Refrigerator compressor noise adaptive suppression method based on deep learning

Abstract

The application discloses a refrigerator compressor noise adaptive suppression method based on deep learning and belongs to the technical field of refrigerator compressor noise control. The method acquires original vibration signals and original noise signals of a compressor; carries out chaotic feature analysis on the vibration signals to extract vibration state sequences; inputs the noise signals into a pre-trained generative adversarial network model to separate out compressor body noise components. A noise suppression intelligent agent is constructed by using a deep deterministic policy gradient network, time-frequency features of the vibration state sequences and the body noise components are taken as environment state inputs, a speed adjustment instruction is taken as an action, and a weighted sum of noise suppression effects and energy consumption indexes is taken as a reward function for training. The trained intelligent agent is deployed into a refrigerator controller to output a speed control signal to a variable frequency drive board in real time. The application realizes adaptive intelligent suppression of the compressor noise by fusing chaotic analysis, adversarial noise separation and deep reinforcement learning.

Description

Technical Field

[0001] This invention relates to the field of refrigerator compressor noise control technology, specifically a deep learning-based adaptive noise suppression method for refrigerator compressors. Background Technology

[0002] Refrigerator compressors generate aerodynamic, mechanical, and electromagnetic noise during refrigeration operation, directly impacting the user's quiet and comfortable experience. Currently, common techniques for suppressing compressor noise include: passive noise reduction structures such as laying sound-absorbing and insulating materials inside the refrigerator; fixing the compressor speed according to a preset operating condition table to avoid resonance points; or using a feedforward active noise reduction method based on filtering and inverse synthesis. Passive noise reduction methods have limited effectiveness in the low-frequency range and increase the overall weight and manufacturing cost of the refrigerator. Fixed-speed control strategies cannot respond to load changes, fluctuations in refrigeration demand, and noise characteristic drift caused by mechanical wear, and may sacrifice refrigeration efficiency or generate additional noise under certain operating conditions. Existing active noise reduction solutions require a large number of reference microphones, are sensitive to ambient noise, and when the refrigerator is placed in different acoustic spaces, the reference signal is contaminated by ambient noise, making it difficult for the system's output inverse noise signal to accurately track the compressor's own noise, significantly reducing the noise reduction effect.

[0003] Optimizing compressor speed using reinforcement learning has become a research hotspot. However, conventional methods directly use raw noise signals or simple sound pressure levels as state inputs, failing to distinguish between compressor-radiated noise and noise from other components such as refrigerator fans and refrigerant flow. Furthermore, they struggle to capture the dynamic characteristics reflecting the compressor's nonlinear operating state from vibration signals, resulting in insufficient generalization ability of the learned control strategies in real-world home acoustic environments. In addition, most existing reinforcement learning-based noise reduction schemes only minimize noise sound pressure levels as a single optimization objective, ignoring the energy cost of compressor speed changes. This can easily lead to unreasonable strategies that excessively reduce speed, extending operating time and increasing total power consumption. Simultaneously, conventional experience playback mechanisms use uniform sampling, resulting in inefficient use of scarce samples with higher learning value, drastic reward fluctuations during training, and slow convergence.

[0004] How to accurately separate the compressor's inherent noise from complex acoustic and vibration mixed signals, and extract state features that can characterize the compressor's chaotic dynamic behavior, so as to construct an adaptive control strategy that can balance noise reduction and energy consumption, is a problem that needs to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based adaptive noise suppression method for refrigerator compressors, which accurately separates the compressor body noise and extracts the dynamic features of vibration state in complex acoustic environments, thereby achieving an adaptive trade-off between noise suppression and operating energy consumption, and improving the environmental adaptability and training efficiency of the control strategy.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a deep learning-based adaptive noise suppression method for refrigerator compressors, the method comprising: The process involves acquiring raw vibration and noise signals generated during compressor operation; performing chaotic feature analysis on the raw vibration signals to extract compressor vibration state sequences that characterize the nonlinear dynamic behavior of the compressor; inputting the raw noise signals into a pre-trained generative adversarial network noise separation model, which separates the compressor body noise components from the raw noise signals; constructing a compressor noise suppression agent using a deep deterministic policy gradient network, with the time-frequency characteristics of the compressor vibration state sequences and the compressor body noise components serving as the environmental state input, compressor speed adjustment commands as the action output, and a weighted sum of noise suppression effect and energy consumption index as the reward function; and deploying the trained compressor noise suppression agent into the refrigerator controller, which outputs compressor speed control signals to the inverter drive board in real time, thereby adaptively reducing the perceived noise level while ensuring cooling effect.

[0007] As a preferred embodiment of the present invention, the process of performing chaotic feature analysis on the original vibration signal includes: performing phase space reconstruction processing on the original vibration signal, and determining appropriate embedding dimensions and time delay parameters respectively; calculating the correlation dimension of each phase point in the reconstructed phase space based on the embedding dimension and time delay parameters to obtain the geometric features of the chaotic attractor of the vibration signal; and extracting the recursion rate, determinism, and average diagonal length using a recursive graph analysis method based on the geometric features of the chaotic attractor, which together constitute the compressor vibration state sequence. This state extraction method can fully explore the subtle change patterns of compressor vibration, providing a highly discriminative state representation for subsequent intelligent agents.

[0008] Preferably, the embedding dimension in the phase space reconstruction process is determined by the spurious nearest neighbor method, and the time delay parameter is calculated by the mutual information method, so as to preserve the dynamic information of the original vibration signal to the greatest extent during the reconstruction process.

[0009] In one specific embodiment of the present invention, the separation process of the compressor body noise component by the generative adversarial network noise separation model includes: performing frame-by-frame windowing processing on the original noise signal to obtain multiple noise signal frames; inputting each frame of noise signal into the generator network of the generative adversarial network noise separation model, wherein the generator network adopts an encoder-decoder structure, wherein the encoder extracts deep features of the noise signal through convolutional layers and downsampling layers, and the decoder reconstructs the signal through deconvolutional layers and upsampling layers; sending the reconstructed signal output by the generator network and a pre-collected environmental background noise reference signal into the discriminator network of the generative adversarial network noise separation model, wherein the discriminator network outputs the probability that the reconstructed signal belongs to the compressor body noise, and adjusts the parameters of the generator network in reverse based on the probability until the discriminator network stably determines that the reconstructed signal is the compressor body noise signal, thereby obtaining a high-purity compressor body noise component. Through the above adversarial training mechanism, the model can accurately recover the acoustic characteristics of the compressor itself from strong background noise in the absence of pure noise labels.

[0010] In a further preferred embodiment, a skip connection is introduced between the encoder and decoder in the generator network to directly transmit the feature maps of each layer of the encoder to the corresponding layer of the decoder for splicing, thereby preserving more detailed information during signal reconstruction and improving the fidelity of the body noise separation.

[0011] As the core scheme for constructing a compressor noise suppression agent in this invention, the construction process of the deep deterministic policy gradient network includes: constructing an actor network to map the time-frequency features of the input compressor vibration state sequence and compressor body noise components to the mean value of compressor speed adjustment actions; constructing a critic network to evaluate the action value of performing the compressor speed adjustment action under a given environmental state; dividing the actor network and critic network into a current network and a target network respectively, with the parameters of the target network being copied from the parameters of the current network using a soft update method to suppress oscillations during training; and introducing a priority experience replay mechanism in the training of the critic network, i.e., non-uniform sampling is performed from the experience pool according to the absolute value of the time difference error, and higher priority experience samples are learned more frequently, thereby significantly improving sample efficiency and policy convergence speed.

[0012] Based on the above scheme, the sampling probability of the experience sample in the priority experience playback mechanism is proportional to the absolute value of the time difference error, and the specific priority calculation adopts a hybrid priority strategy that combines proportional priority and sorting priority, taking into account both sampling efficiency and deviation control.

