A method for regulating acoustic environment comfort

Abstract

The application discloses a sound environment comfort regulation method. Sound data is converted into a time-frequency feature matrix, which is input into a convolutional neural network to extract deep features and obtain an acoustic feature vector. A physiological vector is input into a long short-term memory network to model a physiological state sequence. The acoustic feature vector and the physiological state sequence are input into a cross-modal attention gate unit to obtain a correlation weight. Coordinates, environmental sensor labels and activity intensity indexes are input into a fully connected network to obtain a scene probability distribution. A fusion feature is obtained, which is input into an inflated causal convolution layer to obtain a multi-scale feature, and the multi-scale feature is input into a multi-head self-attention module to obtain a key frame. The key frame is mapped into a comfort score distribution. The fusion feature is input into a time series anomaly detector. An early warning trigger function is obtained through an expression, an anomaly intensity is calculated, and a corresponding regulation strategy is obtained, so that a sound environment comfort regulation system with multi-source heterogeneous data fusion, scene adaptive weighting and double-flow collaborative decision is realized.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and intelligent sound environment control technology, and in particular to a method for regulating sound environment comfort. Background Technology

[0002] The acoustic environment is a core element of a healthy campus environment, directly impacting students' perceptions. Numerous studies have shown that prolonged exposure to noise can induce autonomic nervous system stress responses, manifesting as abnormally elevated heart rate, increased skin conductance, and fluctuations in blood oxygen saturation. It can also negatively affect students' mental health. Therefore, constructing a multi-source data fusion-based acoustic comfort evaluation system to reveal the pathway of action—from "physical acoustic stimulation → physiological and psychological response → evaluation feedback"—has become a key scientific issue for improving a healthy campus environment. Summary of the Invention

[0003] This invention provides a method for regulating acoustic environment comfort, and constructs an acoustic environment comfort regulation system based on multi-source heterogeneous data fusion, scene adaptive weighting, and dual-stream collaborative decision-making.

[0004] This invention provides a method for regulating acoustic environment comfort, comprising:

[0005] Acquire acoustic data and convert the acoustic data into a time-frequency feature matrix S;

[0006] The time-frequency feature matrix S is input into a convolutional neural network for deep feature extraction to obtain the acoustic feature vector. The expression is ;in, The convolution kernel weight matrix is... Here, ReLU is the corresponding bias vector, and ReLU is the activation function.

[0007] Constructing physiological vectors , the physiological vector Inputting the data into a long short-term memory network for feature modeling yields a sequence of physiological states. The expression is ;in, This refers to the parameter set of the Long Short-Term Memory network;

[0008] The acoustic feature vector and the physiological state sequence Inputting the cross-modal attention gating unit yields the association weights. The expression is ;in, For the first Frame acoustic features and the first Unnormalized correlation scores of physiological sampling points , and For trainable projection matrices, For attention vectors, For the first Acoustic feature vectors of time frames, For the first Physiological feature vectors of each sampling point For the first Frame acoustic features and the first Unnormalized correlation scores of each physiological sampling point;

[0009] Through expressions Obtain the aligned physiological characteristics ;

[0010] User coordinates Environmental sensor tags and activity intensity index Input a three-layer fully connected network to obtain the scene probability distribution. The expression is ;in, The hidden layer feature vectors of the three fully connected network are... , and These are the weight matrices for the first and second fully connected layers, respectively. and These are the bias vectors for the first and second layers, respectively;

[0011] Through expressions Obtain the fused feature vector ;in, For objective weighting, , Subjective weighting, , It is the Sigmoid activation function. , For trainable weight vectors, , For bias vectors, The probability distribution of the scene The scene encoding vector corresponding to the maximum value;

[0012] The fused feature vector Inputting dilated causal convolutional layers yields multi-scale features. The expression is ;in, For convolution kernel, The coefficient of thermal expansion is 1 / 3. For time step The weighted fusion feature vector;

