves, adding damping materials, and optimizing the structure. Then, combining the specific results obtained from feature quantization analysis of the target noise identification signal and the system's preset noise reduction targets, the selected basic strategy is refined and customized to determine specific execution parameters. Finally, noise reduction operations are performed on the transformer based on these specific parameters, achieving highly customized and target-oriented transformer noise reduction. While meeting the noise reduction targets, it maximizes resource utilization efficiency and improves the transformer's noise reduction effect.

[0021] In one possible implementation, the noise separation model, constructed based on a deep learning model and trained according to several historical mixed noise signals, includes: An initial noise separation model is obtained by constructing a deep learning model; During a preset first historical time period, several single noise signals corresponding to each of the historical noise sources of the transformer are acquired. During a preset second historical time period, each corresponding marker signal is applied to each of the historical noise sources according to the signal modulation parameter set to obtain several historical mixed noise signals of the transformer. A training dataset is constructed based on each of the historical mixed noise signals and each of the single noise signals; The initial noise separation model is trained according to the preset loss function and the training dataset to obtain the noise separation model.

[0022] This application provides a training method for a noise separation model, which is fundamental to ensuring high separation accuracy in practical applications. The key step lies in constructing the training dataset: First, several single noise signals from the transformer are collected within a first historical time period, forming the "labels" required for supervised learning. Then, in a second historical time period, using a pre-set set of signal modulation parameters, corresponding labeled signals are applied to these known historical noise sources, artificially synthesizing "historical mixed noise signals" simulating actual operating conditions. This method of controllable mixing based on real signals generates a dataset that closely approximates the complexity of real-world noise while possessing accurate source signals as ground truth references. Finally, the initial model is trained using this dataset and a pre-set loss function, enabling the model to identify and separate individual noise signals from the mixed noise, thus obtaining the final noise separation model. This training paradigm allows the model to learn the inherent mapping relationship between each source signal from a mixed signal with known modulation labels, greatly improving the model's generalization ability and robustness under real and variable operating conditions. This provides a core guarantee for the reliability of the entire noise reduction method and improves the noise reduction effect on the transformer.

[0023] Secondly, embodiments of this application provide a transformer noise reduction system based on deep learning, including an acquisition module, a modulation module, a noise separation module, an analysis module, and a noise reduction module; The acquisition module is used to acquire the real-time noise signal of the transformer during a preset time period and input the real-time noise signal into a preset noise identification model to identify and determine several real-time noise sources currently existing in the transformer and the corresponding noise source identification information. The noise identification model is constructed based on several historical noise signals of the transformer. The modulation module is used to apply corresponding marker signals to each of the real-time noise sources according to the noise source identification information and the preset signal modulation parameter set to obtain a mixed noise signal. The signal modulation parameter set is set according to the noise source identification information of all historical noise sources of the transformer. The noise separation module is used to input the mixed noise signal into a preset noise separation model, so as to perform noise identification and noise separation on the mixed noise signal according to each marker signal in the mixed noise signal, and obtain the noise identification signal corresponding to each of the real-time noise sources. The noise separation model is constructed based on a deep learning model and trained based on several historical mixed noise signals. The analysis module is used to perform feature quantization analysis on each of the noise identification signals, and determine the current dominant noise mode of the transformer based on the feature quantization analysis results and preset classification rules. The noise reduction module is used to generate a corresponding noise reduction strategy based on the dominant noise mode and the preset noise reduction target, and to reduce the noise of the transformer according to the noise reduction strategy. Attached Figure Description

[0024] Figure 1 A flowchart illustrating a deep learning-based transformer noise reduction method provided in this application embodiment; Figure 2 This is a schematic diagram of a transformer noise reduction system based on deep learning, provided as an embodiment of this application. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0026] It should be noted that the step numbers in this document are only for the convenience of explaining the specific embodiments and are not intended to limit the order in which the steps are performed. In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0027] Example 1: like Figure 1 As shown, Embodiment 1 provides a transformer noise reduction method based on deep learning, including steps S1-S5: Step S1: Obtain the real-time noise signal of the transformer during a preset time period, and input the real-time noise signal into a preset noise identification model to identify and determine several real-time noise sources currently existing in the transformer and the corresponding noise source identification information. The noise identification model is constructed based on several historical noise signals of the transformer. Step S2: Apply corresponding marker signals to each of the real-time noise sources according to the noise source identification information and the preset signal modulation parameter set to obtain a mixed noise signal. The signal modulation parameter set is set according to the noise source identification information of all historical noise sources of the transformer. Step S3: Input the mixed noise signal into a preset noise separation model to identify and separate noise from the mixed noise signal based on each marker signal in the mixed noise signal, and obtain the noise identification signal corresponding to each of the real-time noise sources. The noise separation model is constructed based on a deep learning model and trained based on several historical mixed noise signals. Step S4: Perform feature quantization analysis on each of the noise identification signals, and determine the current dominant noise mode of the transformer based on the feature quantization analysis results and preset classification rules; Step S5: Generate a corresponding noise reduction strategy based on the dominant noise mode and the preset noise reduction target, and perform noise reduction on the transformer according to the noise reduction strategy.

[0028] This application provides a deep learning-based transformer noise reduction method. By actively adding identification signals to each current noise source, it achieves accurate identification and separation of mixed noise signals, and then targets the dominant noise mode for targeted noise reduction, effectively improving the noise reduction effect of the transformer. Traditional noise reduction methods often treat the overall noise of the transformer as a single object of processing, lacking fine differentiation of different internal noise sources, resulting in insufficient targeting of noise reduction measures and poor noise reduction effect. In contrast, this embodiment first analyzes the real-time noise signal through a preset noise identification model, thereby accurately identifying each current noise source and its characteristics, laying the foundation for subsequent targeted processing. Then, using the identified noise source identification information and a preset signal modulation parameter set, a labeling signal is applied to each real-time noise source to generate a mixed noise signal, and a deep learning-based noise separation model is used to separate the mixed signal, realizing the physical decomposition of composite noise. This step overcomes the limitation of traditional methods in handling noise aliasing, enabling various noise sources to be analyzed and processed independently. Finally, by performing feature quantization analysis on the separated noise identification signals, the dominant noise mode is determined, and a noise reduction strategy is generated and executed accordingly, achieving precision and adaptability in the noise reduction process. Overall, this embodiment demonstrates the ability to dynamically adapt to noise changes under different operating conditions, significantly improving the accuracy and efficiency of transformer noise reduction, thereby effectively reducing environmental noise pollution from transformers.

[0029] Furthermore, in step S1, constructing the noise identification model based on several historical noise signals of the transformer includes: Several historical noise signals of the transformer are acquired through several sensors; Based on the preset positions of each sensor, source component analysis is performed on each historical noise signal to obtain several noise source type samples and corresponding historical noise source identification information. The historical noise source identification information includes the noise generation location, noise propagation path, and noise dominant frequency distribution range. The noise recognition model is constructed based on the correspondence between each of the historical noise signals, each of the noise source type samples, and each of the historical noise source identification information.

