Wireless data compression transmission method based on data characteristics

Abstract

The application provides a wireless data compression transmission method based on data characteristics, and the method comprises the following steps: a node terminal device collects multi-source monitoring data and constructs an initial data matrix; a segmented aggregate approximation is used for time domain dimension reduction, and an aggregate matrix and a fluctuation characteristic matrix are extracted; Chebyshev polynomial approximation is used for frequency domain characteristic modeling of a spectrum diagram data, and a spectrum characteristic matrix is constructed; multi-source data fusion and importance scoring are performed on the three matrices, and high-value data and low-value data are marked according to the comparison of the score and a dynamically selected threshold value; adaptive dictionary learning sparse coding is used for sparse representation of the high-value data and the must-transmit data; asymmetric autoencoder compression coding is used; the number of quantization bits and the coding rate are adaptively determined according to the signal-to-noise ratio and the available bandwidth, and are packaged and sent to a platform terminal device for reconstruction. The application greatly reduces the data transmission amount, saves the bandwidth and the node energy consumption under the premise of ensuring the analysis accuracy.

Description

Technical Field

[0001] This invention relates to the field of low-altitude wireless signal monitoring and data transmission technology, and in particular to a wireless data compression and transmission method based on data characteristics. Background Technology

[0002] Low-altitude wireless signal monitoring systems typically consist of node-end equipment (spectrum detection devices mounted on UAVs) and platform-end equipment (ground data processing platforms). During flight, the node-end equipment continuously collects various types of monitoring data, including spectrum data, GPS trajectory data, signal strength time-series data, and subcarrier status data. This data needs to be transmitted to the platform-end equipment via a wireless link to perform tasks such as UAV target detection, localization, tracking, signal analysis, and interference investigation. Existing technologies include data compression schemes for wireless sensor networks, such as time-series dimensionality reduction based on piecewise aggregation approximation, spectrum approximation compression based on Chebyshev polynomials, compressed sensing sparse basis construction based on adaptive dictionary learning, and feature selection transmission based on channel semantic importance assessment.

[0003] However, the aforementioned existing technologies still have the following shortcomings. First, existing methods typically only compress a single type of data, failing to fully utilize the complementary information between multi-source heterogeneous data in low-altitude wireless signal monitoring systems. The inherent correlations between spectrogram data, GPS trajectory data, and signal strength time-series data in the time, frequency, and spatial domains are not incorporated into the compression decision. Second, existing methods employ fixed compression strategies, failing to differentiate processing based on the actual information value of the data. High-value signal components and low-value environmental noise are treated equally, resulting in inefficient use of wireless transmission bandwidth. Third, there is a lack of effective coordination and feedback mechanisms between different compression algorithms. Time-domain compression and frequency-domain compression operate independently, failing to form a closed loop of "compression-evaluation-feedback-optimization." Summary of the Invention

[0004] This invention provides a wireless data compression and transmission method based on data features. By extracting features at multiple levels, the original monitoring data is decomposed into components with different information values, and selective transmission is performed according to the information value density, thereby significantly reducing the amount of data transmission while ensuring the accuracy of analysis.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: Step S1: Collect various types of raw monitoring data at multiple times, and construct an initial data matrix based on the raw monitoring data. The raw monitoring data includes spectrum data, signal strength time series data, and subcarrier state data. The subcarrier state data is organized into a subcarrier state data matrix. Step S2: The segmented aggregation approximation method is used to perform time-domain dimensionality reduction on each column of the initial data matrix to obtain the aggregation matrix. The fluctuation amount between adjacent time windows is extracted based on the aggregation matrix to obtain the fluctuation feature matrix. Step S3: The Chebyshev polynomial approximation method is used to perform frequency domain nonlinear feature modeling on the spectrum data to obtain the spectrum feature matrix; Step S4: The aggregation matrix, fluctuation feature matrix and spectrum feature matrix are fused. The importance score of the average signal strength and signal strength time series data in the spectrum graph within each time window in the aggregation matrix is ​​calculated to obtain the comprehensive importance score. Based on the comprehensive importance score and the available bandwidth of the current wireless link, the data to be transmitted is selected, and the remaining data is discarded. Step S5: Adaptive dictionary learning sparse coding method is used to perform sparse representation and compression coding on the aggregated value corresponding to the transmitted data, the GPS trajectory aggregated value of the aggregated matrix, the subcarrier state vector of the subcarrier state data matrix and the mean vector of the spectrum feature matrix to obtain the sparse coefficient vector. Step S6: The sparse coefficient vector is compressed and encoded using an asymmetric autoencoder to output a latent space encoded vector, and the latent space encoded vector is decoded and restored. Step S7: Organize the latent space coding vector corresponding to the transmitted data and the statistical summary of the discarded data into a transmitted data vector, and encapsulate the transmitted data vector into a data packet after adaptive encoding.