[0013] The multi-scale features Inputting into a multi-head self-attention module yields features. Attention weight matrix The expression is ;in, , , and Let be the projection matrix. For attention head dimension;

[0014] The attention weight matrix After residual connection and layer normalization, the output is mapped to a comfort score distribution through a fully connected layer. The expression is ;in, This is the weight matrix. Given the corresponding bias vector, we obtain the expected value of the comfort index. ;

[0015] The fused feature vector Input a timing anomaly detector based on a gated loop unit to obtain the gated loop unit. At time step Hidden state vector The expression is ;in, For time step The weighted fusion feature vector, For the gated loop unit At time step The hidden state vector, For the gated loop unit The set of parameters;

[0016] Through expressions Get the warning trigger function ;in, For indicator functions, For time step The corresponding gated loop unit The hidden state moving mean vector, For comparison thresholds;

[0017] When the warning trigger function At that time, through the expression The anomaly intensity was calculated. ;in, For the gated loop unit At time step The hidden state vector, For time step The corresponding gated loop unit The hidden state moving mean vector;

[0018] According to the different abnormal intensities Obtain the corresponding regulatory strategies.

[0019] Specifically, through expressions The comparison threshold is calculated. ;in, and To set coefficients, For the gated loop unit The standard deviation of the hidden state within the warning window, , For the gated loop unit At time step The hidden state vector.

[0020] Specifically, it also includes:

[0021] Attention weight matrix from frame i to frame j Gaussian filtering is performed to obtain the filtered attention weight matrix. The expression is ;in, This represents the time decay coefficient of the Gaussian filter. The sequence number of the current frame. The sequence number of the target frame;

[0022] Through expressions Calculate the reset door ;in, and These are the weight matrix and bias vector of the reset gate, respectively, which force the gated loop unit to... Reset door Push it close to zero.

[0023] Specifically, it also includes:

[0024] Through expressions For the fused feature vector Modulation is performed to obtain the modulated fused feature vector. ;in, This represents element-wise multiplication. Empirical modulation coefficients Abnormal intensity The preset normalization upper limit;

[0025] The modulated fused feature vector Inputting into a dilated causal convolutional layer yields updated features. The updated features The input is processed by the multi-head self-attention module until a more accurate expected value of the comfort index is obtained.

[0026] Specifically, based on the different anomaly intensities Obtain the corresponding regulatory strategies, including:

[0027] Define the sound pressure gradient matrix The dominant noise component is separated by singular value decomposition, expressed as follows: , Maximum singular value The corresponding right singular vector Indicates the direction vector of the main noise source;

[0028] Through expressions Obtain the noise feature vector ;in, Noise type;

[0029] The regulation policy network model is optimized using a dual-delay deep deterministic policy gradient algorithm. The expression is ;in, These are the initial parameters of the control strategy network model. The parameters of the updated control strategy network model are... For learning rate, The starting parameter gradient operator, Let the action value function be... Here is the entropy regularization coefficient. For state space, , For action space;

[0030] The abnormal intensity The corresponding control strategy is obtained by inputting it into the control strategy network model.

[0031] Specifically, it also includes:

[0032] Through expressions The regulation benefit function was calculated. ;in, To control the delay in the effective date, This is the variance sensitivity coefficient;

[0033] like A set value is applied to the parameters of the control strategy network model. Regularization, and the corresponding regularization loss function for ;in, As the attenuation factor, The decay start time, This refers to the starting point of the control action.

[0034] Specifically, it also includes:

[0035] like Subjective comfort rating Comparison value, updated by parameter increment. Fine-tune the weights of the network model parameters of the aforementioned control strategy; wherein, , , To fine-tune the step size, For parameter mask, Let the mean squared error loss function be . For batch size, For the control strategy network model, the first Expected comfort index value for each sample For the first The actual human-labeled comfort scores for each sample.

