Noise map dynamic deduction method and system based on multi-source heterogeneous data fusion

Abstract

The invention relates to the technical field of noise maps, in particular to a noise map dynamic deduction method and system based on multi-source heterogeneous data fusion, and the method comprises the steps: carrying out the space-time alignment of a multi-cluster low-dimension feature vector in an edge calculation node, and obtaining a multi-cluster low-dimension feature vector; generating a three-dimensional sound field reconstruction parameter of the first target data sequence and a spectrum energy distribution parameter of the second target data sequence; based on the three-dimensional sound field reconstruction parameter and the spectrum energy distribution parameter, dynamically controlling the sensor cluster to sleep and generating a space-time difference parameter; hierarchical encryption is carried out on the time-space difference parameters and the two types of data sequences in the dormancy period, and a basic ciphertext sequence and a feature ciphertext sequence for verification are generated; according to the method, the noise deduction map with space-time continuity is generated, high-precision dynamic deduction and energy efficiency optimization of the noise map are achieved, the evolution rule of noise along with time and space is effectively reflected, the real-time performance and accuracy of noise data are improved, and the actual situation of the noise can be reflected more truly, efficiently and in real time.

Description

Technical Field

[0001] The present invention relates to the technical field of noise maps, and more specifically, to a method and system for dynamically deriving a noise map based on the fusion of multi-source heterogeneous data. Background Art

[0002] Traditional methods for constructing noise maps mostly rely on single-type sensor data (such as sound pressure level data) and lack the ability to efficiently fuse multi-source heterogeneous data (such as the spatial distribution of the sound field, spectral characteristics, etc.). Especially in complex scenarios such as industrial parks, it is difficult to accurately align the spatio-temporal heterogeneous data streams generated by a large number of noise sensor nodes, resulting in low accuracy and poor real-time performance of sound field reconstruction.

[0003] For example, in an industrial park, noise sources may simultaneously include mechanical vibrations (dominated by low frequencies) and gas emissions (high-frequency bursts). Traditional methods only rely on the calculation of the sound pressure level mean and cannot distinguish the spectral characteristic differences, resulting in misjudgment of noise sources. When the data of nodes in a certain sensor cluster is delayed due to network congestion and is directly aligned with the data of adjacent clusters using linear interpolation during data fusion, it causes phase errors in the sound wave propagation path during sound field reconstruction (typical error > 10%).

[0004] For example, the cluster head node of a high-density sensor cluster (such as 500 nodes / cluster) needs to undertake the task of pre-fusing the original data. However, its computing resources are limited (such as a single-core ARM processor), and it cannot complete the compression of multi-channel data within a 10 ms time window, resulting in queue overflow and data loss at the edge nodes. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the deficiencies of the prior art and provide a method and system for dynamically deriving a noise map based on the fusion of multi-source heterogeneous data, realizing high-precision dynamic derivation and energy efficiency optimization of the noise map.

[0006] To solve the above technical problems, the basic concept of the technical solution adopted by the present invention is as follows: In a first aspect, a method for dynamically deriving a noise map based on the fusion of multi-source heterogeneous data includes: In edge computing nodes, performing spatio-temporal alignment processing on multi-cluster low-dimensional feature vectors to generate three-dimensional sound field reconstruction parameters of a first target data sequence and spectral energy distribution parameters of a second target data sequence; Based on the three-dimensional sound field reconstruction parameters and spectral energy distribution parameters, dynamically controlling the dormancy of sensor clusters and generating spatio-temporal difference parameters; Performing hierarchical encryption on the spatio-temporal difference parameters and two types of data sequences during the dormancy period to generate a basic ciphertext sequence and a feature ciphertext sequence for verification; Based on the basic ciphertext sequence and the feature ciphertext sequence, performing data correction and rendering control to generate a noise derivation map with spatio-temporal continuity.

[0007] In a second aspect, a noise map dynamic deduction system based on multi-source heterogeneous data fusion includes: A generation module, configured to perform spatio-temporal alignment processing on multi-cluster low-dimensional feature vectors in edge computing nodes to generate three-dimensional sound field reconstruction parameters of a first target data sequence and spectral energy distribution parameters of a second target data sequence; A fusion module, configured to dynamically control the dormancy of sensor clusters and generate spatio-temporal difference parameters based on the three-dimensional sound field reconstruction parameters and the spectral energy distribution parameters; An encryption module, configured to perform hierarchical encryption on the spatio-temporal difference parameters and two types of data sequences during the dormancy period to generate a basic ciphertext sequence and a feature ciphertext sequence for verification; A correction module, configured to perform data correction and rendering control based on the basic ciphertext sequence and the feature ciphertext sequence to generate a noise deduction map with spatio-temporal continuity.

[0008] After adopting the above technical solutions, the present invention has the following beneficial effects compared with the prior art: By performing spatio-temporal alignment processing on multi-source heterogeneous data, cross-domain fusion of low-dimensional features is achieved, the accuracy of three-dimensional sound field reconstruction and spectral energy analysis is improved, more reliable basic parameters are provided for noise modeling, and the dormancy of sensor clusters is dynamically controlled. While ensuring the integrity of data spatio-temporal features, device energy consumption and data redundancy are reduced, and the resource utilization efficiency of edge computing nodes is optimized. The hierarchical encryption mechanism differentially processes basic data and feature data, realizes lightweight encryption at the edge, generates a noise map with spatio-temporal continuity, supports real-time / near-real-time dynamic rendering of noise distribution, effectively reflects the evolution law of noise over time and space, improves the timeliness and accuracy of noise data, and can more realistically, efficiently, and timely reflect the actual situation of noise. Description of the Drawings

[0009] Figure 1 is a schematic diagram of the noise map dynamic deduction system based on multi-source heterogeneous data fusion of the present invention.

[0010] Figure 2 is a schematic diagram of the flow of the noise map dynamic deduction method based on multi-source heterogeneous data fusion of the present invention. Detailed Embodiments

[0011] The following embodiments of the present application take the noise map dynamic deduction system based on multi-source heterogeneous data fusion as an example to detail the solution of the present application, but this embodiment does not limit the protection scope of the present application.

[0012] As Figure 1 shown, the present invention provides a noise map dynamic deduction method based on multi-source heterogeneous data fusion, including: Step S1: In the edge computing node, perform spatio-temporal alignment processing on multi-cluster low-dimensional feature vectors to generate three-dimensional sound field reconstruction parameters of the first target data sequence and spectral energy distribution parameters of the second target data sequence; Step S2: Based on the three-dimensional sound field reconstruction parameters and spectral energy distribution parameters generated in Step S1, dynamically control the sensor clusters to sleep and generate spatio-temporal difference parameters; Step S3: Perform hierarchical encryption on the spatio-temporal difference parameters generated in Step S2 and the two types of data sequences during the sleep period to generate a basic ciphertext sequence and a feature ciphertext sequence for verification; Step S4: Based on the ciphertext sequences encrypted in Step S3, perform data correction and rendering control to generate a noise deduction map with spatio-temporal continuity.

[0013] In the embodiment of the present invention, by performing spatio-temporal alignment processing on multi-source heterogeneous data (Step S1), cross-domain fusion of low-dimensional features is achieved, the accuracy of three-dimensional sound field reconstruction and spectral energy analysis is improved, and more reliable basic parameters are provided for noise modeling. Dynamically control the sensor clusters to sleep (Step S2), while ensuring the integrity of the spatio-temporal features of the data, reducing device energy consumption and data redundancy, and optimizing the resource utilization efficiency of the edge computing node. The hierarchical encryption mechanism (Step S3) differentially processes the basic data and feature data, realizes lightweight encryption at the edge, not only meets the security transmission and storage requirements of noise data, but also avoids the high computational overhead of traditional full-scale encryption. Generate a noise deduction map with spatio-temporal continuity (Step S4), support real-time / near-real-time dynamic rendering of the noise distribution, and effectively reflect the evolution law of the noise over time and space.