[0013] Regarding the soft update of the target network, this invention employs the following method: A first decay coefficient is set to control the update rate of the target network parameters for the actor network; a second decay coefficient is set to control the update rate of the target network parameters for the critic network; in each training step, the result of multiplying the current network parameters of the actor network by one and subtracting the first decay coefficient is added to the result of multiplying the target network parameters of the actor network by the first decay coefficient, resulting in the updated target network parameters for the actor network; similarly, the result of multiplying the current network parameters of the critic network by one and subtracting the second decay coefficient is added to the result of multiplying the target network parameters of the critic network by the second decay coefficient, resulting in the updated target network parameters for the critic network. This slow, following parameter update strategy keeps the target network relatively stable, providing a consistent optimization direction for the current network.

[0014] In the specific design of the reward function in this invention, sound pressure levels at multiple locations inside the refrigerator are collected in real time, and their average value is calculated as the current noise level. The current noise level is compared with a preset noise target threshold, and the negative value of the difference between the two is calculated as a noise suppression effect score. Simultaneously, the real-time compressor speed is acquired, and the current energy consumption score is calculated based on a preset mapping relationship between compressor speed and energy consumption. Finally, the result of multiplying the noise suppression effect score by a first weighting coefficient and the energy consumption score by a second weighting coefficient is added to obtain the value of the reward function, and the sum of the first weighting coefficient and the second weighting coefficient remains constant. By weighting and compromising noise and energy consumption, the strategy learned by the agent can meet the user's quiet requirements while avoiding excessive compressor speed reduction leading to insufficient cooling or abnormally high energy consumption.

[0015] Preferably, during the process of applying the speed control signal output by the compressor noise suppression intelligent agent to the variable frequency drive board, the compressor noise suppression intelligent agent outputs the target speed value of the compressor, compares the target speed value with the current actual speed value of the compressor to obtain the speed deviation, and then inputs the speed deviation to the proportional-integral-derivative controller. The proportional-integral-derivative controller outputs the duty cycle of the pulse width modulation signal, and finally outputs the pulse width modulation signal to the variable frequency drive board to drive the frequency converter to generate three-phase AC power of the corresponding frequency, thereby realizing smooth and precise adjustment of the compressor speed.

[0016] As a further improvement to enhance training robustness, the method includes a step of monitoring the stability of the training process after executing the training compressor noise suppression agent: after each training round, the average reward value of the most recent fixed number of rounds is calculated; the reward value of the current training round is compared with the average reward value to calculate the reward fluctuation amplitude; when the reward fluctuation amplitude exceeds a preset fluctuation threshold for a preset number of consecutive times, the learning rate adjustment operation of the deep deterministic policy gradient network is triggered, multiplying the current learning rate by a decay factor between zero and one to obtain the updated learning rate; then training continues with the updated learning rate until the reward fluctuation amplitude stabilizes within the preset fluctuation threshold. This mechanism can promptly detect training divergence trends and automatically mitigate the update step size, ensuring stable policy convergence.

[0017] More specifically, the learning rate adjustment operation involves: comparing the reward fluctuation amplitude with a preset first fluctuation threshold and a second fluctuation threshold, where the first fluctuation threshold is less than the second fluctuation threshold; when the reward fluctuation amplitude is between the first and second fluctuation thresholds, multiplying the learning rate by a first decay factor; when the reward fluctuation amplitude is greater than the second fluctuation threshold, multiplying the learning rate by a second decay factor, where the second decay factor is less than the first decay factor; and simultaneously recording the timestamp and adjustment amplitude of each learning rate adjustment to generate a learning rate adjustment log, which facilitates the analysis and traceability of the training process.

[0018] As another preferred embodiment of the present invention, after deploying the trained compressor noise suppression agent to the refrigerator controller, the method further includes an online adaptive fine-tuning step for the deployed agent: during normal refrigerator operation, compressor noise and vibration signals are collected within a preset time window; it is determined whether the average sound pressure level of the compressor noise signal within the time window exceeds a preset noise comfort threshold; if so, the parameters of the bottom convolutional layers of the actor network in the compressor noise suppression agent are frozen, allowing only the fully connected layer parameters of the actor network to continue updating; the online training mode of the compressor noise suppression agent is started, using the data within the preset time window as training samples, and a preset number of gradient descent steps are executed to update the fully connected layer parameters, resulting in a fine-tuned compressor noise suppression agent. This online fine-tuning process enables the agent to track the noise characteristic drift caused by wear, condenser blockage, and other factors after long-term compressor operation, continuously maintaining good noise suppression performance, and avoiding damage to the already stably extracted bottom-level features.

[0019] The technical effects and advantages provided by the present invention in the above technical solution are as follows: The original vibration signal is reconstructed in phase space. The embedding dimension is determined using the spurious nearest neighbor method, and the time delay parameter is calculated using the mutual information method. Then, the geometric features of the chaotic attractor, such as the correlation dimension, of the reconstructed phase space are extracted. Recursion rate, determinism, and average diagonal length are obtained through recursive graph analysis, forming a compressor vibration state sequence. Chaotic feature analysis can capture the nonlinear dynamic state changes caused by friction, load fluctuations, and other factors during compressor operation. This state sequence is highly sensitive to subtle operating condition drifts and can more fundamentally reflect the instantaneous operating mode transitions of the compressor than traditional time-domain statistical features. On the noise side, a generative adversarial network noise separation model extracts deep features of the noise signal through the encoder-decoder structure of the generator. The reconstructed signal and the environmental background noise reference signal are fed together into the discriminator, forcing the generator to output a reconstructed signal that approximates the pure compressor body noise. A skip connection is introduced between the encoder and decoder to fuse shallow waveform details with deep semantic features. The reconstructed compressor body noise component can effectively remove interference from environmental reflections and noise from other components, providing noise characterization information unaffected by acoustic environmental pollution for subsequent control decisions. By combining the above compressor vibration state sequence with the time-frequency characteristics of the compressor body noise component as the environmental state, the reinforcement learning agent can simultaneously perceive the nonlinear vibration nature inside the compressor and the time-frequency structure of the body noise. This allows it to predict the potential impact of noise reduction operation on vibration energy during frequency conversion adjustment, resulting in more precise speed adjustment actions and avoiding erroneous adjustments due to incomplete input information.

[0020] A compressor noise suppression agent constructed using a deep deterministic policy gradient network separates the actor network and the critic network into a current network and a target network, respectively. The target network uses a soft update method to copy the parameters of the current network with a lag, preventing divergence in value estimation during training. During training, non-uniform sampling is performed from the experience pool based on the absolute value of the time difference error. The sampling probability is proportional to the absolute value of the time difference error, and a hybrid priority strategy combining proportional priority and ranking priority is adopted. This ensures that high-value samples are frequently replayed while avoiding overfitting to a single pattern caused by ranking priority under extreme errors, making the learning process smooth and efficient. The reward function consists of a weighted sum of noise suppression effect score and energy consumption score. The average sound pressure level at multiple locations inside the refrigerator is collected in real time and compared with the target threshold. At the same time, the energy cost is obtained according to the preset mapping relationship between compressor speed and energy consumption. This design allows the agent to automatically search for a speed control strategy that can suppress noise within a quiet and comfortable range without generating excessive cooling power consumption, achieving a dynamic balance between noise reduction and energy efficiency. After deployment, the target speed output by the intelligent agent is converted into a pulse width modulation signal by a proportional-integral-derivative controller to drive the frequency converter, thereby smoothly and steplessly regulating the speed of the compressor and effectively suppressing the additional vibration and noise impact caused by speed step jumps. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0022] Figure 1 This is a flowchart of a deep learning-based adaptive noise suppression method for refrigerator compressors. Figure 2 This is a flowchart of compressor vibration state sequence extraction based on chaotic feature analysis; Figure 3 This is a flowchart of compressor body noise component separation based on generative adversarial networks; Figure 4 It is a curve showing the change of compressor vibration state sequence recursion characteristics over time; Figure 5 It is the curve showing the changes in cumulative reward and average action value during the training of a deep deterministic policy gradient network; Figure 6 It is based on the reward function and score change curve of the compressor speed adjustment action; Figure 7 It is a monitoring curve of the training reward fluctuation amplitude of a deep deterministic policy gradient network. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] See Figure 1This invention provides a deep learning-based adaptive noise suppression method for refrigerator compressors, comprising: acquiring the original vibration signal and original noise signal during compressor operation; performing chaotic feature analysis on the original vibration signal to extract the compressor vibration state sequence; inputting the original noise signal into a pre-trained generative adversarial network noise separation model to separate the compressor body noise component; constructing a compressor noise suppression agent using a deep deterministic policy gradient network, taking the time-frequency features of the compressor vibration state sequence and the compressor body noise component as the environmental state input, taking the compressor speed adjustment command as the action output, and taking the weighted sum of noise suppression effect and energy consumption index as the reward function to train the compressor noise suppression agent; and deploying the trained compressor noise suppression agent into the refrigerator controller to output the compressor speed control signal to the inverter drive board in real time.