[0036] Specifically, it also includes:

[0037] After implementing the aforementioned control strategy, if Less than the comparison threshold or improvement rate of heart rate variability For non-positive growth, a restart control strategy is explored by adding Gaussian noise, and the expression is: ;in, Gaussian noise, rolling back Output from historical strategy network ; for The change in equivalent sound pressure level over time. After regulation The equivalent sound pressure level at any given moment. To regulate the start time The equivalent sound pressure level, The noise barrier position adjustment amount output by the control strategy network model is the amount of noise barrier position adjustment. The sound attenuation coefficient of the sound barrier is... , After regulation Standard deviation of heart rate variability over time To regulate the start time The standard deviation of heart rate variability, This serves as the baseline reference value for heart rate variability.

[0038] Specifically, it also includes:

[0039] After implementing the aforementioned control strategy, if Less than the comparison threshold or improvement rate of heart rate variability If the growth is not positive, update the parameters of the control strategy network model again. ;in, for The change in equivalent sound pressure level over time. , After regulation The equivalent sound pressure level at any given moment. To regulate the start time The equivalent sound pressure level, The noise barrier position adjustment amount output by the control strategy network model is the amount of noise barrier position adjustment. The sound attenuation coefficient of the sound barrier is... , After regulation Standard deviation of heart rate variability over time To regulate the start time The standard deviation of heart rate variability, This serves as the baseline reference value for heart rate variability.

[0040] One or more technical solutions provided in this invention have at least the following technical effects or advantages:

[0041] 1. By integrating physical acoustic parameters, real-time physiological indicators, and contextualized subjective weights, and designing a cross-modal temporal alignment mechanism to address the asynchronous nature of multi-source data, this study ensures the causal relationship between acoustic stimuli and physiological responses. An assessment-early warning dual-stream network is established to achieve collaborative optimization of quantifiable comfort scoring and proactive intervention in abnormal states. This realizes a sound environment comfort regulation system based on multi-source heterogeneous data fusion, scene-adaptive weighting, and dual-stream collaborative decision-making. Specifically, addressing the spatiotemporal mismatch problem of acoustic, physiological, and subjective evaluation data, a cross-modal attention alignment model is constructed based on biological temporal constraints. Trainable spatiotemporal correlation weights are used to dynamically calibrate acoustic frames and physiological sampling points, ensuring a causal temporal relationship where acoustic stimuli precede physiological responses, thus forming a biologically plausible fusion feature expression method. Based on the characteristics of campus functional zoning, a scene-driven dynamic weight generation and feature transformation architecture is designed. A scene recognition network is constructed to map environmental parameters into scene encoding vectors, and dynamic weighting coefficients of physical acoustic features and physiological and psychological features are generated accordingly. Through feature zoning weighting operations, a scene-dependent spatial transformation matrix is ​​constructed to realize a differentiated evaluation paradigm that emphasizes physical indicators in high-concentration scenes such as classrooms and emphasizes physiological responses in high-activity scenes such as playgrounds.

[0042] 2. An evaluation-early warning dual-channel network was also established. Temporal convolutional networks and attention mechanisms were used to achieve quantitative evaluation of comfort. At the same time, a dynamic threshold early warning mechanism was constructed based on gated recurrent units. A feature distillation channel was designed to realize the transfer of attention weights from the evaluation branch to the early warning branch, as well as the inverse modulation of the evaluation feature extraction by the early warning anomaly score.

[0043] 3. A closed-loop management system for source tracing and feedback has also been established. Combining noise source localization, reinforcement learning strategy optimization, and acoustic-physiological dual verification, a "diagnosis-intervention-verification" control method has been formed to ensure the system's continuous self-optimization capability. Attached Figure Description

[0044] Figure 1 A flowchart of the acoustic environment comfort control method provided in the embodiments of the present invention. Detailed Implementation

[0045] This invention provides a method for regulating acoustic environment comfort, and constructs an acoustic environment comfort regulation system based on multi-source heterogeneous data fusion, scene adaptive weighting, and dual-stream collaborative decision-making.