[0014] In a preferred embodiment of the present invention, before Step S1, it further includes: Perform clustered topology management on no less than 500 noise sensor nodes in the industrial park, dynamically divide the sensor clusters according to the node signal strength, and generate the intra-cluster communication link topology; Set the cluster head node based on the intra-cluster communication link topology, perform pre-fusion calculation on the original noise data stream by the cluster head node to generate redundant-reduced compressed data; convert the compressed data into low-dimensional feature vectors by the cluster head node, and transmit the low-dimensional feature vectors to the edge computing node through the edge gateway as the input of the multi-cluster low-dimensional feature vectors.

[0015] In the embodiments of the present invention, first, each noise sensor node periodically broadcasts signal strength detection packets (such as RSSI signals) to collect in real time the signal interaction data of no less than 50 nodes in the industrial park. The density-based clustering algorithm (such as DBSCAN) or the adaptive signal strength threshold method is used. With the physical location (latitude and longitude coordinates) of the nodes as the spatial dimension and the signal strength as the connection weight, the sensor clusters are dynamically divided. The division principle is that the average signal strength between nodes within the same cluster is ≥ the preset threshold (such as -70dBm), and the number of nodes within the cluster is controlled within 50 - 100 to balance the computing load. A star-shaped or mesh communication link topology is generated for each cluster, and the communication paths between the nodes within the cluster and the cluster head node are clarified (such as establishing a point-to-point connection through the ZigBee or LoRa protocol).

[0016] The distributed election algorithm (such as the improved version of the LEACH protocol) is adopted to dynamically elect the cluster head node by comprehensively considering the remaining energy of the node, computing power (CPU main frequency, memory capacity), and location centrality (near the geometric center of the cluster). The election is re-conducted every 5 minutes to avoid excessive energy consumption of a single node. After the initialization of the cluster head node, a two-way authentication mechanism (such as symmetric key encryption) is established with the member nodes within the cluster to ensure the security of data transmission. The member nodes within the cluster collect the noise time-domain signals (voltage amplitude sequences) at a fixed frequency (such as 10Hz) and transmit them to the cluster head node through the communication link. The cluster head node adopts a sliding window redundancy removal algorithm (such as deleting duplicate data with fluctuations < 5% within three consecutive cycles), combines the discrete Fourier transform (DFT) to extract the frequency-domain features (such as octave sound pressure level), and then compresses the multi-dimensional features to 3 - 5 dimensions through principal component analysis (PCA) to generate redundant compressed data (the data volume compression ratio reaches 80% - 90%).

[0017] The cluster head node maps the compressed data into a low-dimensional feature vector (such as a 5-dimensional vector including the time-domain mean, frequency-domain peak value, and spatial coordinates). The vector elements adopt the floating-point data format (single-precision floating-point, about 20 bytes for each vector). The feature vector is normalized (such as standardized to the [-1, 1] interval) to avoid the influence of the dimension on subsequent processing by the edge computing nodes. Each cluster head node is connected to the edge gateway through wired (RJ45) or wireless (Wi-Fi) means. The edge gateway adopts the publish-subscribe mode (such as the MQTT protocol) to collect the low-dimensional feature vectors of each cluster.

[0018] The edge gateway synchronizes the timestamps of the multi-cluster feature vectors (accurate to the millisecond level), encapsulates them into a unified data format (such as ProtocolBuffers), and transmits them to the edge computing node through the Ethernet or 5G network to form the "multi-cluster low-dimensional feature vector" input required in step S1.

[0019] In the embodiments of the present invention, dynamic clustering is adopted to reduce the ineffective communication between nodes. The star topology within the cluster shortens the single-node communication distance by 30%-50%, reduces the energy consumption of sensor nodes by more than 40%, extends the device life. The nodes within the same cluster are relatively close in spatial distribution, and the signal transmission is highly consistent in terms of the influence of industrial electromagnetic interference, reducing the interference of heterogeneous noise on data synchronization. Redundancy compression reduces the original data volume by more than 85%, significantly reducing the input data scale of the edge computing node and improving the subsequent spatio-temporal alignment processing speed (the processing time is reduced by 60%). Through frequency domain conversion and PCA dimensionality reduction, the key features of the noise signal (such as sound pressure level, main frequency distribution) are retained, filtering out high-frequency noise interference, providing a purer feature input for the three-dimensional sound field reconstruction in step S1. The data volume of the single-cluster feature vector transmission is reduced from the KB level of the original time-domain signal to the byte level, and the throughput of the edge gateway is increased by more than 5 times, and the network delay is reduced to within 50 ms. The timestamp synchronization mechanism of the edge gateway ensures that the time consistency error of multi-cluster data is <1 ms, laying a high-precision time benchmark for the spatio-temporal alignment processing in step S1 and avoiding the reconstruction parameter error caused by clock deviation.

[0020] In a preferred embodiment of the present invention, step S1: In the edge computing node, perform spatio-temporal alignment processing on the multi-cluster low-dimensional feature vectors to generate the three-dimensional sound field reconstruction parameters of the first target data sequence and the spectral energy distribution parameters of the second target data sequence, including: Based on the multi-cluster low-dimensional feature vectors, perform spatial alignment on the multi-cluster feature vectors through the convolutional layer of the lightweight hybrid neural network to generate spatially aligned feature vectors; use the long short-term memory network layer to perform temporal alignment on the spatially aligned feature vectors to generate a continuous data stream; Divide the continuous data stream into data blocks with a time window length of 10 milliseconds, and construct a three-dimensional sound field grid model based on the spatial coordinates of each node within the sensor cluster; for the discrete data points in the three-dimensional sound field grid model, fill the sound field intensity values through the inverse distance weighted interpolation algorithm to generate the three-dimensional sound field reconstruction parameters of the first target data sequence; Perform a fast Fourier transform on the 10-millisecond data blocks, extract the 1 / 3 octave spectral energy, and fuse the multi-sensor data through the energy superposition algorithm to generate the spectral energy distribution parameters of the second target data sequence.

[0021] In the embodiments of the present invention, for spatial alignment and feature extraction, multiple clusters of low-dimensional feature vectors (including spatial coordinates, time-domain means, frequency-domain peaks, etc. of each sensor node) are first input into the convolutional layer of a lightweight hybrid neural network. The network uses a simplified two-dimensional convolutional kernel (such as a 3×3 size), with the planar coordinates (X, Y axes) of the sensor nodes as the input dimension, scans the spatial distribution features of different clusters through a sliding window, and automatically identifies the spatial correlation relationships between adjacent sensor clusters (such as the position mapping between noise sources and monitoring nodes in a factory workshop). The spatially aligned feature vectors output by the convolutional layer contain the relative position weights of each cluster node in three-dimensional space and the signal strength correlation features.

[0022] For time alignment and sequence generation, the spatially aligned feature vectors are input into a long short-term memory network (LSTM) layer. The gating mechanism of the LSTM is used to capture the long-term dependencies in the time series (such as the periodic pattern of noise changing with production shifts). The network input is a sequence of feature vectors sorted by timestamp (with a time resolution accurate to the millisecond level). Through the transfer of the hidden layer state, a continuous data stream with timestamp bias removed is output, enabling seamless connection of data at different times and from different clusters in the time dimension.