[0025] Example 1

[0026] See Figure 2 In specific implementation, the process of performing chaotic feature analysis on the original vibration signal and extracting the compressor vibration state sequence includes the following steps.

[0027] The original vibration signal undergoes phase space reconstruction to determine the embedding dimension and time delay parameters. The embedding dimension is determined using the false nearest neighbor method. One implementation of the false nearest neighbor method is as follows: starting with the smallest embedding dimension, such as 1, the one-dimensional time series of the original vibration signal is reconstructed into phase space. In the reconstructed phase space, the nearest neighbor is found for each phase point, and the rate of change of distance between two adjacent phase points is calculated as the embedding dimension increases. If the rate of change of distance exceeds a preset false nearest neighbor determination threshold, the nearest neighbor is marked as a false nearest neighbor. The proportion of false nearest neighbors at each embedding dimension is statistically analyzed. As the embedding dimension increases, when the proportion of false nearest neighbors decreases below a preset threshold, and this proportion no longer changes significantly with further increases in the embedding dimension, the embedding dimension at this point is determined as the final embedding dimension used for phase space reconstruction. The threshold for identifying false nearest neighbors is set to a range of 10 to 15 times the root mean square value of the original vibration signal, with a preset threshold of 5%. This is based on the fact that in chaotic time series analysis, when the proportion of false nearest neighbors is less than 5%, the trajectory of the dynamic system has been fully unfolded in phase space. The time delay parameter is calculated using the mutual information method. For the original vibration signal... Set time delay The range of values ​​for which the original vibration signal is divided into two sequences for each candidate time delay value is determined. and Calculate the joint probability distribution and marginal probability distribution of the two sequences, and then calculate the mutual information function. :

[0028] in, This represents the mutual information value between the original vibration signal and its delayed sequence. Indicates the original vibration signal at time [time]. amplitude, This indicates a delay. The amplitude of the signal after that, For the delay time variable, and They are respectively and marginal probability density, for and The joint probability density is given, and the summation symbol represents traversing all possible discrete amplitude intervals. Based on the mutual information function curve, the time delay corresponding to the first minimum value of the mutual information function is taken as the time delay parameter for phase space reconstruction. If the mutual information function has no obvious minimum value, then the time delay is taken when the mutual information function value decreases to its maximum value. The corresponding delay is used as the time delay parameter. The embedding dimension determined by the false nearest neighbor method and the time delay parameter determined by the mutual information method constitute the parameter set for phase space reconstruction.

[0029] Based on the embedding dimension and the time delay parameter, the correlation dimension of each phase point in the reconstructed phase space is calculated to obtain the chaotic attractor geometric features of the vibration signal. In a specific implementation, the determined embedding dimension is used... and time delay parameters For the original vibration signal sequence Phase space reconstruction is performed to obtain the phase point vector sequence in the reconstructed phase space. subscript Indicates the first There are n phase points, and the vector dimension is 1. The reconstructed phase vector is represented as... Select several scale distances. Calculate the correlation integral based on the value. The meaning of correlation integral is that the distance between phase points in the reconstructed phase space is less than the scaling distance. The proportion of the number of phase pairs to the total number of phase pairs can be estimated using the following formula:

[0030] in, The subscript indicates the total number of phase points in the reconstructed phase space. and subscript The phase point index is and satisfies , and They represent the first The and the first A vector of phase points, For scaling distance variables, For unit step function, symbol " "" represents the norm of a vector, either the 1-norm or the infinity norm. Distances on different scales After calculating the correlation integral, plot it on a logarithmic coordinate axis. about The curve is approximated by least-squares fitting of the approximate straight line segment. The slope of the fitted line is the correlation dimension. Correlation dimension As a geometric feature of the chaotic attractor, the local correlation dimension of each phase point is calculated in the above manner for all phase points or sampled phase points. Alternatively, the overall fitting slope can be used as the global correlation dimension, which can be used as the geometric feature of the chaotic attractor of the vibration signal.

[0031] Based on the geometric characteristics of the chaotic attractor, a recursive graph analysis method is used to extract the recursion rate, determinism, and average diagonal length as the compressor vibration state sequence. After obtaining the geometric characteristics of the chaotic attractor, the phase point vector sequence in the reconstructed phase space is used. Expand the recursive graph analysis. Set a distance threshold. Distance threshold The selection method can be based on the correlation integral. The distance is determined by a multiple of the distance required to reach a specified value, or it can be set as a certain percentage of the average distance between phase points based on the statistical distribution of the distances between phase points. For all phase point pairs... Construct a recursive matrix The rule for the elements of the recursive matrix is: when the phase point... With phase point The distance between them is less than or equal to the distance threshold. When the recursive matrix is ​​in its normal state, the element value is 1; otherwise, it is 0.

[0032] in, For unit step function, The norm is represented. The visualization of a recurrence graph is a two-dimensional representation of the recurrence matrix, but in the compressor's automatic analysis, quantitative features are directly calculated based on the recurrence matrix. Based on the recurrence matrix, the recurrence rate is calculated, defined as the proportion of elements with a value of 1 in the recurrence matrix out of the total number of elements.

[0033] Determinism is calculated, defined as the proportion of recursive points that form a diagonal structure with a length exceeding a preset minimum diagonal length out of the total number of recursive points in the recursion matrix. First, the diagonal length distribution in the recursion matrix is ​​extracted. For each diagonal length... The statistical length is Frequency of diagonal structures The diagonal structure corresponds to a sequence of consecutive non-zero recursive matrix elements. The deterministic calculation formula is:

[0034] in, The minimum diagonal length is preset, typically set to 2, to exclude pseudo-diagonal structures formed by isolated recursive points. The average diagonal length is calculated; the average diagonal length is the length of all elements satisfying the condition that is not less than... Weighted average length of the diagonal structure:

[0035] The calculated recursion rate, determinism, and average diagonal length are combined sequentially to form a sequence characterizing the compressor's vibration state, i.e., the compressor vibration state sequence. During the actual operation of the refrigerator, for continuously acquired raw vibration signals, a sliding time window is used for the aforementioned chaotic feature analysis. Within each time window, a compressor vibration state sequence consisting of the recursion rate, determinism, and average diagonal length is generated. Adjacent time windows may retain partial overlap to ensure the temporal continuity of the sequence. The compressor vibration state sequence serves as part of the environmental state input for the subsequent compressor noise suppression agent, used to sense the dynamic changes in compressor operation.

