Air conditioner and control method and device thereof, storage medium and computer program product

Abstract

The invention provides an air conditioner, a control method and device of the air conditioner, a storage medium and a computer program product. The method comprises the steps that original noise signals generated when the air conditioner operates are collected; the collected original noise signals are input into a pre-trained noise recognition model to generate noise reduction signals, the noise recognition model is a Swinin-Conv-UNet mixed model, and the noise reduction signals comprise noise source and / or type information and corresponding noise intensity information; according to the generated noise reduction signal, the current noise source and / or type and the corresponding noise intensity of the air conditioner are / is determined; according to the determined current noise source and / or type of the air conditioner and the corresponding noise intensity, a noise reduction control strategy of the air conditioner is obtained; and a corresponding control instruction is generated according to the obtained noise reduction control strategy, and the air conditioner is controlled to execute. According to the scheme, the air conditioner noise can be remarkably reduced.

Description

Technical Field

[0001] The present invention relates to the field of control, and particularly to an air conditioner and its control method, device, storage medium, and computer program product. Background Art

[0002] With the rapid development of technology and the continuous improvement of people's living standards, air conditioning equipment has been widely used in various places such as homes, offices, shopping malls, etc., providing a comfortable environment for people's lives and work. However, the noise problem generated during the operation of air conditioning equipment also follows. Especially in places such as bedrooms where a quiet environment is required, the air conditioner noise has a serious impact on people's sleep quality and physical and mental health. Therefore, how to effectively control air conditioner noise and improve people's sleep quality and living quality has become a hot and difficult issue in current research. The noise monitoring methods in related technologies often have problems such as poor real-time performance and low accuracy, and it is difficult to meet the high requirements for noise control in modern homes. Summary of the Invention

[0003] The main purpose of the present invention is to overcome the defects of the above-mentioned related technologies, and provide an air conditioner and its control method, device, storage medium, and computer program product to solve the problems of poor real-time performance and low accuracy in the noise monitoring methods in related technologies.

[0004] On the one hand, the present invention provides a control method for an air conditioner, including: collecting the original noise signal when the air conditioner is running; inputting the collected original noise signal into a pre-trained noise recognition model to generate a noise reduction signal, the noise recognition model being a Swin-Conv-UNet hybrid model, and the noise reduction signal containing noise source and / or type information and corresponding noise intensity information; determining the current noise source and / or type of the air conditioner and the corresponding noise intensity according to the generated noise reduction signal; obtaining a noise reduction control strategy for the air conditioner according to the determined current noise source and / or type of the air conditioner and the corresponding noise intensity; generating a corresponding control instruction according to the obtained noise reduction control strategy, and controlling the air conditioner to execute.

[0005] Optionally, the noise recognition model includes: an input layer, a Swin Transformer layer, a convolutional neural network, a U-Net symmetric structure, and an output layer; inputting the collected original noise signal into a pre-trained noise recognition model to generate a noise-reduced signal, including: inputting the collected original noise signal into the noise recognition model through the input layer; capturing the global dependencies in the input original noise signal through the Swin Transformer layer to generate a feature map; performing local feature extraction on the generated feature map through the convolutional neural network; segmenting and reconstructing the extracted local features through the U-Net symmetric structure to generate a spectrogram or a Mel spectrogram containing noise source and / or type information and corresponding noise intensity information; outputting the spectrogram or Mel spectrogram through one or more output layers, and converting the output spectrogram or Mel spectrogram into a time-domain signal to obtain a noise-reduced signal containing noise source and / or type information and corresponding noise intensity information.

[0006] Optionally, before inputting the collected original noise signal into a pre-trained noise recognition model to generate a noise-reduced signal, it further includes: preprocessing the collected original noise signal to obtain a processed noise signal; and / or generating a spectrogram or a Mel spectrogram according to the collected original noise signal.

[0007] Optionally, preprocessing the collected original noise signal includes: signal conversion, noise reduction processing, and / or filtering processing; the signal conversion includes: converting the original noise signal from an analog signal to a digital signal; and / or the filtering processing includes: low-pass filtering and / or band-pass filtering.

[0008] Optionally, the noise recognition model further includes: a fully connected layer for identifying the source and / or type of noise and the corresponding noise intensity according to the features extracted by the convolutional neural network.

[0009] Optionally, the noise recognition model is trained in the following manner: obtaining original noise samples obtained by pre-classifying and labeling the noise signals generated when different air conditioners operate under different environments and different operating states; using a pre-trained noise synthesis model to synthesize the original noise samples into new noise samples, and integrating the original noise samples and the new noise samples into a training dataset; constructing a Swin-Conv-Unet hybrid model structure, and using the training dataset for model training; evaluating the model and tuning the hyperparameters of the model to obtain a final noise recognition model.

[0010] Optionally, using a pre-trained noise synthesis model to synthesize a new noise sample from the original noise sample, including: inputting the original noise sample into the pre-trained noise synthesis model to synthesize a new noise sample; performing data augmentation on the synthesized new noise sample; the data augmentation includes at least one of adding background noise, changing the volume, time stretching, and frequency offset.

[0011] Optionally, determining the current noise source and / or type of the air conditioner according to the generated noise reduction signal, including: synthesizing the generated noise reduction signal and the original noise signal to obtain a synthesized signal; determining the current noise source and / or type of the air conditioner according to the obtained synthesized signal.

[0012] Optionally, receiving a parameter adjustment instruction issued through a user interface; adjusting the noise generated during the operation of the air conditioner according to the received parameter adjustment instruction.

[0013] On the other hand, the present invention provides a control device for an air conditioner, including: a collection unit for collecting an original noise signal during the operation of the air conditioner; an identification unit for inputting the original noise signal collected by the collection unit into a pre-trained noise identification model to generate a noise reduction signal, the noise identification model being a Swin-Conv-UNet hybrid model, and the noise reduction signal includes noise source and / or type information and corresponding noise intensity information; a determination unit for determining the current noise source and / or type of the air conditioner and the corresponding noise intensity according to the noise reduction signal generated by the identification unit; an acquisition unit for acquiring a noise reduction control strategy of the air conditioner according to the current noise source and / or type of the air conditioner and the corresponding noise intensity determined by the determination unit; a control unit for generating a corresponding control instruction according to the noise reduction control strategy acquired by the acquisition unit and controlling the air conditioner to execute.

[0014] Optionally, the noise recognition model includes: an input layer, a Swin Transformer layer, a convolutional neural network, a U-Net symmetric structure, and an output layer; the recognition unit inputs the collected original noise signal into a pre-trained noise recognition model to generate a noise reduction signal, including: inputting the collected original noise signal into the noise recognition model through the input layer; capturing the global dependencies in the input original noise signal through the Swin Transformer layer to generate a feature map; performing local feature extraction on the generated feature map through the convolutional neural network; performing segmentation and reconstruction on the extracted local features through the U-Net symmetric structure to generate a spectrogram or a Mel spectrogram containing noise source and / or type information and the corresponding noise intensity information; outputting the spectrogram or the Mel spectrogram through one or more output layers, and converting the output spectrogram or Mel spectrogram into a time-domain signal to obtain a noise reduction signal containing noise source and / or type information and the corresponding noise intensity information.

[0015] Optionally, it further includes: a processing unit, configured to preprocess the collected original noise signal to obtain a processed noise signal before the recognition unit inputs the collected original noise signal into a pre-trained noise recognition model to generate a noise reduction signal; and / or, generate a spectrogram or a Mel spectrogram according to the collected original noise signal.

[0016] Optionally, the processing unit preprocesses the collected original noise signal, including: signal conversion, noise reduction processing, and / or filtering processing; the signal conversion includes: converting the original noise signal from an analog signal to a digital signal; and / or, the filtering processing includes: low-pass filtering and / or band-pass filtering.

[0017] Optionally, the noise recognition model further includes: a fully connected layer, configured to identify the source and / or type of noise and the corresponding noise intensity according to the features extracted by the convolutional neural network.

[0018] Optionally, the noise recognition model is trained in the following manner: obtaining original noise samples obtained by pre-classifying and labeling the noise signals generated by different air conditioners operating in different environments and different operating states; using a pre-trained noise synthesis model to synthesize the original noise samples into new noise samples, and integrating the original noise samples and the new noise samples into a training dataset; constructing a Swin-Conv-Unet hybrid model structure, and training the model using the training dataset; evaluating the model, and tuning the hyperparameters of the model to obtain a final noise recognition model.

[0019] Optionally, using a pre-trained noise synthesis model to synthesize a new noise sample from the original noise sample, including: inputting the original noise sample into the pre-trained noise synthesis model to synthesize a new noise sample; performing data augmentation on the synthesized new noise sample; the data augmentation includes at least one of adding background noise, changing the volume, time stretching, and frequency shift.

[0020] Optionally, the determining unit determines the current noise source and / or type of the air conditioner according to the noise reduction signal generated by the identifying unit, including: performing signal synthesis on the generated noise reduction signal and the original noise signal to obtain a synthesized signal; determining the current noise source and / or type of the air conditioner according to the obtained synthesized signal.

[0021] Optionally, it further includes: a receiving unit for receiving a parameter adjustment instruction issued through a user interface; an adjusting unit for adjusting the noise generated during the operation of the air conditioner according to the parameter adjustment instruction received by the receiving unit.

[0022] Another aspect of the present invention provides a storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of any of the foregoing methods are implemented.

[0023] Another aspect of the present invention provides an air conditioner, including a processor, a memory, and a computer program stored on the memory and operable on the processor, and when the processor executes the program, the steps of any of the foregoing methods are implemented.

[0024] Another aspect of the present invention provides an air conditioner, including any of the foregoing control devices.

[0025] Another aspect of the present invention provides a computer program product, including a computer program, and when the computer program is executed by a processor, the steps of any of the foregoing methods are implemented.

[0026] According to the technical solution of the present invention, it is not only possible to monitor the noise during the operation of the air conditioner in real time, but also to automatically adjust the operation parameters of the air conditioner according to the monitoring data, effectively reducing the noise, thereby significantly improving the user's sleep quality and living comfort.