[0023] For data block segmentation and grid modeling, the continuous data stream is segmented into independent data blocks with a time window of 10 milliseconds (for example, each data block contains multi-cluster features of 100 time points). At the same time, based on the three-dimensional spatial coordinates (X, Y, Z axes, where the Z axis represents height, such as different floors of a factory building) of the nodes within each sensor cluster, the target area is divided into regular three-dimensional grids (such as cubic units of 0.5 m × 0.5 m × 0.5 m), and each grid point corresponds to a sound field intensity calculation position to be filled.

[0024] For sound field reconstruction parameter generation, for each discrete data point in the three-dimensional grid, data of several sensor nodes closest to this point are collected (such as taking the 8 nodes with the closest Euclidean distance). The inverse distance weighting method is adopted: the closer the node is, the higher the contribution weight to the sound field intensity of the grid point. The sound field intensity value of each grid point is calculated by weighted average, and finally, three-dimensional sound field grid data covering the target area, that is, the three-dimensional sound field reconstruction parameters of the first target data sequence, are generated.

[0025] For spectral energy parameter generation, a fast Fourier transform is performed on each 10-millisecond data block to convert the time-domain noise signal into a frequency-domain energy distribution. The energy values in the 1 / 3 octave band (such as common noise analysis frequency bands like 63 Hz, 80 Hz, 100 Hz, etc.) are extracted. Then, the energy values in the same frequency band of multiple sensor clusters within the same time window are superimposed and averaged (such as taking the arithmetic mean or weighted mean) to eliminate the accidental errors of individual sensors, and the spectral energy distribution parameters of the second target data sequence are generated, reflecting the distribution characteristics of noise energy in different frequency bands in space.

[0026] In the embodiment of the present invention, the lightweight hybrid neural network combines the spatial feature extraction of the convolutional layer and the time series modeling of the LSTM to achieve spatio-temporal alignment with low computational complexity at the edge computing node. Compared with the traditional interpolation method, the recognition accuracy of spatial position association is increased by 30%, and the continuity error of the time series is reduced to less than 5%. A three-dimensional grid model is constructed based on the actual spatial coordinates of the sensors, and the sound field intensity is filled by inverse distance weighted interpolation, effectively solving the problem of sound field blank between discrete sensor nodes. The reconstructed sound field parameters can accurately reflect the noise gradient changes on different floors and near equipment in the industrial plant (for example, the sound field intensity error within 5 meters from the noise source < 3%); The 1 / 3 octave band energy extraction and multi-sensor fusion algorithm is adopted, which can not only retain the key frequency domain features of the noise signal (such as the abnormal energy in a specific frequency band of mechanical vibration), but also reduce the single-point measurement noise through data superposition, increasing the signal-to-noise ratio of the spectral energy distribution parameters by 40%.

[0027] In a preferred embodiment of the present invention, based on the multi-cluster low-dimensional feature vectors, the convolutional layer of the lightweight hybrid neural network is used to perform spatial alignment on the multi-cluster feature vectors to generate spatially aligned feature vectors; The long short-term memory network layer is used to perform time alignment on the spatially aligned feature vectors to generate a continuous data stream, including: The spatial features of the multi-cluster low-dimensional feature vectors are extracted by using the multi-layer convolutional kernels of the convolutional layer, and the coordinate data of different sensor clusters are normalized to the same spatial reference system through the feature mapping matrix; The long short-term memory network layer is used to perform a sliding window analysis on the time series features of the continuous data stream, and the time offset of adjacent data blocks is dynamically compensated through the gating mechanism to achieve millisecond-level time synchronization.

[0028] In the embodiment of the present invention, for the spatial feature extraction of the multi-layer convolutional kernels, the multi-cluster low-dimensional feature vectors (including the three-dimensional coordinates X / Y / Z, signal intensity, equipment type, etc. of each sensor node) are first input into the convolutional layer of the lightweight neural network. The convolutional layer uses 2-3 two-dimensional convolutional kernels of different sizes (for example, the first layer 5×5 kernel captures the spatial association in a larger range, and the second layer 3×3 kernel focuses on the local node relationship), and takes the plane coordinates (X, Y) of the sensor nodes as the input dimension, and scans the spatial distribution of different sensor clusters layer by layer.

[0029] The first-layer convolutional kernel calculates the spatial distance weights of adjacent nodes through a sliding window to identify the spatial dependence of inter-cluster signal propagation (for example, nodes within 10 meters are considered strongly correlated); the second-layer convolutional kernel extracts the spatial density features of intra-cluster nodes (such as the number of nodes within every 20 square meters), generating a preliminary feature map containing spatial position associations and signal attenuation rules. Coordinate normalization and spatial reference system unification address possible differences in local coordinate systems for different sensor clusters (such as different factories using independent coordinate origins), and perform coordinate transformation through a preset feature mapping matrix: First, collect the physical coordinates of all sensor nodes (obtain absolute coordinates through GPS or UWB positioning systems), and determine the global coordinate system of the industrial park (with the park center as the origin, the X-axis due east, the Y-axis due north, and the Z-axis for vertical height); For the coordinate data of each sensor cluster (such as the local coordinates with a corner as the origin in a certain workshop), through translation (X' = X + offset), rotation (correct the angle deviation of the coordinate system), and scale scaling (unify the unit to meters) operations, normalize it to the global coordinate system to ensure that the spatial data of different clusters are comparable at the same scale.

[0030] Sliding window time-series feature analysis sorts the spatially aligned feature vectors by timestamp and inputs them into the long short-term memory network (LSTM) layer. A fixed-length sliding window (such as a 500ms window containing 50 10ms data points) is used to process the data stream segment by segment. The window slides 10ms each time, ensuring that adjacent data blocks have an overlapping area of 490ms to capture the continuous features of the time series.

[0031] The data within the window contains the signal strength changes of each sensor cluster at consecutive time points (such as noise mutations caused by the start and stop of night shift machines). LSTM passes through the hidden layer state to remember the feature dependence relationships at different time points (such as the impact of the noise peak in the previous 100ms on the current moment).

[0032] The gating mechanism dynamically compensates for time offsets for possible clock synchronization errors in different sensor clusters (such as time stamp deviations of ±5ms for some nodes due to communication delays), and uses the gating mechanism of LSTM (input gate, forget gate, output gate) for time calibration: Input gate: Calculate the offset between the data at the current time point and the theoretical time stamp (such as the difference between the actual reception time and the time reported by the sensor), and generate an offset weight factor; Forget gate: Reduce the weight of data points with a time offset exceeding 3ms to suppress the influence of outdated information (such as ignoring abnormal data with a delay exceeding 5ms); Output gate: Dynamically adjust the output value at the current moment according to the time continuity of adjacent data blocks (for example, compensate for missing intermediate values through interpolation of data at the previous and next moments), and finally generate a continuous data stream with a time stamp error < 1ms.

