A method for predicting noise in a heat pump air conditioner compressor and related products
By using a conditionally aligned residual network model to extract features and align the distribution of multi-source sensor signals from a heat pump air conditioner compressor, the problem of rapid and accurate prediction of second harmonic noise of the heat pump air conditioner compressor under different operating conditions is solved. This improves the model's generalization ability and prediction accuracy, and reduces the dependence on the anechoic chamber.
Patent Information
- Application Number
- CN202611413662.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-11
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies struggle to quickly and accurately predict the second harmonic noise of heat pump air conditioner compressors under different operating conditions, and existing methods rely on anechoic chamber testing, which is costly and lacks generalization ability across operating conditions.
A conditional alignment residual network model is adopted. High-dimensional latent features of multi-source sensor signals are extracted through a residual feature encoder. The conditional maximum mean difference regularization module is used to align the conditional feature distributions of different working conditions in a continuous label space. Noise prediction is then performed in combination with a regression prediction head.
It enables rapid and accurate prediction of compressor second harmonic noise under varying operating conditions, improves the model's generalization ability and prediction accuracy, avoids negative migration, and reduces dependence on anechoic chambers.
Smart Images

Figure CN123116707A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of air conditioning noise prediction technology, specifically to a method for predicting the noise of a heat pump air conditioner compressor and related products. Background Technology
[0002] When a heat pump air conditioner operates in cooling and heating modes, the compressor's periodic intake and exhaust processes generate torque pulsations at twice the rotational frequency. These pulsations are transmitted through the piping to the compressor casing and the sheet metal parts of the outdoor unit, easily inducing structural resonance and radiating high-intensity second-frequency discrete noise into the environment. Unlike broadband noise such as fan blade noise, compressor second-frequency noise has a high energy proportion and prominent peak value, significantly impacting user hearing. With increasing competition in the air conditioning market, user demands for acoustic comfort have evolved from low noise to high sound quality. Effective prediction and control of compressor second-frequency noise has become one of the core challenges in the development of heat pump air conditioning systems.
[0003] Currently, two main technical approaches are used to predict air conditioner compressor noise. The first is the sound power level testing method based on semi-anechoic chamber testing. This method relies on large-area, high-investment free-field microphone arrays or sound intensity probe systems. A single complete operating condition test cycle is long, and it is extremely sensitive to background vibration, reflected sound, and meteorological conditions. More importantly, each time the compressor support topology is optimized, the piping layout is adjusted, or the sound insulation structure is modified, a new prototype must be prepared and queued for retesting in the anechoic chamber, resulting in high R&D cycle time costs for design, verification, and improvement. The second approach is the statistical regression method based on data fitting. This method establishes a noise prediction model by fitting test data under limited operating conditions. While it reduces dependence on the anechoic chamber environment to some extent, the statistical regression model established by the correlation analysis lacks cross-condition generalization ability and struggles to cope with the dynamic characteristics of compressor second-harmonic noise, which fluctuates significantly by tens of decibels with changes in environmental heat load, compressor operating frequency, and refrigerant state parameters. In addition, although existing vibration and noise assessment methods based on mechanism simulation have clear physical interpretability, their calculation cycle is long and the simulation accuracy is limited by a large number of parameters that are difficult to obtain accurately, such as damping and joint surface stiffness. Therefore, they cannot meet the needs of rapid and accurate prediction of compressor second harmonic noise in short-cycle, multi-variety R&D scenarios.
[0004] Therefore, how to quickly and accurately predict the second harmonic noise of the heat pump air conditioner compressor under different operating conditions has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] The purpose of this invention is to provide a method for predicting the noise of a heat pump air conditioner compressor and related products, so as to overcome the problem that the existing prediction model has insufficient generalization ability under different operating conditions and is difficult to cope with the problem of large fluctuations in second harmonic noise with operating conditions.
[0006] The present invention solves the above-mentioned technical problems through the following technical solution: This invention provides a method for predicting the noise of a heat pump air conditioner compressor, comprising the following steps: The system acquires multi-source sensor signals from the heat pump air conditioner compressor under operating conditions and inputs these signals into a conditionally aligned residual network model. The conditionally aligned residual network model includes a residual feature encoder, a conditional maximum mean difference regularization module, and a regression prediction head. High-dimensional latent features are obtained by extracting features from multi-source sensor signals using a residual feature encoder. By using the conditional maximum mean difference regularization module, the conditional feature distribution of high-dimensional latent features under different working conditions is aligned in the continuous label space to obtain domain-invariant feature representations. By using a regression prediction head and performing regression prediction based on domain-invariant feature representation, the predicted value of the second harmonic noise of the heat pump air conditioner compressor is obtained.
[0007] A further improvement of this invention lies in that, through a conditional maximum mean difference regularization module, the conditional feature distributions of high-dimensional latent features under different operating conditions are aligned in a continuous label space to obtain a domain-invariant feature representation, specifically including: Calculate the pairwise Euclidean distance between the noise labels of all samples in the training batch, and use the radial basis kernel function to calculate the affinity weight matrix, where the affinity weight matrix is used to characterize the similarity between the noise labels of two samples. Construct a cross-domain indicator function matrix, which is used to distinguish sample pairs in different operating domains from sample pairs in the same operating domain; Based on the affinity weight matrix and the cross-domain indicator function matrix, the feature distance matrix is weighted, the conditional distribution alignment loss is calculated, and the parameters of the residual feature encoder are updated according to the conditional distribution alignment loss to obtain the domain-invariant feature representation.