[0046] The technical solutions in the embodiments of the present invention are designed to achieve the above-mentioned technical effects, and the overall concept is as follows:

[0047] This invention provides a method for evaluating and controlling the comfort of a campus acoustic environment based on multi-source data. By constructing a multimodal dynamic weighting and temporal early warning network, and following the pathway of "physical acoustic stimulus → physiological and psychological response → evaluation feedback," it integrates physical acoustic data, student physiological parameters, and subjective comfort scores. It dynamically allocates subjective and objective weights based on scene type to achieve synergistic optimization of acoustic comfort level assessment and anomaly early warning. The core technology lies in using a cross-modal temporal alignment mechanism to address the heterogeneity of multi-source data sampling frequencies, enhancing the model's generalization ability through scene-adaptive weighting, and establishing a dual-stream network to jointly optimize comfort assessment and early warning feedback, ultimately forming a source-tracing feedback mechanism. Through noise source localization, reinforcement learning control strategies, and online model calibration, it achieves accurate diagnosis, proactive intervention, and continuous optimization of campus acoustic environment problems, providing a systematic solution for intelligent acoustic environment governance.

[0048] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0049] like Figure 1 As shown, the acoustic environment comfort control method provided in this embodiment of the invention includes:

[0050] Step S110: Acquire acoustic data and convert the acoustic data into a time-frequency feature matrix S;

[0051] This step is explained in detail: acquiring acoustic data and converting it into a time-frequency feature matrix S includes:

[0052] Acquire acoustic data, segment the acoustic data time series into 125-millisecond frames, and generate a time-frequency feature matrix S through Mel spectrum transformation.

[0053] Step S120: Input the time-frequency feature matrix S into the convolutional neural network for deep feature extraction to obtain the acoustic feature vector. The expression is ;in, The convolution kernel weight matrix is... is the corresponding bias vector, and ReLU is the activation function to enhance nonlinear expressive power.

[0054] Step S130: Construct physiological vectors , physiological vector Inputting the data into a long short-term memory network for feature modeling yields a sequence of physiological states. It effectively captures physiological rhythm characteristics such as heart rate variability, and the expression is: ;in, This refers to the parameter set of the Long Short-Term Memory network; specifically, for physiological monitoring data (body temperature)... Heart rate Blood oxygen saturation Constructing three-dimensional physiological vectors .

[0055] Step S140: Convert the acoustic feature vector and physiological state sequence Inputting the cross-modal attention gating unit yields the association weights. The expression is ;in, For the first Frame acoustic features and the first The unnormalized correlation score of the th physiological sampling point represents the th Frame acoustic features and the first The strength of the biophysical association between physiological sampling points , and For trainable projection matrices, For attention vectors, For the first Acoustic feature vectors of time frames, For the first Physiological feature vectors of each sampling point For the first Frame acoustic features and the first The unnormalized correlation score of each physiological sampling point is used to dynamically calculate the association weight between the acoustic frame and the physiological sampling point through a cross-modal alignment mechanism to achieve accurate alignment.

[0056] Step S150: Through the expression Obtain the aligned physiological characteristics Weight Quantifying acoustics Frame and Physiology The correlation of sampling points is used to generate aligned physiological features. .

[0057] Step S160: Set user coordinates Environmental sensor tags (Four-dimensional coding represents classrooms, playgrounds, dormitories, libraries, etc.) and activity intensity index (Quantized from motion sensors) Input into a three-layer fully connected network to obtain the scene probability distribution. The expression is ;in, These are the hidden layer feature vectors of a three-layer fully connected network. , and These are the weight matrices for the first and second layers, respectively. and These are the bias vectors for the first and second fully connected networks, respectively. The Dropout layer prevents overfitting. Based on the acoustic characteristics and human response patterns of different functional areas on campus, the contribution weights of physical acoustic features and physiological and psychological features in comfort evaluation are dynamically adjusted.