[0030] This application details the construction process of a noise identification model. Historical noise signals are acquired through multiple sensors, and source component analysis is performed based on the preset locations of the sensors. This yields noise source type samples and corresponding historical noise source identification information, fully utilizing multi-source sensor data and spatial location information to more comprehensively capture the spatiotemporal distribution characteristics of transformer noise. Through in-depth analysis of historical noise signals, this embodiment can learn the typical acoustic characteristics of different noise sources (such as core magnetostriction, winding electromagnetic force, cooling systems, etc.) and their correlation with noise generation locations, propagation paths, and dominant frequency distributions. This model, built based on historical data, not only improves the accuracy of noise identification but also enhances its adaptability to different transformer types and operating conditions. Furthermore, by establishing the correspondence between noise signals, noise source types, and identification information, the model can more effectively generalize to new real-time noise signal processing processes. This provides a more reliable basis for subsequent noise separation and denoising strategy generation, ensuring the accuracy of subsequent noise identification and separation, and improving the noise reduction effect of the transformer.

[0031] In a preferred embodiment, under normal historical operation or specific test conditions of the transformer, acoustic and vibration sensors are used to collect the mixed noise and vibration signals emitted by the transformer to obtain prior samples of noise sources. The acoustic performance of the transformer under different loads, temperatures, and health conditions is recorded. Then, source component analysis is performed on the collected prior samples of noise sources to obtain corresponding noise source type samples, which include at least core magnetostriction noise, winding electromagnetic force noise, and structural enclosure vibration noise. Among them, core magnetostriction noise is generated by the periodic expansion and contraction deformation of the transformer core under the action of an alternating magnetic field, and is characterized by electromagnetic noise with power frequencies such as 50Hz / 100Hz and their harmonics. Winding electromagnetic force noise is generated by the periodic Lorentz force generated by the current in the winding under the action of the leakage magnetic field, which causes the winding to vibrate. Its frequency components are related to the load current and the resonant frequency of the winding structure. Structural enclosure vibration noise is the noise radiated from the surface of the tank caused by the vibration of the core and windings transmitted to the tank wall through the mechanical structure. Its spectrum is often wider and may include the structural resonant frequency.

[0032] Then, for each type of noise source in the noise source type sample, characteristic attributes that can characterize its essence are extracted to determine the noise source identification information, including the noise generation location, noise propagation path, and noise dominant frequency distribution range. The noise generation location is used to clarify the physical source location of this type of noise, such as iron core, high-voltage winding, low-voltage winding, oil tank sidewall, etc.; the noise propagation path describes the transmission path of vibration or sound waves from the source to the measurement point, such as transmission through insulating oil and structural components or direct radiation through air; the noise dominant frequency distribution range is used to determine the concentration range of this type of noise energy in the frequency domain, for example, the dominant frequency of iron core noise is concentrated in 100Hz~600Hz, and structural noise may be concentrated in 500Hz~2000Hz. Then, based on the mapping relationship between each type of noise source and the corresponding noise source identification information, a noise identification model is obtained. This model can infer the most likely known noise source type to which the noise signal belongs and provide a physical characteristic description based on new noise observation data and its measurement point location.

[0033] Further, in step S2, applying corresponding marker signals to each of the real-time noise sources according to the noise source identification information and the preset signal modulation parameter set to obtain a mixed noise signal includes: Based on the noise source identification information and the signal modulation parameter set, determine the modulation mode type, modulation frequency, modulation amplitude and modulation input corresponding to each real-time noise source; A modulation signal corresponding to each of the real-time noise sources is generated based on each of the modulation scheme types, each of the modulation frequencies, and each of the modulation amplitudes. According to the power requirements of each modulation input, each modulation signal is amplified to generate a corresponding marker signal; According to each of the modulation inputs, a corresponding marker signal is applied to each of the real-time noise sources to obtain the mixed noise signal.

[0034] This application provides a method for applying a marker signal. First, based on the identification information, the modulation method, frequency, amplitude, and modulation input of each noise source are determined, ensuring the targeting and specificity of the marker signal. After generating the modulation signal, it is amplified according to the power requirements of each modulation input, ensuring that the marker signal can be effectively embedded into the original noise without being submerged. Finally, the amplified marker signal is precisely applied to the corresponding position on the transformer, forming a mixed noise signal. This embodiment actively introduces an identifiable marker signal, providing clear "clues" or "guidance" for subsequent noise separation. Compared to traditional blind source separation methods that rely entirely on signal statistical characteristics, this embodiment utilizes prior modulation information, greatly reducing the difficulty of noise separation and improving separation accuracy under strong background noise and dynamic operating conditions. Simultaneously, this controllable marker signal application method enhances the ability to independently analyze different noise sources, creating favorable conditions for subsequent feature extraction and noise reduction strategy formulation, and improving the noise reduction effect of the transformer.

[0035] In a preferred embodiment, a signal modulation module is provided to convert the signal modulation parameter set into a real physical signal applied to the operating transformer and act on the transformer. The signal modulation module includes a signal generator and power amplifier, an actuator / injection device, and a safety isolation and protection circuit. The signal generator and power amplifier generate a corresponding low-power electrical signal, such as a sine wave of a specific frequency or a coded sequence, according to the modulation mode type, modulation frequency, and modulation amplitude in the signal modulation parameter set, and amplify it to a power level that can effectively act on the target modulation entry. The actuator / injection device converts the amplified electrical signal into a corresponding physical excitation according to the modulation entry determined in the signal modulation parameter set. For example, for an electrical entry, the signal is injected into the grid side or a specific terminal of the transformer through an isolation transformer or coupler; for a mechanical / acoustic entry, vibration or acoustic signals are applied to a specific location in the tank through a piezoelectric ceramic actuator or exciter. The safety isolation and protection circuit ensures that the injected marking signal will not pose a threat to the safe operation of the transformer or the power quality of the grid under any circumstances. The signal modulation module receives and parses the signal modulation parameter set, controls the signal generator, amplifier and actuator to work together, accurately outputs the target marking signal, and monitors the amplitude and frequency of the injected signal in real time to ensure that it strictly conforms to the preset parameters, and automatically shuts down in case of abnormality.

[0036] After the marker signal modulation is completed, a multi-sensor array deployed around and at key locations of the transformer is used to acquire the transformer's mixed noise signal in real time. This includes all inherent noise and the total sound / vibration field signal of actively added markers. The sensor types include high-precision acoustic sensors and vibration sensors, which are arranged in a linear, circular, or area array according to the needs of sound source localization and separation to capture the spatial orientation and propagation characteristics of the noise. All sensors are connected to a synchronous data acquisition component to ensure that the data from all channels are strictly aligned in time and are continuously acquired at a sampling rate much higher than twice the highest frequency of the target noise. This yields the transformer's mixed noise signal, which includes all inherent noise generated by the transformer itself, such as that from the core, windings, and cooler, as well as the marker signal generated by the signal modulation module according to a known set of marker signal modulation parameters.

[0037] In one possible implementation, step S2, which involves setting and obtaining the signal modulation parameter set based on the noise source identification information of all historical noise sources of the transformer, includes: The generation location and propagation path of each historical noise source are determined based on the noise source identification information, and then the modulation input corresponding to each historical noise source is determined based on the generation location and propagation path. Based on the identification information of each noise source, a candidate modulation parameter set corresponding to each historical noise source is determined. The candidate modulation parameter set includes several candidate modulation methods, several candidate modulation frequencies, and several candidate modulation amplitudes. For each historical noise source, several candidate parameter combinations are generated in a random combination manner based on the corresponding candidate modulation parameter set, and modulation separation evaluation is performed on each candidate parameter combination to obtain each corresponding evaluation result. Then, the optimal parameter combination is determined based on each evaluation result, and the signal modulation parameter combination of the noise source is generated based on the optimal parameter combination and the modulation input of the historical noise source. The signal modulation parameter set is constructed by combining the signal modulation parameters of each of the historical noise sources.