[0033] In the embodiment of the present invention, cross-cluster coordinate unification converts the scattered local coordinate systems into a global coordinate system through a feature mapping matrix, completely solving the problem of inconsistent spatial references of multi-source sensors (such as converting the monitoring data of different workshops from "workshop local coordinates" to "park global coordinates"), improving the spatial positioning accuracy of subsequent three-dimensional sound field modeling to within 0.5 meters, and avoiding the distortion of sound field reconstruction caused by coordinate deviation (such as the spatial positioning error of more than 10 meters that may be caused by traditional unnormalized methods). The spatial correlation intelligent recognition multi-layer convolution kernel automatically captures the spatial distribution law of sensor nodes (such as the node group densely deployed along the production line and the nodes sparsely distributed at the edge of the factory area), generates spatial alignment features including distance weights and signal attenuation characteristics, and provides a more accurate spatial correlation basis for subsequent inverse distance weighted interpolation (for example, correctly identifying the high-weight contribution of nodes near the noise source to the grid points). Compared with the method of manually setting the distance threshold, the spatial feature extraction efficiency is increased by 40%. The millisecond-level time synchronization ability combines a sliding window with a gating mechanism to effectively handle the problem of asynchronous sensor clocks, compressing the timestamp error of multi-cluster data from ±10 ms of traditional NTP time synchronization to within ±1 ms, ensuring the synchronous display of data in different clusters for the same noise event (such as the moment of machine start and stop), and avoiding misjudgment of the noise evolution law caused by time dislocation (for example, correctly capturing the millisecond-level time correspondence between the noise peak and the equipment action). The continuity of time series features enhances the memory characteristics of LSTM to retain the noise change trend on a long time scale (such as the periodic fluctuation of noise within an 8-hour shift), and at the same time maintains the feature continuity between data blocks through an overlapping sliding window (avoiding signal breaks at the window boundary), enabling the continuous data stream to fully reflect the gradual change process of noise over time (such as the smooth decline curve of noise during equipment maintenance at noon). The lightweight network design (such as limiting the depth of the convolutional layer to 3 layers and the dimension of the LSTM hidden layer to 64 dimensions) reduces the single-node processing power consumption by 60%, and controls the inference delay on edge computing nodes (such as NVIDIA Jetson AGX Orin) within 15 ms, meeting the strict requirements of industrial noise real-time monitoring for low latency and low energy consumption. Compared with the traditional cloud full-volume data processing solution, the response speed is increased by more than 3 times, and the network transmission load is reduced by 70%.

[0034] In a preferred embodiment of the present invention, a fast Fourier transform is performed on the 10-millisecond data block, the 1 / 3 octave band spectral energy is extracted, and multi-sensor data is fused through an energy superposition algorithm to generate the spectral energy distribution parameters of the second target data sequence, including: The time-domain signal of each 10-millisecond data block is windowed with a Hanning window. After eliminating spectral leakage, the frequency-domain energy distribution is obtained through a fast Fourier transform; Based on the standard 1 / 3 octave band division rule, the frequency domain energy distribution is segmented by center frequency, and the energy values within each band are integrated to generate initial energy distribution parameters indexed by the band. Based on the spatial distribution density of each node within the sensor cluster, the initial energy distribution parameters are spatially weighted and fused to generate the spectral energy distribution parameters of the second target data sequence.

[0035] In the embodiment of the present invention, the Hanning window function is applied to the time-domain noise signal (voltage amplitude sequence) of each 10-millisecond data block for windowing. The window width is the same as the data block length, and spectral leakage is reduced by smoothing the signal edges (to avoid the spread of spectral energy to adjacent frequencies). The fast Fourier transform (FFT) is performed on the windowed signal to convert the time-domain signal into a frequency domain energy distribution, with the output frequency range covering 20 Hz - 20 kHz (the audible range of the human ear), and the frequency resolution being approximately 100 Hz (determined by the number of FFT points).

[0036] According to the 1 / 3 octave center frequencies defined by international standards (such as ISO266) (such as 31.5 Hz, 40 Hz, 50 Hz... 16 kHz), the continuous spectrum output by the FFT is divided into 30 bands (the bandwidth of each band is approximately 23% of the current center frequency). The discrete spectral energy values within each band are integrated (i.e., the energies of all frequency points within the band are accumulated) to generate initial energy distribution parameters indexed by the 1 / 3 octave center frequency, with each parameter representing the noise energy intensity of a specific band.

[0037] Calculate the spatial distribution density of each node within the sensor cluster: with each node as the center, a circular area with a radius of 2 meters is demarcated, and the number of other nodes within the area is counted as the density index; weights are assigned to each node according to the density index: nodes in high-density areas (such as near the central equipment in the workshop) have lower weights (such as 0.8), and nodes in low-density areas (such as the edge of the factory area) have higher weights (such as 1.2) to balance the monitoring contributions of different areas. The initial energy parameters of multiple nodes in the same band are weighted and averaged to generate the final spectral energy distribution parameters, reflecting the spatial distribution of the noise frequency characteristics of the entire monitoring area.

[0038] In the embodiments of the present invention, the Hamming window windowing process effectively suppresses spectral leakage, making the energy boundaries of each frequency band clearer (for example, the error of the energy leakage from the 500 Hz frequency band to the 400 Hz frequency band is reduced from 15% of the traditional rectangular window to 3%), improving the frequency resolution. The 1 / 3 octave analysis is more refined than the traditional 1 octave (for example, three frequency bands of 800 Hz, 1 kHz, and 1.25 kHz can be distinguished near 1 kHz), and it can more accurately identify the characteristic frequencies of the equipment (such as the 1000 Hz electromagnetic noise of the motor and the 1250 Hz mechanical noise of the gearbox). The spatial weighted fusion algorithm avoids the overfitting problem of data in high-density regions, making the noise characteristics in the edge region (such as the noise at the factory boundary) account for a reasonable proportion in the final parameters, and fully presenting the overall spatial distribution of the noise source. The energy superposition process of multi-sensor data eliminates the accidental error of single-point measurement (such as random environmental noise interference), making the spectral energy distribution parameters closer to the true sound field characteristics (such as the energy fluctuation range of a certain frequency band is reduced from ±5 dB to ±1.5 dB). The 10-millisecond processing window balances the time resolution (capturing sudden noise events) and the frequency resolution (meeting the requirements of 1 / 3 octave analysis), and is applicable to the dual scenarios of the start and stop of industrial equipment (the time scale is at the 100 ms level) and steady-state noise monitoring. The edge end performs FFT and energy calculation, and the data processing delay <5 ms, meeting the requirements of real-time noise warning (such as abnormal frequency energy mutation alarm), and the response speed is increased by 90% compared with the cloud processing solution.

[0039] In a preferred embodiment of the present invention, step S2: Based on the three-dimensional sound field reconstruction parameters and spectral energy distribution parameters generated in step S1, dynamically control the sensor cluster to sleep and generate spatio-temporal difference parameters, including: Calculate the activity score of each sensor cluster according to the sound field intensity value in the three-dimensional sound field reconstruction parameters and the frequency band energy ratio in the spectral energy distribution parameters; if the activity score is lower than the preset threshold, send a sleep instruction to the corresponding sensor cluster, and record the sleep start timestamp and the spatio-temporal coordinates of the missing data during sleep; Based on the recorded sleep start timestamp and spatio-temporal coordinates, obtain the three-dimensional sound field reconstruction parameters before sleep, and obtain the three-dimensional sound field reconstruction parameters after sleep after sleep ends, and calculate the sound field intensity difference matrix between the two; Input the sound field intensity difference matrix into the spatio-temporal Kalman filtering algorithm, and predict the dynamic sound field intensity change amount in the sleep area by fusing the historical sound field intensity change rate and the spatio-temporal distribution characteristics of the current sensor cluster, and generate spatio-temporal difference parameters including timestamp compensation and spatial interpolation weight.

[0040] In the embodiments of the present invention, based on the sound field intensity value in the three-dimensional sound field reconstruction parameters, the average sound pressure level (such as A-weighted sound level) of the monitoring area of each sensor cluster is calculated. Combining the frequency band energy ratio in the spectral energy distribution parameters, the characteristic frequency bands that contribute greatly to the total sound pressure are identified (for example, industrial equipment noise is usually concentrated in the range of 500 Hz - 4 kHz). Considering the stability of the sound field intensity and the characteristic frequency band energy (such as a fluctuation coefficient < 5% is considered stable), an activity score (0 - 100 points) is generated for each sensor cluster.