[0008] A further improvement of this invention is that, before processing the multi-source sensor signals through the conditionally aligned residual network model, the following steps are included: Acquire training sample sets under multiple operating conditions. The training sample set under each operating condition includes sample labels and multi-source sensor signals. Domain-balanced sampling is performed on the training sample set. In each iteration step, the sampler extracts a predetermined number of samples from each of all working domains in parallel and independently, and then concatenates the samples extracted from each working domain into a global training batch. The conditionally aligned residual network model is trained based on global training batches.
[0009] A further improvement of this invention lies in acquiring multi-source sensor signals of the heat pump air conditioner compressor under operating conditions, specifically: Noise signals are collected by a microphone array installed on the heat pump air conditioner compressor; Vibration signals are collected by an accelerometer installed on the heat pump air conditioner compressor; Fast Fourier transform is performed on the noise and vibration signals to extract the root mean square value in the frequency band near the second harmonic of the compressor's operating frequency, which is then used as the feature value of the multi-source sensor signal.
[0010] A further improvement of the present invention is that the conditional maximum mean difference regularization module uses a radial basis function kernel to calculate the affinity weight matrix, and the scale parameter of the radial basis function kernel is the median of the Euclidean distance between the noise labels of all samples in the current training batch.
[0011] A further improvement of this invention is that the joint optimization objective when training the conditionally aligned residual network model is a weighted sum of the regression task loss and the conditional distribution alignment loss, and the regression task loss adopts the mean squared error loss function. The weights of the regression task loss and the conditional distribution-aligned loss are determined through a hyperparameter grid search.
[0012] The present invention also provides a noise prediction system for a heat pump air conditioner compressor, comprising: The first module is used to acquire multi-source sensor signals of the heat pump air conditioner compressor under operating conditions and input the multi-source sensor signals into the conditional aligned residual network model. The conditional aligned residual network model includes a residual feature encoder, a conditional maximum mean difference regularization module, and a regression prediction head. The second module is used to extract features from multi-source sensor signals through a residual feature encoder to obtain high-dimensional latent features; The third module is used to align the distribution of conditional features of high-dimensional latent features under different working conditions in the continuous label space through the conditional maximum mean difference regularization module, so as to obtain the domain-invariant feature representation. The fourth module is used to perform regression prediction based on the domain-invariant feature representation through the regression prediction head to obtain the predicted value of the second harmonic noise of the heat pump air conditioner compressor.
[0013] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described heat pump air conditioner compressor noise prediction method.
[0014] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described heat pump air conditioner compressor noise prediction method.
[0015] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described heat pump air conditioner compressor noise prediction method.
[0016] Compared with the prior art, the positive and progressive effects of the present invention are as follows: The noise prediction method for heat pump air conditioner compressors provided by this invention employs a conditionally aligned residual network model comprising a residual feature encoder, a conditional maximum mean difference regularization module, and a regression prediction head. First, the residual feature encoder performs deep and stable feature extraction from multi-source sensor signals to obtain high-dimensional latent features. Then, the conditional maximum mean difference regularization module aligns the distribution of conditional features under different operating conditions within a continuous label space using label information, thereby obtaining a domain-invariant feature representation. This effectively filters out interference strongly correlated with operating conditions but unrelated to noise, while retaining causal features related to noise mechanisms, thus solving the prediction inaccuracy problem caused by data distribution drift. Finally, the regression prediction head performs prediction based on these domain-invariant features. This method can improve the model's generalization ability and prediction accuracy under varying operating conditions by relying solely on multi-source sensor signals and without relying on an anechoic chamber, effectively avoiding negative transfer.
[0017] Furthermore, by calculating the pairwise Euclidean distance between noisy labels within a batch and constructing an affinity weight matrix using a radial basis function kernel, precise quantification of label similarity between samples is achieved within a continuous label space. Simultaneously, by constructing a cross-domain indicator function matrix, the alignment penalty is precisely limited to cross-domain sample pairs. Based on this, a weighted processing of the feature distance matrix is applied to calculate the conditional distribution alignment loss, which drives the model to selectively bring samples from different operating conditions but with similar noise values closer together in the feature space, while allowing samples with different noise values to retain their differences. This refined alignment strategy can more effectively suppress interference from irrelevant domain information, further improving the quality of domain-invariant feature representation and the model's resistance to negative transfer.
[0018] Furthermore, by introducing a domain-balanced sampling strategy during the model training phase, the problem of imbalanced work condition distribution in training batches caused by the uneven amount of data in different work condition domains is solved. In each iteration step, the sampler extracts an equal amount of samples from all work condition domains and concatenates them into a global training batch. This ensures the balance and unbiasedness of cross-domain aligned gradient directions during model training, preventing the model from tilting towards work condition domains with larger data volumes. This guarantees that the conditional distribution alignment module can fairly learn the domain-invariant features of all work conditions, improving the stability of model training and the final generalization performance.
[0019] Furthermore, by setting the scaling parameter of the radial basis kernel function to the median of the Euclidean distance between all sample noise labels in the current training batch, the affinity weight calculation becomes highly adaptable to different data batches and label distribution ranges, eliminating the need for repeated manual parameter adjustments and ensuring the stability and rationality of numerical calculations during conditional distribution alignment.