[0006] In summary, this invention has at least the following beneficial effects: First, it performs joint time-domain dimensionality reduction on multi-source heterogeneous monitoring data using a piecewise aggregation approximation method, compressing the original data from N time points into aggregated data for K time windows. Second, it models the frequency-domain nonlinear characteristics of spectrogram data using a Chebyshev polynomial approximation method, compressing the spectral data from P frequency sampling points into Q+1 Chebyshev approximation coefficients. Third, it employs an importance scoring mechanism based on multi-source data fusion, comprehensively considering three dimensions—fluctuation amplitude, statistical significance, and spectral correlation—to differentiate the scoring of each data type within each time window, dynamically adjusting the selection threshold based on available bandwidth. Fourth, it utilizes adaptive dictionary learning sparse coding, achieving a compact representation of high-value data through linear combinations of a few atoms in an overcomplete dictionary. Fifth, it fully leverages the differences in computing power between the node-side encoding and platform-side decoding using an asymmetric autoencoder. Attached Figure Description

[0007] Figure 1 This is a flowchart illustrating a wireless data compression and transmission method based on data characteristics.

[0008] Figure 2 This is a schematic diagram of the multi-source data fusion and importance scoring process involved in this invention.

[0009] Figure 3 This is a schematic diagram of the adaptive dictionary learning sparse coding process involved in this invention.

[0010] Figure 4 This is a schematic diagram of the asymmetric self-encoder and data encapsulation and transmission process involved in this invention. Detailed Implementation

[0011] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0012] like Figure 1 As shown, this embodiment provides a wireless data compression and transmission method based on data characteristics.

[0013] In step S1, the node-end equipment in the low-altitude wireless signal monitoring system simultaneously collects multiple types of raw monitoring data. The spectrum analysis module on the node-end equipment scans the airborne wireless signal at a fixed sampling period to obtain spectrum data. The GPS receiver module on the node-end equipment records the spatial coordinates of the node-end equipment at the same sampling period to obtain GPS trajectory data, which includes longitude, latitude, and altitude components. The signal strength detection module on the node-end equipment records the signal strength values ​​at each frequency point at the same sampling period to obtain signal strength time-series data. The subcarrier status monitoring module on the node-end equipment records the occupancy status and signal quality parameters of each subcarrier at the same sampling period to obtain subcarrier status data. All four acquisition modules share the same clock source, provided by the system clock of the node-end equipment. The sampling trigger signal is generated by the main control module of the node-end equipment, which has a built-in hardware timer with a timer period of... A trigger interrupt is generated. When the trigger interrupt arrives, the main control module simultaneously sends latch enable signals to the data latch registers of each of the four acquisition modules. The four acquisition modules complete their respective data latching operations within the same clock cycle. The latching time accuracy is determined by the clock source of the hardware timer, and the accuracy is better than [previous value]. Microseconds. After latching is complete, each module sequentially reads the latched data into the main control module's memory buffer and organizes it into a data vector. Sampling period. Values Instant Between seconds, a typical value is Second.