[0038] This application provides a method for constructing a signal modulation parameter set. Based on the identification information of all historical noise sources, a systematic parameter optimization process is used to determine the optimal modulation parameter combination for each noise source. First, the modulation entry point is determined according to the noise source's generation location and propagation path, ensuring that the marker signal can be effectively coupled to the target noise. Second, a candidate modulation parameter set is defined for each historical noise source, and multiple candidate schemes are generated through random combinations. Then, modulation separability is evaluated for each candidate parameter combination, and the optimal combination is selected. This evaluation and optimization process is crucial because it ensures sufficient distinguishability between marker signals designed for different noise sources, enabling accurate identification and separation in subsequent mixed signal separation. By integrating the optimal parameter combinations of all historical noise sources, a systematic signal modulation parameter set is finally constructed, laying an important data foundation for subsequent marker signal application and noise identification, and improving the performance and reliability of the entire noise reduction system.

[0039] Furthermore, for any of the candidate parameter combinations, the modulation separation evaluation of each candidate parameter combination to obtain the corresponding evaluation results includes: Random candidate parameter combinations for each of the other historical noise sources are obtained by random sampling. A first marker signal is generated in the simulation environment based on the candidate parameter combination, and a first index of the candidate parameter combination is determined based on the signal-to-noise ratio of the marker signal. Based on each of the random candidate parameter combinations, generate corresponding second marker signals; The correlation or orthogonality between the first marker signal and each of the second marker signals is calculated respectively, and then the second index of the candidate parameter combination is determined based on the correlation or orthogonality. The first indicator and the second indicator are weighted and summed according to preset weight parameters to obtain the evaluation result of the candidate parameter combination.

[0040] This application further refines the specific steps of modulation separability evaluation, providing a quantitative evaluation system for selecting the optimal modulation parameter combination. The evaluation process first generates a first labeled signal corresponding to the candidate parameter combination through simulation, and determines a first index based on its signal-to-noise ratio (SNR), reflecting the clarity and anti-interference capability of the labeled signal itself. Simultaneously, candidate parameter combinations from other noise sources are randomly selected to generate corresponding second labeled signals, and the correlation or orthogonality between the first labeled signal and each second labeled signal is calculated, thereby determining a second index. The second index directly measures the separability between labeled signals from different noise sources and is crucial for ensuring successful subsequent noise separation. Finally, the two indices are weighted and summed using preset weights to obtain a comprehensive evaluation result. This dual-index evaluation mechanism takes into account both the individual quality of the labeled signals and the differences between groups, avoiding the bias that may arise from considering only a single factor. Through the modulation separability evaluation provided in this embodiment, modulation parameters that perform well in both SNR and separation can be selected, thereby maximizing the effectiveness of the labeled signals and providing a high-quality input data foundation for subsequent deep learning models to achieve high-precision noise source separation, thus improving the noise reduction effect of the transformer.

[0041] In a preferred embodiment, the modulation entry point is determined based on the generation location and propagation path of each type of noise in the transformer. For example, for core magnetostrictive noise, the modulation entry point may be the transformer's excitation circuit, which affects the magnetostriction of the core by injecting specific current harmonics; for winding electromagnetic force noise, the modulation entry point may be the current circuit of a specific winding; for structural enclosure vibration noise, the modulation entry point may be the tank wall or a specific structural component, which applies weak vibration excitation through an additional piezoelectric actuator. The modulation entry point is used to define the optimal point and manner of application for physically modulating this type of noise with a marker signal. Then, the modulation parameters are defined, which determine the key technical parameters of the marker signal used to modulate the noise. These include the modulation type, modulation frequency, and modulation amplitude. The modulation type may include amplitude modulation (AM), frequency modulation (FM), specific sequence coding modulation, or direct injection of a sinusoidal / sweep excitation at a specific frequency. The modulation frequency refers to the frequency component of the marker signal itself, such as injecting a 250Hz sinusoidal signal or using a 5Hz signal to amplitude modulate the original noise. The modulation amplitude refers to the strength of the marker signal, which must be balanced between being detectable and not affecting the normal operation of the equipment.

[0042] Then, based on the modulation parameter terms for each type of noise source, different modulation methods, modulation frequencies, and modulation amplitudes are combined at their physically feasible modulation inputs to generate a corresponding modulation candidate parameter group (i.e., the candidate parameter combination) containing multiple possible labeled signal schemes. For example, for iron core magnetostrictive noise, the candidate parameter group may include injected sinusoidal current, frequency 245Hz, amplitude of 0.1% of rated current, injected sinusoidal current, frequency 255Hz, amplitude of 0.2% of rated current, amplitude modulation, modulation frequency 5Hz, and modulation depth 5%. Based on the noise dominant frequency distribution range, the modulation separation performance of each modulation candidate parameter group is evaluated. This assesses whether adding the marker signal makes the target noise source easier to separate and identify in a mixed signal. Evaluation metrics include modulation feature signal-to-noise ratio (SNR), modulation feature cross-correlation coefficient (CCR), and modulation feature recognition accuracy. The SNR is the ratio of the energy of the marker signal to the energy of the background noise, used to evaluate the detectability of the modulation feature in mixed noise in a simulation environment. A higher SNR value indicates easier detection of the marker signal. The CCR calculates the correlation or orthogonality between marker signals from different noise sources. A lower CCR indicates stronger marker uniqueness and less susceptibility to confusion. The modulation feature recognition accuracy calculates the probability of correctly identifying the corresponding noise source based on the marker, evaluating the stability of the modulation feature under different operating conditions.

[0043] Finally, a weighted comprehensive index is calculated using preset weights to evaluate the modulation feature signal-to-noise ratio, modulation feature cross-correlation coefficient, and modulation feature recognition accuracy. This index is used to quantify the overall separation potential of each candidate parameter group. The preset weights are configured based on prior data and feature importance to obtain the modulation separation performance evaluation result. Based on the modulation separation evaluation result, the optimal signal modulation parameter combination for each type of noise source is selected from each modulation candidate parameter group to construct the signal modulation parameter set. For example, at the transformer neutral point, a sinusoidal current signal with an amplitude of 0.15% of the rated current and a frequency of 248Hz is continuously injected as a marker signal to separate the magnetostrictive noise of the iron core, thereby ensuring high-precision noise separation.

[0044] Furthermore, based on the modulation separation evaluation results, several screening parameter groups that are greater than the preset separation index threshold can be selected from each of the modulation candidate parameter groups; then, transformer operation interference evaluation is performed on the several screening parameter groups to obtain interference evaluation results, including a weighted comprehensive index of electrical performance impact, thermal stability impact, and mechanical life impact; finally, based on the interference evaluation results, signal modulation parameter combinations corresponding to each type of noise source are generated from the screening parameter groups.