[0036] See Figure 4 In the figure, the horizontal axis represents the time window number, and the vertical axis represents the corresponding chaotic characteristic value. The curves in the figure represent the three chaotic characteristics of the compressor vibration state sequence: recursion rate, determinism, and average diagonal length.

[0037] The recurrence rate curve is represented by solid dot marks. Its value remains between approximately 0.25 and 0.35, showing slight fluctuations. This indicates that the recurrence frequency of phase points in the reconstructed phase space is stable, reflecting that the similarity distribution of the compressor vibration signal is relatively uniform and without drastic changes.

[0038] The deterministic curve, marked by dashed squares, maintains a value between approximately 0.65 and 0.75, exhibiting a slight fluctuation trend. This indicates that the proportion of diagonal structures in the recursive graph is stable, suggesting that the vibration dynamics exhibit strong determinism and regularity, and the vibration state sequence is highly predictable.

[0039] The average diagonal length curve is represented by dotted triangles. Its characteristic value fluctuates between 11.5 and 13.0. The curve shows a slight upward trend, especially in the time window number range of 400 to 500, indicating that the average length of the diagonal structure in the recursive graph has increased, reflecting that the dynamic structure of the vibration signal tends to be more continuous and stable.

[0040] Overall, none of the three curves showed drastic fluctuations or abrupt changes, indicating that the phase space reconstruction parameters determined by the spurious nearest neighbor method and mutual information method can stably extract the chaotic features of the compressor vibration signal. The recursion rate and determinism remained relatively stable, while the average diagonal length slightly increased, reflecting a relatively stable and somewhat regular dynamic evolution of the compressor vibration state within the monitoring time range. These chaotic feature sequences provide effective environmental state input information for the subsequent compressor noise suppression agent.

[0041] Example 2

[0042] See Figure 3 In practice, the process of inputting the original noise signal into a pre-trained generative adversarial network noise separation model and separating the compressor body noise component includes the following steps.

[0043] The original noise signal is processed by framing and windowing to obtain multiple noise signal frames. The original noise signal is a discrete-time signal sequence. The frame length and frame shift are determined based on the frequency domain characteristics of the refrigerator compressor noise. With a sampling rate of 16kHz, a frame length of 1024 points is selected, corresponding to a time length of 64 milliseconds, which meets the resolution requirements for the low-frequency and mid-frequency components of the compressor. The frame shift is set to 512 points, meaning there is a 50% overlap between adjacent frames to ensure signal continuity between frames. The windowing process uses a Hamming window function, which is multiplied point-by-point with each frame signal to reduce spectral leakage. The expression for the Hamming window is:

[0044] in, Indicates frame length, Indicates the intra-frame sample number. Indicates the Hamming window in the serial number The weights at each position are determined. After framing and windowing, a sequence of noise signal frames indexed by frame number is obtained, with each frame serving as a processing unit.

[0045] Each frame of noise signal is input into the generator network of the Generative Adversarial Network (GAN) noise separation model. The overall architecture of the generator network is an encoder-decoder structure. The encoder consists of multiple stacked encoder units, each containing a convolutional layer and a downsampling layer. The convolutional layer extracts local time-frequency patterns by filtering the input feature map. The kernel size, stride, and number of output channels of the convolutional layer vary layer by layer according to the dimension of the time-domain signal. The convolutional layer of the first encoder unit processes the input single-channel noise signal frame and outputs a multi-channel feature map containing shallow features; the convolutional layers of subsequent encoder units extract deeper features with a larger receptive field. The downsampling layer in each encoder unit downsamples the feature map output by the convolutional layer. The downsampling operation is implemented using a convolution with a stride of 2, which halves the time dimension of the feature map layer by layer, prompting the network to focus on the abstract representation of the signal. The decoder consists of multiple stacked decoder units, each containing an upsampling layer and a deconvolutional layer. The upsampling layer restores the temporal dimension to twice its original value by performing nearest-neighbor interpolation or subpixel convolution on the input feature map. Subsequently, the deconvolution layer performs convolution operations to gradually reconstruct the detailed components of the signal. The final deconvolution layer of the decoder outputs a single-channel signal, which is the reconstructed signal of the same length as the input noise signal frame.

[0046] Skip connections are introduced between the encoder and decoder of the generator network. Specifically, the feature map output by each encoder unit is adjusted in the temporal dimension to have the same length as the input feature map of the corresponding decoder unit by copying or truncating it. Then, the two are concatenated in the channel dimension, so that the number of input channels of the deconvolution layer of the decoder unit is the sum of the number of channels of the original decoder feature map and the number of channels of the encoder feature map passed by the skip connection. Through skip connections, the decoder integrates the shallow temporal fine-structure information and deep semantic information of the encoder during reconstruction, thereby improving the ability of the reconstructed signal to express the noise components of the compressor itself.

[0047] The reconstructed signal output from the generator network and the acquired ambient background noise reference signal are simultaneously input into the discriminator network of the generative adversarial network noise separation model. The ambient background noise reference signal is acquired by using the same acoustic sensor at the same location during the refrigerator compressor's shutdown period to collect a signal superimposed from other noise sources in the refrigerator (such as fan and refrigerant flow noise). This signal serves as the ambient noise benchmark, in addition to the compressor's own noise. The reconstructed signal and the ambient background noise reference signal are concatenated along the channel dimension to form a two-channel input data, which is then fed into the discriminator network. The discriminator network consists of multiple stacked discriminator convolutional layers. Each discriminator convolutional layer performs stride convolution on the input to compress the time dimension and gradually increases the number of channels. The feature map output from the last discriminator convolutional layer is then global average pooled and fed into a fully connected layer. The output of the fully connected layer is mapped to a scalar value between 0 and 1 using a sigmoid activation function. This scalar value represents the probability that the reconstructed signal contains compressor noise. The closer the probability value is to 1, the more the discriminator network identifies the reconstructed signal as being consistent with the characteristics of the compressor's own noise; the closer the probability value is to 0, the closer the discriminator network identifies the reconstructed signal as being closer to the ambient background noise or other non-compressor noise components.

[0048] The parameters of the generator network are adjusted inversely based on the probability output of the discriminator network until the discriminator network determines that the reconstructed signal is the compressor body noise signal, thus obtaining the compressor body noise component. During the separation phase, the parameters of the discriminator network in the generative adversarial network noise separation model remain fixed, and only the parameters of the generator network are updated. For the currently processed frame of noise signal, the parameter update loss of the generator network is defined as:

[0049] in, This represents the current input frame of noise signal. Indicates that the generator network is for The output reconstructed signal, This represents the ambient background noise reference signal. This represents the probability output by the discriminator network when the input is the reconstructed signal and the ambient background noise reference signal. This represents the loss value of the generator network. (Symbol) This represents the forward computation function of the discriminator network. Minimize the loss. This is equivalent to maximizing the probability that the discriminator network identifies the reconstructed signal as compressor noise. The generator network's parameters are updated using gradient descent, with Adam as the optimizer and a learning rate of 0.0001. For each incoming noise signal frame, a fixed number of iterations are performed, e.g., 5 parameter update steps, with the same signal frame used for each update. and ambient background noise reference signal Calculate the gradient. The iteration terminates when the probability output by the discriminator network exceeds a preset threshold (set to 0.9), at which point the reconstructed signal is considered successfully separated as a compressor body noise component. Alternatively, the iteration terminates after reaching the maximum number of iterations, even if the probability does not reach the threshold, and the final reconstructed signal is output as the compressor body noise component corresponding to that frame. Perform the above operation on each frame of noise signal to obtain the compressor body noise component sequence for all frames. The frame sequences are then synthesized into a complete temporal compressor body noise signal component using an overlap-addition method.

[0050] Example 3

[0051] In practice, the process of constructing a compressor noise suppression agent using a deep deterministic policy gradient network includes the following steps.