[0041] Sleep decision: If the score of a certain sensor cluster is lower than the preset threshold (such as 40 points) for three consecutive cycles (such as 30 seconds), it is determined that the sound field in this area is stable, and a sleep command is sent to it. The start timestamp of sleep (accurate to milliseconds) and the spatio-temporal coordinates (three-dimensional grid positions) of the missing data during sleep are recorded.

[0042] Calculation of the sound field intensity difference matrix, data acquisition: Before the sensor cluster goes to sleep (such as within 100 ms before sleep), the sound field intensity values in the three-dimensional sound field reconstruction parameters of this area are extracted to form a pre-sleep intensity matrix. After the sensor cluster wakes up (such as within 50 ms after waking up), the sound field intensity values of the same area are immediately collected to form a post-sleep intensity matrix.

[0043] Difference calculation: The difference between the post-sleep and pre-sleep intensity matrices is calculated point by point to generate a sound field intensity difference matrix. The matrix elements represent the change in sound pressure level of each grid point during sleep (such as ΔdB = post-sleep dB - pre-sleep dB).

[0044] Spatio-temporal Kalman filter prediction and differential parameter generation, historical feature extraction: Analyze the change rate of the sound field intensity in this area from historical data (such as the change trend of sound pressure level per hour), identify the diurnal periodic pattern (such as the noise reduction caused by equipment shutdown at night), and combine the spatial distribution density of the sensor cluster to determine the sound field change sensitivity coefficients at different positions (such as the change rate is higher in the area closer to the noise source).

[0045] Spatio-temporal fusion prediction: The sound field intensity difference matrix is input into the spatio-temporal Kalman filter algorithm, which fuses the historical change rate and the spatial sensitivity coefficient. The algorithm dynamically adjusts the prediction weight according to the sleep duration (such as the longer the sleep time, the higher the weight of the historical trend), and outputs the predicted change in sound field intensity.

[0046] Differential parameter generation: Spatio-temporal differential parameters including timestamp compensation values (such as aligning the prediction result to the current time) and spatial interpolation weights (such as the closer the grid point is to the sleep area, the higher the weight) are generated for subsequent data correction.

[0047] In the embodiments of the present invention, through sleep control, the power consumption of sensors in non-critical areas is reduced by more than 60% (for example, sensors that originally worked for 24 hours become intermittent), frequent data transmission is reduced, the network bandwidth occupancy is reduced by 50%, the service life of edge gateways and communication devices is extended, the spatio-temporal Kalman filtering algorithm effectively fills the data gaps during sleep, the sound field intensity prediction error is controlled within ±2 dB (the allowable error of industrial noise standards is ±3 dB), the timestamp compensation mechanism ensures the time continuity of dynamic deduction, avoids the phenomenon of noise map jumps caused by data interruption, reduces the real-time processing load of edge computing nodes (for example, the number of active sensor clusters processed simultaneously is reduced by 30%), releases computing resources for more complex noise propagation modeling, the sleep strategy enables the system to concentrate computing power on high-activity areas (such as near production lines where equipment starts and stops frequently), improves the monitoring accuracy of critical areas, the spatio-temporal difference parameters reflect the dynamic change trend of the noise field (such as the sound pressure level rise rate when equipment starts), the noise map update frequency is increased to 10 Hz (the traditional method is usually 1 Hz), and the response speed to sudden noise events (such as high-frequency noise generated by pipeline leaks) is shortened from the second level to the millisecond level, meeting the timeliness requirements of industrial safety warnings.

[0048] In a preferred embodiment of the present invention, step S3: Perform hierarchical encryption on the spatio-temporal difference parameters generated in step S2 and the two types of data sequences during sleep to generate a basic ciphertext sequence and a feature ciphertext sequence for verification, including: Based on the timestamp compensation and spatial interpolation weight in the spatio-temporal difference parameters, perform data integrity verification on the sound pressure values and three-dimensional coordinates of the first target data sequence during sleep. After the verification passes, use the national secret SM4 algorithm for encryption to generate a basic ciphertext sequence; For the spectral energy values and frequency band labels of the second target data sequence during sleep, adjust the energy distribution ratio according to the spatial interpolation weight in the spatio-temporal difference parameters, and use the homomorphic encryption algorithm to encrypt the adjusted energy values and frequency band labels to generate a feature ciphertext sequence; Input the basic ciphertext sequence and the feature ciphertext sequence into the watermark embedding module, generate a digital watermark identifier based on the identifier of the sensor cluster and the edge node number, and perform hash binding on the digital watermark identifier and the synchronization timestamp in the spatio-temporal difference parameters to generate an encrypted data block for verification.

[0049] In the embodiments of the present invention, the timestamp compensation value in the spatio-temporal difference parameter is extracted and compared with the timestamps of the first target data sequence (sound pressure value and three-dimensional coordinates) during the sleep period to verify data continuity. The sound pressure value is reconstructed and calculated using the spatial interpolation weight and compared with the original recorded value. Data points with a deviation exceeding ±3% are marked as suspicious. The data that passes the verification is grouped and encrypted using the national secret SM4 algorithm with a key length of 128 bits. The encryption mode uses the CBC (Cipher Block Chaining) mode to enhance security. The initialization vector (IV) is dynamically generated based on the sensor cluster ID and the current timestamp to ensure the uniqueness of each encryption.

[0050] According to the spatial interpolation weight in the spatio-temporal difference parameter, the second target data sequence (spectrum energy value and frequency band label) is weighted and adjusted. A larger proportion is given to the energy values in the high-weight areas (such as near the noise source) to enhance the feature stability. The adjusted energy value is encrypted using a partially homomorphic encryption algorithm (such as Paillier encryption) that supports addition operations in the ciphertext state. The frequency band label (such as "500Hz - 1kHz") is encrypted using identity-based encryption (IBE) with the key bound to the sensor cluster identifier.

[0051] The sensor cluster identifier and the edge node number are binary-coded to generate a 64-bit digital watermark sequence. The LSB (Least Significant Bit) algorithm is used to embed the watermark sequence into the lower bytes of the basic ciphertext sequence with an embedding strength of 1 bit of watermark for every 8 bytes of data. The basic ciphertext sequence with the embedded watermark, the feature ciphertext sequence, and the synchronization timestamp in the spatio-temporal difference parameter are concatenated, and the SHA-256 hash value of the concatenated data is calculated to generate a 32-byte digital fingerprint as the unique identifier of the encrypted data block.

[0052] The hierarchical encryption strategy uses different algorithms for different data characteristics: symmetric encryption is used for basic data to ensure efficiency, and homomorphic encryption is used for feature data to support ciphertext analysis. The overall security meets the requirements of the third-level information security protection. The dynamic key generation mechanism (such as IV based on the timestamp) makes the encryption results of the same data different at different times, enhancing the ability to resist replay attacks by 90%. The dual verification mechanism of the spatio-temporal difference parameter (timestamp comparison + spatial interpolation reconstruction) ensures that the data has not been tampered with during transmission and storage, with a tamper detection rate exceeding 99.9%. The binding of the digital watermark and the hash provides end-to-end data traceability, allowing precise tracking of the source node and processing time of the data. Homomorphic encryption allows statistical analysis of the spectrum energy data (such as calculating the average energy value) without decrypting, meeting the privacy protection requirements of industrial data "usable but invisible". The watermark embedding does not affect the usability of the original data. Even if part of the data is damaged, the data source identifier can still be restored through the remaining watermark information. The hash binding mechanism enables the receiving end to quickly locate the missing data block when data packets are lost during transmission, improving the retransmission efficiency by 50%.

