A low-voltage power supply environment noise tracing edge computing method and system

Abstract

The application discloses a low-voltage power supply environment noise tracing edge computing method and system, comprising: synchronously collecting data through a multi-modal sensing module, constructing a model to predict historical noise space-time distribution, combining statistical anomaly judgment and environment noise limit to divide multi-class noise events, and dynamically scheduling local lightweight models and cloud high-precision models according to event priority to complete tracing analysis; the system comprises an edge device layer and a cloud server layer, and realizes resource optimization configuration through local and cloud collaborative architecture. Through the dual-mode power supply architecture, multi-dimensional dynamic analysis and intelligent computing power scheduling mechanism, the application significantly improves the detection rate of implicit noise events, reduces the cloud load, and supports accurate classification of five categories of noise such as nature and traffic, and fine-grained sub-classes, thereby providing an efficient, low-power and high-precision technical solution for environment noise control.

Description

Technical Field

[0001] This invention relates to the field of environmental noise identification technology, specifically to a method and system for edge computing to trace the source of low-voltage power supply environmental noise. Background Technology

[0002] Currently, noise control faces challenges such as insufficient monitoring coverage and delayed response. To ensure the timeliness and accuracy of noise control and to build a closed loop of real-time assessment of acoustic environment quality and precise noise source tracing, there is an urgent need for intelligent monitoring equipment with rapid deployment capabilities and support for efficient multi-dimensional analysis. However, existing noise source tracing equipment still faces three core contradictions that urgently need to be addressed when meeting practical needs: First, there is a mismatch between power supply mode and deployment flexibility. Current equipment is mainly divided into two categories: fixed and mobile. Although fixed equipment has the advantage of stable power supply, it is limited by the conditions of mains power access and is difficult to deploy quickly in temporary or power-free areas. Mobile equipment can be deployed flexibly, but it is limited by the low-voltage power supply environment and cannot support the high computing power unit required for high-precision noise source tracing, resulting in limited monitoring depth. There is an urgent need for equipment that can quickly respond to complaints to meet the requirements of noise pollution prevention and control (China Noise Pollution Prevention and Control Report (2025)).

[0003] Second, there is a disconnect between static threshold determination and the elements of noise pollution. Existing noise source tracing and monitoring systems all use static environmental noise limits as the judgment benchmark. For example, Song Weihua et al.'s "Noise Monitoring Data Analysis Method and Device Based on Threshold Linkage Trigger Strategy" (CN 118329188B) only performs source tracing analysis on audio exceeding the threshold, ignoring the elements of noise pollution that may exist during the compliance period and disturb others' normal life, work, and study. For example, nighttime construction noise may be misjudged as a compliant event because it does not exceed the threshold, or sudden noise may be ignored by the system because it does not trigger the fixed threshold. This judgment mechanism, lacking multi-dimensional dynamic analysis, makes it difficult to comprehensively assess the noise pollution situation, not only reducing the accuracy of source tracing but also potentially masking potential pollution sources and weakening the precision of noise control.

[0004] Third, there is an imbalance between the demand for computing power and the allocation of resources. Currently, noise source identification mainly relies on high-precision cloud models or local lightweight models. Although cloud models are highly accurate, with the establishment of the automatic monitoring network for the acoustic environment quality of functional areas, the contradiction between the real-time processing demand for massive data and the bottleneck of cloud computing power has become prominent. Taking Guangzhou as an example, 900,000 recordings of noise exceeding the standard were collected from January to November 2024. The amount of data to be analyzed is large, and the server is under great pressure (Guangzhou Daily. Guangzhou has set up 20 monitoring points throughout the city to fully build an automatic monitoring network for the acoustic environment of functional areas [EB / OL]. (2024-11-04).). Although local models can efficiently process conventional noise such as traffic noise at fixed times, they are difficult to distinguish specific noise sources in complex scenarios such as mixed noise from multiple sources and sudden noise due to insufficient computing power, and the reliability of identification has decreased significantly (Li Xia. Analysis of the current situation and improvement measures of noise pollution prevention and control in Guangdong Province [J]. China Resources Comprehensive Utilization, 2024, 42(10):159-161.). Summary of the Invention

[0005] This invention provides a method and system for edge computing to trace the source of noise in low-voltage power supply environments. Through multi-dimensional data analysis and dynamic computing power scheduling mechanism, it effectively solves the core problems in the prior art, such as inflexible deployment, incomplete source analysis, and uneven distribution of computing power.

[0006] The present invention is achieved by at least one of the following technical solutions.

[0007] A method for noise source edge calculation in low-voltage power supply environments includes the following steps: Step S1: Collect multimodal noise data; Step S2: Preprocess the collected multimodal noise data to obtain the noise temporal characteristics; Step S3: Input the time-series noise features into the trained historical noise distribution prediction model to predict the historical noise distribution; Step S4: Classify dynamic noise events based on the predicted historical noise distribution; Step S5: Perform intelligent computing power scheduling and source tracing for the divided noise events.

[0008] Furthermore, the preprocessing includes: cleaning the raw data; deleting samples with missing values ​​exceeding a set threshold; interpolating and filling in missing values ​​for the remaining data using the daily mean interpolation method; normalizing the interpolated continuous data; and converting the normalized data into a supervised learning format. A time-series sliding window method is used to construct input-output pairs usable for supervised learning from the normalized time-series data. By creating lag terms with specific time steps for each feature, an input feature matrix containing historical time-series information is formed, and the actual observations at the corresponding times are used as the output targets. This transforms the continuous time-series data into supervised learning data with a clear feature-label correspondence.