[0052] An actor network is constructed to map the time-frequency features of the input compressor vibration state sequence and compressor body noise components to the mean value of the compressor speed adjustment actions. The input layer of the actor network has a receiving dimension of [missing information]. The environmental state vector is composed of the current and historical values ​​of the compressor vibration state sequence and the time-frequency characteristics of the compressor body noise component. The compressor vibration state sequence contains three features: recursion rate, determinism, and average diagonal length. The values ​​at the current time and the two previous time points are extracted, forming a total of 9-dimensional features. The time-frequency characteristics of the compressor body noise component are obtained by performing a short-time Fourier transform on the compressor body noise component, extracting the Mel frequency cepstral coefficients and their first and second order differences. After mean normalization, the first 13 coefficients are selected to form a 13-dimensional feature. The two are then concatenated to obtain the final environmental state vector. A 3D environment state vector. The actor network consists of two fully connected layers. The first fully connected layer has 256 neurons and uses a linear rectified function (RCF) for activation. The second fully connected layer has 128 neurons and also uses an RCF for activation. The output of the second fully connected layer is connected to the output layer, which contains one neuron and uses a hyperbolic tangent activation function. The output value ranges from 1 to 1. The average compressor speed adjustment action is obtained by multiplying the output of the hyperbolic tangent activation function by half the speed adjustment range and then summing it with the median of the speed adjustment range. The speed adjustment range of the compressor speed adjustment action is preset according to the actual operating range of the compressor. revolutions per minute, therefore the mean is from Given, among which The input is a linear combination of the output layer neurons. Indicates the environmental state The mean of the output is used. During the training phase, exploration noise is superimposed on the mean. The exploration noise is generated using an Ornstein-Uhlenbeck process with parameters set to an initial value of 0, a regression rate of 0.15, and a volatility of 0.2, producing time-dependent random noise. The superposition of these noises yields the actual compressor speed adjustment actions. ,in This is random noise. During the deployment phase, the actor network only outputs the mean, without adding exploration noise.

[0053] A critic network is constructed to evaluate the action value of performing a compressor speed adjustment action under given environmental conditions. The critic network contains two input paths; the environmental state input path has a dimension of [missing information]. The environment state vector is input into a fully connected layer with 256 neurons and a rectified linear function (CLM) activation function. The action input path receives the action scalar output from the actor network or the action values ​​stored in the experience pool, and maps them to a 128-dimensional feature space through a fully connected layer with a CLM activation function. The fused feature vector is then fed into another fully connected layer with 128 neurons and a linear rectified function as the activation function. The output of this fully connected layer is connected to an output layer, which contains a single neuron that outputs a scalar action value without using an activation function. , Indicates the environmental state Next action The expected long-term cumulative rewards.

[0054] The current networks of both the actor network and the critic network are separated from the target network. The current network of the actor network is represented as follows: ,in The parameter set of the current network represents the actor network; the target network of the actor network is represented as... , The parameter set representing the actor network and the target network. The critic network is represented as the current network. , The parameter set of the current network represents the critic network; the target network of the critic network is represented as... , This represents the parameter set of the critic network and the target network. The parameters of the actor network (current network) and the critic network (current network) are initialized orthogonally, while the parameters of the target network are directly copied from the corresponding parameters of the current network during initialization.

[0055] The target network parameters are copied from the current network parameters using a soft update method with a lag. A first attenuation coefficient is set. This is used to control the parameter update rate of the target network in the actor network; a second attenuation coefficient is set. This is used to control the parameter update rate of the target network in the commentator network. In each training step, the current network parameters of the actor network are multiplied by one and subtracted from a first decay coefficient, then added to the target network parameters of the actor network multiplied by the first decay coefficient to obtain the updated actor network target network parameters; similarly, the current network parameters of the commentator network are multiplied by one and subtracted from a second decay coefficient, then added to the target network parameters of the commentator network multiplied by the second decay coefficient to obtain the updated commentator network target network parameters. The first decay coefficient... Second attenuation coefficient All values ​​are set to a constant of 0.005. This value is selected based on the stable convergence experience of deep deterministic policy gradient networks in continuous control tasks, so that the target network parameters gradually approach the current network parameters with a small step size, thus maintaining training stability.

[0056] During the training of the critic network, a priority experience replay mechanism is introduced. Non-uniform sampling is performed from the experience pool based on the absolute value of the temporal difference error, allowing higher-priority experience samples to be learned more frequently. The experience samples stored in the experience pool are tuples. subscript This is a globally unique index for the empirical sample. For the first The environmental state corresponding to each experience sample For the first Actions performed in a sample of experience For the first The instant reward corresponding to each experience sample The next environmental state to transition to after performing an action. A Boolean value, indicating Is it in a terminated state? Time difference error. The calculation method is as follows:

[0057] in, This represents the discount factor, set to 0.99, signifying the importance of future rewards in the current decision-making process. For the current state of the critic network and actions The value of the output action; For the actor network target network, according to the next state The target action to be output; The target action value output by the critic network for the next state and target action. The absolute value of the temporal difference error. This indicates the learning value of the empirical sample for the current network parameters.

[0058] The sampling probability of empirical samples is proportional to the absolute value of the time difference error, and a hybrid priority strategy combining proportional priority and ranking priority is adopted. Proportional priority... ,in For a value The positive constant is used to ensure that all empirical samples have a non-zero sampling probability, avoiding permanent exclusion from sampling due to zero time difference error. The sorting priority is based on... The descending order is determined if the empirical samples Its position in the sorting is The sorting priority is defined as follows: The mixed priority is calculated as follows:

[0059] in, This represents the mixed weighting coefficient, with a value range of [value range missing]. The value was set to 0.7, based on a balanced approach selected after testing performance in steps of 0.1 between 0.5 and 0.9 using a grid search. This balances the need for enhanced attention to high-error samples with the smoothing effect of ranking priority on extreme values. This is a mixed priority across all empirical samples. After obtaining the data, calculate the sampling probability of each empirical sample. :

[0060] in, Indicates the total number of experience samples currently stored in the experience pool, subscript. This is the traversal index for all samples in the experience pool. This is a priority index used to adjust the strength of the impact of priorities. Priority Index The range of values ​​is ,when Degenerates into uniform sampling when Sampling is strictly followed according to priority; settings Based on the initial Based on this, the probability increases linearly with the number of training steps until it reaches 0.6 and then remains constant, allowing for a focus on sample diversity in the early stages of training and a focus on high-priority samples in the later stages. This is based on sampling probability. Sampling with replacement is performed, and the probability of each empirical sample being selected is related to its... Proportional.

[0061] See Figure 5In the graph, the horizontal axis represents the number of training rounds, ranging from 0 to approximately 520 rounds, while the vertical axis represents the numerical range, covering approximately -20 to 140. The solid curve in the graph represents the trend of cumulative reward value, and the dashed curve represents the trend of average action value Q. Overall, both cumulative reward and average action value Q show a process of initially rising rapidly and then tending to fluctuate and stabilize.

[0062] In the early stages of training (approximately 0 to 50 rounds), both the cumulative reward and the average action value Q rapidly increased from low levels. The cumulative reward started from near zero or even negative values ​​and gradually climbed to around 60, while the average action value Q also increased synchronously. This indicates that the compressor noise suppression agent of the deep deterministic policy gradient network significantly improved policy performance through continuous learning, and the long-term reward expectation of action selection was enhanced.

[0063] In the mid-term (approximately 50 to 250 rounds), both curves continue to show a slow upward trend, but the fluctuation range increases significantly. The cumulative reward fluctuates between 60 and 110, and the average action value Q also fluctuates in a similar range, indicating that the training process has entered the strategy refinement stage. The agent seeks a better balance between noise suppression effect and energy consumption trade-off. There is a certain degree of exploration and adjustment in the learning process, resulting in short-term fluctuations in reward and action value.