[0045] Specifically, based on the benchmark separation performance analysis results, the minimum quality requirements of the input signal, and the maximum tolerable marker strength constraint, a preset separation index threshold is configured to eliminate candidate modulation schemes that cannot provide sufficiently strong recognition signals. Based on the comprehensive modulation separation evaluation index in the modulation separation evaluation results, modulation candidate parameter groups with values ​​greater than the preset separation index threshold are selected from the modulation candidate parameter group as screening parameter groups to ensure the effectiveness of the modulation candidate parameter group as a marker signal. Then, a transformer operation interference assessment is performed on the selected parameter groups. This involves evaluating the potential negative impacts on the transformer's normal operation when applied as an external excitation signal. This includes assessing the impact on electrical performance, thermal stability, and mechanical life. Electrical performance impact assesses whether the injected marker signal causes excessive distortion of the current / voltage waveform, malfunctions / misjudgments of protective relays, or interference with the normal operation of online monitoring equipment within the transformer. Thermal stability impact assesses whether the additional copper losses in the windings and iron losses in the core caused by the marker signal lead to local or overall temperature rise exceeding allowable limits, affecting insulation life and load capacity. Mechanical life impact assesses whether the marker signal induces or exacerbates mechanical resonance in the windings, core, or fasteners, accelerating material fatigue or generating harmful cumulative stress that affects the long-term stability of the mechanical structure. The interference assessment result is a weighted composite index of electrical performance, thermal stability, and mechanical life impacts, used to quantify the intrusiveness of the modulation parameter group on transformer operation. Finally, based on the interference assessment results, the signal modulation parameter combination with the least interference and the safest operation is selected from each selected parameter group for each type of noise source, thus constructing the signal modulation parameter set.

[0046] In one possible implementation, in step S3, the noise separation model includes a first feature encoding sub-network, a second feature encoding sub-network, a feature joint network, an association attention network, and a separation decoding network. The step of inputting the mixed noise signal into the preset noise separation model to perform noise identification and noise separation on the mixed noise signal based on each labeled signal in the mixed noise signal, obtaining noise identification signals corresponding to each of the real-time noise sources, includes: The mixed noise signal is input into the noise separation model; The first feature encoding subnetwork extracts mixed noise features related to acoustic properties from the mixed noise signal; Based on the signal modulation parameter set, the second feature coding subnetwork encodes several signal modulation parameters corresponding to the mixed noise signal into mixed noise signal modulation parameter features. The feature combination network is used to combine the mixed noise features and the mixed noise signal modulation parameter features to obtain joint feature samples. Based on the signal modulation parameter set, the noise source weight distribution of the joint feature samples is calculated through the associated attention network; The joint feature samples are decoded by the separation decoding network according to the noise source weight distribution to obtain each noise identification signal.

[0047] This application clarifies the data processing flow within the noise separation model. A first feature encoding sub-network extracts deep features related to acoustic characteristics from the original mixed noise signal, capturing the essential time-frequency domain information of the noise. Simultaneously, a second feature encoding sub-network encodes known signal modulation parameters as features, providing the model with prior knowledge about the "identity" of the noise sources. Then, a feature joint network fuses these two types of features to form joint feature samples containing original signal characteristics and manually labeled information, greatly enriching the information dimensions available to the model. Based on this, an association attention network calculates the weight distribution of different noise sources in the joint features. Its core function is to simulate and strengthen the model's differentiated allocation of attention to each noise source, thereby highlighting the target source in the aliased signal. Finally, a separation decoding network decodes the joint features based on this weight distribution, reconstructing each independent noise identification signal. This end-to-end network architecture incorporating an attention mechanism significantly improves the model's separation performance and generalization ability in complex aliasing scenarios, providing a technical guarantee for achieving accurate physical separation of noise sources and improving the noise reduction effect of transformers.

[0048] In a preferred embodiment, a noise separation model is used to separate noise sources from the mixed noise signal. Specifically, the feature encoding network of the noise separation model includes a first feature encoding subnetwork and a second feature encoding subnetwork. The first feature encoding subnetwork extracts mixed noise features that characterize the acoustic properties of the mixed noise signal through multiple convolutional or transformer layers. These features may correspond to different frequency bands, transient events, spatial correlation patterns, etc. The second feature encoding subnetwork encodes the signal modulation parameter set corresponding to the current mixed signal into mixed noise signal modulation parameter features through an embedding layer or a fully connected network. The feature joint network receives the features output by the first and second feature encoding subnetworks and combines them into a joint feature sample through concatenation, addition, or fusion operations.

[0049] Then, the associative attention network calculates the noise source weight distribution of the joint feature samples based on the signal modulation parameter set. Specifically, the associative attention network uses the features encoded by the signal modulation parameter set as queries and the joint feature samples as keys and values. It evaluates the correlation between each part of the joint feature samples and the given label parameters through attention calculation. The feature parts with high correlation indicate that the feature at that point is likely contributed by the labeled target noise source. The network then outputs the noise source weight distribution, which clearly indicates the time point and frequency component where the target noise source dominates in the joint feature samples. For example, based on the addition of the 248Hz label, it identifies all feature regions that are highly correlated with 248Hz in complex mixed feature samples and considers them to belong to iron core magnetic expansion noise. The separation decoding network decodes joint feature samples based on the noise source weight distribution. That is, it uses the noise source weight distribution as a dynamic mask or conditional parameter to selectively filter and transform the joint features. For example, the weight distribution is applied to the time-frequency representation of the joint features to suppress regions related to non-target sources and enhance regions related to target sources. Finally, the separation decoding network decodes the weighted features back to the time domain or time-frequency domain through operations such as deconvolution or upsampling, and outputs a relatively clean noise source signal that has been estimated and separated. This process is repeated for each type of noise source to obtain multiple separated noise signals.

[0050] Furthermore, in step S4, the step of performing feature quantization analysis on each of the noise identification signals and determining the current dominant noise mode of the transformer based on the feature quantization analysis results and preset classification rules includes: For each of the noise identification signals, the noise energy characteristics of the noise identification signal are determined by analyzing the sound power or vibration energy of the noise identification signal; By analyzing the main frequency amplitude and the amplitude of each harmonic of the noise identification signal, the main frequency and harmonic distribution characteristics of the noise identification signal are determined. The temporal stability characteristics of the noise identification signal are determined by calculating the variance of the energy or main frequency amplitude fluctuation of the noise identification signal over several preset time periods. By analyzing the peak-to-peak value and fluctuation rate of the instantaneous amplitude of the noise identification signal, the signal amplitude fluctuation characteristics of the noise identification signal are determined; Based on the noise energy characteristics, harmonic distribution characteristics, time stability characteristics, signal amplitude fluctuation characteristics, and preset classification rules of each noise identification signal, the current dominant noise mode of the transformer is determined.

[0051] This application provides a method for multi-dimensional feature quantization analysis of separated noise identification signals. The feature quantization analysis covers four core dimensions: determining noise energy characteristics through sound power or vibration energy, directly reflecting the total intensity of the noise source and its contribution to the overall noise; determining the distribution characteristics of the dominant frequency and harmonics by analyzing the dominant frequency amplitude and subharmonic amplitude, revealing the spectral composition of the noise and helping to identify its physical generation mechanism (such as specific mechanical vibration frequencies or electromagnetic harmonics); determining time stability characteristics by calculating the fluctuation variance of energy or dominant frequency amplitude over multiple time periods, characterizing the time-varying characteristics of the noise, and serving as an important basis for distinguishing steady-state noise from transient and impulsive noise; and determining signal amplitude fluctuation characteristics by analyzing the peak-to-peak value and volatility of the instantaneous amplitude, describing the instantaneous dynamic range of the noise signal. By integrating these four types of characteristics and comprehensively evaluating each noise source from multiple perspectives such as energy, spectrum, and time-domain stability according to preset classification rules, the dominant noise mode that needs to be prioritized can be determined more scientifically and robustly. This provides accurate data support for subsequently developing the most targeted noise reduction strategy and improves the noise reduction effect of transformers.