[0020] Furthermore, the joint optimization objective was clarified as a weighted sum of the regression task loss and the conditional distribution alignment loss. The weights of these two losses were determined through hyperparameter grid search, mathematically defining the convergence direction of model training. The optimal balance between regression accuracy and domain alignment strength was found through grid search. This avoids the problems of underfitting due to over-alignment or poor generalization ability due to insufficient alignment, thus ensuring that the model can maintain both high accuracy and high robustness under unknown conditions. Attached Figure Description
[0021] The accompanying drawings are provided to further understand the invention and constitute a part of this invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0022] Figure 1 This is a schematic diagram of the fitting scatter points for each working condition described in the embodiment; Figure 2 for Figure 1 A comparative diagram of the prediction results for medium-load condition 1; Figure 3 for Figure 1 A comparative diagram of the prediction results for medium-load condition 2; Figure 4 for Figure 1 A comparative diagram of the prediction results for medium-load condition 3; Figure 5 for Figure 1 A comparative diagram of the prediction results for medium-load condition 4; Figure 6 for Figure 1 A comparative diagram of the prediction results for medium-load condition 5; Figure 7 for Figure 1 A comparative diagram of the prediction results for medium-load condition 6; Figure 8 for Figure 1 A schematic diagram comparing the ECDF prediction results for various working conditions; Figure 9 This is a flowchart illustrating the method of the present invention. 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] In the description of this invention, it should be understood that the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.
[0025] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.
[0026] It should be understood that although terms such as first, second, third, etc., may be used in the embodiments of the present invention to describe the preset range, these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from one another. For example, without departing from the scope of the embodiments of the present invention, the first preset range may also be referred to as the second preset range, and similarly, the second preset range may also be referred to as the first preset range.
[0027] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."
[0028] Before detailing the embodiments of the present invention, the technical problem addressed by the present invention will be briefly described. The second harmonic noise of the compressor in the outdoor unit of a variable refrigerant flow multi-split air conditioner fluctuates significantly by tens of decibels depending on the ambient heat load, compressor frequency, and refrigerant state parameters. Test results under a single standard operating condition cannot characterize the actual acoustic performance across the entire operating envelope. Traditional anechoic chamber testing relies on high-cost acoustic environments and equipment, and has a long testing cycle, making it difficult to meet the needs of rapid iterative R&D for short cycles and multiple product types. Conventional single-condition simulation or cross-domain prediction methods are prone to negative migration problems when facing source domains with significantly different label distributions, and cannot achieve rapid and accurate assessment of second harmonic noise across the entire operating range.
[0029] Based on the above background, this invention provides a method for predicting the noise of a heat pump air conditioner compressor. This method acquires multi-source sensor signals from the compressor under operating conditions and processes them using a conditionally aligned residual network model comprising a residual feature encoder, a conditional maximum mean difference regularization module, and a regression prediction head. The residual feature encoder extracts high-dimensional latent features, and the conditional maximum mean difference regularization module aligns the distribution of conditional features under different operating conditions within a continuous label space to obtain a domain-invariant feature representation. Finally, the regression prediction head outputs the predicted value of the second harmonic noise. This method effectively alleviates the data distribution drift problem between different operating conditions and achieves rapid and accurate prediction of compressor second harmonic noise across the entire operating range without relying on an anechoic chamber testing environment.
[0030] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This is an explanation of the present invention and not a limitation thereof.
[0031] See Figure 9 This invention provides a method for predicting the noise of a heat pump air conditioner compressor. The execution process of this method includes the following steps: Step one involves acquiring multi-source sensor signals from the heat pump air conditioner compressor during its operation. These multi-source sensor signals refer to the set of signals reflecting the compressor's operating status and physical response, collected synchronously by various types of sensors during actual operation. This step forms the data foundation for subsequent model processing.
[0032] Step two involves inputting the acquired multi-source sensor signals into a pre-built and trained conditional aligned residual network model. This conditional aligned residual network model is a neural network model with a clearly defined internal structure comprising three core components: a residual feature encoder, a conditional maximum mean difference regularization module, and a regression prediction head. The residual feature encoder receives the raw input, the conditional maximum mean difference regularization module operates on the feature level of the encoder output, and the regression prediction head generates the final result.
[0033] Step three involves feature extraction from the input multi-source sensor signals using a residual feature encoder. The residual feature encoder, based on a deep network structure, can automatically learn and extract deep nonlinear features from the input data. In this step, the original multi-source sensor signals are mapped to a high-dimensional space, resulting in a set of high-dimensional latent features. This set of high-dimensional latent features provides a more compact and abstract representation of the original physical signals.
[0034] Step four involves processing the high-dimensional latent features obtained in the previous step using the conditional maximum mean difference regularization module. This module does not simply align all features indiscriminately; instead, it introduces label information and performs conditional judgments within a continuous noise label space. It uses an algorithm to measure the similarity of noise labels for samples under different operating conditions and, based on this similarity, selectively aligns the distribution of conditional features under different operating conditions. After processing by this module, those variable factors in the high-dimensional latent features that are related to specific operating conditions but unrelated to the noise itself are suppressed, while the essential features truly related to the noise generation mechanism are preserved, thus obtaining a domain-invariant feature representation.
[0035] Step five involves using a regression prediction head to perform regression prediction based on the domain-invariant feature representation obtained in the previous step. The regression prediction head is a mapping network that takes the domain-invariant feature representation as input, performs a nonlinear transformation, and ultimately outputs a specific scalar value, which is the predicted value of the second harmonic noise of the heat pump air conditioner compressor. This completes the end-to-end prediction from multi-source physical signals to noise decibel values.
[0036] Through the steps described above, the method shown in this embodiment can predict compressor second harmonic noise based on easily obtainable multi-source sensor signals without relying on an anechoic chamber. Because the model incorporates a conditional maximum mean difference regularization module, it can align feature distributions under different operating conditions within a continuous label space, learning domain-invariant feature representations. This effectively addresses data distribution drift caused by changes in environmental heat load, compressor frequency, etc., improving the model's prediction accuracy and generalization ability under multiple operating conditions, and avoiding the negative transfer phenomenon that may occur with conventional methods.