[0009] Furthermore, the training steps for the historical noise distribution prediction model include the following: Data preparation and preprocessing: Preprocessing is performed on the collected multimodal noise data; LSTM model architecture design: Construct a multi-layer long short-term memory network (LSTM) architecture, including an input layer, multi-level stacked LSTM hidden layers, and a fully connected output layer. Dynamic capture of temporal features is achieved through a gating mechanism. Hyperparameter optimization system construction: Establish a search space including parameters such as learning rate, batch size, and number of LSTM stacking layers, construct a probabilistic model using Bayesian optimization algorithm, and optimize parameters with the coefficient of determination as the optimization objective function; Model training and optimization: Iterative training of the model is performed within the hyperparameter search space, and a historical noise distribution prediction model with strong generalization ability is output through cross-validation strategy; Historical noise distribution interval calculation: based on model prediction results and 3 The principle is to construct a dynamically updated confidence interval in real time to characterize the reasonable fluctuation range of noise values. This represents the standard deviation of the residuals.

[0010] Furthermore, the historical noise distribution interval is calculated as follows: Residual sequence extraction: Calculate the difference between the true value sequence and the predicted value sequence of the historical noise distribution prediction model on the test set, and construct the residual sequence; Residual statistical analysis: Calculating the standard deviation of residuals based on residual sequences ; Real-time prediction: Input the noise time series characteristic data of the current period into the historical noise distribution prediction model, and output the noise baseline value of the current period. ; Historical noise distribution interval generation: based on the current noise baseline value With 3 Based on the principle, a dynamic confidence interval is constructed to characterize the reasonable fluctuation range of noise values ​​in the current period, providing a data benchmark for the statistical anomaly judgment of subsequent dynamic noise event classification.

[0011] Furthermore, the dynamic noise events are classified, including the following steps: Step 41: Historical noise distribution prediction: Based on the historical noise distribution intervals, construct a dynamic confidence interval that reflects the historical noise distribution in the current time period; Step 42, Statistical Anomaly Detection: By comparing the current noise monitoring values By comparing the noise distribution range with historical data, the statistical characteristics of noise events are determined, and a three-level statistical anomaly weight is defined to reflect the current noise monitoring value. The degree of deviation from the historical noise distribution range is defined as: Abnormally high weight: A weight of 4 indicates that the current noise value significantly exceeds the historical noise distribution; Abnormally low weight: A weight of 3 indicates that the current noise value is significantly lower than the historical noise distribution. Regular weights: A weight of 1 indicates that the current noise value conforms to the historical noise distribution; Step 43, Determining Environmental Noise Limits: According to the environmental noise limits defined in the environmental noise quality standards Determine the compliance of noise events and define limit judgment weights to reflect the current noise level. Whether it complies with the corresponding environmental noise standard is defined as: Excess weight: A weight of 2 indicates that the current noise level exceeds the standard and needs to be addressed promptly. Target weight: A weight of 1 indicates that the current noise value meets the standard. Step 43, Joint Event Classification: By combining the dual criteria of statistical anomaly detection and environmental noise limit determination, and dynamically determining the event processing priority through the product of the two weights, noise events are classified as follows: The abnormally high exceedance event T1 has a weighted product of 4 × 2 = 8. This is characterized by a current noise value that is significantly higher than the historical noise distribution and a sound pressure level that exceeds the standard, which is a high-priority event. For an abnormally low exceedance event T2: the weighted product is 3 × 2 = 6. This is characterized as a current noise value that is significantly lower than the historical noise distribution, but the sound pressure level still exceeds the standard, and it belongs to the second priority event. The abnormally high pass rate for event T3: the weighted product is 4 × 1 = 4. This is characterized by noise values ​​that are significantly higher than the historical noise distribution but do not exceed the standard, which is a potential risk event that needs further verification. The abnormally low pass rate for event T4 has a weighted product of 3 × 1 = 3. This is characterized by noise values ​​that are significantly lower than historical noise distributions and meet the standards, which is an abnormal event that requires further verification. For a typical out-of-limit event T5, the weighted product is 1 × 2 = 2. and This is characterized by noise values ​​that conform to the historical noise distribution but continuously exceed the standard, which is a regular exceedance event; For a typical achievement event T6, the weighted product is 1 × 1 = 1. and This indicates that the noise value conforms to the historical noise distribution and meets the standard, which is a normal event.

[0012] Furthermore, step S5 involves constructing a dual-end collaborative strategy driven by the noise event classification results, based on the classification of noise events. This strategy dynamically allocates computing resources and performs differentiated processing according to event characteristics. The specific strategy is as follows: The abnormally high exceedance event T1 has the dual characteristics of noise value significantly higher than the historical noise distribution and sound pressure level exceeding the standard. To ensure the accuracy of source tracing, it is uploaded to the cloud server first, and the high-precision cloud model is used for noise source classification and source tracing analysis. T2 is an abnormally low-priority event: the noise value is significantly lower than the historical noise distribution but the sound pressure level still exceeds the standard. It belongs to the secondary priority event. First, it is initially classified by the local lightweight model. If the confidence of the local model is lower than the threshold, it is uploaded to the cloud during the idle period to use the high-precision model for secondary verification. T3, an event with an abnormally high noise level: The noise value is significantly higher than the historical noise distribution but does not exceed the standard. It is a potential risk event that needs further verification. First, trend analysis is performed through the local model. If a continuous upward trend or the characteristics of sporadic noise sources are detected, the cloud model is triggered to perform source tracing analysis. T4, an event with an abnormally low compliance rate: The noise value is significantly lower than the historical noise distribution and meets the standard. It is an abnormal event that needs further verification. A comprehensive judgment is made based on the context data. If it is determined to be an occasional event, it is archived locally. If it is determined to be a potential anomaly, the local model is used for source tracing analysis. Typical out-of-limit events T5: Noise values ​​conform to historical noise distribution but continue to exceed limits. These are usually interpretable, routine noise, and the noise source classification and recording are completed directly by the local lightweight model. T6 is a normal event that meets the standard. The noise value is consistent with the historical noise distribution and meets the standard. It does not require real-time analysis and processing. It is only uploaded to the cloud for backup during system idle periods for long-term trend analysis.

[0013] Furthermore, the source analysis results are stored in the database in real time, and a visualization interface is built using Vue3+Fastapi to dynamically display key information such as the proportion of noise sources and the distribution of sound pressure levels.