[0014] Assume that the node-end equipment collects a total of [number] data within a complete monitoring cycle. Monitoring data at time point 1. For the 1st... At that moment The collected data vector is represented as follows: ; in Indicates the first At the [time]th moment Scalar representative values ​​for various monitoring data types. For time index, Index for data types, The total number of time points within the monitoring period. Total monitoring duration The relationship between them is superscript This represents the vector transpose operation. For the first A complete data vector at each time point. Total number of data types. Values Data type index The mapping relationship is as follows: Corresponding spectrum data, The longitude component corresponding to GPS trajectory data, The latitude component corresponds to the GPS trajectory data. The elevation component corresponding to GPS trajectory data. Corresponding signal strength time series data, This corresponds to the subcarrier state data. This mapping relationship remains consistent in the importance score of step S4 and the data type identification of step S7. Specifically, Take the first The spectrum data at each time point in the entire The average signal strength value at each frequency sampling point, in dBm; Take the first The longitude value of the GPS trajectory data at each moment, in degrees; Take the first The latitude value of the GPS trajectory data at each moment, in degrees; Take the first The elevation values ​​of the GPS trajectory data at each moment, in meters; Take the first The average signal strength value of the time-series signal strength data at each moment over a preset set of monitoring frequencies, in dBm; Take the first The number of occupied subcarriers in the subcarrier state data at each time point is dimensionless. It should be noted that this scalar value is only used for the existence marker in the "Data Type Identifier" of the data packet header in step S7 (i.e., to indicate to the platform that subcarrier state data exists in this period). It does not participate in the temporal aggregation in step S2, the importance scoring in step S4, or the construction of the vector to be encoded in step S5. The complete information of the subcarrier state data (i.e....) The vector is still compressed and transmitted using the vector to be encoded in step S5. Meanwhile, the... The complete spectrum data at each time point is saved as a spectrum vector. ,in For the first At the [time]th moment The signal strength values ​​at each frequency sampling point, p=1,2,...,P, This represents the total number of frequency sampling points. (All) The spectrum vectors at each time point form the spectrum data matrix. . No. The complete subcarrier state data at each time point is saved as a subcarrier state vector. ,in For the first At the [time]th moment The state coding values ​​of each subcarrier, r=1,2,...,R, This represents the total number of subcarriers. (All) The subcarrier state vectors at each time point form the subcarrier state data matrix. Total number of subcarriers This is determined during system deployment and remains unchanged during system operation; adjustments may be necessary due to hardware replacement or changes in the deployment scenario. Then, it is necessary to recalibrate the fixed normalization parameters, reinitialize the overcomplete dictionary and the asymmetric autoencoder. In actual deployment, this is achieved by setting a monitoring cycle. The lower limit of the value is This is to avoid situations where there is insufficient data.

[0015] Subcarrier state coding value It is a floating-point number, and its value range is... , indicating the first Each subcarrier at time The signal quality normalized evaluation value. This coded value is obtained by the subcarrier state monitoring module of the node device through a comprehensive mapping based on the subcarrier occupancy status and the measured signal-to-noise ratio value. The mapping formula is as follows: ,in For the first At the [time]th moment Measured signal-to-noise ratio (in dB) on each subcarrier. , These are the preset mapping parameters. Because... For a numerical continuous variable, a window average is taken in step S5 (i.e., ...). It has a clear physical meaning, representing the average signal quality level of the subcarrier within the time window.

[0016] All The data vectors at each time point are arranged in chronological order to construct the initial data matrix. : ; in express OK The space of real matrices in columns, This is the initial data matrix.

[0017] Initial data matrix Each row corresponds to the scalar representative value of all monitoring data at a given time point. This matrix completely records the scalar representative values ​​of all raw monitoring data within the monitoring period. Initial Data Matrix Spectrum data matrix and subcarrier state data matrix The construction process is completed in the local memory of the node device, and the completed data is stored in the local memory of the node device.