[0058] Step S170: By expression Obtain the fused feature vector ;in, For objective weighting, , Subjective weighting, , It is the Sigmoid activation function. , For trainable weight vectors, , For bias vectors, Scene probability distribution The scene encoding vector corresponding to the maximum value; objective weights With subjective weight Satisfying constraints This operation essentially constructs a scene-dependent feature space transformation matrix. ,in and The identity matrix. Weighted features. It can improve the model's generalization ability.

[0059] Step S180: Combine the fused feature vectors Inputting dilated causal convolutional layers yields multi-scale features. The expression is ;in, For convolution kernel, The coefficient of thermal expansion ( (for the expansion cycle) For time step Weighted fusion feature vectors; ensure output Relying only on historical frames ( ).

[0060] Step S190: Combine multi-scale features Inputting into a multi-head self-attention module yields features. Attention weight matrix The expression is ;in, , , and Let be the projection matrix. For attention head dimension;

[0061] Step S200: Adjust the attention weight matrix After residual connection and layer normalization, the output is mapped to a comfort score distribution through a fully connected layer. The expression is ;in, This is the weight matrix. Given the corresponding bias vector, we obtain the expected value of the comfort index. Among them, the output Given a 5-level comfort probability vector, its expected value is... As the ultimate comfort index.

[0062] Step S210: Combine the fused feature vectors Input a timing anomaly detector based on a gated loop unit to obtain the gated loop unit. At time step Hidden state vector The expression is ;in, For time step The weighted fusion feature vector, Gated loop unit At time step The hidden state vector, Gated loop unit The set of parameters;

[0063] Step S220: By expression Get the warning trigger function ;in, For indicator functions, For time step Corresponding gated loop unit The hidden state moving mean vector, For comparison thresholds;

[0064] Specifically, through expressions The comparison threshold was calculated. ;in, and To set the coefficients, by Dynamic adjustment, library scene Playground scene . Gated loop unit The standard deviation of the hidden state within the warning window, , Gated loop unit At time step The hidden state vector.

[0065] In order to incorporate the attention weight matrix of the comfort assessment branch The Gaussian-filtered prior knowledge is used as the warning branch to achieve knowledge sharing between the comfort assessment branch and the warning branch, and also includes:

[0066] Attention weight matrix from frame i to frame j Perform Gaussian filtering to obtain the filtered attention weights. The expression is ;in, This represents the time decay coefficient of the Gaussian filter. The sequence number of the current frame. The sequence number of the target frame is used; this operation strengthens the correlation between adjacent frames. The original weights are susceptible to transient noise interference. Gaussian filtering smooths random fluctuations while preserving the correlation between key frames, which is consistent with the inertial characteristics of physiological response.

[0067] Through expressions Calculate the reset door ;in, and These represent the weight matrix and bias vector of the reset gate, and the forced gated recurrent unit. Reset door Pushing the value close to zero clears irrelevant historical states to reduce false positive rates and scene response latency.

[0068] Step S230: When the warning trigger function is activated When this occurs, it indicates an abnormal acoustic environment, indicated by the expression. The anomaly intensity was calculated. ;in, Gated loop unit At time step The hidden state vector, For time step Corresponding gated loop unit The hidden state moving mean vector;

[0069] In order to output the abnormal score through the early warning branch The input feature intensity of the comfort assessment branch is modulated through a nonlinear scaling mechanism in the feature space. This enhances the sensory channel gain to improve the signal-to-noise ratio when environmental anomalies are detected, thereby improving the accuracy of modulation. The method also includes:

[0070] Through expressions For fused feature vectors Modulation is performed to obtain the modulated fused feature vector. ;in, This represents element-wise multiplication. Empirical modulation coefficients Abnormal intensity The preset normalization upper limit. Abnormal scores compressed to modulation intensity with Monotonically increasing enhances the characteristic expression of abnormal periods.