[0014] The device for implementing the aforementioned method for tracing the source edge of noise in a low-voltage power supply environment includes: Multimodal sensing module: integrates a high-precision digital omnidirectional microphone and sound pressure level sensor, and a multi-parameter meteorological sensor. It achieves millisecond-level spatiotemporal synchronous acquisition of sound pressure level data, meteorological parameters and environmental audio through a time synchronization circuit, and supports parallel processing of multi-channel data streams; Network communication module: Integrates 4G / 5G dual-mode communication module and adopts dynamic bandwidth allocation strategy: monitoring data and local identification results use low bandwidth transmission channel, and automatically upgrade to high-speed channel for abnormally high exceedance event T1 or abnormally low exceedance event T2; the communication module supports breakpoint resume and encrypted data transmission to ensure reliable transmission of environmental audio and analysis results; Real-time display module: includes a capacitive touch screen for integrating a visualization system to display multi-dimensional noise characteristics in real time, including second-level sound pressure level time-series curves, noise source proportion pie charts, and noise event time period information; Edge computing unit: Pre-installed with a dual-modal computing framework, supporting the concurrent deployment of local lightweight models and historical noise distribution prediction models to meet real-time analysis needs; Power management module: Includes hot-swappable lithium battery pack and intelligent power management unit, with mains power priority power supply strategy and battery life optimization algorithm.

[0015] The system for implementing the aforementioned method for tracing the source edge of noise in a low-voltage power supply environment includes: Edge device layer: As a data sensing and decision-making unit deployed at the monitoring site, it is used to collect data in real time and make rapid decisions and responses. Based on the noise event classification results, it responds and processes routine events in real time; for complex events, it automatically initiates collaborative analysis requests with the cloud server layer and completes in-depth judgment with the help of high-precision models in the cloud. Cloud server layer: Built on the cloud platform, including: running high-precision analysis models and handling complex tracing tasks; continuously optimizing the model algorithm and iterating the version, and actively pushing the optimized model to the edge.

[0016] The present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0017] Compared with the prior art, the present invention can achieve at least the following beneficial effects: 1. Dual-mode power supply architecture breaks through deployment bottlenecks: It adopts hot-swappable power management modules to build a two-level hybrid power supply system of "mains power + energy storage". The equipment can be flexibly deployed in complex scenarios such as municipal roads, construction sites, and community squares. The single device can last for up to 48 hours and supports continuous monitoring and source analysis of sudden noise events. In addition, the modular design avoids the duplication of mains power infrastructure construction and reduces deployment costs.

[0018] 2. Multidimensional Dynamic Analysis Enhances Source Tracing Accuracy: A three-dimensional analysis model integrating noise, meteorology, and spatiotemporal data is constructed. By merging environmental noise monitoring data, meteorological parameters (wind speed, humidity, temperature), and spatiotemporal characteristics (time period, location), a dynamic noise pollution assessment system is established. The system employs deep learning algorithms to mine the spatiotemporal distribution patterns of urban noise, achieving a shift from single-threshold judgment to multidimensional noise event classification. Testing and verification have shown that in scenarios such as nighttime construction noise and sudden noise during holidays, the system improves the detection rate of latent noise events by 45% and reduces the false alarm rate to below 8%.

[0019] 3. Dynamic Scheduling Mechanism Optimizes Computing Power Allocation: An intelligent computing power scheduling algorithm based on event classification is designed to construct a collaborative processing architecture of "local lightweight model + cloud high-precision model". The system dynamically allocates computing resources according to the type of noise event: for simple scenarios such as regular traffic noise, the local model achieves millisecond-level response; for complex scenarios such as multi-source mixing and sudden noise, edge computing triggers cloud model collaborative processing to achieve accurate source tracing of abnormal events. Test data shows that under the same computing power conditions, the system resource utilization rate is improved by 70%, the cloud load is reduced by 50%, the response time for abnormal event handling is shortened to within 3 seconds, and the overall system performance is improved by 55%. Attached Figure Description

[0020] Figure 1 This is a flowchart of the steps of a method for edge calculation of noise source tracing in a low-voltage power supply environment provided by an embodiment of the present invention.

[0021] Figure 2 This is a flowchart of noise timing feature preprocessing in an embodiment of the present invention.

[0022] Figure 3 This is a performance evaluation diagram of the historical noise distribution prediction model in this embodiment of the invention.

[0023] Figure 4 This is a flowchart illustrating the calculation of the historical noise distribution range in an embodiment of the present invention.

[0024] Figure 5 This is a schematic diagram of the structure of an edge computing device for tracing noise sources in a low-voltage power supply environment provided in this embodiment.

[0025] Figure 6This is a front view of an edge computing device for tracing noise sources in a low-voltage power supply environment, according to an embodiment.

[0026] Figure 7 This is a schematic diagram of the architecture of an edge computing system for tracing noise sources in a low-voltage power supply environment, provided in another embodiment of the present invention. Detailed Implementation

[0027] 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0028] Example 1 Please see Figures 1-2 The method for edge calculation of noise source tracing in a low-voltage power supply environment according to this embodiment specifically includes the following steps: Step S1: Multimodal noise data acquisition.

[0029] In this embodiment, ambient sound pressure level data is collected by a high-precision digital sound level meter, meteorological parameters such as wind speed and rainfall are collected synchronously by a multi-parameter meteorological sensor, and ambient audio is recorded using an omnidirectional microphone at a sampling rate of 48kHz. The sound pressure level data, meteorological parameters and ambient audio are synchronized at the millisecond level through a time synchronization circuit.

[0030] Step S2: Preprocessing of noise temporal features.

[0031] In this embodiment, the multimodal noise data collected in step S1 is preprocessed to extract noise temporal features. The noise temporal features include time features (month, day, hour, week, whether it is a weekend), environmental noise features (Leq, Lmax, Lmin, L5, L10, L50, L90, L95, effective acquisition rate), and meteorological features (wind speed, rainfall).

[0032] In this embodiment, the data undergoes the following preprocessing.

[0033] S201. Data cleaning: Remove samples with a missing rate greater than 30%.