[0037] For a specific implementation of aligning conditional feature distributions using the conditional maximum mean difference regularization module, in one embodiment, the process can be broken down into the following calculation steps: First, the pairwise Euclidean distances between the noise labels of all samples in the current training batch are calculated. Next, an affinity weight matrix is calculated using a radial basis function (RBF) kernel. Each element in this matrix represents the similarity between the noise labels of two samples within the batch. In one optional implementation, the scale parameter of the RDF kernel is taken as the median of the Euclidean distances between the noise labels of all samples in the current training batch. Using the median intra-batch distance as the scale parameter allows the affinity weights to be adaptive to different label distribution ranges, ensuring the numerical stability of the calculation process.
[0038] Simultaneously, a cross-domain indicator function matrix is constructed. This matrix is a binary matrix, used to distinguish sample pairs from different operating conditions from those from the same operating condition. For example, if two samples come from the same operating condition, the corresponding matrix element is set to 0; if they come from different operating conditions, it is set to 1. This definition explicitly limits the feature alignment penalty to cross-domain sample pairs, avoiding interference with the natural distribution of data within the same operating condition.
[0039] Based on the above calculations, the feature distance matrix obtained from the high-dimensional latent features is weighted using the affinity weight matrix and the cross-domain indicator function matrix to calculate the conditional distribution alignment loss. This loss quantifies the distribution differences in the feature space of samples with similar noise labels under different operating conditions. Finally, during model training, the network parameters of the residual feature encoder are backpropagated and updated based on this conditional distribution alignment loss. Through iterative optimization of this mechanism, the encoder is forced to learn to bring samples from different operating conditions but with similar noise values closer together in the feature space, thus obtaining a domain-invariant feature representation. This dynamic weighting mechanism acts like a "physical filter," filtering out interfering features that are strongly correlated with the operating conditions but unrelated to the noise, while retaining the causal features that characterize the nature of the noise.
[0040] In one implementation of the training process for the conditionally aligned residual network model before application, a special domain-balanced sampling strategy is employed. Specifically, training sample sets are first acquired across multiple operating domains, each containing sample labels and corresponding multi-source sensor signals. Since the number of samples that can be collected under different operating conditions is often uneven, using conventional random sampling would lead to model training biased towards operating conditions with a larger sample size.
[0041] To address this issue, this embodiment employs domain-balanced sampling of the training sample set. In each iteration step, the sampler extracts a predetermined number of samples from each of all operating domains in parallel and independently. For example, if there are K operating domains, B samples are extracted from each domain in each step. These samples extracted from each operating domain are then concatenated to form a global training batch. The conditional alignment residual network model is trained based on this global training batch, ensuring that each batch contains an equal number of representative samples from all operating conditions. This guarantees the balance of the cross-domain gradient direction in the calculation of the conditional distribution alignment loss and avoids alignment bias caused by imbalance in the distribution of operating conditions within a batch.
[0042] The joint optimization objective for model training is defined as the weighted sum of the regression task loss and the conditional distribution alignment loss. The regression task loss uses the mean squared error loss function to constrain the accuracy between the predicted values and the true labels. The balance between the two losses—the weights of the regression task loss and the conditional distribution alignment loss—is determined in a preferred implementation through a hyperparameter grid search. In practice, a search space can be predefined, for example, by setting a set of candidate weight values. By traversing combinations of these candidate values, the model's performance on the validation set is evaluated, and finally, a set of weight coefficients that optimizes the model's generalization ability is selected.
[0043] By employing domain-equalized sampling and joint optimization training, the method in this embodiment can more effectively learn domain-invariant features, further improving the model's predictive robustness and accuracy when dealing with variable conditions and large differences in data distribution.
[0044] This invention proposes a Conditional Alignment Residual Network (CARN) model for predicting second-harmonic noise in VRF multi-unit outdoor units. Specifically, the model combines a residual feature encoder with Conditional Maximum Mean Difference (CMD) regularization. The residual encoder provides deep and stable feature extraction capabilities, while CARN forces the model to learn label-related domain-invariant representations through label-aware cross-domain distribution alignment. This method effectively alleviates the negative transfer problem while maintaining prediction accuracy, and is particularly suitable for scenarios where there are significant differences in label distributions across different source domains.
[0045] CARN consists of three core modules: 1. Residual feature encoder: responsible for mapping multi-source sensor signals to a high-dimensional latent feature space.
[0046] 2. Conditional Maximum Mean Difference Regularization Module: Aligns the distribution of conditional features for different operating conditions within a continuous label space.
[0047] 3. Regression Prediction Head: Outputs the final prediction scalar based on the aligned features.
[0048] Specifically: 1. Residual Feature Encoder Due to input ( The characteristics of physical sensors across modalities vary greatly in their dimensions and physical meanings. CARN constructs a residual bottleneck structure based on a multilayer perceptron, projecting the original signal onto a compact latent feature manifold. ( In the network. For the first Each residual module, in order to suppress gradient vanishing and stabilize the operator norm while deepening the network, has its forward propagation defined as: (1) (2) In the formula, This is the output of the residual block. It is a Gaussian error linear unit, used to introduce smooth nonlinear expression capabilities.
[0049] 2. Conditional Maximum Mean Difference Module In regression tasks, labels It is a continuous scalar. Traditional edge distribution alignment attempts to minimize the overall feature distribution distance between all operating conditions, which destroys the feature distribution. With the goal The intrinsic mapping relationship between them. CARN introduces the conditional MMD strategy, the core idea of which is: only when the target labels of two samples are similar, will their feature distributions under different operating conditions be brought closer. The specific calculation process consists of the following three steps: I. Gaussian kernel affinity matrix of continuous labels First, calculate the labels of all samples in the batch. Pairwise Euclidean distance To map absolute distance to relative "label similarity", we use the radial basis function (RBF kernel) to calculate affinity weights. : (3) in: It is the square of the Euclidean distance between the two sample true labels. The scaling parameter is used; to ensure numerical stability, this model selects the median of the label distances of all samples within a batch as the scaling parameter. And add a minimum value To prevent the denominator from being zero (i.e.) ). When the labels are closer, the weight increases. The closer it gets to 1.