[0018] In step S2, a piecewise aggregation approximation method is used to approximate the initial data matrix. The data in columns 1 to 5 undergo time-domain dimensionality reduction to extract feature sequences that characterize the overall fluctuation trend of the data; the data in column 6 is not included in time-domain aggregation. The core operation of piecewise aggregation approximation is to segment the data into columns of length 1 to 5. Time series are divided into A time window of equal length, in which The preset number of time windows, Each time window contains One original data point, operator Indicates floor function. Preset time window length. Values to Integers between [a certain range]. Values ​​based on monitoring cycle and preset time window length Sure, The actual number of data points contained in each time window. Values The actual duration of the time window is... ,when Second, Coverage per time window Seconds of monitoring data, when Second, Coverage per time window Monitoring data in seconds. The value is determined based on the time-varying characteristics of the actual monitored signal; a smaller value is taken for scenarios where the signal changes rapidly. To retain higher temporal resolution, a larger value is taken for scenarios where the signal changes slowly. To achieve a higher compression ratio. ,but If the conditions for performing segmented aggregation approximation are not met at this point, the method terminates and returns an error code. The node device records log information indicating that the monitoring period is too short to perform segmented aggregation approximation in its local memory, and temporarily stores all raw monitoring data in its local memory until more data is accumulated in the next monitoring period before processing. If the node device does not have enough time windows in its local storage to perform the fluctuation feature extraction log information, then the method will terminate and return an error code. and Hengyou The above termination conditions will not be triggered in actual operation. Cannot be Divisible by, then the last digit If any data point is not covered by any time window, these uncovered data points will be discarded and will not participate in subsequent feature extraction and compression transmission. The number of discarded data points is less than [number missing]. One, accounting for The proportion is less than .

[0019] For the Time series of various data types ( ), No. Time window The aggregate value is all values ​​within this window. Arithmetic mean of the data points: ; in Indicates the first The first type of data type Aggregated values ​​within a time window For the time-series index of data points within the window, The number of raw data points contained in each time window.

[0020] All The aggregated values ​​of the nth time window are arranged in chronological order to obtain the nth time window. Aggregated sequences of various data types Perform the above aggregation operation on data types 1 through 5 respectively to obtain the dimensionality-reduced aggregation matrix. : ; Aggregation Matrix Dimensions Much smaller than the initial data matrix Dimensions Aggregation matrix Each element The physical meaning of is the first The first type of data type The average value within a time window.

[0021] In obtaining the aggregation matrix Based on this, we further extract the fluctuation characteristics of each data type between adjacent time windows. For the first... Types of data ( ), define the first The first time window and the first Fluctuation between time windows for: ;in For the first The first type of data type The time window is relative to the first The fluctuation amount within a time window.

[0022] For all The fluctuation amount is calculated using adjacent windows to obtain the fluctuation characteristic matrix. : ; Fluctuation characteristic matrix Each row of elements reflects the magnitude of change in various monitoring data within that time window relative to the previous window. (Aggregation Matrix) and fluctuation characteristic matrix All are stored in the local memory of the node-side device.

[0023] Both steps S2 and S3 use the initial data matrix output in step S1. and spectrum data matrix Since there is no data dependency between the two steps (S2 and S3) as inputs, they can be executed in parallel provided the computing resources of the node device allow. When the node device has a multi-core processor or a hardware architecture that supports instruction-level parallelism, the parallel execution of steps S2 and S3 can shorten the overall processing time. When the node device has a single-core processor or limited computing resources, steps S2 and S3 are executed sequentially without affecting the final output. Regardless of whether parallel or serial execution is used, step S4 waits for steps S2 and S3 to complete before starting its execution.

[0024] In step S3, for the specific data type of spectrogram data, the Chebyshev polynomial approximation method is used to model its nonlinear characteristics in the frequency domain, using a small number of polynomial coefficients to characterize the overall outline of the spectrogram. The highest order of the Chebyshev polynomial... This is a preset fixed value, set during system deployment based on the typical spectral complexity of signals within the monitoring frequency band. The value ranges from 5 to 15 and remains unchanged during system operation. Adjustments may be necessary if significant changes occur due to changes in the monitoring scenario. If so, the complete dictionary and asymmetric autoencoder need to be reinitialized.