[0034] S202. Complete the remaining data using interpolation, filling missing values ​​with the daily average. Assuming the environmental noise characteristics are stable throughout the day, filling missing values ​​with the daily average will not introduce significant bias, and the calculation is simple and easy to implement, making it suitable for completing time-series data such as environmental noise and meteorological data. The calculation formula is:

[0035] in: This is the arithmetic mean of the environmental noise characteristics for the day, reflecting the typical level of that characteristic on that day; The non-missing values ​​of the same feature on that day. This represents the number of non-missing samples.

[0036] S203. Normalize the interpolated continuous data by using Max-Min Scaling to map the data to the [0,1] interval. The formula is as follows:

[0037] in: These are the normalized eigenvalues, with a range of [0,1]. These are the original feature values ​​(such as the Leq value at a certain moment). The maximum value in the sample. It is the minimum value in the sample.

[0038] S204. The normalized data is converted to a supervised learning format. The input time step is set to 24 and the output time step to 1. The time series sliding window method is used to construct the normalized time series data into an input-output pair that can be used for supervised learning. By creating lag terms of a specific time step for each feature (t-1, t-2, ..., t-n_in, where n_in is the input time step, indicating how many past time steps of features are used as input; t-n_in represents the feature values ​​of the previous n_in time steps of the current time step t, such as t-1 representing the features of the previous 1 hour, t-2 representing the features of the previous 2 hours, and so on), an input feature matrix containing historical time series information is formed. The actual observation value at the corresponding time is used as the output target, thereby converting the continuous time series data into supervised learning data with a clear feature-label correspondence.

[0039] Step S3: Use the trained model to predict the historical noise distribution.

[0040] S301. Construction and training of historical noise distribution prediction model.

[0041] In this embodiment, a historical noise distribution prediction model is constructed. This step includes data preparation and preprocessing, LSTM model architecture design, hyperparameter optimization system construction, model training and optimization, and calculation of historical noise distribution intervals.

[0042] The data preparation and preprocessing process involves selecting historical data from noise monitoring stations and executing steps S201-S204 to construct an environmental noise time-series feature dataset.

[0043] The LSTM model architecture employs a multi-layer Long Short-Term Memory (LSTM) network. Each LSTM layer includes an input layer, multiple stacked LSTM hidden layers (integrating forget gates, input gates, and output gates), and a fully connected output layer. This gating mechanism enables dynamic capture of temporal features. The final LSTM layer outputs only the feature vector from the last time step, and a linear layer maps the LSTM output to the predicted value. The LSTM forward propagation formula is as follows:

[0044]

[0045]

[0046]

[0047]

[0048] In the formula, the input gate Determine the current moment's information input information The screening, the forgetting gate Decide on the information entered at the previous moment Screening, output gate Determine the information to output at the current moment The sieving, , , , , , , , , , All parameters are obtained through training. This is a convolution operation; , for Hidden states and memory units at any given moment; This refers to the hidden state and memory unit of the previous time step; These are candidate states obtained through nonlinear functions; for The training set input at each time step; All are three-dimensional tensors, with hidden layers. These are the basic parameters for storing and transmitting time information; for function; It is Hadamaji.

[0049] The historical noise distribution prediction model has an L-layer LSTM architecture, and the final output hidden state... for:

[0050] In the formula, T is the last time step of the sequence. This represents the hidden state at the last time step.

[0051] The formula for calculating the linear layer mapping prediction value is as follows:

[0052] In the formula, For the predicted environmental noise value, For the weights of the linear layer, This is the bias for the linear layer.

[0053] Among them, the hyperparameter optimization system is constructed using the Bayesian optimization algorithm to build a probabilistic model, with the coefficients of determination... To optimize the objective function and perform parameter optimization, a search space containing the following parameters is established: The number of neurons in the LSTM hidden layer (hidden_size), an integer ranging from 128 to 512, with a step size of 16; The number of LSTM stack layers (num_layers), an integer ranging from 1 to 3, with a step size of 1; Dropout rate, a continuous range from 1e-6 to 0.1; Batch size (batch_size), an integer range from 64 to 512, with a step size of 32; The learning rate is a continuous range from 1e-6 to 1e-3. The objective function is defined as follows: Using the coefficient of determination As a performance evaluation metric for the model, the objective function is computed on the validation set. The optimization objective is to maximize :

[0054] in, The sum of squares of the residuals ( It reflects the total amount of model prediction error; The total sum of squares ( It reflects the overall variability of the data. and The calculation formula is as follows:

[0055]

[0056] in, For the true value, For predicted values, This is the mean of the true values.

[0057] The Bayesian optimization process includes: use ( The algorithm is used as a sampling method; The number of optimization trials was set to 100, with the first 10 being random explorations; In each trial, the model is trained using the current hyperparameter combination and computed on the validation set. value; The model training and optimization process includes performance evaluation and calculation. , , , Performance indicators, etc.

[0058] in, Mean absolute error ( ), Root mean square error (RMSE) ), Normalized average error ( The formulas for calculating these indicators are as follows:

[0059]

[0060]

[0061] in, The total number of samples, For the true value, For predicted values, This is the mean of the true values.

[0062] Please see Figure 3 , Figure 3 (a), (b), and (c) show that the historical noise distribution prediction model of this embodiment exhibits excellent fitting, good generalization ability, and ideal error characteristics in three typical noise monitoring sites (industrial scenario, street scenario, and main road scenario). The model achieves stable and reliable prediction performance in all three monitoring sites. Specifically: ① High model prediction accuracy: Regression analysis of the three sites shows high determination coefficients. The values ​​reached 0.94, 0.90, and 0.90 respectively (all ≥0.9, please refer to [reference]). Figure 3 The values ​​of a1, b1, and c1 indicate that the model can explain the vast majority of data variation. The comparison graph of the actual and predicted values ​​also visually demonstrates that the prediction curve closely follows the fluctuation trend of the actual noise level (see [link to graph]). Figure 3 (a2, b2, c2).

[0063] ② The model training was effective and no overfitting occurred: As can be seen from the loss convergence curves at each site, both the training loss and the validation loss decreased synchronously with the increase of the number of iterations, and eventually tended to stabilize. Moreover, the two curves converged and the gap was very small (see [link to relevant documentation]). Figure 3 (a3, b3, c3), which strongly demonstrates that the model was adequately trained and did not overfit.