[0064] In the later stages (approximately 250 to 520 rounds), the cumulative reward and average action value Q curves tend to stabilize, with values ​​oscillating within the range of 80 to 120 without a significant downward trend. This indicates that the strategy has converged to a better state, and the agent can effectively adapt to the environmental input and output reasonable compressor speed adjustment actions to achieve the goal of noise adaptive suppression.

[0065] Overall, this figure reflects the training process of the compressor noise suppression agent based on a deep deterministic policy gradient network as described in Example 3. The effectiveness and stability of the training strategy are verified through the simultaneous increase and stable fluctuation of cumulative rewards and average action values. The similar trends in reward values ​​and action values ​​during training further illustrate the accuracy of the critic network's assessment of the actor network's action quality, providing strong support for agent strategy optimization.

[0066] Example 4

[0067] In practice, the process of using the weighted sum of noise suppression effect and energy consumption index as the reward function includes the following steps.

[0068] The sound pressure level (SPL) at multiple locations inside the refrigerator is collected in real time, and the average of all SPL levels is calculated as the current noise level. Omnidirectional microphones are installed at several pre-set acoustic sampling locations inside the refrigerator. Each microphone synchronously collects SPL signals at a consistent sampling frequency. After A-weighted filtering, the SPL level at each location is calculated in decibels (dB). The arithmetic mean of the SPL values ​​collected at all locations at the same time is taken to obtain the average SPL inside the refrigerator at that moment, denoted as _____. , as the current noise level.

[0069] The current noise level is compared with a preset noise target threshold, the difference between the two is calculated, and the negative value of the difference is used as the noise suppression effect score. Noise target threshold Based on the requirement for quiet operation of the refrigerator, the noise level was predetermined at 28 decibels. This value was determined according to the comfortable perception boundary of compressor noise for the human ear in a quiet environment. Noise suppression effect score. The calculation method is as follows When the current noise level is higher than the noise target threshold, the difference is positive and the noise suppression effect score is negative, indicating a penalty; when the current noise level is lower than the noise target threshold, the noise suppression effect score is positive, indicating a reward.

[0070] The system acquires the real-time compressor speed and calculates the current energy consumption score based on a preset mapping relationship between compressor speed and energy consumption. This mapping relationship is established through offline calibration experiments, measuring the compressor's electrical power under different speed conditions and generating a lookup table or piecewise function after smoothing and fitting. During refrigerator operation, the compressor speed value fed back from the inverter drive board is read in real time and input into the mapping relationship to obtain the corresponding energy consumption value. To match the noise suppression effect score in terms of dimensions and numerical range, the energy consumption value is normalized: subtracting the baseline energy consumption from the energy consumption value, dividing by the energy consumption scale coefficient, and then multiplying by negative one yields the current energy consumption score. The baseline energy consumption is set as the steady-state power of the compressor at its minimum allowable operating speed, and the energy consumption scale factor is set as the power difference between the compressor at its maximum and minimum allowable operating speeds. Therefore, when the compressor speed is low and energy consumption is low, the energy consumption score is high; when the compressor speed increases and energy consumption increases, the energy consumption score decreases.

[0071] The reward function is obtained by adding the noise suppression effect score multiplied by the first weighting coefficient and the energy consumption score multiplied by the second weighting coefficient. The first weighting coefficient is denoted as... The second weighting coefficient is denoted as The sum of the first weighting coefficient and the second weighting coefficient is a constant. It is set to 1.0. The value is set to 0.7. The value of is set to 0.3. This setting is based on a trade-off point obtained through Pareto front analysis during the algorithm simulation phase, ensuring that the overall reward prioritizes noise suppression effectiveness while also considering energy consumption performance. The formula for calculating the reward function is:

[0072] in, This represents the value of the reward function. This represents the first weighting coefficient. The score indicates the noise suppression effect. This represents the second weighting coefficient. This represents the current energy consumption score. Through this reward function, the deep deterministic policy gradient network automatically learns a speed adjustment strategy that achieves a reasonable balance between noise level and energy consumption during training, with the goal of maximizing cumulative reward.

[0073] In practice, the process of outputting the compressor speed control signal to the frequency converter drive board in real time includes the following steps.

[0074] The compressor noise suppression agent outputs the target compressor speed value. The actor network deployed in the refrigerator controller calculates the value of the output neuron through forward propagation based on the real-time input environmental state vector. After scaling and offset processing, the target compressor speed value is obtained, which is limited to the upper and lower limits of the compressor's allowable safe operating speed.

[0075] The target compressor speed is compared with the current actual compressor speed to calculate the speed deviation. The current actual compressor speed is provided in real time by the inverter drive board through the speed feedback interface. The speed deviation is obtained by subtracting the current actual compressor speed from the target compressor speed. The speed deviation is signed; a positive value indicates that the speed needs to be increased, and a negative value indicates that the speed needs to be decreased.

[0076] The speed deviation is input to a proportional-integral-derivative (PID) controller, which outputs the duty cycle of a pulse-width modulation (PWM) signal. The PID controller internally stores the proportional gain coefficient. Integral gain coefficient and differential gain coefficient These three gain coefficients were tuned using a frequency domain correction method based on the transfer function characteristics of the compressor and the variable frequency drive system. Set to 0.8, Set to 0.05, Set to 0.02. The proportional-integral-derivative controller uses the speed deviation and its accumulated integral and the current rate of change to calculate the duty cycle adjustment by weighting and summing the proportional, integral and derivative terms respectively. Then, the duty cycle adjustment is superimposed on the current duty cycle base value to obtain the updated pulse width modulation signal duty cycle. The duty cycle value is limited to the range of 10% to 90%.

[0077] The pulse width modulation (PWM) signal is output to the inverter drive board to drive the inverter to generate three-phase AC power of the corresponding frequency to adjust the compressor speed. After receiving the PWM signal, the inverter drive board converts the DC bus voltage into an equivalent DC voltage corresponding to the duty cycle through the power drive circuit. Then, the space vector pulse width modulation algorithm controls the inverter bridge to generate three-phase AC power of the required frequency and amplitude, driving the compressor motor to smoothly transition to the new speed, thereby completing one noise adaptive suppression control cycle.

[0078] See Figure 6 The horizontal axis of the graph represents the compressor speed adjustment action, in revolutions per minute (rpm), ranging from -320 to 320. The vertical axis represents the corresponding score values, including the noise suppression effect score, the current energy consumption score, and the overall bonus value. The dashed curve represents the noise suppression effect score, which generally shows a decreasing trend as the speed adjustment action changes from negative (deceleration) to positive (increase). Initially, the noise suppression effect score fluctuates between 5 and 6, then stabilizes and drops below -5, reflecting that the positive contribution of speed reduction to noise suppression gradually weakens as the speed increases, eventually leading to a penalty. The dotted-line curve represents the current energy consumption score, which shows a relatively flat trend with a slight decrease, and the score value remains basically around 0, indicating that the current energy consumption score has low sensitivity to speed adjustment actions and the energy consumption change is small. The solid curve represents the overall reward value. This curve closely follows the changing trend of the noise suppression effect score. Initially, the reward value is approximately 4.5. As the speed adjustment action gradually increases from negative to positive, the reward value approaches 0 and continues to change towards negative values. This indicates that the noise suppression effect has a high weight in the reward function, while the energy consumption score has a relatively small impact on the reward. Overall, this figure intuitively illustrates the trade-off between the compressor speed adjustment action and the noise suppression effect and energy consumption index based on the reward function defined in Example 4. It shows that a higher noise suppression effect and a better overall reward can be achieved within the negative speed adjustment range, while the positive speed adjustment action gradually reduces the reward value, suggesting that the system tends to take speed reduction measures to achieve an optimal balance between adaptive noise suppression and energy consumption control.