[0027] According to the technical solution of the present invention, the Swin-Conv-UNet hybrid model is applied to the air conditioner noise control to realize the real-time monitoring of the air conditioner noise. By real-time monitoring and analyzing the noise data generated during the operation of the air conditioner, it is not only possible to quickly identify and locate the noise source and / or type based on the noise data, but also to dynamically adjust the operation parameters of the air conditioner according to the actual situation, thereby achieving the effect of significantly reducing the noise. Description of the Drawings

[0028] The accompanying drawings described herein are used to provide a further understanding of the present invention and form a part of the present invention. The illustrative embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0029] Figure 1 is a schematic diagram of a method of an embodiment of the control method of an air conditioner provided by the present invention;

[0030] Figure 2 is a flowchart of the step of inputting the collected original noise signal into a pre-trained noise recognition model to generate a noise reduction signal according to the present invention;

[0031] Figure 3 is a flowchart of the training step of the noise recognition model according to the present invention;

[0032] Figure 4 is a flowchart of the training step of the noise synthesis model according to the present invention;

[0033] Figure 5 is a structural diagram of air conditioner noise reduction control of a specific embodiment of the control method of an air conditioner provided by the present invention;

[0034] Figure 6 is a flowchart of air conditioner noise reduction control of a specific embodiment of the control method of an air conditioner provided by the present invention;

[0035] Figure 7 is a block diagram of an embodiment of the control device of an air conditioner provided by the present invention. Detailed Embodiment

[0036] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0037] It should be noted that the terms "first", "second", etc. in the description, claims and above-mentioned drawings of the present invention are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present invention described here can be implemented in an order other than those illustrated or described here. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0038] The air conditioner noise control methods in the related art mainly include physical noise reduction, installation of sound insulation materials, adjustment of wind speed, etc. These methods can reduce the air conditioner noise to a certain extent, but the effect is limited and there are certain limitations. For example, the physical noise reduction method requires modification of the air conditioner equipment, with high costs and unstable effects; although installing sound insulation materials can reduce noise transmission, it will affect the heat dissipation effect of the air conditioner equipment; although adjusting the wind speed can reduce noise, it will affect the cooling or heating effect of the air conditioner equipment.

[0039] The air conditioner noise control methods in the related art often have problems such as poor real-time performance and low accuracy, and it is difficult to meet the high requirements for noise control in modern homes. Most of the air conditioner noise control methods in the related art rely on fixed preset parameters and rules, lacking intelligence and self-adaptability. The air conditioner noise control methods in the related art are difficult to meet the personalized needs of different users, affecting the user experience.

[0040] The present invention provides a control method for an air conditioner.

[0041] Figure 1 It is a schematic diagram of the method of an embodiment of the control method for an air conditioner provided by the present invention.

[0042] As Figure 1 shown, according to an embodiment of the present invention, the air conditioner control method at least includes step S110, step S120, step S130, step S140 and step S150.

[0043] Step S110, collect the original noise signal when the air conditioner is running.

[0044] In a specific embodiment, the original noise signal can be collected through a microphone array. For example, select a microphone array with high sensitivity and wideband response to ensure the quality of the collected audio data. The number and arrangement of microphones can be set according to the actual application scenario to cover as large a space range as possible and capture the original audio signal.

[0045] Step S120: Input the collected original noise signal into a pre-trained noise recognition model to generate a noise-reduced signal.

[0046] The noise-reduced signal contains noise source and / or type information and corresponding noise intensity information. Preferably, before inputting the collected original noise signal into a pre-trained noise recognition model to generate a noise-reduced signal, preprocess the collected original noise signal to obtain a processed noise signal. In a specific embodiment, the preprocessing may specifically include: signal conversion, noise reduction processing, and / or filtering processing.

[0047] After sampling and quantization processing, the noise signal is converted from an analog signal to a digital signal. To improve the quality of the spectrogram, perform noise reduction and filtering processing on the sound signal to ensure signal quality. Sampling is to discretize the sound signal at a certain time interval. Quantization is to discretize the signal amplitude obtained by sampling, usually converted into an integer form. Noise reduction is to remove background noise or other interference signals. Filtering is to eliminate unwanted frequency components.

[0048] The signal conversion may specifically include: converting the original noise signal from an analog signal to a digital signal. The noise signal usually exists in the form of an analog signal, so it needs to go through a series of processing steps to be converted into a digital signal to make it suitable for subsequent digital processing and analysis. Specifically, convert the continuous analog signal into a discrete digital signal, and the conversion can be achieved through an analog-to-digital converter (ADC).

[0049] During the acquisition process, the noise signal is often interfered by various noises, such as electromagnetic interference, environmental noise, etc. A filter can be used to remove unwanted frequency components, or signal enhancement techniques can be used to improve the signal-to-noise ratio (SNR). The filtering processing includes: low-pass filtering and / or band-pass filtering. For example, use a low-pass filter to remove high-frequency noise and a band-pass filter to retain useful signals within a specific frequency band.

[0050] Preferably, before inputting the collected original noise signal into a pre-trained noise recognition model to generate a noise-reduced signal, generate a spectrogram or a Mel spectrogram according to the collected original noise signal.

[0051] In addition to signal conversion and noise processing, the processed digital signal also needs to be formatted into a format that can be processed by the noise recognition model. For example, convert it into a spectrogram or a Mel spectrogram. These visualized representations help the model better understand the characteristics of the sound data. After preprocessing the collected original noise signal to obtain a processed noise signal, generate a spectrogram or a Mel spectrogram according to the obtained processed noise signal.

[0052] The generation of the spectrogram depends on the Short-Time Fourier Transform (STFT). The sound signal is segmented into multiple short-time windows, and the Fourier transform is performed on each window to obtain the frequency information at each time point. Specifically, a suitable window function (such as the Hann window, Hamming window, etc.) is selected, and the window size and overlap ratio are determined. The window size determines the balance between time resolution and frequency resolution, while the overlap ratio affects the smoothness of the spectrogram. The Fast Fourier Transform (FFT) is performed on the signal within each window to obtain the spectral information of that window. Through the sliding window technique, a continuous spectrogram is generated, reflecting the frequency distribution of the sound signal over time. The spectrogram is a two-dimensional image, with the horizontal axis representing time and the vertical axis representing frequency, and the color indicating the energy intensity. By adjusting the color mapping, the change in the frequency of the sound signal over time can be more intuitively displayed.

[0053] The Mel spectrogram is a method of representing sound based on human auditory perception. It converts the frequency axis from linear frequency to the Mel frequency scale, which is closer to the perceptual characteristics of the human auditory system for sound frequencies. The Mel frequency scale expands the frequency intervals in the low-frequency part and compresses the frequency intervals in the high-frequency part to better reflect the non-linear characteristics of human hearing. To generate the Mel spectrogram, first perform the Short-Time Fourier Transform (STFT) on the sound signal to obtain the spectrogram; based on the spectrogram, apply a Mel filter bank (i.e., a set of band-pass filters) to evenly distribute the spectrogram response on the Mel frequency scale; perform a logarithmic transformation on the Mel spectrogram to enhance the perceptibility of the low-energy part and at the same time compress the dynamic range to make it more suitable for input to machine learning models.

[0054] Preferably, to ensure that the data has the same scale when input into the model and avoid affecting the model performance due to signal intensity differences, the spectrogram or Mel spectrogram can also be normalized. Normalization methods can include, for example: (1) Min-Max normalization - scaling the minimum and maximum values of the data to the range [0,1]. (2) Z-score standardization - converting the data to a distribution with a mean of 0 and a standard deviation of 1.

[0055] Preferably, the normalized spectrogram or Mel spectrogram is converted to a preset format and then input into the noise recognition model, that is, converted to a format suitable for input to a machine learning model. For example, the processed spectrogram or Mel spectrogram is converted to a format suitable for input to the Swin-Conv-UNet model. In a specific implementation, converting the normalized spectrogram or Mel spectrogram to a preset format can specifically include:

[0056] (1) Data dimension adjustment - Adjust the dimension of the data according to the input requirements of the model. For example, some models may require a three-dimensional tensor (such as a deep learning model), and the two-dimensional spectrogram needs to be extended to three-dimensional data.

[0057] (2) Channel number expansion - If the model requires multi-channel input, the demand can be met by expanding the channel dimension (e.g., expanding from a single channel to three channels).

[0058] (2) Data format conversion - Convert the processed data into a format suitable for model input. For example, it can be at least one of a NumPy array and a TensorFlow tensor.

[0059] In a specific embodiment, the noise recognition model is a Swin-Conv-UNet hybrid model. Specifically, the noise recognition model may include: an input layer, a Swin Transformer layer, a convolutional neural network (CNN), a U-Net symmetric structure, and an output layer.

[0060] Figure 2 It is a flowchart of the step of inputting the collected original noise signal into a pre-trained noise recognition model to generate a noise reduction signal according to the present invention. As Figure 2 shown, the step of inputting the collected original noise signal into a pre-trained noise recognition model to generate a noise reduction signal may specifically include the following steps S121 to S125.

[0061] Step S121, input the collected original noise signal into the noise recognition model through the input layer.

[0062] Specifically, input the processed spectrogram or mel spectrogram into the input layer. These signals can be data in the form of images, audio, or others, which contain unwanted noise components.

[0063] Step S122, capture the global dependencies in the input original noise signal through the Swin Transformer layer to generate a feature map.

[0064] Specifically, in the Swin-Conv-UNet model, the Swin Transformer layer can effectively capture the global dependencies in the input data and process complex features through its unique hierarchical and moving window self-attention mechanism. Specifically for air conditioner noise data, its global dependencies mainly include the following aspects:

[0065] 1. Global dependency in time

[0066] There is an association between the sound signals at different time points. For example, the noise level at a certain moment may be affected by the operating state of the device at the previous moment. This temporal continuity needs to be captured by the model to accurately predict future noise changes.

[0067] 2. Global dependency in frequency

[0068] There are also interactions between different frequency components of the noise signal. For example, there may be an inherent connection between low-frequency noise and high-frequency noise, and this frequency-dependent relationship is crucial for a comprehensive understanding of the noise characteristics.