[0052] In a preferred embodiment, each noise signal representing an independent noise source is subjected to refined feature analysis and quantification. This involves extracting its noise energy, dominant frequency and harmonic distribution, time stability characteristics, and signal amplitude fluctuation characteristics to determine noise source-level characteristic parameters. Noise energy is calculated by measuring the sound power or vibration energy of the noise source per unit time, quantifying its overall intensity. Dominant frequency and harmonic distribution accurately analyze the concentrated location of the noise source's energy in the frequency domain and the amplitude of its integer multiples of frequency, directly corresponding to its physical generation mechanism and indicating the core frequency target points that noise reduction measures should focus on. Time stability characteristics assess the smoothness of the noise signal's intensity and frequency changes over time, such as continuous stability, periodic fluctuations with load, or random bursts. This can be obtained by calculating the variance of the signal's energy or dominant frequency amplitude fluctuations over multiple consecutive time segments. Signal amplitude fluctuation characteristics analyze the instantaneous amplitude variation patterns of the signal, such as peak-to-peak value and volatility, reflecting the impact or modulation characteristics of the noise, thereby identifying abnormal states such as structural loosening or discharge. Then, the dominant noise mode is determined based on the noise source-level characteristic parameters of each separated noise signal. Specifically, dominant noise mode labels are defined, including low-frequency electromagnetic dominant type, mid-frequency structural resonance type, broadband flow-induced noise type, and modulation howling type. Based on the pre-trained classification rules of historical noise classification cases, where historical noise classification cases include cases of each known dominant noise mode label and noise source-level characteristic sample parameters representing independent noise sources, a judgment rule is established based on the noise source-level characteristic sample parameters and the dominant noise mode type to identify the most important noise problem type under the current operating conditions. For example, in the low-frequency electromagnetic dominant type, the core noise energy is significantly higher than other sources and the main frequency is concentrated in the low-frequency band; in the mid-frequency structural resonance type, the enclosure vibration noise is prominent and has a peak at a specific narrow-band frequency, indicating the presence of structural resonance; in the broadband flow-induced noise type, the noise energy of the cooling fan or oil pump is distributed over a wide frequency band; in the modulation howling type, a noise source exhibits obvious amplitude or frequency modulation characteristics, which may indicate potential mechanical or electrical problems.

[0053] Furthermore, the step of generating a corresponding noise reduction strategy based on the dominant noise mode and a preset noise reduction target, and then performing noise reduction on the transformer according to the noise reduction strategy, includes: The corresponding noise reduction strategy is determined from a preset strategy matching library based on the dominant noise pattern. Based on the feature quantization analysis results of the target noise identification signal corresponding to the dominant noise mode and the noise reduction target, the specific execution parameters of the noise reduction strategy are determined. The transformer is subjected to noise reduction according to the specific execution parameters of the noise reduction strategy.

[0054] This application illustrates how to generate and execute noise reduction strategies based on a determined dominant noise mode and preset noise reduction targets, achieving a closed loop from "analysis and diagnosis" to "precise intervention." First, a corresponding basic noise reduction strategy is retrieved from a preset strategy matching library based on the dominant noise mode. This strategy library integrates expert knowledge and historical successful experiences for different noise modes, such as adjusting operating parameters, applying reverse sound waves, adding damping materials, and optimizing the structure. Then, combining the specific results obtained from feature quantization analysis of the target noise identification signal and the system's preset noise reduction targets, the selected basic strategy is refined and customized to determine specific execution parameters. Finally, noise reduction operations are performed on the transformer based on these specific parameters, achieving highly customized and target-oriented transformer noise reduction. While meeting the noise reduction targets, it maximizes resource utilization efficiency and improves the transformer's noise reduction effect.

[0055] In a preferred embodiment, in step S5, a noise reduction strategy is solved with a preset noise reduction effect as the target, generating an optimal noise reduction strategy for the identified dominant noise mode. The preset noise reduction effect is a pre-defined quantized target, such as "reducing the A-weighted sound pressure level by 10 dB" or "suppressing the vibration amplitude in the 250Hz frequency band to below 50%". A strategy matching library is constructed to store typical effective strategies corresponding to various noise modes. For example, for low-frequency electromagnetic dominant noise, the noise reduction strategy might be to activate active noise cancellation at specific frequencies and generate anti-phase sound waves for the dominant frequency and harmonics; for mid-frequency structural resonance... The noise reduction strategy might involve controlling actuators installed at specific locations within the enclosure to apply damped vibrations to disrupt resonance, or suggesting the addition of constraints or damping materials at specific locations. Then, based on the currently extracted precise feature parameters, the execution parameters of the matched noise reduction strategy are dynamically optimized. While ensuring noise reduction effectiveness, constraints such as energy consumption, equipment burden, and cost are considered to perform multi-objective optimization, generating corresponding multi-channel noise reduction strategies for multiple separated noise signals. This ensures improved overall noise reduction accuracy. For example, when the noise energy of a certain separated noise source-level feature parameter is 72dB and the main frequency is 100... The harmonics are clearly distributed at 200 Hz and 300 Hz, with time stability fluctuations of less than 1 dB. The signal amplitude fluctuation is characterized by periodic fluctuations, specifically manifested as concentrated main frequency, prominent noise energy, and small fluctuations in time stability and signal amplitude. It can be determined that the dominant noise mode is low-frequency electromagnetic dominant. For low-frequency electromagnetic dominant, excitation current waveform optimization can be selected as a noise reduction method. A preset noise reduction strategy is defined to reduce the noise from 72 dB to less than or equal to 65 dB. The execution parameters of the noise reduction method are optimized according to the preset noise reduction target to obtain the third harmonic injection amplitude of the excitation current and the phase shift of the excitation current, and output the specific noise reduction strategy for each separated noise.

[0056] In one possible implementation, the noise separation model, constructed based on a deep learning model and trained according to several historical mixed noise signals, includes: An initial noise separation model is obtained by constructing a deep learning model; During a preset first historical time period, several single noise signals corresponding to each of the historical noise sources of the transformer are acquired. During a preset second historical time period, each corresponding marker signal is applied to each of the historical noise sources according to the signal modulation parameter set to obtain several historical mixed noise signals of the transformer. A training dataset is constructed based on each of the historical mixed noise signals and each of the single noise signals; The initial noise separation model is trained according to the preset loss function and the training dataset to obtain the noise separation model.