[0071] Modulated fused feature vector Inputting into a dilated causal convolutional layer yields updated features. The updated features The input is processed by a multi-head self-attention module until a more accurate expected value of the comfort index is obtained. By enhancing the representation of features during abnormal periods, the sensitivity of the comfort score to sudden changes in the acoustic environment is improved.

[0072] Step S240: Based on different anomaly intensities Obtain the corresponding control strategy. For example, for level 1, send a sound pressure level optimization command to the environmental control system. ( (For level 2, activate the user terminal vibration alert and push soothing audio; for level 3, link the building management system to activate the sound barrier).

[0073] This step will be explained in detail, depending on the different anomaly intensities. Obtain the corresponding regulatory strategies, including:

[0074] First, noise source identification and feature decoupling are performed, and the main noise sources are located using spatiotemporal analysis of multi-channel acoustic signals. A sound pressure gradient matrix is ​​then defined. The dominant noise component is separated by singular value decomposition, expressed as follows: , Maximum singular value The corresponding right singular vector Indicates the direction vector of the main noise source;

[0075] Through expressions Obtain the noise feature vector It is used to guide precise regulation; among them, The noise type; specifically, based on scene information. Decoupling noise types. Among them, mechanical noise is... ( (referring to the reference direction of the air conditioning equipment), voice interference is (Extract the signal-to-noise ratio of the 300-3400Hz resonant peak).

[0076] The regulation policy network model is optimized using a dual-delay deep deterministic policy gradient algorithm. The expression is ;in, To adjust the initial parameters of the policy network model, The parameters of the updated regulation strategy network model, For learning rate, For initial parameters gradient operator, Let the action value function be... Here is the entropy regularization coefficient. For state space, , For the action space, including physical control (sound barrier position) Air conditioner fan speed ) and physiological regulation (audio gain) );

[0077] abnormal intensity The corresponding control strategy is obtained by inputting it into the control strategy network model.

[0078] To implement the model parameter weight decay update mechanism, thereby achieving dynamic calibration of model parameters and ensuring that the model maintains high accuracy and robustness during long-term operation, the following measures are also included:

[0079] Through expressions The regulation benefit function was calculated. ;in, To control the delay in the effective date, This is the variance sensitivity coefficient;

[0080] like Set values ​​are applied to the parameters of the control strategy network model. Regularization, and the corresponding regularization loss function for ;in, As the attenuation factor, The decay start time, This refers to the starting moment of the control action. Specifically, when Time regularization is enhanced. By penalizing large weights, the model complexity is reduced, the risk of overfitting is suppressed, and the generalization ability of scene recognition is ensured.

[0081] To implement the model parameter weight enhancement update mechanism, triggering incremental learning to achieve dynamic calibration of model parameters and further ensure that the model maintains high accuracy and robustness in long-term operation, the following measures are also included:

[0082] like Subjective comfort rating Comparison value, updated by parameter increment. The weights of the network model parameters for fine-tuning the control strategy; where, , , To fine-tune the step size, As a parameter mask, only the last dimension of the fully connected layer is updated; is the mean squared error loss function, used to quantify the deviation between the model's predicted comfort level and the actual human annotations. For batch size, For the regulation strategy network model, the first Expected comfort index value for each sample For the first The comfort scores were manually labeled based on real samples. This addresses model prediction drift caused by sudden environmental changes. By fine-tuning key parameters using small-step gradient descent, prediction errors can be quickly reduced, avoiding the resource consumption of global retraining.