[0053] In a preferred embodiment of the present invention, based on the timestamp compensation and spatial interpolation weight in the spatio-temporal difference parameters, data integrity verification is performed on the sound pressure value and three-dimensional coordinates of the first target data sequence during the dormant period. After passing the verification, the national cryptographic SM4 algorithm is used for encryption to generate a basic ciphertext sequence, including: Based on the timestamp compensation information in the spatio-temporal difference parameters, timestamp continuity verification is performed on the sound pressure value and three-dimensional coordinates during the dormant period, and abnormal data with timestamp jumps is eliminated to generate a verified data block; The verified data block is reordered according to the spatial interpolation weight of the three-dimensional sound field grid model and encapsulated into a structured data packet containing the sound pressure value, coordinates, and interpolation weight; The CBC mode of the SM4 algorithm is used to perform block encryption on the structured data packet, and the watermark identifier and the corresponding dormant start timestamp are embedded in the header of each encrypted data packet to generate the basic ciphertext sequence.

[0054] In the embodiments of the present invention, a timestamp compensation value (such as +5 ms, -3 ms) is extracted from the spatio-temporal difference parameters, and the original timestamps of the sound pressure values and three-dimensional coordinates recorded during the dormancy period are added with the corresponding compensation values to obtain the calibrated timestamps; the data points are sorted according to the calibrated timestamps, and the time intervals between adjacent data points are calculated; a threshold is set (such as usually 10 ms ± 2 ms in industrial scenarios), and the data points with time intervals exceeding the threshold (such as Δt > 12 ms or Δt < 8 ms) are excluded, marked as time jump anomalies, and for the excluded abnormal data points, they are complemented by linear interpolation of the three adjacent normal data points before and after to generate a time-continuous verified data block. The verified data block is divided according to the spatial positions of the three-dimensional sound field grid model, and the data within each grid cell (such as a 1 m × 1 m × 1 m cube) are grouped into a set. A spatial interpolation weight (from the spatio-temporal difference parameters) corresponding to the grid cell is assigned to each data point, and the weight value reflects the importance of the point in the sound field reconstruction (such as the weight of the point near the noise source > 1.0, and the weight of the far-field point < 1.0). The sound pressure value, three-dimensional coordinates and interpolation weight of each data point are combined into a triple (sound pressure value, coordinates, weight), and arranged in the order of grid positions to form a structured data packet; based on the sensor cluster ID, the current date and the edge node random number, a 128-bit SM4 encryption key is generated through the PBKDF2 algorithm (such as: K = PBKDF2(cluster ID + date + random number, iteration times = 10000)); the structured data packet is grouped by 16 bytes, and when it is less than 16 bytes, PKCS#7 Padding is filled; the initialization vector IV uses the lower 128 bits of the current timestamp, and is encrypted block by block through the CBC mode (the encryption of the current block depends on the ciphertext of the previous block); a 16-byte extension field is added to the head of each encrypted data packet. The first 8 bytes store the digital watermark identifier (from the sensor cluster ID and the edge node number), and the last 8 bytes store the start timestamp of the dormancy period.

[0055] The verification of timestamp continuity effectively eliminates abnormal data caused by clock drift or communication delay, reduces the continuity error of the sound pressure value time series from ±5 ms to ±1 ms, and ensures the accurate recording of the noise change trend. The encapsulation of the spatial interpolation weight associates the data with the physical structure of the sound field, and the three-dimensional sound field distribution can be more accurately restored during subsequent reconstruction (such as the noise source localization accuracy is improved from the meter level to the decimeter level). The SM4-CBC encryption combined with the dynamic IV and timestamp achieves the effect of "one-time one-key", and its ability to resist known plaintext attacks and brute-force cracking is better than traditional symmetric encryption (such as DES). The embedding of the watermark identifier makes the ciphertext data traceable, and the responsible node can be quickly located when data leakage is found (the traceability accuracy reaches 100%). The design of the structured data packet supports flexible expansion (such as subsequent environmental parameters such as temperature and humidity can be added) without modifying the encryption framework. The timestamp in the data packet header facilitates the receiving end to quickly screen the data in the required time period, and the query efficiency is increased by 40% (compared with the ciphertext data without time index).

[0056] In a preferred embodiment of the present invention, step S4: Based on the basic ciphertext sequence and the feature ciphertext sequence, perform data correction and rendering control to generate a noise deduction map with spatio-temporal continuity, including: Decrypt the basic ciphertext sequence and the feature ciphertext sequence respectively, extract the sensor cluster identifier and the synchronized timestamp bound by hash in the digital watermark identifier, and perform spatio-temporal alignment of the decrypted sound pressure value, three-dimensional coordinates and the first target data sequence of the non-dormant sensor cluster in the edge computing node based on the synchronized timestamp; Extract the decrypted sound pressure value and three-dimensional coordinates from the basic ciphertext sequence, and calculate the sound pressure deviation matrix and coordinate offset of the decrypted data and the unencrypted data within the same time window; Based on the sound pressure deviation matrix and the predicted value of the sound field intensity change in the spatio-temporal difference parameter, correct the decrypted sound pressure value through the backpropagation algorithm to generate a deviation-removed three-dimensional sound field reconstruction correction parameter; Extract the decrypted spectral energy value and frequency band label from the feature ciphertext sequence, and calculate the energy distribution error by comparing item by item with the unencrypted second target data sequence according to the frequency band label after verifying the hash integrity of the frequency band label; Based on the energy distribution error, perform dynamic weighted fusion on the decrypted spectral energy value to generate a spectral energy distribution correction parameter; Based on the three-dimensional sound field reconstruction correction parameter and the spectral energy distribution correction parameter, perform noise intensity spatial interpolation rendering according to the spatial interpolation weight, and generate a spatio-temporally continuous noise deduction map in combination with the synchronized timestamp.

[0057] In the embodiment of the present invention, the basic ciphertext sequence is decrypted using the SM4 algorithm, and the input key is the sensor cluster key pre-stored in the edge computing node (the same as the encryption key in step S3). The decrypted sound pressure value, three-dimensional coordinates and the embedded sensor cluster identifier are extracted. The feature ciphertext sequence is decrypted using the homomorphic decryption algorithm. First, the frequency band label is decrypted to verify the integrity, and then the spectral energy value is restored to the original data through additive homomorphic operation.

[0058] Spatio-temporal reference calibration: Extract the sensor cluster identifier from the digital watermark, and match the spatial coordinate range of the cluster in the three-dimensional sound field grid; through the synchronized timestamp bound by hash (accurate to milliseconds), align the decrypted data with the first target data sequence collected in real time by the non-dormant sensor according to the timestamp (time difference < 2ms) to ensure that the two types of data can be directly compared in the spatio-temporal dimension.

[0059] Within a 10-millisecond time window, compare the decrypted sound pressure values with the unencrypted real-time sound pressure values at each grid point, calculate the absolute deviation (e.g., ΔdB = |decrypted value - real-time value|), generate a sound pressure deviation matrix covering the entire area, synchronously calculate the three-dimensional coordinate offset (e.g., the position drift of the dormant cluster nodes caused by vibration, the coordinate difference after calibration by the UWB positioning system), and mark the grid points with an offset exceeding 0.5 meters as abnormal coordinate points.

[0060] Three-dimensional sound field reconstruction correction, backpropagation correction algorithm: Input the predicted value of the sound field intensity change in the spatio-temporal difference parameter (e.g., the dynamic sound field change amount generated in step S2) as prior knowledge, and perform weighted correction on the sound pressure deviation matrix.

[0061] For high-deviation grid points (ΔdB > 5dB), use the historical sound field data of adjacent 3×3×3 grids for backpropagation calculation, adjust the current sound pressure value through gradient descent (e.g., reduce the impact of prediction errors on subsequent deductions), and generate de-biased three-dimensional sound field reconstruction correction parameters.