[0050] II. Construction of Cross-Domain Mask Matrix To avoid penalizing sample differences within the same working condition, the model introduces a cross-domain indicator function, which is a binary Boolean mask matrix.
[0051] (4) in, Representative sample The operating condition domain attribute.
[0052] III. Conditional Distribution Alignment Penalty Calculation Finally, in the feature space In this process, by performing a Hadamard product between the Feature Distance Matrix and the two matrices mentioned above, an unbiased estimate of the conditional alignment loss is obtained. Finally, in the feature space... In this process, by performing a Hadamard product between the feature distance matrix and the two matrices mentioned above, an unbiased estimate of the conditional alignment loss is obtained: (5) in: This represents the total number of samples in the current training batch. and These are the deep features extracted from the samples by the ResNet encoder. Let L2 be the norm of the vector in the feature space.
[0053] The essence of this formula is a dynamically weighted metric learning process. If two samples come from different working conditions ( ), and their actual noise values are extremely close ( Then a huge penalty is imposed, forcing the network to shrink in the feature space. Conversely, if the noise values are different ( If this condition is met, the model is allowed to preserve feature differences. This mechanism makes the model act like a "physical filter": filtering out environmental noise that is strongly correlated with the operating conditions but unrelated to the noise itself, while retaining causal features that purely characterize the noise mechanism. 3. Prediction Head and Joint Optimization Objective The top layer of the model is the regression prediction head. It consists of a multilayer perceptron (MLP) and is used to align deep features. Mapped to the final predicted value : (6) During the model training phase, CARN employs a multi-task joint optimization paradigm. The basic regression task uses mean squared error (MSE) as the loss function. (7) Ultimately, CARN's global joint optimization objective (Total Loss) is defined as the weighted sum of the regression task loss and the conditional alignment regularization loss: (8) in: This is the penalty weight, used to balance the accuracy of the regression task with the strength of cross-domain feature alignment. The optimal weight can be determined for different datasets through hyperparameter grid search. The value is determined to achieve optimal generalization performance.
[0054] The training and optimization strategy of this invention: Since cross-domain alignment loss needs to be calculated in the joint optimization objective, traditional hybrid random sampling can lead to severe imbalance in the distribution of working conditions within a batch due to the inconsistent data volume of each working condition, resulting in deviations in the direction of the alignment gradient. To address this, our framework constructs a domain-balanced infinite iterator. In each iteration step, the sampler will sample data in parallel and independently from... Each source domain has an extraction size of [size]. The micro-batches are then concatenated into a single global training batch: (9) During the parameter update phase, CARN employs the Adam Optimizer, an adaptive moment estimator with driving forces, and introduces weight decay. Regularization is used to limit model complexity and prevent overfitting. (10) in, This is the weight decay coefficient.
[0055] Joint optimization objective of the CARN model In the middle, manifold penalty weights The selection is crucial. If If the value is too small, the model degenerates into an ERM (Extremely Small Model) and cannot withstand load condition shifts. If... Excessive and forced feature stretching can disrupt the inherent causal mapping between the input and the noisy target, leading to underfitting.
[0056] To find the optimal game equilibrium, the training stream introduces a rigorous hyperparameter grid search, defining the search space as: (11) Among them, the basic learning rate set Set of penalty coefficients .
[0057] Traditional random cross-validation can compromise the independence of operating condition data, leading to data traversal problems and failing to accurately reflect the model's performance under entirely new operating conditions. Therefore, this invention strictly adheres to the standard protocol for domain generalization and employs a leave-one-domain cross-validation strategy.
[0058] Assuming the entire set of data contains An independent operating domain In the During round testing: Target Domain: Select a domain As a pure test set This domain is completely unseen by the model during the training phase.
[0059] Source Domains: The remaining ones Each domain forms the training source domain set. .
[0060] This study conducted simultaneous vibration and noise measurements on a VRF outdoor unit in a fully anechoic chamber. The tested unit has a rated cooling capacity of 18kW and is driven by a dual-rotor jet enthalpy-increasing compressor. The jet mass flow rate can be precisely controlled by adjusting the jet valve opening, thereby flexibly switching the compressor's thermodynamic state. Vibration and noise signals were acquired using the Siemens SCADAS Mobile system. Following company internal standards, eight microphones were arranged in a circular array at a uniform height of 1m around the unit. For structural vibration, a combination of single-axis and triaxial accelerometers was used, with a total of 39 acceleration channels. The sensor locations covered key vibration transmission paths and radiation surfaces, including the compressor housing, intake and exhaust pipes, and the outdoor unit panel.
[0061] The test operation was set to the rated heating mode. The compressor operated at its rated frequency, and the main electronic expansion valve opening was fixed at the rated value. The thermodynamic cycle state of the system was changed by adjusting the steam injection valve opening. A set of operating conditions covering different injection states was designed, and each condition was maintained for 1–2 hours after the system reached thermal equilibrium. During the test, the wet-bulb temperature inside the silencer chamber was at… The temperature naturally fluctuates between 1°C and 5°C, altering the outdoor unit's heat load and dynamically affecting the system's thermodynamic state. Therefore, this environmental factor is recorded synchronously as an independent variable. Data acquisition is triggered at 1-minute intervals, with each acquisition lasting 10 seconds. Vibration and acoustic signals are sampled at a 20kHz sampling rate, and signal averaging is employed to improve the signal-to-noise ratio. For quasi-static thermodynamic parameters such as temperature, the arithmetic mean within each 10-second window is taken as the representative characteristic value of the corresponding sample.