[0025] Suppose that the spectrum data collected at a certain moment includes Each frequency sampling point corresponds to a frequency value. The signal strength value is Frequency value , , The unit for all values ​​is Hertz (Hz). In practice, the frequency values ​​are recorded in Hz by the spectrum analysis module and stored in the local memory. The aforementioned frequency and signal strength values ​​are acquired by the spectrum analysis module of the node device in step S1 and stored in the local memory. Step S3 reads the aforementioned data from the local memory.

[0026] First, the frequency values ​​are linearly mapped to the interval. The normalized frequency values ​​are obtained from the above. : ; in For the first The normalized frequency value corresponding to each frequency sampling point For the first The original frequency values ​​of each frequency sampling point For all The minimum frequency among the frequency sampling points For all The maximum frequency value among each frequency sampling point For frequency sampling point index, This represents the total number of frequency sampling points.

[0027] Chebyshev polynomials In the interval Regarding the weight function Orthogonal, its recurrence relation is: ;in For the first Chebyshev polynomials in independent variable The value at that location, Let be the order of the polynomial. The normalized frequency independent variable is denoted as .

[0028] Using Chebyshev polynomials to analyze spectral data Order approximation. Coefficients Calculated using the discrete Chebyshev transform: ; in For the first Approximation coefficients of the Chebyshev polynomial, For polynomial order index, The highest order of the predefined polynomial. For the first Signal strength values ​​at each frequency sampling point For the first Chebyshev polynomials in normalized frequency values The value at that location.

[0029] Will The spectral feature vector is composed of several approximation coefficients. ,in For spectral feature vectors, express A real vector space. For all samples collected in step S1... The spectrum data at each time point were subjected to the Chebyshev polynomial approximation operation described above, resulting in... spectral feature vectors ,in Indicates the first The spectral feature vector at each time step. Use time indexes. Arrange these feature vectors in chronological order to construct a spectral feature matrix. : ; in Indicates the first spectral data at time n Chebyshev approximation coefficients This is the spectral characteristic matrix. It is stored in the local memory of the node device.

[0030] In step S4, the aggregation matrix output in step S2 is processed. With the fluctuation characteristic matrix and the spectral feature matrix output in step S3 The data is fused and scored according to its importance across different data types and time windows. Based on the scoring results, high-value data (i.e., the data to be transmitted) is dynamically selected for priority transmission. Among these, the GPS trajectory data consists of three components (…). Subcarrier status data ( and the mean vector of spectral features for each time window. All data is mandatory and is not included in the high / low value labeling; only the average signal strength of the spectrum is considered. ) and signal strength time series data ( Participate in high-value / low-value selection. Those deemed low-value... or The original time data corresponding to the data (i.e., the discarded data). (From the initial data matrix in step S1) After step S4 is completed and before step S7 is executed, the data is temporarily stored in the local memory of the node device, pending the completion of step S7 to determine the variance. It can be released after calculation; the mean value corresponding to data judged as low value. It is then continuously stored in local memory for step S7 to read and transmit as a statistical summary.

[0031] Definition of the first The first time window Local importance score of various data types ,in The score integrates information from three dimensions: the fluctuation range of the data within the window, the statistical significance of the data type over the entire monitoring period, and the correlation between the data and spectral characteristics.

[0032] Volatility Contribution The fluctuation characteristic matrix is ​​directly taken from step S2. The corresponding absolute value of the fluctuation: ;in For the first The first time window The contribution value of the fluctuation range of each data type. This indicates the absolute value operation. For ,definition .

[0033] Statistical significance contribution Measure the first The dispersion of this data type throughout the entire monitoring period, using all of this data type Calculation of the standard deviation of aggregated values ​​for each window: ; in For the first The statistical significance contribution values ​​of each data type For the first All data types The average of aggregated values ​​across windows, To represent the square root operation, The aggregate value defined in step S2, The range of values ​​is .