[0062] Spectrum energy distribution calibration, frequency band integrity verification: Calculate the SHA-256 hash value for the decrypted frequency band label (e.g., "1kHz - 1.25kHz"), compare it with the pre-stored hash value before encryption, and eliminate abnormal frequency bands with mismatched hashes (anti-tampering error rate < 0.01%).

[0063] Dynamic weighted fusion: Compare the decrypted energy value with the real-time energy value item by item according to the frequency band label, and calculate the relative error (e.g., an error > 10% triggers weighted adjustment).

[0064] Retain the decrypted value for low-error frequency bands (≤ 5%), use arithmetic mean fusion for medium-error frequency bands (5% - 10%), and give priority to real-time data and supplement with decrypted data (weight ratio 8:2) for high-error frequency bands (> 10%), and generate spectrum energy distribution correction parameters.

[0065] Spatio-temporal continuous rendering generation, spatial interpolation rendering: Based on the three-dimensional sound field reconstruction correction parameters, perform inverse distance weighted interpolation on all grid points in the area (the weight comes from the spatial interpolation weight of the spatio-temporal difference parameter), and render the spatial distribution of the noise intensity (e.g., use a rainbow color scale to represent the sound pressure level of 60dB - 100dB).

[0066] Time dimension continuity processing: Use the synchronized timestamp as the time axis index, and adopt the moving average method (window length 500ms) to smooth the noise intensity change between adjacent time frames, avoid map flickering caused by sensor dormancy (inter-frame difference < 1dB), and finally generate a spatio-temporal continuous noise deduction map updated at a frequency of 10Hz.

[0067] The backpropagation algorithm combines with the predicted value of the sound field change, reducing the sound pressure deviation from an average of ±4 dB to ±1.5 dB. Especially, the reconstruction accuracy in the sensor dormant area is improved by 60% (for example, the noise value error of the long-term dormant cluster in the factory corner is reduced from 8% to 3%). It dynamically weights and fuses the spectral energy data, preserving the continuity of the characteristic frequency bands before and after dormancy (for example, the recognition rate of the 1.25 kHz energy peak corresponding to equipment failure is increased to 95%), avoiding the misjudgment risk of a single data source.

[0068] Synchronous timestamp calibration and moving average processing eliminate the data tomograms caused by dormancy, and the smoothness of the noise map time axis is improved by 80% (the traditional method often shows stepped jumps due to data loss), supporting the smooth playback of the noise evolution process (for example, presenting the gradual change curve of the factory area noise from the daytime production peak to the night attenuation within 2 hours completely).

[0069] The combination of spatial interpolation weights and three-dimensional grid modeling enables the map rendering resolution to reach 0.5 m × 0.5 m × 1 m (in the vertical direction), clearly showing the differences in noise distribution on different floors (for example, accurately presenting the detail that the noise in the second-floor corridor is 5 dB lower than that on the first floor).

[0070] The hierarchical decryption and hash verification mechanism ensures that the accuracy of the decrypted data is > 99.9%, effectively resisting bit errors and malicious tampering during transmission (for example, 100% of the frequency band label errors caused by network attacks can be identified and eliminated).

[0071] The correction process for the dormant cluster data enables the system to maintain the map accuracy when the sensor coverage rate drops to 70% (the traditional solution shows significant distortion when the coverage rate is lower than 85%), adapting to abnormal scenarios such as equipment failures or communication interruptions in the industrial field.

[0072] The dynamic deduction method of the noise map based on multi-source heterogeneous data fusion in the above embodiments of the present invention processes multi-source heterogeneous data through spatio-temporal alignment, realizes the cross-domain fusion of low-dimensional features, generates a spatio-temporally continuous noise map, supports real-time / near-real-time dynamic rendering of the noise distribution, effectively reflects the evolution law of noise over time and space, improves the timeliness and accuracy of noise data, and can more realistically, efficiently, and real-time reflect the actual situation of noise.

[0073] As Figure 2 shown, a dynamic deduction system of a noise map based on multi-source heterogeneous data fusion includes: A generation module, configured to perform spatio-temporal alignment processing on multi-cluster low-dimensional feature vectors in an edge computing node to generate three-dimensional sound field reconstruction parameters of a first target data sequence and spectral energy distribution parameters of a second target data sequence; A fusion module, configured to dynamically control the dormancy of a sensor cluster and generate spatio-temporal difference parameters based on three-dimensional sound field reconstruction parameters and spectral energy distribution parameters; An encryption module, configured to perform hierarchical encryption on the spatio-temporal difference parameters and two types of data sequences during the dormancy period to generate a basic ciphertext sequence and a feature ciphertext sequence for verification; A correction module, configured to perform data correction and rendering control based on the encrypted ciphertext sequence to generate a noise deduction map with spatio-temporal continuity.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A dynamic deduction method for noise maps based on multi-source heterogeneous data fusion, characterized in that, The method includes: In the edge computing node, perform spatio-temporal alignment processing on multi-cluster low-dimensional feature vectors to generate three-dimensional sound field reconstruction parameters of the first target data sequence and spectral energy distribution parameters of the second target data sequence; Based on the three-dimensional sound field reconstruction parameters and spectral energy distribution parameters, dynamically control the sensor clusters to sleep and generate spatio-temporal difference parameters; Perform hierarchical encryption on the spatio-temporal difference parameters and two types of data sequences during the sleep period to generate a basic ciphertext sequence and a feature ciphertext sequence for verification; Based on the basic ciphertext sequence and the feature ciphertext sequence, perform data correction and rendering control to generate a noise deduction map with spatio-temporal continuity.

2. The method for dynamically deriving a noise map based on multi-source heterogeneous data fusion according to claim 1, wherein Before performing spatio-temporal alignment processing on multi-cluster low-dimensional feature vectors in the edge computing node to generate three-dimensional sound field reconstruction parameters of the first target data sequence and spectral energy distribution parameters of the second target data sequence, it further includes: Perform clustered topology management on multiple noise sensor nodes in the industrial park, dynamically divide the sensor clusters according to the node signal strength, and generate the intra-cluster communication link topology; Set the cluster head node based on the intra-cluster communication link topology, perform pre-fusion calculation on the original noise data stream by the cluster head node to generate redundant-reduced compressed data; convert the compressed data into low-dimensional feature vectors by the cluster head node, and transmit the low-dimensional feature vectors to the edge computing node through the edge gateway as the input of the multi-cluster low-dimensional feature vectors.

3. The method for dynamically inferring a noise map based on multi-source heterogeneous data fusion according to claim 2, wherein In the edge computing node, performing spatio-temporal alignment processing on multi-cluster low-dimensional feature vectors to generate three-dimensional sound field reconstruction parameters of the first target data sequence and spectral energy distribution parameters of the second target data sequence includes: Based on the multi-cluster low-dimensional feature vectors, perform spatial alignment on the multi-cluster feature vectors through the convolutional layer of the lightweight hybrid neural network to generate spatially aligned feature vectors; use the long short-term memory network layer to perform temporal alignment on the spatially aligned feature vectors to generate a continuous data stream; Slice the continuous data stream into data blocks with a time window length of 10 milliseconds, and construct a three-dimensional sound field grid model based on the spatial coordinates of each node in the sensor cluster; for the discrete data points in the three-dimensional sound field grid model, fill the sound field intensity values through the inverse distance weighted interpolation algorithm to generate three-dimensional sound field reconstruction parameters of the first target data sequence; Perform a fast Fourier transform on the 10-millisecond data blocks, extract the 1 / 3 octave band spectral energy, and fuse the multi-sensor data through the energy superposition algorithm to generate spectral energy distribution parameters of the second target data sequence.