[0064] ③ The residual analysis results are ideal: the residual density maps of each station all show a bell-shaped curve centered at zero and approximately symmetrical, with a shape close to a normal distribution (see [link]). Figure 3 (a4, b4, c4). This indicates that the model error is random and there is no obvious systematic bias, further verifying the reliability of the model.

[0065] S302, Calculation of Historical Noise Distribution Range (see also) Figure 4 ).

[0066] The optimal model trained in step S301 is used to predict historical noise distribution. The dynamic confidence interval that reflects the historical noise distribution is calculated. The calculation steps are as follows: (1) Residual sequence extraction: Calculate the true value sequence of the historical noise distribution prediction model on the test set. With predicted value sequence The difference between them is used to construct a residual sequence. .in: True value sequence It contains n true values ​​for the test set, denoted as follows: , This represents the true value of the test set at time n; Predicted value sequence It contains n model predictions, denoted as follows: , This is represented as the model prediction value at time n; Residual sequence for and The sequence obtained by pointwise subtraction contains n residual values, denoted as follows: , This is represented as the residual value at time n. The residual at time i is... The calculation formula is:

[0067] in, Let be the true value of the test set at time i. Let be the model prediction value at time i.

[0068] (2) Residual statistical analysis: Calculate the standard deviation of the residuals The calculation formula is:

[0069] in, The total length of the test set. It is a residual sequence. The residual mean is calculated using the following formula:

[0070] (3) Real-time prediction: Input the noise time series characteristic data of the current period into the historical noise distribution prediction model, and output the noise baseline value of the current period. .

[0071] (4) Generation of historical noise distribution intervals: based on the current noise baseline value With 3 The principle is to construct dynamic confidence intervals: This interval represents the reasonable fluctuation range of noise values ​​in the current period, providing a data benchmark for the statistical anomaly determination of subsequent dynamic noise event classification.

[0072] Step S4: Classification of dynamic noise events.

[0073] S401, Historical noise distribution prediction.

[0074] Based on the historical noise distribution interval calculated in step S302, a dynamic confidence interval that reflects the historical noise distribution in the current time period is constructed: .

[0075] S402, Statistical Anomaly Detection.

[0076] By comparing the current noise monitoring values By comparing the noise distribution range with historical data, the statistical characteristics of noise events are determined, and a three-level statistical anomaly weight is defined to reflect the current noise monitoring value. Compared with historical noise distribution range The degree of deviation is specifically defined as: Abnormally high weight: A weight of 4 (high priority) indicates that the current noise value significantly exceeds the historical noise distribution; Abnormally low weight: A weight of 3 (medium priority) indicates that the current noise value is significantly lower than the historical noise distribution. Regular weights: A weight of 1 (low priority) indicates that the current noise value conforms to the historical noise distribution.

[0077] S403, Determination of environmental noise limits.

[0078] According to the environmental noise limits defined in the "Environmental Noise Quality Standard" (GB 3096-2008) Determine the compliance of noise events and define limit judgment weights to reflect the current noise level. Whether it complies with the corresponding environmental noise standard, specifically defined as: Excess weight: A weight of 2 indicates that the current noise level exceeds the standard and needs to be addressed promptly. Target weight: A weight of 1 indicates that the current noise value meets the standard.

[0079] S404, Joint Classification of Events.

[0080] By combining the dual criteria of statistical anomaly detection and environmental noise limit determination, and dynamically determining the event processing priority through the product of the two weights, noise events are classified as follows: The abnormally high exceedance event T1 has a weighted product of 4 × 2 = 8. This is characterized by a current noise value that is significantly higher than the historical noise distribution and a sound pressure level that exceeds the standard, which is a high-priority event. For an abnormally low exceedance event T2: the weighted product is 3 × 2 = 6. This is characterized as a current noise value that is significantly lower than the historical noise distribution, but the sound pressure level still exceeds the standard, and it belongs to the second priority event. The abnormally high pass rate for event T3: the weighted product is 4 × 1 = 4. This is characterized by noise values ​​that are significantly higher than the historical noise distribution but do not exceed the standard, which is a potential risk event that needs further verification. The abnormally low pass rate for event T4 has a weighted product of 3 × 1 = 3. This is characterized by noise values ​​that are significantly lower than historical noise distributions and meet the standards, which is an abnormal event that requires further verification. For a typical out-of-limit event T5, the weighted product is 1 × 2 = 2. and This is characterized by noise values ​​that conform to the historical noise distribution but continuously exceed the standard, which is a regular exceedance event; For a typical achievement event T6, the weighted product is 1 × 1 = 1. and This indicates that the noise value conforms to the historical noise distribution and meets the standard, which is a normal event.

[0081] Step S5: Tracing the source of intelligent computing power scheduling.

[0082] For the six categories of environmental noise events identified in step S4, a collaborative strategy driven by the noise event classification results is constructed, involving both local monitoring equipment and a cloud server. This strategy dynamically allocates computing resources and performs differentiated processing based on event characteristics. The specific strategy is as follows: Event T1: It has the dual characteristics of noise value significantly higher than the historical noise distribution and sound pressure level exceeding the standard. To ensure the accuracy of source tracing, it is uploaded to the cloud server first, and a high-precision model is used for noise source classification and source tracing analysis. T2 event: The noise value is significantly lower than the historical noise distribution, but the sound pressure level still exceeds the standard. It is a secondary priority event. First, a preliminary classification is performed using a local lightweight model. If the confidence of the local model is lower than the threshold of 0.8, it is uploaded to the cloud during idle time for secondary verification using a high-precision model. T3 event: The noise value is significantly higher than the historical noise distribution but does not exceed the standard. It is a potential risk event that needs further verification. First, trend analysis is performed through the local model. If a continuous upward trend is detected or there are characteristics of occasional noise sources, the cloud model is triggered to perform source tracing analysis. T4 event: The noise value is significantly lower than the historical noise distribution and meets the standard. It is an abnormal event that needs further verification. A comprehensive judgment is made in combination with contextual data (such as weather, time, and historical noise distribution). If it is determined to be an occasional event, it is archived locally. If it is determined to be a potential anomaly, the local model is used for source tracing analysis. T5 event: The noise value conforms to the historical noise distribution but continues to exceed the standard. It is usually an interpretable regular noise (such as traffic noise at fixed times). The noise source classification and recording are completed directly by the local lightweight model. T6 event: The noise value conforms to the historical noise distribution and meets the standard. It is a normal event and does not require real-time analysis and processing. It is only uploaded to the cloud for archiving and backup during system idle periods for long-term trend analysis.