[0057] This application provides a training method for a noise separation model, which is fundamental to ensuring high separation accuracy in practical applications. The key step lies in constructing the training dataset: First, several single noise signals from the transformer are collected within a first historical time period, forming the "labels" required for supervised learning. Then, in a second historical time period, using a pre-set set of signal modulation parameters, corresponding labeled signals are applied to these known historical noise sources, artificially synthesizing "historical mixed noise signals" simulating actual operating conditions. This method of controllable mixing based on real signals generates a dataset that closely approximates the complexity of real-world noise while possessing accurate source signals as ground truth references. Finally, the initial model is trained using this dataset and a pre-set loss function, enabling the model to identify and separate individual noise signals from the mixed noise, thus obtaining the final noise separation model. This training paradigm allows the model to learn the inherent mapping relationship between each source signal from a mixed signal with known modulation labels, greatly improving the model's generalization ability and robustness under real and variable operating conditions. This provides a core guarantee for the reliability of the entire noise reduction method and improves the noise reduction effect on the transformer.

[0058] In a preferred embodiment, a deep neural network is used to perform deep learning on a signal modulation parameter set to construct a noise separation model for simultaneously processing the original noise signal and known label information. The noise separation model includes a feature encoding network, a feature joint network, an association attention network, and a separation decoding network. The feature encoding network is a dual-stream input encoder, containing a first feature encoding sub-network and a second feature encoding sub-network, which are responsible for automatically extracting deep time-frequency features or spatial features from the original mixed noise signal collected by multiple sensors, processing the signal modulation parameter set corresponding to the current mixed signal, and encoding it into a feature vector that the model can understand. The feature joint network is used to fuse the encoded features of the first feature encoding sub-network and the second feature encoding sub-network to generate a joint feature representation, which includes the current mixed sound and the pre-added label information. The association attention network calculates the attention weight distribution based on the input signal modulation parameter set, which is used to emphasize or focus on the feature parts of the mixed signal associated with a specific label, thereby guiding the model to focus on specific target noise sources at the feature level. The separation decoding network decodes and maps the joint features based on the weights provided by the association attention network, and finally outputs the estimated signals after multi-path separation, with each signal corresponding to an independent noise source.

[0059] Regarding the training dataset, a training sample set containing cause-effect correspondences is constructed, including real noise signal samples from known noise sources and input noise signal samples modulated by labeled signals. The real noise signal samples from known noise sources are clean source signals and serve as training targets. These may be single noise source signals excited and collected individually in a laboratory environment, various noise source signals generated through simulation models, or source signals approximately separated through precise measurement and signal processing under specific operating conditions of the equipment. The input noise signal samples modulated by labeled signals serve as training inputs. By mixing multiple clean source signals in a certain proportion to simulate real mixed noise, according to the specifications in the signal modulation parameter set, one or more source signals are selectively modulated to generate a modulated mixed noise signal and corresponding parameter labels used to modulate the corresponding source signals.

[0060] During training, a loss function is introduced to calculate the noise separation error of the noise separation model's output based on real noise signal samples from known noise sources. This allows the deep neural network to learn to reconstruct individual noise source signals from the mixed signal based on modulation labels. The loss function uses a function that measures the similarity of audio signals, such as scale-invariant signal-to-noise ratio or a time-frequency domain variant of mean square error. Specifically, in each training iteration, the modulated mixed signal and its corresponding modulation parameter label are read in. After forward propagation, multiple estimated source signals are output. The loss function calculates the sum of differences between each estimated signal output by the model and the corresponding clean real source signal in the training samples. Through backpropagation and the Adam gradient descent optimizer, multiple parameter values ​​in the neural network are automatically adjusted, causing the loss function value to continuously decrease. This process is iterated repeatedly on a large number of training samples until the separation error no longer decreases significantly and performs well on independent validation datasets. Training is then complete, resulting in a noise separation model that can accurately separate labeled noise source signals from the corresponding mixed noise signals based on modulation parameter labels.

[0061] Example 2: like Figure 2 As shown, Embodiment 2 provides a transformer noise reduction system based on deep learning, including an acquisition module 10, a modulation module 20, a noise separation module 30, an analysis module 40, and a noise reduction module 50. The acquisition module 10 is used to acquire the real-time noise signal of the transformer during a preset time period and input the real-time noise signal into a preset noise identification model to identify and determine several real-time noise sources currently existing in the transformer and the corresponding noise source identification information. The noise identification model is constructed based on several historical noise signals of the transformer. The modulation module 20 is used to apply corresponding marker signals to each of the real-time noise sources according to the noise source identification information and the preset signal modulation parameter set to obtain a mixed noise signal. The signal modulation parameter set is set according to the noise source identification information of all historical noise sources of the transformer. The noise separation module 30 is used to input the mixed noise signal into a preset noise separation model, so as to perform noise identification and noise separation on the mixed noise signal according to each marker signal in the mixed noise signal, and obtain the noise identification signal corresponding to each of the real-time noise sources. The noise separation model is constructed based on a deep learning model and trained according to several historical mixed noise signals. The analysis module 40 is used to perform feature quantization analysis on each of the noise identification signals, and determine the current dominant noise mode of the transformer based on the feature quantization analysis results and preset classification rules. The noise reduction module 50 is used to generate a corresponding noise reduction strategy based on the dominant noise mode and the preset noise reduction target, and to reduce the noise of the transformer according to the noise reduction strategy.

[0062] Furthermore, the step of constructing the noise identification model based on several historical noise signals of the transformer includes: Several historical noise signals of the transformer are acquired through several sensors; Based on the preset positions of each sensor, source component analysis is performed on each historical noise signal to obtain several noise source type samples and corresponding historical noise source identification information. The historical noise source identification information includes the noise generation location, noise propagation path, and noise dominant frequency distribution range. The noise recognition model is constructed based on the correspondence between each of the historical noise signals, each of the noise source type samples, and each of the historical noise source identification information.

[0063] Furthermore, the modulation module 20 applies corresponding marker signals to each of the real-time noise sources according to the noise source identification information and the preset signal modulation parameter set to obtain a mixed noise signal, including: Based on the noise source identification information and the signal modulation parameter set, determine the modulation mode type, modulation frequency, modulation amplitude and modulation input corresponding to each real-time noise source; A modulation signal corresponding to each of the real-time noise sources is generated based on each of the modulation scheme types, each of the modulation frequencies, and each of the modulation amplitudes. According to the power requirements of each modulation input, each modulation signal is amplified to generate a corresponding marker signal; According to each of the modulation inputs, a corresponding marker signal is applied to each of the real-time noise sources to obtain the mixed noise signal.

[0064] In one possible implementation, obtaining the signal modulation parameter set based on the noise source identification information of all historical noise sources of the transformer includes: The generation location and propagation path of each historical noise source are determined based on the noise source identification information, and then the modulation input corresponding to each historical noise source is determined based on the generation location and propagation path. Based on the identification information of each noise source, a candidate modulation parameter set corresponding to each historical noise source is determined. The candidate modulation parameter set includes several candidate modulation methods, several candidate modulation frequencies, and several candidate modulation amplitudes. For each historical noise source, several candidate parameter combinations are generated in a random combination manner based on the corresponding candidate modulation parameter set, and modulation separation evaluation is performed on each candidate parameter combination to obtain each corresponding evaluation result. Then, the optimal parameter combination is determined based on each evaluation result, and the signal modulation parameter combination of the noise source is generated based on the optimal parameter combination and the modulation input of the historical noise source. The signal modulation parameter set is constructed by combining the signal modulation parameters of each of the historical noise sources.