[0069] 3. Global Dependence in Space

[0070] In the noise data of multiple devices or multiple locations, there may be interactions between the sound signals of different devices or locations. For example, when multiple air-conditioning devices are operating simultaneously, their noises may be superimposed or interfered with each other, forming a complex noise environment.

[0071] 4. Dependence between Operating State and Noise

[0072] There is a close relationship between the operating state of the air-conditioning device (such as load, temperature setting, operating time, etc.) and the noise level it generates. The model needs to capture the dependence between these device state parameters and the noise data in order to perform noise prediction and analysis more accurately.

[0073] The Swin Transformer layer can learn these global dependence relationships at different levels through its hierarchical structure and moving window mechanism. The hierarchical structure allows the model to capture features at different scales, while the moving window mechanism enables the model to effectively handle long-range dependence relationships while maintaining computational efficiency. Combined with the convolutional layer, the Swin-Conv-UNet model can fully utilize the advantages of both to achieve efficient capture and processing of complex features in the air-conditioning noise data.

[0074] Step S123, perform local feature extraction on the generated feature map through a convolutional neural network.

[0075] Specifically, the Swin Transformer layer captures the global dependence relationships in the sound signal through its hierarchical structure and moving window self-attention mechanism, generating a feature map. The feature map processed by the Swin Transformer is passed to a convolutional neural network (CNN), and the convolutional layer is used to extract local features, that is, to capture multi-scale features in the sound signal through multi-scale convolutions (convolution kernels of different sizes). A residual connection can be introduced between the Swin Transformer layer and the convolutional layer to enhance the feature propagation ability. For example, the output of the Swin Transformer layer is added to the output of the convolutional layer to form a new feature map. Through the use of multiple neural network layers and activation functions in the convolutional neural network, the model can learn more detailed information about the noise and the signal.

[0076] Preferably, the noise recognition model further includes: a fully connected layer, configured to identify the source and / or type of the noise and the corresponding noise intensity according to the features extracted by the convolutional neural network. Specifically, after feature extraction, one or more fully connected layers are used as classifiers to locate and identify the source and / or type of the noise and the corresponding noise intensity according to the extracted features.

[0077] Step S124, segment and reconstruct the extracted local features through a U-Net symmetric structure to generate a spectrogram or mel spectrogram containing information about the source and / or type of the noise.

[0078] The U-Net symmetric structure can learn the difference between the background noise and the useful signal in the air conditioner noise, so as to generate a denoised spectrogram or mel spectrogram. Specifically, the U-Net symmetric structure is used to reconstruct a clear signal from the features obtained from the convolutional neural network (CNN) layer. U-Net is a network with an encoder-decoder structure. The encoder part gradually reduces the spatial dimension of the feature map, and the decoder part gradually restores the dimension. At the same time, the combination of low-level and high-level features is maintained through skip connections.

[0079] The decoder part is responsible for reconstructing the feature map extracted by the encoder into an optimized spectrogram or mel spectrogram, combining the decoder architecture of U-Net and skip connections: upsample through a transposed convolutional layer or an interpolation method to gradually restore the spatial resolution of the feature map; and introduce skip connections to combine the low-level features of the encoder part with the current feature map to enhance the reconstruction ability. Use convolutional layers to finely adjust the feature map to further optimize the reconstruction result; capture the detailed features in the sound signal through multi-scale convolutions. Introduce residual connections between the convolutional layers in the decoder part to enhance the propagation ability of the features. For example, add the output of the current convolutional layer to the output of the previous layer to form a new feature map.

[0080] To improve the training effect and feature propagation ability of the model, residual connections and skip connections are introduced in the encoder and decoder parts: Residual connections are introduced between the convolutional layers of the encoder and decoder to form residual blocks; the output of the residual block is the sum of the output of the current layer and the input feature map, which helps to alleviate the problem of gradient disappearance. Skip connections are introduced between the encoder and decoder to combine the low-level features of the encoder part with the high-level features of the decoder part; skip connections help to retain the detailed information in the sound signal and improve the accuracy of reconstruction. The skip connection transfers the high-resolution features in the encoder stage to the decoder stage, ensuring that more detailed information is retained during the reconstruction process, improving the reconstruction quality and robustness of the model, and also promoting the fusion of features at different levels, enabling the model to better capture multi-scale features and improve the generalization ability of the model. After the feature output of the encoder, it can be connected to a fully connected layer to identify the source and / or type of noise and the corresponding noise intensity, that is, the fully connected layer locates and identifies the source and / or type of noise and the corresponding noise intensity according to the features output by the encoder.

[0081] Step S125, output the spectrogram or mel spectrogram through more than one output layer, and convert the output spectrogram or mel spectrogram into a time-domain signal to obtain a noise reduction signal containing information about the source and / or type of noise and the corresponding noise intensity.

[0082] Specifically, the result output by the decoder of the U-Net symmetric structure is processed by one or more output layers to generate a final noise reduction signal. The output layer adjusts the number of channels of the feature map output by the decoder through a convolution operation to make it consistent with the number of channels of the input spectrogram. Preferably, the output layer uses a linear activation function to ensure that the amplitude information of the spectrogram is retained and avoid distorting the signal amplitude by non-linear transformation.

[0083] If the output layer outputs a spectrogram or mel spectrogram, it is converted into a time-domain signal. For example, it is converted into a time-domain signal through the inverse short-time Fourier transform iSTFT. The implementation of iSTFT needs to consider parameters such as the window function and frame shift to ensure the reconstruction quality of the time-domain signal. Optionally, the time-domain signal generated by the inverse short-time Fourier transform iSTFT is normalized to ensure that the signal amplitude is within an appropriate range and avoid sound quality distortion.

[0084] The Swin-Conv-UNet hybrid model combines the global perception ability of the Swin Transformer, the local feature extraction ability of the CNN, and the effective reconstruction mechanism of the U-Net to achieve efficient and accurate noise feature extraction, noise source identification, and noise reduction signal generation. This makes it perform excellently in processing various noisy signals, whether it is visual images or audio data.

[0085] Figure 3It is a flowchart of the training steps of the noise recognition model according to the present invention. As Figure 3 shown, in a specific embodiment, the noise recognition model is trained through the following steps S1 to S4:

[0086] Step S1, obtain the original noise samples obtained by pre-labeling the noise signals generated when different air conditioners operate under different environments and different operating states.

[0087] Specifically, pre-collect the noise signals generated when different air conditioners operate under different environments and different operating states; and label the collected noise signals generated when different air conditioners operate under different environments and different operating states as the original noise samples.

[0088] Collect the noise data when the air conditioner equipment operates under different environmental conditions and label the data. The collection of noise data can use a microphone array with high sensitivity and wide-band response to ensure the quality of the collected audio data. The number and arrangement of microphones can be optimized according to the actual application scenario to cover as large a spatial range as possible. Use professional recording equipment to ensure that the collected audio data is distortion-free and free of noise interference.

[0089] The different environments can specifically include different indoor environment types, outdoor environment types, temperature, and / or humidity. The indoor environment type can specifically be the room type, such as a bedroom, a living room, an office; the outdoor environment type can include, for example: a courtyard and / or a street. Collecting data under different humidity and / or temperature conditions can ensure the applicability of the model under various environmental conditions.

[0090] The method of collecting the noise signals generated when different air conditioners operate under different environments and different operating states can specifically include: performing continuous collection for a preset duration under different environments and different operating states. That is, perform long-term continuous collection in each environment and record the sounds of the air conditioner under different operating states. Collect noise data under different operating modes (such as cooling, heating, dehumidifying) to ensure that the model can process the noise under various working states. Different air conditioners can specifically include air conditioners of different brands and / or models. Collecting the noise data when different brands and models of air conditioner equipment operate can improve the versatility of the model.

[0091] The annotation information may specifically include: the noise source and / or type and the corresponding noise intensity. The noise source may specifically be the equipment, component, and / or location where the noise is generated. For example, it may include at least one of compressor noise, evaporator noise, and fan noise; the noise type may specifically be a description of the noise characteristics. For example, it may include at least one of mechanical noise (the vibration sound of mechanical components during air conditioner operation), wind noise (the airflow sound generated when the air conditioner blows air), electronic noise (the noise generated when internal electronic components of the air conditioner work), and external noise (background noise in the environment such as traffic noise, human voices, etc.).

[0092] When annotating the noise signals generated during the operation of different air conditioners collected under different environments and different operating states, the noise source and / or type can be classified first, and then for different noise sources and / or types, professional equipment is used to measure the noise decibel values under different noise sources and / or types to ensure the accuracy of the annotation. In addition, the intensity of the noise can be evaluated through human subjective auditory tests to ensure the comprehensiveness of the annotation. Manual annotation can be performed through professional audio annotation software, or combined with automated tools and manual review to improve the annotation efficiency and accuracy.

[0093] Step S2: Use the pre-trained noise synthesis model to synthesize new noise samples from the original noise samples, and integrate the original noise samples and the new noise samples into a training data set.

[0094] The noise synthesis model is a generative model used to generate new noise data based on existing noise samples. Select a suitable noise synthesis model according to the task requirements and data characteristics. For example, GANs are suitable for generating high-quality noise samples, and VAEs are suitable for learning the distribution of the latent space. Input the original noise samples into the pre-trained noise synthesis model to synthesize new noise samples.

[0095] Preferably, after inputting the original noise samples into the pre-trained noise synthesis model to synthesize new noise samples, data augmentation can also be performed on the synthesized new noise samples. Specifically, after inputting the original noise samples into the pre-trained noise synthesis model to generate new noise samples, it is then determined whether additional data augmentation is required based on the generation ability of the model and the diversity of the generated samples. If the generated samples already have sufficient diversity, the additional data augmentation step can be omitted; if the generated samples are relatively single, data augmentation can be considered after generating new noise samples to enrich the data set and improve the generalization ability of the model. The data augmentation may specifically include at least one of adding background noise, changing the volume, time stretching, and frequency offset.