[0062] This invention aims to predict the second harmonic (2f) component related to the compressor's operating frequency. The data preprocessing process is as follows: First, a fast Fourier transform is performed on the original vibration and noise signals to extract the root mean square value within a ±3Hz range around the 2f frequency, and this value is converted to decibels (dB). Second, the label for each sample is defined as the arithmetic mean of the 2f sound pressure level measured by eight microphones, and the input features include the 2f amplitude from 39 acceleration channels. Six representative operating conditions with different steam injection valve openings were tested. After removing non-steady-state data during condition switching and data samples affected by sensor anomalies, the number of samples under each condition ranged from 40 to 120, and the 2f noise label distribution ranged from 40 dB to 55 dB (a span of 15 dB). Due to the combined influence of injection valve adjustment and ambient temperature fluctuations, the data distribution of different condition groups exhibited significant and irregular statistical variations.
[0063] To verify the effectiveness of the CARN model proposed in this invention, it was compared with Empirical Risk Minimization (ERM), Correlation Alignment (CORAL), Joint Maximum Mean Discrepancy (JMMD), and Local Maximum Mean Discrepancy (LMMD). All methods were trained under the same conditions, with a batch size of 32 and a learning rate of 0.001.
[0064] Evaluation metrics included Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Coefficient of Determination (R²), Concordance Correlation Coefficient (CCC), and accuracy with an error less than 3 dB. Experiments were conducted in a unified hardware environment, primarily configured with an Intel Core i7-14700F processor and an NVIDIA GeForce RTX 4060 graphics card.
[0065] Comparison results: The evaluation metrics for all operating conditions are shown in Table 1. As can be seen from the table, CARN performs best across all metrics. Its mean RMSE is 0.725 and mean MAE is 0.495, both the lowest in terms of error rate and with small standard deviations, indicating high prediction accuracy and stable performance. R² reaches 0.829 and CCC is 0.909, both significantly higher than other methods, with standard deviations of ±0.118 and ±0.065 respectively, showing good consistency and robustness. ERM and CORAL have negative R² values of -3.711 and -4.239 respectively, with standard deviations exceeding 10, indicating serious prediction bias and instability in these two methods. JMMD and LMMD effectively improved R² through distribution alignment, reaching 0.553 and 0.723 respectively, but still lag behind CARN. Regarding accuracy less than 3dB, all methods maintain a high level, with CARN achieving an accuracy of 98.5%, similar to CORAL, JMMD, and LMMD. Overall, CARN achieves a balance of low error, high correlation, and stability in cross-condition prediction, and its generalization ability is significantly better than the comparison methods.
[0066] Table 1. Overall evaluation indicators of the above method under all working conditions
[0067] The scatter plots of the fitting results for each method are as follows: Figure 1 As shown, to evaluate the prediction performance and robustness of different domain generalization methods under varying operating conditions, this invention compares the core evaluation metrics and prediction distributions of ERM, CORAL, JMMD, LMMD, and CARN models. As shown in the table and figure, the accuracy of ERM and CORAL methods with errors less than 3 dB is higher than 94%, but their coefficient of determination R0 is lower. 2 The values were negative, -3.711 and -4.239 respectively. Their predicted values were generally below the ideal baseline of y=x, exhibiting a negative migration phenomenon. In contrast, the JMMD and LMMD methods, which introduce distribution alignment mechanisms, mitigated extreme prediction bias by extracting domain-invariant features, increasing R² to 0.553 and 0.723 respectively. The CARN method performed better in terms of prediction consistency and cross-domain feature alignment, with the lowest RMSE (0.725) and MAE (0.495); the highest R² (0.829) and CCC (0.909), and smaller standard deviation of the index fluctuations. Figure 1 As can be seen, the CARN model's predicted scatter points are closely distributed near the diagonal in all six operating conditions, and the marginal kernel density estimation curves basically coincide with the true value distribution. These results demonstrate that the CARN method can effectively handle multi-operating-condition distribution shifts and possesses good global generalization ability.
[0068] Figure 2 The comparison of prediction results for Condition 1 is shown. In the figure, solid dots represent the true values of the target variable, dashed squares represent the predicted values of each model, and the semi-transparent filled area between them reflects the cumulative magnitude of the absolute prediction error. From the scatter plot distribution of the ERM model's predictions, it can be seen that without domain feature alignment, the model exhibits a significant global prediction shift. Both the CORAL and CARN models demonstrate good dynamic tracking capabilities, responding quickly after a sudden change in the target variable and closely adhering to the new high-level stable phase. It is worth noting that although CORAL achieved a locally optimal quantification index under this specific condition, with an RMSE of 1.410, R 2 With an RMSE of 0.940, the CARN model also demonstrated competitive predictive performance, with an RMSE of 1.460, an R² of 0.935, and an accuracy of 92.1% for less than 3 dB.