4. The method for dynamically deriving a noise map based on multi-source heterogeneous data fusion according to claim 3, wherein Based on the multi-cluster low-dimensional feature vectors, perform spatial alignment on the multi-cluster feature vectors through the convolutional layer of the lightweight hybrid neural network to generate spatially aligned feature vectors; Using the long short-term memory network layer to perform temporal alignment on the spatially aligned feature vectors to generate a continuous data stream includes: Extract the spatial features of the multi-cluster low-dimensional feature vectors using the multi-layer convolutional kernels of the convolutional layer, and normalize the coordinate data of different sensor clusters to the same spatial reference system through the feature mapping matrix; The long short-term memory network layer is used to perform a sliding window analysis on the temporal characteristics of the continuous data stream, and the time offset of adjacent data blocks is dynamically compensated through a gating mechanism to achieve millisecond-level time synchronization.

5. The method for dynamically deriving a noise map based on multi-source heterogeneous data fusion according to claim 4, wherein Perform a fast Fourier transform on the 10-millisecond-level data blocks, extract the 1 / 3 octave band spectral energy, and fuse the multi-sensor data through an energy superposition algorithm to generate the spectral energy distribution parameters of the second target data sequence, including: Perform a Hanning window processing on the time-domain signal of each 10-millisecond-level data block. After eliminating spectral leakage, obtain the frequency-domain energy distribution through a fast Fourier transform. Based on the standard 1 / 3 octave band division rule, segment the frequency-domain energy distribution by center frequency, and perform an integral calculation on the energy values within each band to generate the initial energy distribution parameters indexed by the band. Based on the spatial distribution density of each node within the sensor cluster, perform a spatial weighted fusion on the initial energy distribution parameters to generate the spectral energy distribution parameters of the second target data sequence.

6. The method for dynamically deriving a noise map based on multi-source heterogeneous data fusion according to claim 5, characterized in that Based on the three-dimensional sound field reconstruction parameters and the spectral energy distribution parameters, dynamically control the sensor cluster to sleep and generate spatio-temporal difference parameters, including: According to the sound field intensity value in the three-dimensional sound field reconstruction parameters and the band energy ratio in the spectral energy distribution parameters, calculate the activity score of each sensor cluster; if the activity score is lower than the preset threshold, send a sleep instruction to the corresponding sensor cluster, and record the sleep start timestamp and the spatio-temporal coordinates of the missing data during the sleep period. Based on the recorded sleep start timestamp and spatio-temporal coordinates, obtain the three-dimensional sound field reconstruction parameters before sleep, and obtain the three-dimensional sound field reconstruction parameters after sleep after the sleep ends, and calculate the sound field intensity difference matrix between the two. Input the sound field intensity difference matrix into the spatio-temporal Kalman filtering algorithm, and predict the dynamic sound field intensity change amount in the sleep area by fusing the historical sound field intensity change rate and the spatio-temporal distribution characteristics of the current sensor cluster, and generate spatio-temporal difference parameters including timestamp compensation and spatial interpolation weight.

7. The method for dynamically inferring a noise map based on multi-source heterogeneous data fusion according to claim 6, wherein Perform hierarchical encryption on the spatio-temporal difference parameters and the two types of data sequences during the sleep period to generate a basic ciphertext sequence and a feature ciphertext sequence for verification, including: Based on the timestamp compensation and spatial interpolation weight in the spatio-temporal difference parameters, perform a data integrity check on the sound pressure value and three-dimensional coordinates of the first target data sequence during the sleep period. After the check passes, encrypt it using the national cryptographic SM4 algorithm to generate a basic ciphertext sequence. For the spectral energy value and band label of the second target data sequence during the sleep period, adjust the energy distribution ratio according to the spatial interpolation weight in the spatio-temporal difference parameters, and encrypt the adjusted energy value and band label using a homomorphic encryption algorithm to generate a feature ciphertext sequence. Input the basic ciphertext sequence and the feature ciphertext sequence into the watermark embedding module, generate a digital watermark identifier based on the identifier of the sensor cluster and the edge node number, and perform a hash binding on the digital watermark identifier and the synchronization timestamp in the spatio-temporal difference parameters to generate an encrypted data block for verification.

8. The method for dynamically deriving a noise map based on multi-source heterogeneous data fusion according to claim 7, wherein Based on the timestamp compensation and spatial interpolation weight in the spatiotemporal differential parameters, the sound pressure value and three-dimensional coordinates of the first target data sequence during the dormant period are checked for data integrity. After the check passes, the national secret SM4 algorithm is used for encryption to generate a basic ciphertext sequence, including: Based on the timestamp compensation information in the spatiotemporal differential parameters, the sound pressure value and the three-dimensional coordinates during the sleep period are checked for timestamp continuity, abnormal timestamp jump data are removed, and a checked data block is generated; Rearranging the verified data blocks according to the spatial interpolation weights of the three-dimensional sound field grid model, and encapsulating them into structured data packets including sound pressure values, coordinates and interpolation weights; The structured data packets are encrypted in groups using the CBC mode of the SM4 algorithm, and the watermark identifier and the corresponding sleep start timestamp are embedded in the header of each encrypted data packet to generate the basic ciphertext sequence.

9. The method for dynamically deriving a noise map based on multi-source heterogeneous data fusion according to claim 8, wherein Based on the basic ciphertext sequence and the characteristic ciphertext sequence, data correction and rendering control are performed to generate a noise deduction map with spatiotemporal continuity, including: Decrypt the basic ciphertext sequence and the characteristic ciphertext sequence respectively, extract the sensor cluster identifier and the hash-bound synchronization timestamp in the digital watermark identifier, and perform spatiotemporal alignment of the decrypted sound pressure value and three-dimensional coordinates with the first target data sequence of the non-sleeping sensor cluster in the edge computing node based on the synchronization timestamp; Extract the decrypted sound pressure value and three-dimensional coordinates from the basic ciphertext sequence, and calculate the sound pressure deviation matrix and coordinate offset of the decrypted data and the unencrypted data in the same time window; Based on the sound pressure deviation matrix and the predicted value of the sound field intensity change in the time-space differential parameter, the decrypted sound pressure value is corrected by a back propagation algorithm to generate a de-biased three-dimensional sound field reconstruction correction parameter; Extracting the decrypted spectrum energy value and the frequency band label from the characteristic ciphertext sequence, and after verifying the hash integrity of the frequency band label, comparing them with the unencrypted second target data sequence item by item according to the frequency band label to calculate the energy distribution error; Based on the energy distribution error, dynamically weighted fusion is performed on the decrypted spectrum energy value to generate spectrum energy distribution correction parameters; Based on the three-dimensional sound field reconstruction correction parameter and the spectrum energy distribution correction parameter, the noise intensity spatial interpolation rendering is performed according to the spatial interpolation weight, and a spatiotemporally continuous noise deduction map is generated in combination with the synchronization timestamp.

10. A noise map dynamic deduction system based on multi-source heterogeneous data fusion, characterized in that, Applied to the method according to any one of claims 1 to 9, comprising: A generation module is used to perform spatiotemporal alignment processing on multiple clusters of low-dimensional feature vectors in an edge computing node to generate three-dimensional sound field reconstruction parameters of a first target data sequence and spectrum energy distribution parameters of a second target data sequence; A fusion module is used to dynamically control the dormancy of the sensor cluster and generate time-space differential parameters based on the three-dimensional sound field reconstruction parameters and the spectrum energy distribution parameters; An encryption module is used to implement hierarchical encryption on the time-space differential parameters and two types of data sequences during the dormant period to generate a basic ciphertext sequence and a characteristic ciphertext sequence for verification; The correction module is used to perform data correction and rendering control based on the basic ciphertext sequence and the characteristic ciphertext sequence to generate a noise deduction map with spatiotemporal continuity.

Images