[0096] Add background noise, that is, add different types of background noise to the generated noise samples to simulate the noise situation in a complex environment. Change the volume, that is, adjust the volume of the generated noise to simulate the sound intensity at different distances and in different environments. Time stretching, that is, change the time length of the noise samples to simulate the sound changes under different device operating states. Frequency shift, that is, adjust the frequency of the noise samples to simulate the frequency characteristics under different devices or environments. Use audio processing libraries (such as Librosa, SoundFile, etc.) to implement the above data augmentation techniques to generate diverse training data.

[0097] Figure 4 is a flowchart of the training steps of the noise synthesis model according to the present invention. As Figure 4 shown, in a specific embodiment, the noise synthesis model can be specifically trained through the following steps S21 to step S23:

[0098] Step S21, obtain pre-annotated noise samples and preprocess the noise samples.

[0099] Organize the annotated noise samples into a format suitable for model input, for example, spectrogram or mel spectrogram. Preprocess the noise samples, including normalization, standardization, etc., to ensure that the data has the same scale when input into the model.

[0100] Step S22, design the network architecture of the noise synthesis model according to the pre-determined model type.

[0101] According to the selected model type, design a suitable network architecture, including a generator and a discriminator (such as GANs) or an encoder and a decoder (such as VAEs).

[0102] Step S23, use the noise samples to train the noise synthesis model and optimize the model parameters.

[0103] Use the annotated noise samples to train the noise synthesis model and optimize the model parameters so that the generator can generate realistic noise samples.

[0104] Generating diverse training data through the noise synthesis model can further improve the generalization ability and anti-interference ability of the model, ensuring the performance of the model in practical applications. Increase the diversity and complexity of the data, and additionally perform data augmentation by adding background noise, changing the volume, time stretching, and frequency shift, etc., to further improve the robustness of the model. By generating diverse training data, the model can exhibit stable performance under different environmental conditions. Diverse data can cover more application scenarios and improve the generalization ability of the model. In addition, the model is exposed to various types of noise during the training process, enabling it to better handle complex noise environments in practical applications and improve the anti-interference ability.

[0105] Integrate the original noise samples and the new noise samples into a training dataset, and use the integrated training dataset to train the Swin-Conv-UNet hybrid model, which can improve the noise reduction and recognition capabilities of the model in different noise environments.

[0106] Step S3: Construct the Swin-Conv-Unet hybrid model structure and use the training dataset to train the model.

[0107] The Swin-Conv-Unet hybrid model structure may specifically include: an input layer, a Swin Transformer layer, a convolutional neural network (CNN), a U-Net symmetric structure, and an output layer.

[0108] The Swin Transformer layer can effectively capture local features by dividing the input spectrogram into multiple non-overlapping windows and performing self-attention mechanisms within each window. This mechanism reduces the computational amount and memory consumption while maintaining the learning ability for long-range dependencies. The Swin Transformer can capture features at different scales, from local details to global structures, providing rich feature representations for the model. The convolutional neural network CNN captures local features in the spectrogram, such as edges and textures. Through multi-layer convolution and pooling operations, the CNN can gradually extract higher-level features, and the convolution operations of the CNN have efficient computational performance and are suitable for processing large-scale data. The fully connected layer identifies the source and / or type of the noise and the corresponding noise intensity based on the features extracted by the convolutional neural network; the UNet gradually reduces the spatial resolution of the spectrogram through the encoder, increases the number of channels of the feature map, and captures deep-level features; while the decoder gradually restores the spatial resolution of the spectrogram and reduces the number of channels of the feature map to achieve high-quality reconstruction.

[0109] When training the model, first select appropriate initial model parameters, such as including weight initialization and / or activation functions; then use the Adam optimizer to adjust the model parameters through the backpropagation algorithm; then, through the learning rate decay strategy, gradually reduce the learning rate to prevent oscillations during the training process and improve the convergence speed of the model. Finally, select an appropriate batch size to balance computational resources and training efficiency and ensure that the model is trained within a reasonable time.

[0110] Optimize the model using L1 or L2 regularization techniques before model training. By adding a regularization term to the loss function, prevent the model parameters from being too large, reduce the complexity of the model, and prevent overfitting. During the training process, monitor the performance of the validation set and stop training early when the performance of the validation set no longer improves to avoid overfitting. After training, perform data augmentation by adding background noise, changing the volume, time stretching, and frequency shifting, etc. on the basis of the previous training, and then conduct the next model training. At this time, the diversity and complexity of the training data will be increased, and the generalization ability of the model will be improved.

[0111] Step S4, evaluate the model and tune the hyperparameters of the model to obtain the final noise recognition model.

[0112] The evaluation metrics should comprehensively consider both objective and subjective aspects to ensure the actual application effect of the model. At the same time, at least one of the signal-to-noise ratio (SNR), peak signal-to-noise ratio (PSNR), and subjective auditory test can be used to evaluate the noise reduction effect of the model. For tuning the hyperparameters, methods such as grid search and Bayesian optimization can be used to find the best combination of hyperparameters to further improve the performance of the model. After model evaluation and parameter tuning, use the cross-validation method to evaluate the generalization ability of the model to ensure the stable performance of the model on different datasets.

[0113] Collect the original noise signal during the operation of the air conditioner and input it into the noise recognition model, which can identify the noise source and / or type and the corresponding noise intensity, and output a noise reduction signal, where the noise reduction signal contains the identified noise source and / or type and the corresponding noise intensity.

[0114] Step S130, determine the current noise source and / or type and the corresponding noise intensity of the air conditioner according to the generated noise reduction signal.

[0115] Specifically, after being processed by the noise recognition model, it can identify the noise source and / or type and the corresponding noise intensity, and output a noise reduction signal, where the noise reduction signal contains the identified noise source and / or type and the corresponding noise intensity. Identifying the noise source and / or type and the noise intensity in the noise reduction signal is the current noise source and / or type and the current noise intensity of the air conditioner.

[0116] Preferably, synthesize the generated noise reduction signal with the original noise signal to obtain a synthesized signal; determine the current noise source and / or type of the air conditioner according to the obtained synthesized signal.

[0117] Generate a noise-reduced signal through the trained Swin-Conv-UNet hybrid model, and then synthesize the noise-reduced signal with the original noise signal through audio processing technology to obtain a synthesized signal. Before signal synthesis, make the phases of the noise-reduced signal and the original noise signal consistent to avoid distortion caused by inconsistent phases; during signal synthesis, use a smooth transition technology to avoid auditory discomfort caused by sudden signal changes; finally, perform a weighted average on the noise-reduced signal and the original noise signal to generate the final synthesized signal. By adjusting the weights, noise can be effectively removed while preserving the original audio information. The synthesized signal combines the noise-reduced signal and the original noise signal, retains the useful information in the original signal, and avoids information loss caused by excessive noise reduction. By ensuring that the phases of the noise-reduced signal and the original noise signal are consistent, sound distortion caused by inconsistent phases is avoided, and the sound quality is improved.

[0118] In a specific implementation, the source and / or type of noise and the noise intensity in the noise-reduced signal or the synthesized signal can be identified through a pre-trained machine learning model. That is, input the noise-reduced signal into the model, and output the source and / or type of noise and the noise intensity.

[0119] Step S140, according to the determined current noise source and / or type of the air conditioner and the corresponding noise intensity, obtain the noise reduction control strategy of the air conditioner.

[0120] Specifically, different noise sources and / or types and the corresponding noise intensities have different noise reduction control strategies. In a specific implementation, according to the magnitude of the corresponding noise intensity exceeding the corresponding preset threshold, determine the adjustment value of the air conditioner control parameter, or according to the magnitude of the corresponding noise intensity exceeding the corresponding preset threshold, combined with the deviation of the current environmental parameter from the air conditioner set parameter (for example, the deviation of the indoor environmental temperature from the set temperature of the air conditioner), determine the adjustment value of the air conditioner control parameter. When the current environmental parameter has reached the set parameter, determine the adjustment value of the corresponding control parameter according to the magnitude of the difference between the current environmental parameter and the set parameter and the magnitude of the corresponding noise intensity exceeding the corresponding preset threshold.

[0121] For example, if it is compressor noise and the corresponding noise intensity exceeds the corresponding preset threshold, the compressor frequency can be reduced. The reduction value of the compressor frequency can be determined according to the magnitude of the corresponding noise intensity exceeding the corresponding preset threshold, or according to the magnitude of the corresponding noise intensity exceeding the corresponding preset threshold, combined with the deviation of the current environmental parameters relative to the air conditioner set parameters (for example, the deviation of the indoor environmental temperature relative to the set temperature of the air conditioner), to determine the reduction value of the compressor frequency. For example, when the current environmental parameters have reached the set parameters, the reduction value of the corresponding compressor frequency is determined according to the magnitude of the difference between the current environmental parameters and the set parameters and the magnitude of the corresponding noise intensity exceeding the corresponding preset threshold.

[0122] Also for example, if it is fan noise and the corresponding noise intensity exceeds the corresponding preset threshold, the fan speed can be reduced. The reduction value of the fan speed can be determined according to the magnitude of the corresponding noise intensity exceeding the corresponding preset threshold, or according to the magnitude of the corresponding noise intensity exceeding the corresponding preset threshold, combined with the deviation of the current environmental parameters relative to the air conditioner set parameters, to determine the reduction value of the fan speed. For example, when the current environmental parameters have reached the set parameters, the reduction value of the corresponding fan speed is determined according to the magnitude of the difference between the current environmental parameters and the set parameters and the magnitude of the corresponding noise intensity exceeding the corresponding preset threshold.

[0123] Step S150, generate corresponding control instructions according to the obtained noise reduction control strategy, and control the air conditioner to execute.

[0124] For example, generate corresponding control instructions according to the adjustment value of the air conditioner control parameters, and send the generated control instructions to the corresponding actuator for execution. The actuator adjusts the operating state of the air conditioner according to the received parameters.

[0125] Optionally, the method may further include: collecting the current environmental data and / or the operating data of the air conditioner, and feeding back the collected current environmental data and / or the operating data of the air conditioner, as well as the identified current noise source and / or type information and the corresponding noise intensity information of the air conditioner to the air conditioner monitoring interface. For example, feeding back to the air conditioner monitoring interface in the user UI interface module, and the air conditioner status update can be synchronized to the air conditioner monitoring interface in the user UI interface module in real time, and the user can manually adjust the operating parameters of the air conditioner, such as temperature, humidity, and noise intensity.