[0069] Figure 3 A comparison of the prediction results for working condition 2 is presented. It can be seen that the baseline model exhibits significant tracking lag and systematic bias. The ERM model performs the worst; its prediction curve fails to follow the decreasing trend of the true value in the later stages of the sample, resulting in a large area of error. Its RMSE is 1.656, and its coefficient of determination is... R 2 The R² value of only 0.067 indicates that the model has a weak ability to explain the fluctuations of the target variable. CORAL and LMMD outperform ERM in terms of overall trend, but systematic biases still exist in local intervals, with noticeable error shading. In contrast, JMMD and CARN predict values that match the true values better, with significantly reduced error imputation area. Specifically, JMMD achieved the best local fit on this sample segment, with an RMSE of 0.694 and an R² of 0.836; the CARN model also performed well, with an RMSE of 0.875. R 2 The value is 0.739, which not only accurately follows the overall downward trend, but also reflects subtle local fluctuations, such as in the sample range of 20 to 40.
[0070] Figure 4 A comparison of prediction results for operating condition 3 is presented. With the introduction of the domain alignment strategy, the prediction curves of JMMD and LMMD show significant improvements in smoothness and fit, with a significant reduction in the error region and R² increasing to 0.749 and 0.710, respectively. The CARN model achieves the best overall performance under this operating condition. In the rapid rise of the peak and the dense oscillation of the trough, its prediction curve closely follows the trajectory of the true value, with an RMSE as low as 0.445 and an R² of 0.821, both of which are the best among all methods.
[0071] Figure 5The comparison of prediction results for operating condition 4 is presented. As can be seen from the joint probability distribution plot of the ERM model, its predicted values deviate significantly from the diagonal of y=x, with an average R² of -2.066. Both ERM and CORAL exhibit significant systematic prediction drift, with the prediction curves shifting downwards overall, deviating from the true value baseline and forming a large error region, with local R² dropping to -25.018 and -28.773, respectively. This indicates that the ERM and CORAL methods struggle to accurately calibrate the baseline level of the target variable when operating conditions change drastically. JMMD eliminates the aforementioned magnitude of deviation, pulling the predicted values back to near the true values, but its R² remains low. 2 The value remains -0.841, indicating that its ability to explain local variance is still weaker than directly predicting the mean. LMMD further increases the R-squared value. 2 The error range was significantly reduced by increasing the value to 0.440. In comparison, the CARN model demonstrated better cross-domain robustness and dynamic tracking ability. In the time series plot, when other models had large prediction deviations, the CARN prediction curve still closely followed the changes in the true value, eliminating the mean shift caused by cross-domain errors and reflecting high-frequency fluctuations in the signal.
[0072] Figure 6 The comparison of prediction results for condition 5 is presented. The time series plot shows that the target variable undergoes a significant step change near sample number 40, with the value surging from 45 to 60. Under such drastic state transitions, the baseline ERM model exhibits fitting lag in the low-level oscillation region before the abrupt change, demonstrating insufficient robustness to disturbances, a large error range, and an accuracy drop to 94.4%. After introducing a domain feature alignment strategy, the model's responsiveness is significantly improved. CORAL, LMMD, and CARN all respond rapidly at the moment of the step change and closely follow the high-level stable state after the abrupt change. Among them, the determination coefficient R of CARN and CORAL is significantly higher. 2 Both reached 0.984, indicating that the domain generalization method has good robustness and state reset capability when dealing with sudden changes in working conditions in industrial scenarios.
[0073] Figure 7 A comparison of the prediction results for operating condition 6 is presented. The ERM exhibits a significant system tracking deviation during the initial rapid ascent phase. R 2 With an RMSE of only 0.271, it is difficult to characterize the early dynamic features. After entering the high-level plateau period, although JMMD and LMMD can follow the overall trend, there is still a certain steady-state deviation. In contrast, CARN shows higher fitting accuracy throughout the entire period, converges rapidly in the transient rise phase, has a smaller steady-state deviation in the high-level plateau phase, and can reflect the high-frequency micro-oscillations on the plateau; its RMSE is as low as 0.436. R 2The predicted value reached an optimal value of 0.827, with the predicted value closely matching the actual value. This indicates that CARN has good performance in both macro-trend alignment and micro-fluctuation feature analysis.
[0074] Further analysis was conducted using the Empirical Cumulative Distribution Function (ECDF) of the absolute step error. The figure above shows the ECDF curves of the five domain generalization methods under six representative operating conditions, intuitively reflecting the overall error level and extreme value control capability of each model under different data distributions. Overall, CARN exhibits superior and more stable prediction performance in all operating conditions. As can be seen from the subplots, CARN's ECDF curve is consistently closer to the left vertical axis and rises more steeply, indicating that the prediction error of most samples is concentrated in a smaller range. At the 90% cumulative probability, CARN reaches this threshold first with the smallest absolute error in almost all operating conditions, demonstrating high prediction accuracy. In contrast, the performance of traditional baseline models fluctuates more significantly. ERM and CORAL show significant performance degradation in some operating conditions, especially in subplot d, with the curves shifting to the right and exhibiting a long tail, indicating that these methods struggle to maintain stability when faced with significant domain shifts, resulting in more pronounced extreme prediction errors. LMMD performed second best, closely following CARN in most scenarios, but its convergence speed was slightly slower when the cumulative probability approached 100%. Overall, the ECDF results show that the CARN model not only outperforms CARN in global absolute prediction accuracy but also more effectively suppresses long-tailed anomaly errors. This characteristic of maintaining a steep curve close to the left vertical axis under multiple unknown scenarios reflects CARN's good cross-domain generalization ability and robustness.
[0075] Based on the same inventive concept, this application provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of a method for predicting noise in a heat pump air conditioner compressor. The memory may include main memory, such as high-speed random access memory, or it may also include non-volatile memory, such as at least one disk storage device. The processor, network interface, and memory are interconnected via an internal bus, which may be an industry-standard architecture bus, a peripheral component interconnection standard bus, an extended industry-standard architecture bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. The memory stores the program; specifically, the program may include program code, which includes computer operation instructions. The memory may include main memory and non-volatile memory, and provides instructions and data to the processor.