[0083] In this embodiment, the low-power model deployed on the local monitoring device is the ResNet-18 model, whose technical features include: ① Input features: MelSpectrogram, which extracts the time-frequency features of environmental noise through short-time Fourier transform; ② Number of model parameters: approximately 11.18 million parameters (11.18M), suitable for low-power deployment on edge devices; ③ Classification capability: Supports classification of five major categories of noise: natural, traffic, industrial, social life, and construction noise, with a classification accuracy of ≥85%; ④ Response speed: Single frame inference time ≤ 50ms, meeting the requirements for real-time noise source tracing.

[0084] In this embodiment, the high-precision model deployed on the cloud server is the PANNs-CNN14 model, whose technical features include: ① Input features: MelSpectrogram, which captures noise details through high-resolution time-frequency analysis; ② Number of model parameters: approximately 79.7 million parameters (79.7M), supporting high-precision classification in complex noisy scenes; ③ Classification capability: Building upon the classification capabilities of the local low-power model, fine-grained subclass classification is further implemented, achieving a classification accuracy of ≥92%. While supporting five major noise categories, this model can perform refined identification of noise types. Fine-grained subclass classification includes the following noise categories: Natural noises: birdsong, insect chirping, wind, rain, thunder, etc.; Traffic noise: engine noise, horn noise, brake noise, train noise, motorcycle noise, etc.; Industrial noise: mechanical operating noise, assembly line noise, compressor noise, welding noise, etc.; Social noise: square dance music, fireworks sounds, street music, shopping mall announcements, etc.; Construction noise includes: pile driver noise, angle grinder noise, cutting machine noise, concrete mixing noise, etc.

[0085] Please refer to Tables 1 and 2. The edge computing method for tracing noise sources in low-voltage power supply environments provided by this invention achieves a high-precision noise source tracing capability of 96.67% in a typical nighttime main road scenario. By classifying environmental noise events into six categories, the system reduces cloud processing load to 33.33% (a 66.67% reduction compared to the traditional full-volume analysis mode). While significantly reducing computing power consumption, the overall processing time is optimized from 3.4s in the traditional mode to 2.1s, improving efficiency by 38.24%. Furthermore, the system breaks through the limitations of traditional single-classification, performing refined classification of noise sources (background noise / horn / car), improving accuracy by 1.06%. The results show that this invention maintains source tracing accuracy comparable to full-volume analysis while significantly reducing cloud computing power consumption, achieving an organic unity of refined and low-power operation in the noise monitoring system. By combining the advantages of local low-power models and cloud-based high-precision models through a dual-end collaborative mechanism, the system ensures real-time processing efficiency for routine events while improving the reliability of abnormal event identification through dynamic thresholds and context verification mechanisms. This provides a solution for environmental noise control that balances high efficiency, low power consumption, and high precision.

[0086] Table 1. Examples of Noise Event Source Tracing Analysis in Main Road Scenarios

[0087] Table 2 Performance Comparison Analysis of the Invention and Traditional Methods

[0088] *Note: The accuracy of the prediction results from traditional full-volume analysis is the confidence level of the prediction results for automobiles.

[0089] Step S6: Record and visualize the results.

[0090] The source analysis results are stored in the database in real time, and a visualization interface is built using Vue3+Fastapi to dynamically display key information such as the proportion of noise sources and the distribution of sound pressure levels.

[0091] Example 2 Please see Figure 5 , Figure 6 This embodiment provides a low-voltage power supply environment noise source tracing edge computing system, which includes the following functional modules: Multimodal sensing module: Integrates a high-precision digital omnidirectional microphone and sound pressure level sensor 1, and a multi-parameter meteorological sensor 3. Through a time synchronization circuit, it achieves millisecond-level spatiotemporal synchronous acquisition of sound pressure level data, meteorological parameters, and ambient audio, supporting parallel processing of multi-channel data streams. Specifically, the microphone sampling rate is 48kHz; the sound pressure level sensor has a measurement range of 30-130dB(A) with an accuracy of ±1.5dB; and the meteorological sensor has a wind speed resolution of 0.1m / s.

[0092] Network communication module: Integrates a 4G / 5G dual-mode communication module 2, employing a dynamic bandwidth allocation strategy: monitoring data and local identification results use a 256kbps transmission channel, while T1 / T2 events automatically upgrade to a 1Mbps high-speed channel. The communication module supports breakpoint resumption and encrypted data transmission, ensuring reliable transmission of environmental audio and analysis results.

[0093] Real-time display module: It adopts a 7-inch capacitive touch screen with a resolution of 1280×800, and integrates the Vue3+FastAPI visualization system to display multi-dimensional noise characteristics in real time, including second-level sound pressure level time-series curves, noise source proportion pie charts, and noise event time period information.

[0094] Edge Computing Unit 5: Equipped with an ARM Cortex-A55 quad-core, quad-thread processor (2.0GHz) and dual-channel LPDDR4 memory (4GB), and an eMMC 5.1 storage interface (64GB), it integrates an FPGA coprocessor for model acceleration. The system comes pre-installed with a dual-modal computing framework, supporting concurrent deployment of a local lightweight model (ResNet-18) and a historical noise distribution prediction model. Model loading time is <200ms, and inference latency is <100ms, meeting real-time analysis requirements.