[0126] The environmental data may specifically include: indoor environmental temperature and / or humidity, outdoor environmental temperature and / or humidity. The operating data of the air conditioner may specifically include: at least one of set temperature, set humidity, and set wind speed.

[0127] For example, humidity sensors, temperature sensors, sound sensors, and / or vibration sensors are installed in the air conditioner control system to monitor in real time the ambient humidity, temperature, and / or the original data such as the noise and / or vibration generated during the operation of the air conditioner.

[0128] Optionally, the method may further include: receiving a noise adjustment instruction issued through the user interface; and adjusting the noise generated during the operation of the air conditioner according to the received noise adjustment instruction.

[0129] For example, according to the noise intensity adjusted by the user on the user interface, the noise generated during the operation of the air conditioner is adjusted by adjusting the control parameters of the air conditioner (such as the compressor frequency and the fan speed).

[0130] Optionally, the operation record of adjusting the noise generated during the operation of the air conditioner can also be stored. For example, a data storage service is requested to store the current control operation record of the air conditioner, which is convenient for analyzing the operation situation of the air conditioner through the control operation record of the air conditioner later, and the stored record is synchronized to the air conditioner monitoring interface data in real time.

[0131] To clearly illustrate the technical solution of the present invention, the execution process of the control method of the air conditioner provided by the present invention will be described below with several specific embodiments.

[0132] Figure 5 It is a structural diagram of air conditioner noise reduction control in a specific embodiment of the control method of the air conditioner provided by the present invention. As Figure 5 shown, environmental data and noise data are collected in real time through sensors and / or microphones. The collected data is analyzed by the trained Swin-Conv-UNet hybrid model integrated in the controller to identify the current noise source and / or type and the current noise intensity of the air conditioner. The actuator generates and executes control instructions according to the identified current noise source and / or type and the current noise intensity of the air conditioner by the controller, and the preset noise reduction strategy.

[0133] Figure 6 It is a flowchart of air conditioner noise reduction control in a specific embodiment of the control method of the air conditioner provided by the present invention. As Figure 6 shown, the data acquisition module acquires noise data, the Swin-Conv-UNet hybrid model performs noise feature extraction and noise reduction signal generation, the noise control execution module controls the air conditioner according to the noise reduction signal, adjusts the operation parameters of the air conditioner, and can save the operation record through the data storage service, and the collected environmental data and noise data are displayed in real time through the air conditioner monitoring interface of the user UI interface, and the operation parameters of the air conditioner manually adjusted by the user, including parameters such as temperature and noise, are received through the air conditioner control interface of the user UI interface.

[0134] The present invention also provides a control device for an air conditioner.

[0135] Figure 7 This is a structural block diagram of an embodiment of the control device of the air conditioner provided by the present invention. As Figure 7 shown, the control device 100 includes: a collection unit 110, an identification unit 120, a determination unit 130, an acquisition unit 140, and a control unit 150.

[0136] The collection unit 110 is used to collect the original noise signal when the air conditioner is running.

[0137] In a specific embodiment, the original noise signal can be collected by a microphone array. For example, a microphone array with high sensitivity and wide-band response is selected to ensure the quality of the collected audio data. The number and arrangement of microphones can be set according to the actual application scenario to cover as large a spatial range as possible and capture the original audio signal.

[0138] The identification unit 120 is used to input the original noise signal collected by the collection unit 110 into a pre-trained noise identification model to generate a noise reduction signal.

[0139] The noise reduction signal contains noise source and / or type information and corresponding noise intensity information. Preferably, before inputting the collected original noise signal into a pre-trained noise identification model to generate a noise reduction signal, the collected original noise signal is pre-processed to obtain a processed noise signal. In a specific embodiment, the pre-processing may specifically include: signal conversion, noise reduction processing, and / or filtering processing.

[0140] After sampling and quantization processing, the noise signal is converted from an analog signal to a digital signal. To improve the quality of the spectrogram, noise reduction and filtering processing are performed on the sound signal to ensure signal quality. Sampling is to discretize the sound signal at a certain time interval. Quantization is to discretize the signal amplitude obtained by sampling, usually converted into an integer form. Noise reduction is to remove background noise or other interference signals. Filtering is to eliminate unwanted frequency components.

[0141] The signal conversion may specifically include: converting the original noise signal from an analog signal to a digital signal. The noise signal usually exists in the form of an analog signal, so it needs to go through a series of processing steps to be converted into a digital signal to make it suitable for subsequent digital processing and analysis. Specifically, the continuous analog signal is converted into a discrete digital signal, and the conversion can be achieved through an analog-to-digital converter (ADC).

[0142] Noise signals are often interfered by various noises during the acquisition process, such as electromagnetic interference, environmental noise, etc. Filters can be used to remove unwanted frequency components, or signal enhancement techniques can be used to improve the signal-to-noise ratio (SNR). The filtering process includes: low-pass filtering and / or band-pass filtering. For example, a low-pass filter is used to remove high-frequency noise, and a band-pass filter is used to retain useful signals within a specific frequency band.

[0143] Preferably, before inputting the collected original noise signal into a pre-trained noise recognition model to generate a denoised signal, a spectrogram or a Mel spectrogram is generated according to the collected original noise signal.

[0144] In addition to signal conversion and noise processing, the processed digital signal also needs to be formatted into a format that can be processed by the noise recognition model. For example, it is converted into a spectrogram or a Mel spectrogram. These graphical representations help the model better understand the characteristics of sound data. After preprocessing the collected original noise signal to obtain a processed noise signal, a spectrogram or a Mel spectrogram is generated according to the obtained processed noise signal.

[0145] The generation of a spectrogram depends on the short-time Fourier transform (STFT). The sound signal is divided into multiple short-time windows, and the Fourier transform is performed on each window to obtain the frequency information at each time point. Specifically, a suitable window function (such as a Hann window, a Hamming window, etc.) is selected, and the window size and overlap ratio are determined. The window size determines the balance between time resolution and frequency resolution, and the overlap ratio affects the smoothness of the spectrogram. The fast Fourier transform (FFT) is performed on the signal within each window to obtain the spectral information of that window. Through the sliding window technique, a continuous spectrogram is generated, reflecting the frequency distribution of the sound signal over time. The spectrogram is a two-dimensional image, with the horizontal axis representing time and the vertical axis representing frequency, and the color representing the energy intensity. By adjusting the color mapping, the change in frequency of the sound signal over time can be more intuitively displayed.

[0146] The Mel spectrogram is a method of representing sound based on human auditory perception. It converts the frequency axis from linear frequency to the Mel frequency scale, which is closer to the perceptual characteristics of the human auditory system for sound frequencies. The Mel frequency scale expands the frequency intervals in the low-frequency part and compresses the frequency intervals in the high-frequency part to better reflect the non-linear characteristics of human hearing. To generate a Mel spectrogram, first perform the short-time Fourier transform STFT on the sound signal to obtain a spectrogram; on the basis of the spectrogram, apply a Mel filter bank (i.e., a set of band-pass filters) to evenly distribute the spectrogram response on the Mel frequency scale; perform a logarithmic transformation on the Mel spectrogram to enhance the perceptibility of the low-energy part and at the same time compress the dynamic range to make it more suitable for input to machine learning models.

[0147] Preferably, in order to ensure that the data has the same scale when input into the model and avoid affecting the model performance due to signal strength differences, the spectrogram or mel spectrogram can also be normalized. The normalization methods can include, for example: (1) Min-max normalization - scaling the minimum and maximum values of the data to the range [0,1]. (2) Z-score standardization - converting the data into a distribution with a mean of 0 and a standard deviation of 1.

[0148] Preferably, after converting the normalized spectrogram or mel spectrogram into a preset format, it is input into the noise recognition model, that is, converted into a format suitable for input into the machine learning model. For example, the processed spectrogram or mel spectrogram is converted into a format suitable for input into the Swin-Conv-UNet model. In a specific embodiment, converting the normalized spectrogram or mel spectrogram into a preset format may specifically include:

[0149] (1) Data dimension adjustment - adjusting the dimension of the data according to the input requirements of the model. For example, some models may require a three-dimensional tensor (such as a deep learning model), and the two-dimensional spectrogram needs to be extended to three-dimensional data.

[0150] (2) Channel number extension - if the model requires multi-channel input, the demand can be met by extending the channel dimension (such as extending a single channel to three channels).

[0151] (2) Data format conversion - converting the processed data into a format suitable for model input. For example, it can be at least one of a NumPy array and a TensorFlow tensor.

[0152] In a specific embodiment, the noise recognition model is a Swin-Conv-UNet hybrid model. The noise recognition model may specifically include: an input layer, a Swin Transformer layer, a convolutional neural network (CNN), a U-Net symmetric structure, and an output layer.

[0153] As Figure 2 shown, in a specific embodiment, the recognition unit 120 inputting the original noise signal collected by the collection unit 110 into a pre-trained noise recognition model to generate a noise reduction signal may specifically include the following steps S121 to step S125.

[0154] Step S121, inputting the collected original noise signal into the noise recognition model through the input layer.

[0155] Specifically, inputting the processed spectrogram or mel spectrogram into the input layer. These signals can be data in the form of images, audio, or other forms, which contain unwanted noise components.

[0156] Step S122: Capture the global dependencies in the input original noise signal through the Swin Transformer layer to generate a feature map.

[0157] Specifically, in the Swin-Conv-UNet model, the Swin Transformer layer can effectively capture the global dependencies in the input data and process complex features through its unique hierarchical and shifted window self-attention mechanism. Specifically for air conditioner noise data, its global dependencies mainly include the following aspects:

[0158] 1. Global dependencies in time

[0159] There are correlations between the sound signals at different time points. For example, the noise level at a certain moment may be affected by the operating state of the device at the previous moment. This temporal continuity needs to be captured by the model to accurately predict future noise changes.