[0076] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the steps of the heat pump air conditioner compressor noise prediction method. Specifically, the computer-readable storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. The volatile memory may include RAM (Random Access Memory) and / or cache memory, etc. The non-volatile memory may include ROM (Read-Only Memory), hard disk, flash memory, optical disk, magnetic disk, etc.
[0077] Based on the same inventive concept, this application provides a computer program product, which includes a computer program stored on a computer-readable storage medium. The computer program includes program instructions, which, when executed by a computer device, cause the computer device to perform the steps of the above-described heat pump air conditioner compressor noise prediction method.
[0078] Those skilled in the art will understand that embodiments of the present invention can be provided as methods or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM (Compact Disc Read-Only Memory), optical storage, etc.) containing computer-usable program code.
[0079] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer apparatus or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0080] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer device or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0081] These computer program instructions may also be loaded onto a computer device or other programmable data processing equipment to cause a series of operational steps to be performed on the computer device or other programmable equipment to produce a process implemented by the computer device, thereby providing instructions that execute on the computer device or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0082] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.
[0083] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. A method for predicting the noise of a heat pump air conditioner compressor, characterized in that, Includes the following steps: The system acquires multi-source sensor signals from the heat pump air conditioner compressor under operating conditions and inputs these signals into a conditionally aligned residual network model. The conditionally aligned residual network model includes a residual feature encoder, a conditional maximum mean difference regularization module, and a regression prediction head. High-dimensional latent features are obtained by extracting features from multi-source sensor signals using a residual feature encoder. By using the conditional maximum mean difference regularization module, the conditional feature distribution of high-dimensional latent features under different working conditions is aligned in the continuous label space to obtain domain-invariant feature representations. By using a regression prediction head and performing regression prediction based on domain-invariant feature representation, the predicted value of the second harmonic noise of the heat pump air conditioner compressor is obtained.
2. The method for predicting noise of a heat pump air conditioner compressor according to claim 1, characterized in that, The conditional maximum mean difference regularization module aligns the conditional feature distributions of high-dimensional latent features under different operating conditions within the continuous label space, resulting in a domain-invariant feature representation, specifically including: Calculate the pairwise Euclidean distance between the noise labels of all samples in the training batch, and use the radial basis kernel function to calculate the affinity weight matrix, where the affinity weight matrix is used to characterize the similarity between the noise labels of two samples. Construct a cross-domain indicator function matrix, which is used to distinguish sample pairs in different operating domains from sample pairs in the same operating domain; Based on the affinity weight matrix and the cross-domain indicator function matrix, the feature distance matrix is weighted, the conditional distribution alignment loss is calculated, and the parameters of the residual feature encoder are updated according to the conditional distribution alignment loss to obtain the domain-invariant feature representation.
3. The method for predicting noise of a heat pump air conditioner compressor according to claim 1, characterized in that, Before processing multi-source sensor signals using a conditionally aligned residual network model, the following steps are also included: Acquire training sample sets under multiple operating conditions. The training sample set under each operating condition includes sample labels and multi-source sensor signals. Domain-balanced sampling is performed on the training sample set. In each iteration step, the sampler extracts a predetermined number of samples from each of all working domains in parallel and independently, and then concatenates the samples extracted from each working domain into a global training batch. The conditionally aligned residual network model is trained based on global training batches.
4. The method for predicting noise of a heat pump air conditioner compressor according to claim 1, characterized in that, Acquire multi-source sensor signals from the heat pump air conditioner compressor under operating conditions, specifically: Noise signals are collected by a microphone array installed on the heat pump air conditioner compressor; Vibration signals are collected by an accelerometer installed on the heat pump air conditioner compressor; Fast Fourier transform is performed on the noise and vibration signals to extract the root mean square value in the frequency band near the second harmonic of the compressor's operating frequency, which is then used as the feature value of the multi-source sensor signal.
5. The method for predicting noise of a heat pump air conditioner compressor according to claim 1, characterized in that, The conditional maximum mean difference regularization module uses the radial basis function kernel to calculate the affinity weight matrix. The scale parameter of the radial basis function kernel is the median of the Euclidean distance between the noise labels of all samples in the current training batch.
6. The method for predicting noise of a heat pump air conditioner compressor according to claim 1, characterized in that, The joint optimization objective when training the conditionally aligned residual network model is a weighted sum of the regression task loss and the conditional distribution alignment loss, with the mean squared error loss function used for the regression task loss. The weights of the regression task loss and the conditional distribution-aligned loss are determined through a hyperparameter grid search.
7. A noise prediction system for a heat pump air conditioner compressor, characterized in that, include: The first module is used to acquire multi-source sensor signals of the heat pump air conditioner compressor under operating conditions and input the multi-source sensor signals into the conditional aligned residual network model. The conditional aligned residual network model includes a residual feature encoder, a conditional maximum mean difference regularization module, and a regression prediction head. The second module is used to extract features from multi-source sensor signals through a residual feature encoder to obtain high-dimensional latent features; The third module is used to align the distribution of conditional features of high-dimensional latent features under different working conditions in the continuous label space through the conditional maximum mean difference regularization module, so as to obtain the domain-invariant feature representation. The fourth module is used to perform regression prediction based on the domain-invariant feature representation through the regression prediction head to obtain the predicted value of the second harmonic noise of the heat pump air conditioner compressor.
8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the heat pump air conditioner compressor noise prediction method according to any one of claims 1 to 6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the heat pump air conditioner compressor noise prediction method according to any one of claims 1 to 6.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the heat pump air conditioner compressor noise prediction method according to any one of claims 1 to 6.