[0083] To achieve a closed-loop feedback verification process, the following is also included:

[0084] After implementing the control strategy, if The deviation from the theoretical value of the expected noise reduction effect must be less than the tolerance threshold. This verification is directly related to the acoustic feature extraction results mentioned above. ) and noise source identification results (such as the main noise direction vector) If the deviation exceeds the limit, it indicates that the physical regulation has not achieved the expected results. Or, the improvement rate of heart rate variability. For non-positive growth, a restart control strategy is explored by adding Gaussian noise, and the expression is: ;in, Gaussian noise, rolling back Output from historical strategy network ; for The change in equivalent sound pressure level over time. , After regulation The equivalent sound pressure level at any given moment. To regulate the start time The equivalent sound pressure level, The adjustment amount of the sound barrier position output by the control strategy network model. The sound attenuation coefficient of the sound barrier is... , After regulation Standard deviation of heart rate variability over time To regulate the start time The standard deviation of heart rate variability, This serves as a baseline reference value for heart rate variability. Noise injection is used to ensure exploration efficiency while avoiding significant deviations from the optimized strategy.

[0085] Furthermore, after implementing the control strategy, if Less than the comparison threshold or improvement rate of heart rate variability For non-positive growth, update the parameters of the regulatory strategy network model. .

[0086] In summary, this invention overcomes three major bottlenecks by constructing a complete system encompassing multi-source heterogeneous data fusion, scene-adaptive weighting, dual-stream collaborative decision-making, and closed-loop traceability feedback: First, it utilizes a cross-modal temporal alignment mechanism to accurately correlate acoustic stimuli with physiological responses, solving the problem of asynchronous data and improving assessment accuracy. Second, it achieves real-time weight calibration of physical indicators and physiological characteristics based on dynamic scene recognition, enhancing the model's cross-scene generalization ability. Third, it designs a closed-loop architecture of assessment-early warning dual-stream network and reinforcement learning regulation, realizing a complete optimization chain from anomaly diagnosis to proactive intervention, providing traceable intelligent decision support for campus acoustic environment governance.

[0087] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, 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, optical storage, etc.) containing computer-usable program code.

[0088] 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, and 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 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.

[0089] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer 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.

[0090] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer 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.

[0091] Any aspects of this invention not described in detail in the embodiments are well-known techniques to those skilled in the art. Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this invention and not to limit it. Although this invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of this invention without departing from the spirit and scope of this invention, and all such modifications and substitutions should be covered within the scope of the claims of this invention.

Claims

1. A method for controlling acoustic environment comfort, characterized in that, include: Acquire acoustic data and convert the acoustic data into a time-frequency feature matrix S; The time-frequency feature matrix S is input into a convolutional neural network for deep feature extraction to obtain the acoustic feature vector. ; Constructing physiological vectors , the physiological vector Inputting the data into a long short-term memory network for feature modeling yields a sequence of physiological states. ; The acoustic feature vector and the physiological state sequence Inputting the cross-modal attention gating unit yields the association weights. ; Through expression Obtain the aligned physiological characteristics ;in, For the first Physiological feature vectors of each sampling point; User coordinates, environmental sensor labels, and activity intensity index are input into a three-layer fully connected network to obtain the scene probability distribution. ; Through expression Obtain the fused feature vector ;in, For objective weighting, , Subjective weighting, , It is the Sigmoid activation function. For trainable weight vectors, For bias vectors, The probability distribution of the scene The scene encoding vector corresponding to the maximum value, For the first Acoustic feature vectors of time frames; The fused feature vector Input a timing anomaly detector based on a gated loop unit to obtain the gated loop unit. At time step Hidden state vector ; Through expression Get the warning trigger function ;in, For indicator functions, For comparison thresholds; When the warning trigger function At that time, through the expression The anomaly intensity was calculated. ;in, For the gated loop unit At time step The hidden state vector, For time steps The corresponding gated loop unit The hidden state moving mean vector; According to different abnormal intensities Obtain the corresponding regulatory strategies.

2. The acoustic environment comfort control method as described in claim 1, characterized in that, Through expression The comparison threshold is calculated. ;in, and To set coefficients, For the gated loop unit The standard deviation of the hidden state within the warning window. , For the gated loop unit At time step The hidden state vector, For time steps The corresponding gated loop unit The hidden state moving mean vector.