[0160] 2. Global dependencies in frequency

[0161] There are also mutual influences between different frequency components of the noise signal. For example, there may be some intrinsic connection between low-frequency noise and high-frequency noise. This frequency dependence is crucial for comprehensively understanding the noise characteristics.

[0162] 3. Global dependencies in space

[0163] In the noise data of multiple devices or multiple locations, there may be mutual influences between the sound signals of different devices or locations. For example, when multiple air conditioner devices are running simultaneously, their noises may be superimposed or interfered with each other, forming a complex noise environment.

[0164] 4. Dependencies between operating state and noise

[0165] There is a close relationship between the operating state of the air conditioner device (such as load, temperature setting, operating time, etc.) and the noise level it generates. The model needs to capture the dependencies between these device state parameters and the noise data to perform more accurate noise prediction and analysis.

[0166] The Swin Transformer layer can learn these global dependencies at different levels through its hierarchical structure and shifted window mechanism. The hierarchical structure allows the model to capture features at different scales, while the shifted window mechanism enables the model to effectively handle long-range dependencies while maintaining computational efficiency. Combined with the convolutional layer, the Swin-Conv-UNet model can fully utilize the advantages of both to achieve efficient capture and processing of complex features in air conditioner noise data.

[0167] Step S123, perform local feature extraction on the generated feature map through a convolutional neural network.

[0168] Specifically, the Swin Transformer layer captures the global dependencies in the sound signal through its hierarchical structure and shifted window self-attention mechanism to generate a feature map. The feature map processed by the Swin Transformer is passed to a convolutional neural network (CNN), and convolutional layers are used to extract local features, that is, multi-scale features in the sound signal are captured through multi-scale convolutions (convolution kernels of different sizes). A residual connection can be introduced between the Swin Transformer layer and the convolutional layer to enhance the feature propagation ability. For example, the output of the Swin Transformer layer is added to the output of the convolutional layer to form a new feature map. Through the use of multiple neural network layers and activation functions, the convolutional neural network can learn more detailed information about noise and signals.

[0169] Preferably, the noise recognition model further includes: a fully connected layer for identifying the source and / or type of the noise and the corresponding noise intensity according to the features extracted by the convolutional neural network. Specifically, after feature extraction, one or more fully connected layers are used as classifiers to locate and identify the source and / or type of the noise and the corresponding noise intensity according to the extracted features.

[0170] Step S124, perform segmentation and reconstruction on the extracted local features through a U-Net symmetric structure to generate a spectrogram or mel spectrogram containing information about the source and / or type of the noise.

[0171] The U-Net symmetric structure can learn the difference between the background noise and the useful signal in the air conditioner noise, thereby generating a denoised spectrogram or mel spectrogram. Specifically, the U-Net symmetric structure is used to reconstruct a clear signal from the features obtained from the convolutional neural network (CNN) layer. U-Net is a network with an encoder-decoder structure. The encoder part gradually reduces the spatial dimension of the feature map, and the decoder part gradually restores the dimension. At the same time, the combination of low-level and high-level features is maintained through skip connections.

[0172] The decoder part is responsible for reconstructing the feature maps extracted by the encoder into an optimized spectrogram or mel spectrogram, combining the decoder architecture of U-Net and skip connections: upsample through transposed convolutional layers or interpolation methods to gradually restore the spatial resolution of the feature maps; and introduce skip connections to combine the low-level features of the encoder part with the current feature maps to enhance the reconstruction ability. Use convolutional layers to finely adjust the feature maps to further optimize the reconstruction results; capture the detailed features in the sound signal through multi-scale convolutions. Introduce residual connections between the convolutional layers in the decoder part to enhance the feature propagation ability. For example, add the output of the current convolutional layer to the output of the previous layer to form a new feature map.

[0173] To improve the training effect and feature propagation ability of the model, residual connections and skip connections are introduced in the encoder and decoder parts: introduce residual connections between the convolutional layers of the encoder and decoder to form residual blocks; the output of the residual block is the sum of the output of the current layer and the input feature map, which helps to alleviate the problem of gradient disappearance. Introduce skip connections between the encoder and decoder to combine the low-level features of the encoder part with the high-level features of the decoder part; skip connections help to retain the detailed information in the sound signal and improve the accuracy of reconstruction. The skip connections transfer the high-resolution features in the encoder stage to the decoder stage, ensuring that more detailed information is retained during the reconstruction process, improving the reconstruction quality and robustness of the model, and also promoting the fusion of features at different levels, enabling the model to better capture multi-scale features and improve the generalization ability of the model. After the feature output of the encoder, it can be connected to a fully connected layer to identify the noise source and / or type and the corresponding noise intensity, that is, the fully connected layer locates and identifies the noise source and / or type and the corresponding noise intensity according to the features output by the encoder.

[0174] Step S125, output the spectrogram or mel spectrogram through more than one output layer, and convert the output spectrogram or mel spectrogram into a time-domain signal to obtain a noise-reduced signal containing noise source and / or type information and the corresponding noise intensity information.

[0175] Specifically, the result output by the U-Net symmetric structure decoder is processed by one or more output layers to generate a final noise-reduced signal. The output layer adjusts the number of channels of the feature map output by the decoder through convolutional operations to make it consistent with the number of channels of the input spectrogram. Preferably, the output layer uses a linear activation function to ensure that the amplitude information of the spectrogram is retained and avoid distorting the signal amplitude by non-linear transformation.

[0176] If the output of the output layer is a spectrogram or a Mel spectrogram, it is converted into a time-domain signal. For example, it is converted into a time-domain signal through the inverse short-time Fourier transform (iSTFT). The implementation of iSTFT needs to consider parameters such as the window function and frame shift to ensure the reconstruction quality of the time-domain signal. Optionally, the time-domain signal generated by the inverse short-time Fourier transform (iSTFT) is normalized to ensure that the signal amplitude is within an appropriate range and avoid sound quality distortion.

[0177] The Swin-Conv-UNet hybrid model combines the global perception ability of the Swin Transformer, the local feature extraction ability of the CNN, and the effective reconstruction mechanism of the U-Net to achieve efficient and accurate noise feature extraction, noise source identification, and denoised signal generation. This enables it to perform well in processing various noisy signals, whether visual images or audio data.

[0178] Figure 3 It is a flowchart of the training steps of the noise recognition model according to the present invention. As Figure 3 shown, in a specific embodiment, the noise recognition model is trained through the following steps S1 to step S6:

[0179] Step S1: Obtain the original noise samples obtained by pre-labeling the noise signals generated by different air conditioners operating under different environments and different operating states.

[0180] Specifically, pre-collect the noise signals generated by different air conditioners operating under different environments and different operating states; and label the collected noise signals generated by different air conditioners operating under different environments and different operating states as the original noise samples.

[0181] Collect the noise data of the air conditioner equipment operating under different environmental conditions and label the data. The collection of noise data can use a microphone array with high sensitivity and wide-band response to ensure the quality of the collected audio data. The number and arrangement of microphones can be optimized according to the actual application scenario to cover as large a spatial range as possible. Use professional recording equipment to ensure that the collected audio data is distortion-free and free of noise interference.

[0182] The different environments may specifically include different indoor environment types, outdoor environment types, temperature, and / or humidity. The indoor environment type may specifically be a room type, such as a bedroom, a living room, or an office; the outdoor environment type may include, for example, a courtyard and / or a street. Collecting data under different humidity and / or temperature conditions can ensure the applicability of the model under various environmental conditions.

[0183] The methods for collecting noise signals generated by different air conditioners operating in different environments and different operating states may specifically include: continuous collection for a preset duration in different environments and different operating states. That is, continuous collection for a long time in each environment, and recording the sounds of the air conditioner in different operating states. Noise data collection is carried out in different operating modes (such as cooling, heating, dehumidification) to ensure that the model can process noises in multiple working states. Different air conditioners may specifically include air conditioners of different brands and / or models. Collecting noise data when different brands and models of air conditioner equipment are operating can improve the generality of the model.

[0184] The annotation information may specifically include: the noise source and / or type and the corresponding noise intensity. The noise source may specifically be the equipment, components and / or positions where the noise is generated. For example, it may include at least one of compressor noise, evaporator noise, and fan noise; the noise type may specifically be a description of the noise characteristics. For example, it may include at least one of mechanical noise (vibration sound of mechanical components during air conditioner operation), wind noise (airflow sound generated during air conditioner air supply), electronic noise (noise generated when internal electronic components of the air conditioner are working), and external noise (background noise in the environment such as traffic noise, human voices, etc.).

[0185] To annotate the noise signals generated by the different air conditioners operating in different environments and different operating states, the noise source and / or type can be classified first, and then for different noise sources and / or types, professional equipment is used to measure the noise decibel values under different noise sources and / or types to ensure the accuracy of the annotation. In addition, the intensity of the noise can be evaluated through subjective auditory tests of humans to ensure the comprehensiveness of the annotation. Manual annotation can be carried out through professional audio annotation software, or combined with automated tools and manual review to improve the annotation efficiency and accuracy.

[0186] Step S2, using a pre-trained noise synthesis model, synthesize new noise samples from the original noise samples, and integrate the original noise samples and the new noise samples into a training data set.

[0187] The noise synthesis model is a generative model used to generate new noise data based on existing noise samples. Select a suitable noise synthesis model according to the task requirements and data characteristics. For example, GANs are suitable for generating high-quality noise samples, and VAEs are suitable for learning the distribution of the latent space. Input the original noise samples into the pre-trained noise synthesis model to synthesize new noise samples.

[0188] Preferably, after the original noise samples are input into the pre-trained noise synthesis model to synthesize new noise samples, data augmentation can also be performed on the synthesized new noise samples. Specifically, after the original noise samples are input into the pre-trained noise synthesis model to generate new noise samples, it is determined whether additional data augmentation is required based on the generation ability of the model and the diversity of the generated samples. If the generated samples already have sufficient diversity, the additional data augmentation step can be omitted; if the generated samples are relatively single, data augmentation can be considered after generating new noise samples to enrich the data set and improve the generalization ability of the model. The data augmentation can specifically include at least one of adding background noise, changing the volume, time stretching, and frequency shifting.