[0079] Example 5

[0080] In practice, stability monitoring is performed on the training process of the deep deterministic policy gradient network. After each training epoch, the average reward value of the most recent fixed number of epochs is calculated. The fixed number of epochs is set to 10 epochs. The selection of 10 epochs is based on the fact that in the refrigerator compressor noise suppression task, the duration of a single training epoch is approximately 30 to 60 seconds. 10 epochs correspond to a time span of 5 to 10 minutes, which can smooth out short-term random fluctuations and reflect changes in training trends in a timely manner. The arithmetic mean of the reward values ​​of the most recent 10 epochs is calculated and denoted as . Get the reward value for the current training round. ,Will and The comparison is performed to calculate the reward fluctuation range. The reward fluctuation range quantifies the degree to which the current round's reward value deviates from the recent average level, and is calculated as follows:

[0081] in, Indicates the fluctuation range of rewards. This represents the reward value for the current training round. This represents the average reward value over the most recent fixed number of rounds. To prevent zero constant, set to This is used to avoid calculation overflow caused by a zero denominator when the average reward value is close to zero.

[0082] When the reward fluctuation exceeds a preset fluctuation threshold a certain number of times consecutively, the learning rate adjustment operation of the deep deterministic policy gradient network is triggered. The preset number of times is set to 3. Exceeding the preset fluctuation threshold three times consecutively indicates that the reward fluctuation is not accidental and that there is continuous instability in the training process. The preset fluctuation threshold includes a first fluctuation threshold and a second fluctuation threshold. The first fluctuation threshold is set to 0.15, and the second fluctuation threshold is set to 0.30. The values ​​of the first fluctuation threshold of 0.15 and the second fluctuation threshold of 0.30 are obtained through statistical analysis of the reward fluctuation amplitude during the stable training phase of the deep deterministic policy gradient network in the compressor noise suppression task. The reward fluctuation amplitude during the stable phase is usually no more than 0.1. 0.15 is set as the upper limit of one and a half times, and 0.30 is set as the upper limit of three times, corresponding to slight fluctuation and severe fluctuation states, respectively.

[0083] The reward fluctuation amplitude is compared with preset first fluctuation thresholds and preset second fluctuation thresholds. When the reward fluctuation amplitude is between the first and second fluctuation thresholds, the learning rate of the deep deterministic policy gradient network is multiplied by a first decay factor. The first decay factor is set to 0.5, indicating that the learning rate is halved, reducing the parameter update step size and enhancing training stability. When the reward fluctuation amplitude is greater than the second fluctuation threshold, the learning rate is multiplied by a second decay factor, which is set to 0.2, smaller than the first decay factor, reflecting a larger decay of the learning rate to cope with drastic instability. The second decay factor is smaller than the first decay factor, causing the network to update parameters at a more conservative pace when fluctuations are severe. The learning rate of the deep deterministic policy gradient network includes the actor network learning rate. and the network learning rate of critics Both were initially set to When adjusting the learning rate, the same decay multiplication operation is performed simultaneously on both learning rates.

[0084] The timestamp and adjustment magnitude of each learning rate adjustment are recorded to generate a learning rate adjustment log. The timestamp records the specific training round number in which the learning rate adjustment occurred. The adjustment magnitude records the ratio of the learning rate before adjustment to the learning rate after adjustment, or the specific value of the adjustment factor. The learning rate adjustment log is stored in the training controller's storage unit in the form of structured data. Each log entry includes the training round number, the reward fluctuation magnitude value at the time of triggering, the trigger threshold level, the decay factor used, the actor network learning rate before adjustment, the actor network learning rate after adjustment, the commentator network learning rate before adjustment, and the commentator network learning rate after adjustment.

[0085] Continue training the deep deterministic policy gradient network with the updated learning rate until the reward fluctuation stabilizes within a preset fluctuation threshold. The criterion for stabilization within the preset fluctuation threshold is that the reward fluctuation is lower than the first fluctuation threshold of 0.15 for 10 consecutive rounds. Once this condition is met, the learning rate will no longer decay, and training can proceed by gradually restoring the learning rate to its initial value or maintaining the current learning rate.

[0086] After deploying the trained deep deterministic policy gradient network (i.e., the compressor noise suppression agent) into the refrigerator controller, the deployed compressor noise suppression agent undergoes online adaptive fine-tuning. During normal refrigerator operation, compressor noise and vibration signals are collected within a preset time window. The preset time window is set to 5 minutes, based on the fact that the refrigerator compressor's operating cycle typically includes multiple states such as start-up, stable operation, and shutdown. Five minutes covers at least one complete operating phase, ensuring that the collected samples contain sufficient operational status information. Compressor noise signals are collected by a microphone deployed inside the refrigerator, and vibration signals are collected by an accelerometer mounted on the compressor housing.

[0087] The system determines whether the average sound pressure level of the compressor noise signal exceeds a preset noise comfort threshold. The compressor noise signal collected within a preset time window is divided into frames and windowed. The A-weighted sound pressure level of each frame is calculated, and then the arithmetic mean of the sound pressure levels of all frames is taken to obtain the average sound pressure level. The noise comfort threshold is set to 35 dB, which is the upper limit of the background noise level in a living room environment. Exceeding this level will make the compressor noise noticeably perceptible and affect the user experience.

[0088] When the average sound pressure level exceeds the noise comfort threshold, it indicates a deviation between the current refrigerator operating environment or compressor status and the training data distribution, requiring online fine-tuning of the compressor noise suppression agent. During fine-tuning, the parameters of the bottom convolutional layers in the actor network of the compressor noise suppression agent are frozen, allowing only the fully connected layer parameters to continue updating. The actor network's specific structure includes bottom convolutional layers and upper fully connected layers. The bottom convolutional layers are used to extract local time-frequency patterns by performing two-dimensional convolution operations on the time-frequency features of the input compressor vibration state sequence and compressor body noise components. The fully connected layers are used to flatten the feature maps output by the convolutional layers and map them to the action space. Freezing the bottom convolutional layer parameters means that the weights and biases of these convolutional layers remain unchanged during fine-tuning; only the weights and biases of the fully connected layers are updated with gradients. This is done to preserve the general time-frequency feature extraction capability learned during pre-training, adjusting only the decision mapping part oriented towards specific environmental conditions, and preventing catastrophic forgetting.

[0089] The online training mode of the compressor noise suppression agent is initiated, using data within a preset time window as training samples. Vibration and noise signals collected within the preset time window are processed using the same preprocessing steps as in the training phase to extract environmental states, construct experience samples, and store them in a temporary fine-tuning experience buffer, without entering the main experience pool. A preset number of gradient descent steps are executed, set to 5 steps. The 5-step gradient descent step is chosen because online fine-tuning needs to quickly adapt to the new environment, but should avoid excessive modification of network parameters leading to instability; 5 steps provide adequate correction space. In each gradient descent step, a batch of experience samples is randomly sampled from the fine-tuning experience buffer, with a batch size of 16. The loss function of the critic network is calculated, and the policy gradient of the fully connected layer of the actor network is calculated. The parameters of the fully connected layer of the actor network are updated using the Adam optimizer, with the learner rate set to... The learning rate is set to half of the original rate to ensure careful parameter updates during fine-tuning. After completing the preset 5 gradient descent steps, the online training mode is turned off, resulting in the fine-tuned compressor noise suppression agent. The fine-tuned compressor noise suppression agent continues to receive real-time environmental data and outputs compressor speed control signals.