[0095] Power management module: Includes 6 hot-swappable 12V / 10Ah lithium battery packs (supporting operation from -20℃ to 55℃) and an intelligent power management unit, featuring a mains power priority strategy and battery life optimization algorithm. The system has a battery life of ≥48 hours under full load operation mode.

[0096] Example 3 Please see Figure 7 This invention provides a low-voltage power supply environment noise source tracing edge computing system, comprising an edge device layer and a cloud server layer, jointly constructing a distributed collaborative computing architecture, including: Edge device layer: As the data sensing and decision-making unit deployed at the monitoring site, this layer integrates all the functional modules of Embodiment 2, enabling real-time data acquisition and rapid decision-making and response. Based on noise event classification results, it provides real-time response and processing for routine events; for complex events, it automatically initiates collaborative analysis requests with the cloud server layer, leveraging the high-precision model in the cloud to complete in-depth judgment. Through this decision-making mechanism combining local and cloud-based approaches, it achieves an optimal balance between analytical accuracy and computational power consumption.

[0097] Cloud server layer: This layer is built on the cloud platform and is supported by high-performance server clusters and big data storage systems. It has elastic and scalable parallel computing capabilities. Its main functions include: running high-precision analysis models and handling complex tracing tasks; continuously optimizing the model algorithms and iterating the version, and actively pushing the optimized model to the edge to ensure the timeliness and accuracy of edge device analysis.

[0098] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for calculating the source edge of noise in a low-voltage power supply environment, characterized in that, Includes the following steps: Step S1: Collect multimodal noise data; Step S2: Preprocess the collected multimodal noise data to obtain the noise temporal characteristics; Step S3: Input the time-series noise features into the trained historical noise distribution prediction model to predict the historical noise distribution; Step S4: Classify dynamic noise events based on the predicted historical noise distribution, specifically including: Based on the historical noise distribution intervals, construct a dynamic confidence interval that reflects the historical noise distribution in the current time period; By comparing the current noise monitoring value with the historical noise distribution range, the statistical characteristics of the noise event are determined, and a three-level statistical anomaly weight is defined to reflect the degree of deviation between the current noise monitoring value and the historical noise distribution range. Based on the environmental noise limits defined in the environmental noise quality standards, the compliance of noise events is determined, and the limit determination weights are defined to reflect whether the current noise value meets the corresponding environmental noise standards. The priority of event handling is dynamically determined by combining the dual conditions of joint statistical anomaly judgment and environmental noise limit judgment, and the product of the two weights. Step S5: Perform intelligent computing power scheduling and source tracing for the segmented noise events: Based on the classification of noise events, construct a dual-end collaborative strategy driven by the noise event classification results, dynamically allocate computing resources according to event characteristics, and perform differentiated processing.

2. The method for calculating the source edge of noise in a low-voltage power supply environment according to claim 1, characterized in that, Preprocessing includes: cleaning the raw data; deleting samples with missing values ​​exceeding a set threshold; interpolating and filling in missing values ​​using the daily mean interpolation method; normalizing the interpolated continuous data; and converting the normalized data into a supervised learning format. A time-series sliding window method is used to construct input-output pairs usable for supervised learning from the normalized time series data. By creating lag terms with specific time steps for each feature, an input feature matrix containing historical time-series information is formed, and the actual observations at the corresponding times are used as the output targets. This transforms continuous time series data into supervised learning data with a clear feature-label correspondence.

3. The method for calculating the source edge of noise in a low-voltage power supply environment according to claim 1, characterized in that, The training steps for the historical noise distribution prediction model include the following: Data preparation and preprocessing: Preprocessing is performed on the collected multimodal noise data; LSTM model architecture design: Construct a multi-layer long short-term memory network (LSTM) architecture, including an input layer, multi-level stacked LSTM hidden layers, and a fully connected output layer. Dynamic capture of temporal features is achieved through a gating mechanism. Hyperparameter optimization system construction: Establish a search space including parameters such as learning rate, batch size, and number of LSTM stacking layers, construct a probabilistic model using Bayesian optimization algorithm, and optimize parameters with the coefficient of determination as the optimization objective function; Model training and optimization: Iterative training of the model is performed within the hyperparameter search space, and a historical noise distribution prediction model with strong generalization ability is output through cross-validation strategy; Historical noise distribution interval calculation: based on model prediction results and 3 The principle is to construct a dynamically updated confidence interval in real time to characterize the reasonable fluctuation range of noise values. This represents the standard deviation of the residuals.

4. The method for calculating the source edge of noise in a low-voltage power supply environment according to claim 3, characterized in that, The historical noise distribution interval is calculated as follows: Residual sequence extraction: Calculate the difference between the true value sequence and the predicted value sequence of the historical noise distribution prediction model on the test set, and construct the residual sequence; Residual statistical analysis: Calculating the standard deviation of residuals based on residual sequences ; Real-time prediction: Input the noise time series characteristic data of the current period into the historical noise distribution prediction model, and output the noise baseline value of the current period. ; Historical noise distribution interval generation: based on the current noise baseline value With 3 Based on the principle, a dynamic confidence interval is constructed to characterize the reasonable fluctuation range of noise values ​​in the current period, providing a data benchmark for the statistical anomaly judgment of subsequent dynamic noise event classification.