[0189] Adding background noise means adding different types of background noise to the generated noise samples to simulate the noise situation in a complex environment. Changing the volume means adjusting the volume of the generated noise to simulate the sound intensity at different distances and in different environments. Time stretching means changing the time length of the noise samples to simulate the sound changes under different device operating states. Frequency shifting means adjusting the frequency of the noise samples to simulate the frequency characteristics under different devices or environments. Use audio processing libraries (such as Librosa, SoundFile, etc.) to implement the above data augmentation techniques to generate diverse training data.

[0190] Figure 4 is a flowchart of the training steps of the noise synthesis model according to the present invention. As Figure 4 shown, in a specific embodiment, the noise synthesis model can be specifically trained through the following steps S21 to S23:

[0191] Step S21, obtain pre-labeled noise samples and preprocess the noise samples.

[0192] Organize the labeled noise samples into a format suitable for model input, for example, spectrogram or mel spectrogram. Preprocess the noise samples, including standardization, normalization, etc., to ensure that the data has the same scale when input into the model.

[0193] Step S22, design the network architecture of the noise synthesis model according to the pre-determined model type.

[0194] According to the selected model type, design a suitable network architecture, including a generator and a discriminator (such as GANs) or an encoder and a decoder (such as VAEs).

[0195] Step S23, use the noise samples to train the noise synthesis model and optimize the model parameters.

[0196] Train a noise synthesis model using the labeled noise samples and optimize the model parameters to enable the generator to generate realistic noise samples.

[0197] Generating diverse training data through the noise synthesis model can further improve the generalization ability and anti-interference ability of the model, ensuring its performance in practical applications. Increase the diversity and complexity of the data, and additionally perform data augmentation by adding background noise, changing the volume, time stretching, and frequency shifting, etc., to further improve the robustness of the model. By generating diverse training data, the model can exhibit stable performance under different environmental conditions. Diverse data can cover more application scenarios and improve the generalization ability of the model. In addition, during the training process, the model is exposed to various types of noise, enabling it to better handle complex noise environments in practical applications and improve the anti-interference ability.

[0198] Integrate the original noise samples and the new noise samples into a training dataset, and use the integrated training dataset to train the Swin-Conv-UNet hybrid model, which can enhance the noise reduction and recognition capabilities of the model in different noise environments.

[0199] Step S3, construct the Swin-Conv-Unet hybrid model structure and use the training dataset to train the model.

[0200] The Swin-Conv-Unet hybrid model structure can specifically include: an input layer, a Swin Transformer layer, a convolutional neural network (CNN), a U-Net symmetric structure, and an output layer.

[0201] The Swin Transformer layer can effectively capture local features by dividing the input spectrogram into multiple non-overlapping windows and performing the self-attention mechanism within each window. This mechanism reduces the computational amount and memory consumption while maintaining the ability to learn long-range dependencies. Swin Transformer can capture features at different scales, from local details to global structures, providing rich feature representations for the model. The convolutional neural network CNN captures local features in the spectrogram, such as edges and textures. Through multiple layers of convolution and pooling operations, CNN can gradually extract higher-level features, and the convolution operations of CNN have efficient computational performance and are suitable for processing large-scale data. The fully connected layer identifies the source and / or type of the noise and the corresponding noise intensity based on the features extracted by the convolutional neural network; UNet gradually reduces the spatial resolution of the spectrogram and increases the number of channels of the feature map through the encoder to capture deep-level features; while through the decoder, it gradually restores the spatial resolution of the spectrogram and reduces the number of channels of the feature map to achieve high-quality reconstruction.

[0202] When performing model training, first select appropriate initial model parameters, such as including weight initialization and / or activation functions; then use the Adam optimizer to adjust the model parameters through the backpropagation algorithm; then, through the learning rate decay strategy, gradually reduce the learning rate to prevent oscillations during training and improve the convergence speed of the model. Finally, select an appropriate batch size to balance computing resources and training efficiency, ensuring that the model is trained within a reasonable time.

[0203] Before model training, use L1 or L2 regularization techniques to optimize the model. By adding a regularization term to the loss function, prevent the model parameters from being too large, reduce the complexity of the model, and prevent overfitting; during training, monitor the performance of the validation set, and stop training in advance when the performance of the validation set no longer improves to avoid overfitting; after training, perform data augmentation by adding background noise, changing the volume, time stretching, and frequency shifting, etc. on the basis of the previous training, and then conduct the next model training. At this time, the diversity and complexity of the training data will be increased, and the generalization ability of the model will be improved.

[0204] Step S4, evaluate the model and tune the hyperparameters of the model to obtain the final noise recognition model.

[0205] The evaluation metrics should comprehensively consider both objective and subjective aspects to ensure the actual application effect of the model. At the same time, at least one of the signal-to-noise ratio (SNR), peak signal-to-noise ratio (PSNR), and subjective auditory test can be used to evaluate the noise reduction effect of the model. For hyperparameter tuning, methods such as grid search and Bayesian optimization can be used to find the best combination of hyperparameters to further improve the performance of the model. After model evaluation and parameter tuning, use the cross-validation method to evaluate the generalization ability of the model to ensure the stable performance of the model on different datasets.

[0206] Collect the original noise signal during air conditioner operation and input it into the noise recognition model, which can identify the noise source and / or type and the corresponding noise intensity, and output a noise reduction signal, where the noise reduction signal contains the identified noise source and / or type and the corresponding noise intensity.

[0207] A determination unit 130 is used to determine the current noise source and / or type and the corresponding noise intensity of the air conditioner according to the noise reduction signal generated by the recognition unit.

[0208] Specifically, after being processed by the noise recognition model, it can identify the noise source and / or type and the corresponding noise intensity, and output a noise reduction signal, where the noise reduction signal contains the identified noise source and / or type and the corresponding noise intensity. Identifying the noise source and / or type and the noise intensity in the noise reduction signal is the current noise source and / or type and the current noise intensity of the air conditioner.

[0209] Preferably, the generated noise reduction signal is combined with the original noise signal to obtain a combined signal; based on the obtained combined signal, the current noise source and / or type of the air conditioner are determined.

[0210] The noise reduction signal is generated by a trained Swin-Conv-UNet hybrid model, and then the noise reduction signal and the original noise signal are combined through audio processing technology to obtain a combined signal. Before signal combination, the phases of the noise reduction signal and the original noise signal are made consistent to avoid distortion caused by inconsistent phases; during signal combination, a smooth transition technology is used to avoid auditory discomfort caused by sudden signal changes; finally, the noise reduction signal and the original noise signal are weighted and averaged to generate the final combined signal. By adjusting the weights, noise can be effectively removed while retaining the original audio information. The combined signal combines the noise reduction signal and the original noise signal, retains the useful information in the original signal, and avoids information loss caused by excessive noise reduction. By ensuring that the phases of the noise reduction signal and the original noise signal are consistent, sound distortion caused by inconsistent phases is avoided, and the sound quality is improved.

[0211] In a specific implementation manner, the noise source and / or type and noise intensity in the noise reduction signal or the combined signal can be identified through a pre-trained machine learning model. That is, the noise reduction signal is input into the model, and the noise source and / or type and noise intensity are output.

[0212] An acquisition unit 140 is configured to acquire a noise reduction control strategy of the air conditioner according to the current noise source and / or type of the air conditioner and the corresponding noise intensity determined by the determination unit.

[0213] Specifically, different noise sources and / or types and the corresponding noise intensities have different noise reduction control strategies. In a specific implementation manner, according to the magnitude of the corresponding noise intensity exceeding the corresponding preset threshold, the adjustment value of the air conditioner control parameter is determined, or according to the magnitude of the corresponding noise intensity exceeding the corresponding preset threshold, in combination with the deviation of the current environmental parameter from the air conditioner set parameter (for example, the deviation of the indoor environmental temperature from the set temperature of the air conditioner), the adjustment value of the air conditioner control parameter is determined. When the current environmental parameter has reached the set parameter, according to the magnitude of the difference between the current environmental parameter and the set parameter and the magnitude of the corresponding noise intensity exceeding the corresponding preset threshold, the adjustment value of the corresponding control parameter is determined.

[0214] For example, if it is compressor noise and the corresponding noise intensity exceeds the corresponding preset threshold, the compressor frequency can be reduced. The compressor frequency reduction value can be determined according to the magnitude of the corresponding noise intensity exceeding the corresponding preset threshold, or according to the magnitude of the corresponding noise intensity exceeding the corresponding preset threshold, combined with the deviation of the current environmental parameters relative to the air conditioner set parameters (for example, the deviation of the indoor environmental temperature relative to the set temperature of the air conditioner), to determine the compressor frequency reduction value. For example, when the current environmental parameters have reached the set parameters, the corresponding compressor frequency reduction value is determined according to the magnitude of the difference between the current environmental parameters and the set parameters and the magnitude of the corresponding noise intensity exceeding the corresponding preset threshold.

[0215] For another example, if it is fan noise and the corresponding noise intensity exceeds the corresponding preset threshold, the fan speed can be reduced. The fan speed reduction value can be determined according to the magnitude of the corresponding noise intensity exceeding the corresponding preset threshold, or according to the magnitude of the corresponding noise intensity exceeding the corresponding preset threshold, combined with the deviation of the current environmental parameters relative to the air conditioner set parameters, to determine the fan speed reduction value. For example, when the current environmental parameters have reached the set parameters, the corresponding fan speed reduction value is determined according to the magnitude of the difference between the current environmental parameters and the set parameters and the magnitude of the corresponding noise intensity exceeding the corresponding preset threshold.

[0216] The control unit 150 is configured to generate corresponding control instructions according to the noise reduction control strategy acquired by the acquisition unit, and control the air conditioner to execute.

[0217] For example, generate corresponding control instructions according to the adjustment value of the air conditioner control parameters, and send the generated control instructions to the corresponding actuator for execution. The actuator adjusts the operating state of the air conditioner according to the received parameters.