3. The acoustic environment comfort control method as described in claim 1, characterized in that, Also includes: Attention weight matrix from frame i to frame j Gaussian filtering is performed to obtain the filtered attention weight matrix. The expression is ;in, This represents the time decay coefficient of the Gaussian filter. The sequence number of the current frame. The sequence number of the target frame; Through expression Calculate the reset door ;in, and These are the weight matrix and bias vector of the reset gate, respectively, which force the gated loop unit to... Reset door Push it close to zero.

4. The acoustic environment comfort control method as described in claim 1, characterized in that, Also includes: Through expression For fused feature vectors Modulation is performed to obtain the modulated fused feature vector. ;in, This represents element-wise multiplication. Empirical modulation coefficients Abnormal intensity The preset normalization upper limit; The modulated fused feature vector Inputting into a dilated causal convolutional layer yields updated features. The updated features The input is processed by a multi-head self-attention module until a more accurate expected value of the comfort index is obtained.

5. The acoustic environment comfort control method as described in claim 1, characterized in that, According to different abnormal intensities Obtain the corresponding regulatory strategies, including: Define the sound pressure gradient matrix The dominant noise component is separated by singular value decomposition, expressed as follows: Maximum singular value The corresponding right singular vector Indicates the direction vector of the main noise source; Through expression Obtain the noise feature vector ;in, Noise type; The regulation policy network model is optimized using a dual-delay deep deterministic policy gradient algorithm. The expression is ;in, These are the initial parameters of the control strategy network model. The parameters of the updated control strategy network model are... For learning rate, The starting parameter gradient operator, Let the action value function be... Here is the entropy regularization coefficient. For state space, , For action space; The abnormal intensity The corresponding control strategy is obtained by inputting it into the control strategy network model.

6. The acoustic environment comfort control method as described in claim 5, characterized in that, Also includes: Through expression The regulation benefit function was calculated. ;in, To regulate the delay in the effective date, This is the variance sensitivity coefficient; like A set value is applied to the parameters of the control strategy network model. Regularization, and the corresponding regularization loss function for ;in, As the attenuation factor, The decay start time, This refers to the starting point of the control action.

7. The acoustic environment comfort control method as described in claim 5, characterized in that, Also includes: like Subjective comfort rating Comparison value, updated by parameter increment. Fine-tune the weights of the network model parameters of the control strategy; wherein, , To fine-tune the step size, For parameter mask, Let the mean squared error loss function be . For batch size, For the control strategy network model, the first Expected comfort index value for each sample For the first The actual human-labeled comfort scores for each sample.

8. The acoustic environment comfort control method as described in claim 5, characterized in that, Also includes: After implementing the aforementioned control strategy, if Less than the comparison threshold or improvement rate of heart rate variability For non-positive growth, a restart control strategy is explored by adding Gaussian noise, and the expression is: ;in, Gaussian noise, rolling back Output from historical strategy network ; for The change in equivalent sound pressure level over time. After regulation The equivalent sound pressure level at any given moment. To regulate the start time The equivalent sound pressure level, The noise barrier position adjustment amount is output by the control strategy network model. The sound attenuation coefficient of the sound barrier is... , After regulation Standard deviation of heart rate variability over time To regulate the start time The standard deviation of heart rate variability, This serves as the baseline reference value for heart rate variability.

9. The acoustic environment comfort control method as described in claim 5, characterized in that, Also includes: After implementing the aforementioned control strategy, if Less than the comparison threshold or improvement rate of heart rate variability If the growth is not positive, update the parameters of the control strategy network model again. ;in, for The change in equivalent sound pressure level over time. , After regulation The equivalent sound pressure level at any given moment. To regulate the start time The equivalent sound pressure level, The noise barrier position adjustment amount is output by the control strategy network model. The sound attenuation coefficient of the sound barrier is... , After regulation Standard deviation of heart rate variability over time To regulate the start time The standard deviation of heart rate variability, This serves as the baseline reference value for heart rate variability.

Images