[0065] Furthermore, for any of the candidate parameter combinations, the modulation separation evaluation of each candidate parameter combination to obtain the corresponding evaluation results includes: Random candidate parameter combinations for each of the other historical noise sources are obtained by random sampling. A first marker signal is generated in the simulation environment based on the candidate parameter combination, and a first index of the candidate parameter combination is determined based on the signal-to-noise ratio of the marker signal. Based on each of the random candidate parameter combinations, generate corresponding second marker signals; The correlation or orthogonality between the first marker signal and each of the second marker signals is calculated respectively, and then the second index of the candidate parameter combination is determined based on the correlation or orthogonality. The first indicator and the second indicator are weighted and summed according to preset weight parameters to obtain the evaluation result of the candidate parameter combination.

[0066] In one possible implementation, the noise separation model includes a first feature encoding sub-network, a second feature encoding sub-network, a feature joint network, an association attention network, and a separation decoding network. The noise separation module 30 inputs the mixed noise signal into a preset noise separation model to perform noise identification and noise separation on the mixed noise signal according to each marker signal in the mixed noise signal, obtaining noise identification signals corresponding to each of the real-time noise sources, including: The mixed noise signal is input into the noise separation model; The first feature encoding subnetwork extracts mixed noise features related to acoustic properties from the mixed noise signal; Based on the signal modulation parameter set, the second feature coding subnetwork encodes several signal modulation parameters corresponding to the mixed noise signal into mixed noise signal modulation parameter features. The feature combination network is used to combine the mixed noise features and the mixed noise signal modulation parameter features to obtain joint feature samples. Based on the signal modulation parameter set, the noise source weight distribution of the joint feature samples is calculated through the associated attention network; The joint feature samples are decoded by the separation decoding network according to the noise source weight distribution to obtain each noise identification signal.

[0067] In one possible implementation, the analysis module 40 performs feature quantization analysis on each of the noise identification signals, and determines the current dominant noise mode of the transformer based on the feature quantization analysis results and preset classification rules, including: For each of the noise identification signals, the noise energy characteristics of the noise identification signal are determined by analyzing the sound power or vibration energy of the noise identification signal; By analyzing the main frequency amplitude and the amplitude of each harmonic of the noise identification signal, the main frequency and harmonic distribution characteristics of the noise identification signal are determined. The temporal stability characteristics of the noise identification signal are determined by calculating the variance of the energy or main frequency amplitude fluctuation of the noise identification signal over several preset time periods. By analyzing the peak-to-peak value and fluctuation rate of the instantaneous amplitude of the noise identification signal, the signal amplitude fluctuation characteristics of the noise identification signal are determined; Based on the noise energy characteristics, harmonic distribution characteristics, time stability characteristics, signal amplitude fluctuation characteristics, and preset classification rules of each noise identification signal, the current dominant noise mode of the transformer is determined.

[0068] Furthermore, the noise reduction module 50 generates a corresponding noise reduction strategy based on the dominant noise mode and a preset noise reduction target, and performs noise reduction on the transformer according to the noise reduction strategy, including: The corresponding noise reduction strategy is determined from a preset strategy matching library based on the dominant noise pattern. Based on the feature quantization analysis results of the target noise identification signal corresponding to the dominant noise mode and the noise reduction target, the specific execution parameters of the noise reduction strategy are determined. The transformer is subjected to noise reduction according to the specific execution parameters of the noise reduction strategy.

[0069] In one possible implementation, the noise separation model, constructed based on a deep learning model and trained according to several historical mixed noise signals, includes: An initial noise separation model is obtained by constructing a deep learning model; During a preset first historical time period, several single noise signals corresponding to each of the historical noise sources of the transformer are acquired. During a preset second historical time period, each corresponding marker signal is applied to each of the historical noise sources according to the signal modulation parameter set to obtain several historical mixed noise signals of the transformer. A training dataset is constructed based on each of the historical mixed noise signals and each of the single noise signals; The initial noise separation model is trained according to the preset loss function and the training dataset to obtain the noise separation model.

[0070] This application provides a transformer noise reduction system based on deep learning. By actively adding identification signals to each current noise source, it achieves accurate identification and separation of mixed noise signals, and then targets the dominant noise mode for targeted noise reduction, effectively improving the noise reduction effect of the transformer. Traditional noise reduction methods often treat the overall noise of the transformer as a single object of processing, lacking fine differentiation of different internal noise sources, resulting in insufficient targeting of noise reduction measures and poor noise reduction effect. In contrast, this embodiment first analyzes the real-time noise signal through a preset noise identification model, thereby accurately identifying each current noise source and its characteristics, laying the foundation for subsequent targeted processing. Then, using the identified noise source identification information and a preset signal modulation parameter set, it applies marking signals to each real-time noise source to generate a mixed noise signal, and uses a deep learning-based noise separation model to separate the mixed signal, realizing the physical decomposition of composite noise. This step overcomes the limitation of traditional methods in handling noise aliasing, enabling various noise sources to be analyzed and processed independently. Finally, by performing feature quantization analysis on the separated noise identification signals, the dominant noise mode is determined, and a noise reduction strategy is generated and executed accordingly, achieving precision and adaptability in the noise reduction process. Overall, this embodiment demonstrates the ability to dynamically adapt to noise changes under different operating conditions, significantly improving the accuracy and efficiency of transformer noise reduction, thereby effectively reducing environmental noise pollution from transformers.

[0071] For a more detailed explanation of the working principle and procedures of this embodiment, please refer to the relevant description in Embodiment 1.

[0072] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of this application. It should be understood that the above descriptions are merely specific embodiments of this application and are not intended to limit the scope of protection of this application. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application for those skilled in the art.

Claims

1. A transformer noise reduction method based on deep learning, characterized in that, include: The real-time noise signal of the transformer during a preset time period is acquired, and the real-time noise signal is input into a preset noise identification model to identify and determine several real-time noise sources currently existing in the transformer and the corresponding noise source identification information. The noise identification model is constructed based on several historical noise signals of the transformer. Each of the real-time noise sources is given a corresponding marker signal according to the noise source identification information and the preset signal modulation parameter set to obtain a mixed noise signal. The signal modulation parameter set is set according to the noise source identification information of all historical noise sources of the transformer. The mixed noise signal is input into a preset noise separation model to identify and separate noise based on each marker signal in the mixed noise signal, thereby obtaining the noise identification signal corresponding to each of the real-time noise sources. The noise separation model is constructed based on a deep learning model and trained based on several historical mixed noise signals. The noise identification signals are subjected to feature quantization analysis, and the current dominant noise mode of the transformer is determined based on the feature quantization analysis results and preset classification rules. Based on the dominant noise mode and the preset noise reduction target, a corresponding noise reduction strategy is generated, and the transformer is noise-reduced according to the noise reduction strategy.

2. The transformer noise reduction method based on deep learning as described in claim 1, characterized in that, The step of constructing the noise identification model based on several historical noise signals of the transformer includes: Several historical noise signals of the transformer are acquired through several sensors; Based on the preset positions of each sensor, source component analysis is performed on each historical noise signal to obtain several noise source type samples and corresponding historical noise source identification information; The noise recognition model is constructed based on the correspondence between each of the historical noise signals, each of the noise source type samples, and each of the historical noise source identification information.