[0218] Optionally, the acquisition unit 110 is further configured to: acquire the current environmental data and / or the operating data of the air conditioner; the device 100 further includes a feedback unit, configured to feedback the current environmental data and / or the operating data of the air conditioner acquired by the acquisition unit 110, as well as the identified current noise source and / or type information and the corresponding noise intensity information of the air conditioner to the air conditioner monitoring interface. The air conditioner status update can be synchronized to the air conditioner monitoring interface in the user UI interface module in real time, and the user can manually adjust the operating parameters of the air conditioner, such as temperature, humidity, and noise intensity.

[0219] The environmental data may specifically include: indoor environmental temperature and / or humidity, outdoor environmental temperature and / or humidity. The operating data of the air conditioner may specifically include: at least one of set temperature, set humidity, and set wind speed.

[0220] For example, a humidity sensor, a temperature sensor, a sound sensor, and / or a vibration sensor are installed in the air conditioner control system to monitor original data such as the ambient humidity, temperature, and / or the noise and / or vibration generated during the operation of the air conditioner in real time.

[0221] Optionally, the device 100 further includes: a receiving unit (not shown) for receiving a parameter adjustment instruction issued through the user interface; an adjustment unit (not shown) for adjusting the noise generated during the operation of the air conditioner according to the parameter adjustment instruction received by the receiving unit.

[0222] For example, according to the magnitude of the noise intensity adjusted by the user on the user interface, the noise generated during the operation of the air conditioner is adjusted by adjusting the control parameters of the air conditioner (such as the compressor frequency and the fan speed).

[0223] Optionally, it may further include a storage unit for storing the operation records of adjusting the noise generated during the operation of the air conditioner. For example, request the data storage service to store the current control operation records of the air conditioner, which is convenient for analyzing the operation status of the air conditioner through the control operation records of the air conditioner later, and synchronize the stored records to the air conditioner monitoring interface data in real time.

[0224] The present invention also provides a storage medium corresponding to the control method of the air conditioner, on which a computer program is stored, and when the program is executed by a processor, the steps of any of the foregoing methods are implemented.

[0225] The present invention also provides an air conditioner corresponding to the control method of the air conditioner, including a processor, a memory, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps of any of the foregoing methods are implemented.

[0226] The present invention also provides an air conditioner corresponding to the control device of the air conditioner, including the control device of the air conditioner according to any of the foregoing.

[0227] The present invention also provides a computer program product corresponding to the control method of the air conditioner, including a computer program, and when the computer program is executed by a processor, the steps of any of the foregoing methods are implemented.

[0228] Accordingly, the solution provided by the present invention can not only monitor the noise during the operation of the air conditioner in real time, but also automatically adjust the operation parameters of the air conditioner according to the monitoring data, effectively reducing the noise, thereby significantly improving the user's sleep quality and living comfort.

[0229] According to the technical solution of the present invention, the Swin-Conv-UNet hybrid model is applied to air conditioner noise control to achieve real-time monitoring of air conditioner noise. By real-time monitoring and analyzing the noise data generated during the operation of the air conditioner, not only can the noise source and / or type be quickly identified and located based on the noise data, but also the operating parameters of the air conditioner can be dynamically adjusted according to the actual situation, thereby achieving the effect of significantly reducing noise.

[0230] At the same time, the Swin-Conv-UNet hybrid model has the ability of adaptive learning and can continuously self-optimize according to user feedback and actual effects, gradually adapting to the needs of different users and environmental changes, providing personalized noise control solutions for users, and enhancing the user experience and the adaptability of the system.

[0231] According to the technical solution of the present invention, through the real-time monitoring and intelligent regulation of the Swin-Conv-UNet hybrid model, the noise generated during the operation of the air conditioner can be significantly reduced, providing a more quiet and comfortable living environment for users.

[0232] According to the technical solution of the present invention, by dynamically adjusting the operating parameters of the air conditioner, the operating efficiency of the air conditioner can be improved while ensuring the noise control effect, achieving energy conservation and consumption reduction.

[0233] According to the technical solution of the present invention, by introducing an adaptive learning mechanism, personalized noise control solutions can be provided according to user needs and environmental changes, thereby significantly improving the user experience and satisfaction.

[0234] The functions described herein can be implemented in hardware, software executed by a processor, firmware, or any combination thereof. If implemented in software executed by a processor, the functions can be stored on or transmitted via a computer-readable medium as one or more instructions or codes. Other examples and implementations are within the scope and spirit of the present invention and the appended claims. For example, due to the nature of software, the functions described above can be implemented using software executed by a processor, hardware, firmware, hardwiring, or any combination of these. In addition, each functional unit can be integrated in one processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit.

[0235] In several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are merely illustrative. For example, the division of the units can be a logical function division. In actual implementation, there can be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling or direct coupling or communication connection to each other can be through some interfaces. The indirect coupling or communication connection of units or modules can be in an electrical or other form.

[0236] The units described as separate components may or may not be physically separated. The components serving as control devices may or may not be physical units, that is, they can be located in one place or distributed to multiple units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0237] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the related technology, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), mobile hard disks, magnetic disks, or optical discs that can store program codes.

[0238] The above are only the embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the scope of the claims of the present invention.

Claims

1. A method for controlling an air conditioner, characterized in that: include: Collecting the original noise signal when the air conditioner is running; Inputting the collected original noise signal into a pre-trained noise recognition model to generate a noise reduction signal, wherein the noise recognition model is a Swin-Conv-UNet hybrid model, and the noise reduction signal contains noise source and / or type information and corresponding noise intensity information; Determining the current noise source and / or type of the air conditioner and the corresponding noise intensity according to the generated noise reduction signal; Acquire a noise reduction control strategy for the air conditioner according to the determined current noise source and / or type of the air conditioner and the corresponding noise intensity; Generate corresponding control instructions according to the acquired noise reduction control strategy, and control the air conditioner to execute.

2. The method according to claim 1, characterized in that: The noise recognition model includes: an input layer, a SwinTransformer layer, a convolutional neural network, a U-Net symmetric structure and an output layer; Inputting the collected original noise signal into a pre-trained noise recognition model to generate a noise reduction signal, comprising: Inputting the collected original noise signal into the noise recognition model through an input layer; Capturing the global dependencies in the input raw noise signal through the Swin Transformer layer to generate a feature map; Performing local feature extraction on the generated feature map through a convolutional neural network; The extracted local features are segmented and reconstructed through the U-Net symmetric structure to generate a spectrum or Mel-spectrogram containing noise source and / or type information and corresponding noise intensity information; The spectrogram or Mel-spectrogram is output through one or more output layers, and the output spectrogram or Mel-spectrogram is converted into a time domain signal to obtain a noise reduction signal including noise source and / or type information and corresponding noise intensity information.

3. The method according to claim 2, characterized in that Before inputting the collected original noise signal into a pre-trained noise recognition model to generate a noise reduction signal, the method further includes: Preprocessing the collected original noise signal to obtain a processed noise signal; and / or, A spectrogram or a Mel-spectrogram is generated according to the collected original noise signal.

4. The method according to claim 3, characterized in that Preprocessing the collected original noise signal, including: signal conversion, noise reduction processing and / or filtering processing; The signal conversion includes: converting the original noise signal from an analog signal into a digital signal; and / or, the filtering process includes: low-pass filtering and / or band-pass filtering.

5. The method according to any one of claims 2 to 4, characterized in that: The noise recognition model also includes: a fully connected layer, which is used to identify the source and / or type of noise and the corresponding noise intensity based on the features extracted by the convolutional neural network.

6. The method according to any one of claims 2 to 4, characterized in that: The noise recognition model is trained in the following way: Acquire original noise samples obtained by classifying and labeling noise signals generated by different air conditioners when they are running in different environments and under different operating conditions; Using a pre-trained noise synthesis model, synthesizing the original noise sample into a new noise sample, and integrating the original noise sample and the new noise sample into a training data set; Constructing a Swin-Conv-Unet hybrid model structure and using the training data set to perform model training; The model is evaluated and its hyperparameters are tuned to obtain the final noise recognition model.

7. The method according to claim 6, characterized in that The original noise sample is synthesized into a new noise sample by using a pre-trained noise synthesis model, including: Inputting the original noise sample into the pre-trained noise synthesis model to synthesize a new noise sample; Data enhancement is performed on the synthesized new noise sample; the data enhancement includes: adding background noise, changing the volume, time stretching and frequency shifting at least one of the following.

8. The method according to any one of claims 1 to 4, characterized in that: Determining the current noise source and / or type of the air conditioner according to the generated noise reduction signal includes: Performing signal synthesis on the generated noise reduction signal and the original noise signal to obtain a synthesized signal; The current noise source and / or type of the air conditioner is determined according to the obtained synthetic signal.

9. The method according to any one of claims 1 to 4, characterized in that: Receive a parameter adjustment instruction sent through a user interface; and adjust the noise generated when the air conditioner is running according to the received parameter adjustment instruction.

10. A control device for an air conditioner, characterized in that: include: A collection unit, used for collecting the original noise signal when the air conditioner is running; an identification unit, configured to input the original noise signal collected by the collection unit into a pre-trained noise identification model to generate a noise reduction signal, wherein the noise identification model is a Swin-Conv-UNet hybrid model, and the noise reduction signal includes noise source and / or type information and corresponding noise intensity information; a determination unit, configured to determine a current noise source and / or type of the air conditioner and a corresponding noise intensity according to the noise reduction signal generated by the identification unit; an acquiring unit, configured to acquire a noise reduction control strategy of the air conditioner according to a current noise source and / or type of the air conditioner and a corresponding noise intensity determined by the determining unit; A control unit is used to generate corresponding control instructions according to the noise reduction control strategy acquired by the acquisition unit, and control the air conditioner to execute.

11. A storage medium, characterized in that: A computer program is stored thereon, and when the program is executed by a processor, the steps of any method described in claims 1-9 are implemented.

12. An air conditioner, characterized in that: The method comprises a processor, a memory and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any method described in claims 1 to 9 when executing the program, or comprises the control device described in claim 10.

13. A computer program product, characterized in that The invention comprises a computer program, which implements the steps of any one of the methods of claims 1 to 9 when being executed by a processor.

Images