5. The method for calculating the source edge of noise in a low-voltage power supply environment according to claim 4, characterized in that, Classifying dynamic noise events includes the following steps: Step 41: Historical noise distribution prediction: Based on the historical noise distribution intervals, construct a dynamic confidence interval that reflects the historical noise distribution in the current time period; Step 42, Statistical Anomaly Detection: By comparing the current noise monitoring values By comparing the noise distribution range with historical data, the statistical characteristics of noise events are determined, and a three-level statistical anomaly weight is defined to reflect the current noise monitoring value. The degree of deviation from the historical noise distribution range is defined as: Abnormally high weight: A weight of 4 indicates that the current noise value significantly exceeds the historical noise distribution; Abnormally low weight: A weight of 3 indicates that the current noise value is significantly lower than the historical noise distribution. Regular weights: A weight of 1 indicates that the current noise value conforms to the historical noise distribution; Step 43, Determining Environmental Noise Limits: According to the environmental noise limits defined in the environmental noise quality standards Determine the compliance of noise events and define limit judgment weights to reflect the current noise level. Whether it complies with the corresponding environmental noise standard is defined as: Excess weight: A weight of 2 indicates that the current noise level exceeds the standard and needs to be addressed promptly. Target weight: A weight of 1 indicates that the current noise value meets the standard. Step 44, Joint Event Classification: By combining the dual criteria of statistical anomaly detection and environmental noise limit determination, and dynamically determining the event processing priority through the product of the two weights, noise events are classified as follows: The abnormally high exceedance event T1 has a weighted product of 4 × 2 = 8. This is characterized by a current noise value that is significantly higher than the historical noise distribution and a sound pressure level that exceeds the standard, which is a high-priority event. For an abnormally low exceedance event T2: the weighted product is 3 × 2 = 6. This is characterized as a current noise value that is significantly lower than the historical noise distribution, but the sound pressure level still exceeds the standard, and it belongs to the second priority event. The abnormally high pass rate for event T3: the weighted product is 4 × 1 = 4. This is characterized by noise values ​​that are significantly higher than the historical noise distribution but do not exceed the standard, which is a potential risk event that needs further verification. The abnormally low pass rate for event T4 has a weighted product of 3 × 1 = 3. This is characterized by noise values ​​that are significantly lower than historical noise distributions and meet the standards, which is an abnormal event that requires further verification. For a typical out-of-limit event T5, the weighted product is 1 × 2 = 2. and This is characterized by noise values ​​that conform to the historical noise distribution but continuously exceed the standard, which is a regular exceedance event; For a typical achievement event T6, the weighted product is 1 × 1 = 1. and This indicates that the noise value conforms to the historical noise distribution and meets the standard, which is a normal event.

6. The method for calculating the source edge of noise in a low-voltage power supply environment according to claim 5, characterized in that, Step S5 is based on the classification of noise events. It constructs a two-way collaborative strategy driven by the classification results of noise events, dynamically allocates computing resources according to event characteristics, and performs differentiated processing. The specific strategy is as follows: The abnormally high exceedance event T1 has the dual characteristics of noise value significantly higher than the historical noise distribution and sound pressure level exceeding the standard. To ensure the accuracy of source tracing, it is uploaded to the cloud server first, and the high-precision cloud model is used for noise source classification and source tracing analysis. T2 is an abnormally low-priority event: the noise value is significantly lower than the historical noise distribution but the sound pressure level still exceeds the standard. It belongs to the secondary priority event. First, it is initially classified by the local lightweight model. If the confidence of the local model is lower than the threshold, it is uploaded to the cloud during the idle period to use the high-precision model for secondary verification. T3, an event with an abnormally high noise level: The noise value is significantly higher than the historical noise distribution but does not exceed the standard. It is a potential risk event that needs further verification. First, trend analysis is performed through the local model. If a continuous upward trend or the characteristics of sporadic noise sources are detected, the cloud model is triggered to perform source tracing analysis. T4, an event with an abnormally low compliance rate: The noise value is significantly lower than the historical noise distribution and meets the standard. It is an abnormal event that needs further verification. A comprehensive judgment is made based on the context data. If it is determined to be an occasional event, it is archived locally. If it is determined to be a potential anomaly, the local model is used for source tracing analysis. Typical out-of-limit events T5: Noise values ​​conform to historical noise distribution but continue to exceed limits. These are usually interpretable, routine noise, and the noise source classification and recording are completed directly by the local lightweight model. T6 is a normal event that meets the standard. The noise value is consistent with the historical noise distribution and meets the standard. It does not require real-time analysis and processing. It is only uploaded to the cloud for backup during system idle periods for long-term trend analysis.

7. The method for calculating the source edge of noise in a low-voltage power supply environment according to claim 6, characterized in that, The source analysis results are stored in the database in real time, and a visualization interface is built using Vue3+Fastapi to dynamically display key information such as the proportion of noise sources and the distribution of sound pressure levels.

8. An apparatus for implementing the low-voltage power supply environment noise source edge calculation method according to any one of claims 1 to 7, characterized in that, include: Multimodal sensing module: integrates a high-precision digital omnidirectional microphone and sound pressure level sensor, and a multi-parameter meteorological sensor. It achieves millisecond-level spatiotemporal synchronous acquisition of sound pressure level data, meteorological parameters and environmental audio through a time synchronization circuit, and supports parallel processing of multi-channel data streams; Network communication module: integrates 4G / 5G dual-mode communication module and adopts dynamic bandwidth allocation strategy: monitoring data and local identification results use low bandwidth transmission channel, and automatically upgrade to high-speed channel for abnormally high exceedance event T1 or abnormally low exceedance event T2. The communication module supports breakpoint resume and encrypted data transmission to ensure reliable transmission of environmental audio and analysis results; Real-time display module: includes a capacitive touch screen for integrating a visualization system to display multi-dimensional noise characteristics in real time, including second-level sound pressure level time-series curves, noise source proportion pie charts, and noise event time period information; Edge computing unit: Pre-installed with a dual-modal computing framework, supporting the concurrent deployment of local lightweight models and historical noise distribution prediction models to meet real-time analysis needs; Power management module: Includes hot-swappable lithium battery pack and intelligent power management unit, with mains power priority power supply strategy and battery life optimization algorithm.

9. A system for implementing the low-voltage power supply environment noise source edge calculation method according to any one of claims 1 to 7, characterized in that, include: Edge device layer: As a data sensing and decision-making unit deployed at the monitoring site, it is used to collect data in real time and make rapid decisions and responses. Based on the noise event classification results, it responds and processes routine events in real time; for complex events, it automatically initiates collaborative analysis requests with the cloud server layer and completes in-depth judgment with the help of high-precision models in the cloud. Cloud server layer: Built on the cloud platform, including: running high-precision analysis models and handling complex tracing tasks; We continuously optimize the model's algorithms and iterate its versions, and then proactively push the optimized model to the edge.

10. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1 to 7.

Images