3. The transformer noise reduction method based on deep learning as described in claim 1, characterized in that, The step of applying corresponding marker signals to each of the real-time noise sources according to the noise source identification information and a preset signal modulation parameter set to obtain a mixed noise signal includes: Based on the noise source identification information and the signal modulation parameter set, determine the modulation mode type, modulation frequency, modulation amplitude and modulation input corresponding to each real-time noise source; A modulation signal corresponding to each of the real-time noise sources is generated based on each of the modulation scheme types, each of the modulation frequencies, and each of the modulation amplitudes. According to the power requirements of each modulation input, each modulation signal is amplified to generate a corresponding marker signal; According to each of the modulation inputs, a corresponding marker signal is applied to each of the real-time noise sources to obtain the mixed noise signal.

4. The transformer noise reduction method based on deep learning as described in claim 3, characterized in that, The step of obtaining the signal modulation parameter set based on the noise source identification information of all historical noise sources of the transformer includes: The generation location and propagation path of each historical noise source are determined based on the noise source identification information, and then the modulation input corresponding to each historical noise source is determined based on the generation location and propagation path. Based on the identification information of each noise source, a candidate modulation parameter set corresponding to each historical noise source is determined. The candidate modulation parameter set includes several candidate modulation methods, several candidate modulation frequencies, and several candidate modulation amplitudes. For each historical noise source, several candidate parameter combinations are generated in a random combination manner based on the corresponding candidate modulation parameter set, and modulation separation evaluation is performed on each candidate parameter combination to obtain each corresponding evaluation result. Then, the optimal parameter combination is determined based on each evaluation result, and the signal modulation parameter combination of the noise source is generated based on the optimal parameter combination and the modulation input of the historical noise source. The signal modulation parameter set is constructed by combining the signal modulation parameters of each of the historical noise sources.

5. The transformer noise reduction method based on deep learning as described in claim 4, characterized in that, For any of the candidate parameter combinations, the modulation separation evaluation of each candidate parameter combination to obtain the corresponding evaluation results includes: Random candidate parameter combinations for each of the other historical noise sources are obtained by random sampling. A first marker signal is generated in the simulation environment based on the candidate parameter combination, and a first index of the candidate parameter combination is determined based on the signal-to-noise ratio of the marker signal. Based on each of the random candidate parameter combinations, generate corresponding second marker signals; The separability between the first marker signal and each of the second marker signals is calculated respectively, and then a second index of the candidate parameter combination is determined based on the separability. The first indicator and the second indicator are weighted and summed according to preset weight parameters to obtain the evaluation result of the candidate parameter combination.

6. The transformer noise reduction method based on deep learning as described in claim 1, characterized in that, The noise separation model includes a first feature encoding subnetwork, a second feature encoding subnetwork, a feature joint network, an association attention network, and a separation decoding network. The step of inputting the mixed noise signal into the preset noise separation model to perform noise identification and separation on the mixed noise signal based on each marker signal in the mixed noise signal, and obtaining noise identification signals corresponding to each of the real-time noise sources, includes: The mixed noise signal is input into the noise separation model; The first feature encoding subnetwork extracts mixed noise features related to acoustic properties from the mixed noise signal; Based on the signal modulation parameter set, the second feature coding subnetwork encodes several signal modulation parameters corresponding to the mixed noise signal into mixed noise signal modulation parameter features. The feature combination network is used to combine the mixed noise features and the mixed noise signal modulation parameter features to obtain joint feature samples. Based on the signal modulation parameter set, the noise source weight distribution of the joint feature samples is calculated through the associated attention network; The joint feature samples are decoded by the separation decoding network according to the noise source weight distribution to obtain each noise identification signal.

7. The transformer noise reduction method based on deep learning as described in claim 1, characterized in that, The step of performing feature quantization analysis on each of the noise identification signals, and determining the current dominant noise mode of the transformer based on the feature quantization analysis results and preset classification rules, includes: For each of the noise identification signals, the noise energy characteristics of the noise identification signal are determined by analyzing the sound power or vibration energy of the noise identification signal; By analyzing the main frequency amplitude and the amplitude of each harmonic of the noise identification signal, the main frequency and harmonic distribution characteristics of the noise identification signal are determined. The temporal stability characteristics of the noise identification signal are determined by calculating the variance of the energy or main frequency amplitude fluctuation of the noise identification signal over several preset time periods. By analyzing the peak-to-peak value and fluctuation rate of the instantaneous amplitude of the noise identification signal, the signal amplitude fluctuation characteristics of the noise identification signal are determined; Based on the noise energy characteristics, harmonic distribution characteristics, time stability characteristics, signal amplitude fluctuation characteristics, and preset classification rules of each noise identification signal, the current dominant noise mode of the transformer is determined.

8. The transformer noise reduction method based on deep learning as described in claim 1, characterized in that, The step of generating a corresponding noise reduction strategy based on the dominant noise mode and a preset noise reduction target, and then applying the noise reduction strategy to the transformer, includes: The corresponding noise reduction strategy is determined from a preset strategy matching library based on the dominant noise pattern. Based on the feature quantization analysis results of the target noise identification signal corresponding to the dominant noise mode and the noise reduction target, the specific execution parameters of the noise reduction strategy are determined. The transformer is subjected to noise reduction according to the specific execution parameters of the noise reduction strategy.

9. A transformer noise reduction method based on deep learning as described in any one of claims 1-8, characterized in that, The noise separation model, constructed based on a deep learning model and trained using several historical mixed noise signals, includes: An initial noise separation model is obtained by constructing a deep learning model; During a preset first historical time period, several single noise signals corresponding to each of the historical noise sources of the transformer are acquired. During a preset second historical time period, each corresponding marker signal is applied to each of the historical noise sources according to the signal modulation parameter set to obtain several historical mixed noise signals of the transformer. A training dataset is constructed based on each of the historical mixed noise signals and each of the single noise signals; The initial noise separation model is trained according to the preset loss function and the training dataset to obtain the noise separation model.

10. A transformer noise reduction system based on deep learning, characterized in that, It includes an acquisition module, a modulation module, a noise separation module, an analysis module, and a noise reduction module; The acquisition module is used to acquire the real-time noise signal of the transformer during a preset time period and input the real-time noise signal into a preset noise identification model to identify and determine several real-time noise sources currently existing in the transformer and the corresponding noise source identification information. The noise identification model is constructed based on several historical noise signals of the transformer. The modulation module is used to apply corresponding marker signals to each of the real-time noise sources according to the noise source identification information and the preset signal modulation parameter set to obtain a mixed noise signal. The signal modulation parameter set is set according to the noise source identification information of all historical noise sources of the transformer. The noise separation module is used to input the mixed noise signal into a preset noise separation model, so as to perform noise identification and noise separation on the mixed noise signal according to each marker signal in the mixed noise signal, and obtain the noise identification signal corresponding to each of the real-time noise sources. The noise separation model is constructed based on a deep learning model and trained based on several historical mixed noise signals. The analysis module is used to perform feature quantization analysis on each of the noise identification signals, and determine the current dominant noise mode of the transformer based on the feature quantization analysis results and preset classification rules. The noise reduction module is used to generate a corresponding noise reduction strategy based on the dominant noise mode and the preset noise reduction target, and to reduce the noise of the transformer according to the noise reduction strategy.

Images