[0090] See Figure 7 In the graph, the horizontal axis represents the number of training rounds, ranging from 0 to 520, and the vertical axis represents the reward fluctuation range, ranging from 0 to 0.40. The solid line curve represents the change in reward fluctuation range for each training round. The curve shows a periodic fluctuation trend, with the fluctuation range varying between 0.05 and 0.35, indicating that the reward value fluctuates to some extent relative to the recent average level during training. The dashed line in the graph represents the first fluctuation threshold, with a value of 0.15, as the criterion for judging slight fluctuation; the dotted line represents the second fluctuation threshold, with a value of 0.30, as the criterion for judging severe fluctuation.

[0091] Observing the solid curve, when the reward fluctuation amplitude continuously exceeds the first fluctuation threshold of 0.15, it indicates that the training has a certain degree of instability, which may trigger an adjustment strategy of halving the learning rate. When the reward fluctuation amplitude exceeds the second fluctuation threshold of 0.30, it indicates that the training has experienced large fluctuations, and the system may adopt a larger learning rate decay (multiplied by 0.2) to restore training stability. The figure shows multiple instances where the reward fluctuation amplitude exceeds the first fluctuation threshold, but only a few instances exceed the second fluctuation threshold, indicating that the training process is in a slight fluctuation phase for most of the time, and the overall stability is good.

[0092] This figure illustrates the specific performance of stability monitoring during training of the deep deterministic policy gradient network in Example 5. By comparing the reward fluctuation amplitude with two preset thresholds, the stability of the training process is dynamically judged, and a corresponding learning rate adjustment mechanism is triggered based on the number of consecutive times the threshold is exceeded to ensure stable convergence of the training process. The periodic fluctuations of the curve also indicate that the agent's policy updates involve reasonable exploration and adjustment during consecutive training rounds, without continuous violent oscillations, thus verifying the effectiveness of the stability monitoring and adaptive learning rate adjustment method.

[0093] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A deep learning-based adaptive noise suppression method for refrigerator compressors, characterized in that, The method includes: Acquire the raw vibration and noise signals of the compressor during operation; Chaotic feature analysis was performed on the original vibration signal to extract the compressor vibration state sequence; The original noise signal is input into a pre-trained generative adversarial network noise separation model to separate the compressor body noise component; A compressor noise suppression agent is constructed using a deep deterministic policy gradient network. The time-frequency characteristics of the compressor vibration state sequence and the compressor body noise components are used as environmental state inputs, the compressor speed adjustment command is used as the action output, and the weighted sum of noise suppression effect and energy consumption index is used as the reward function to train the compressor noise suppression agent. The trained compressor noise suppression agent is deployed to the refrigerator controller, and the compressor speed control signal is output to the frequency converter drive board in real time.

2. The deep learning-based adaptive noise suppression method for refrigerator compressors according to claim 1, characterized in that, The process of performing chaotic feature analysis on the original vibration signal to extract the compressor vibration state sequence is as follows: The original vibration signal is subjected to phase space reconstruction processing to determine the embedding dimension and time delay parameters; Based on the embedding dimension and time delay parameter, the correlation dimension of each phase point in the reconstructed phase space is calculated to obtain the chaotic attractor geometric features of the vibration signal; Based on the geometric characteristics of the chaotic attractor, a recursive graph analysis method is used to extract the recursion rate, determinism, and average diagonal length as the vibration state sequence of the compressor.

3. The deep learning-based adaptive noise suppression method for refrigerator compressors according to claim 2, characterized in that, In the phase space reconstruction process, the embedding dimension is determined using the spurious nearest neighbor method, and the time delay parameter is calculated using the mutual information method.

4. The deep learning-based adaptive noise suppression method for refrigerator compressors according to claim 1, characterized in that, The step of inputting the original noise signal into a pre-trained generative adversarial network noise separation model to separate the compressor's inherent noise components specifically involves: The original noise signal is divided into frames and windowed to obtain multiple noise signal frames. Each frame of noise signal is input into the generator network of the generative adversarial network noise separation model. The generator network adopts an encoder-decoder structure, wherein the encoder extracts deep features of the noise signal through convolutional layers and downsampling layers, and the decoder reconstructs the signal through deconvolutional layers and upsampling layers. The reconstructed signal output by the generator network and the collected environmental background noise reference signal are simultaneously input into the discriminator network of the generative adversarial network noise separation model. The discriminator network outputs the probability that the reconstructed signal contains compressor body noise, and the parameters of the generator network are adjusted in reverse based on this probability until the discriminator network determines that the reconstructed signal is a compressor body noise signal, thus obtaining the compressor body noise component.

5. The deep learning-based adaptive noise suppression method for refrigerator compressors according to claim 4, characterized in that, In the generator network of the generative adversarial network noise separation model, a skip connection is introduced between the encoder and decoder to concatenate the feature maps of each layer of the encoder with the corresponding layers of the decoder.

6. The deep learning-based adaptive noise suppression method for refrigerator compressors according to claim 1, characterized in that, The construction of the compressor noise suppression agent using a deep deterministic policy gradient network specifically includes: An actor network is constructed to map the time-frequency characteristics of the input compressor vibration state sequence and the compressor body noise components to the mean value of the compressor speed adjustment action; A network of critics is constructed to evaluate the action value corresponding to performing the compressor speed adjustment action under given environmental conditions; The current network of the actor network and the current network of the critic network are separated from the target network, and the parameters of the target network are copied from the current network parameters in a soft update manner with lag. During the training of the critic network, a priority experience replay mechanism is introduced, which performs non-uniform sampling from the experience pool according to the absolute value of the time difference error, and learns more frequently the experience samples with high priority.

7. The deep learning-based adaptive noise suppression method for refrigerator compressors according to claim 6, characterized in that, In the priority experience replay mechanism, the sampling probability of the experience sample is proportional to the absolute value of the time difference error, and a hybrid priority strategy combining proportional priority and sorting priority is adopted.

8. The deep learning-based adaptive noise suppression method for refrigerator compressors according to claim 6, characterized in that, The separation of the current network of the actor network and the current network of the critic network from the target network specifically involves: Set the first attenuation coefficient to control the parameter update rate of the target network of the actor network; A second attenuation coefficient is set to control the parameter update rate of the target network of the critic network; In each training step, the current network parameters of the actor network are multiplied by one and then subtracted from the first decay coefficient. This result is added to the target network parameters of the actor network multiplied by the first decay coefficient to obtain the updated target network parameters of the actor network. The updated target network parameters of the critic network are obtained by multiplying the current network parameters of the critic network by one and subtracting the second attenuation coefficient, and then adding the result of multiplying the target network parameters of the critic network by the second attenuation coefficient.

9. The deep learning-based adaptive noise suppression method for refrigerator compressors according to claim 1, characterized in that, The reward function is a weighted sum of noise suppression effect and energy consumption index, specifically: The sound pressure level at multiple locations inside the refrigerator is collected in real time, and the average of all sound pressure levels is calculated as the current noise level. The current noise level is compared with a preset noise target threshold, the difference between the two is calculated, and the negative value of the difference is used as the noise suppression effect score. Obtain the real-time compressor speed and calculate the current energy consumption score based on the preset mapping relationship between compressor speed and energy consumption; The value of the reward function is obtained by adding the result of multiplying the noise suppression effect score by the first weighting coefficient and the energy consumption score by the second weighting coefficient, wherein the sum of the first weighting coefficient and the second weighting coefficient is a constant.

10. The deep learning-based adaptive noise suppression method for refrigerator compressors according to claim 1, characterized in that, The real-time output of the compressor speed control signal to the frequency converter drive board is specifically as follows: The compressor noise suppression agent outputs the target compressor speed value; The target compressor speed value is compared with the current actual compressor speed value to calculate the speed deviation. The speed deviation is input to a proportional-integral-derivative (PID) controller, which outputs the duty cycle of a pulse width modulation (PWM) signal. The pulse width modulation signal is output to the frequency converter drive board to drive the frequency converter to generate three-phase AC power of the corresponding frequency, so as to adjust the compressor speed.

Images