[0034] Spectral Correlation Contribution Measure the first Within the first time window The degree of correlation between data types and spectral characteristics. Let the first type be... The original time index range covered by each time window is: ,in For the first The set of time indices covered by the nth time window. The mean of the original data of each data type is the aggregated value in step S2. Its standard deviation within the window Calculate using the following formula: ; in For the first The first type of data type The standard deviation within a time window The first one defined in step S1 At the [time]th moment Scalar representative values ​​for various monitoring data types. The range of values ​​is .

[0035] The spectral feature vector within this window is the first The mean and standard deviation of each component are as follows: ; ; in For the first Within the nth time window, the spectral feature vector is... The mean of each component, For the first At the [time]th moment Chebyshev approximation coefficients For the first The spectral feature vector within the nth window Each component in the whole The standard deviation at each time point.

[0036] No. The first type of data type Within the window and the spectral characteristics Pearson correlation coefficient between the components Defined as: ; in For the first Types of data and spectral characteristics The component in the first The covariance within each window is calculated using the following formula: ; like or Then define Spectral correlation contribution Defined as the correlation coefficient mentioned above in all Absolute value average across each spectral feature dimension: ; in For the first The first time window The spectral correlation contribution value of various data types The highest order of the Chebyshev polynomial defined in step S3. This is the index of the spectral feature vector dimension, with values ​​ranging from 1 to 2. . The physical meaning is: in all On the spectral feature dimension, the first Various data types in the window The average correlation strength between the original data sequence and the spectral feature component sequence.

[0037] By weighting and combining the three contribution factors mentioned above, we obtain the first... The first time window Comprehensive importance score of various data types : ;in For the first The first time window A comprehensive importance score for each data type. The range of values ​​is , The preset weighting coefficients that contribute to the fluctuation range. Preset weighting coefficients for contributions to statistical significance. Preset weighting coefficients contributing to spectral correlation. , For all Each time window and all The maximum contribution of fluctuation amplitude across different data types. For all The maximum statistically significant contribution across different data types. For all Each time window and all The maximum contribution of spectral correlation on each data type. If , or Then the corresponding normalization term directly takes the value. .because And all normalized values ​​are within Within the interval, The range of values ​​is strictly defined as follows: With dynamic selection threshold The range of values ​​is consistent, ensuring that and Comparability between them. , , The values ​​are respectively , , The values ​​in this set were obtained by optimizing historical monitoring data using a grid search method. The optimization objective of the grid search is to maximize the Pearson correlation coefficient between the reconstructed data from the platform and the original data, and the search space is [formula missing]. , , In recent times After iterating through all combinations of historical data for each monitoring period, the set of values ​​that maximizes the average Pearson correlation coefficient is selected. If local storage is insufficient after the initial deployment of the node device or a system reset... For historical data from a monitoring period, the grid search uses the system's preset default weight values. , , It is directly used as a weighting coefficient in the importance score. As the monitoring period accumulates, when the historical data in the local storage reaches... After one monitoring period, grid search automatically starts and overrides the default weight values. Pearson correlation coefficient. Calculate using the following formula: ; in To assess the total number of samples, For the first The original data values ​​of each evaluation sample For the first The reconstructed values ​​of the platform-side equipment for each evaluation sample. For all The mean of the original data values, For all The mean of the reconstructed values. The range of the Pearson correlation coefficient is... The closer This indicates a higher degree of reconstruction fidelity.

[0038] Set a dynamic selection threshold Its value is determined based on the currently available bandwidth. Total amount of data to be transmitted The ratio is adaptively adjusted: The average quantization bit depth actually used in step S7 of the previous monitoring cycle is calculated as follows: Calculation, where For the first The number of quantization bits corresponding to each time window of data; if it is the first monitoring period, then Take the default value . The node-end equipment is estimated using the following formula: ; ; in To dynamically select the threshold, The available bandwidth of the current wireless link. This is the current transmission time slot length. The total amount of data to be transmitted. The preset transmission efficiency coefficient, The value is 0.8. When this data is marked as high-value data (i.e., transmission data), it is included in the transmission candidate set; when At that time, the data was marked as low-value data (i.e., discarded data) and was discarded in this transmission, but the mean value corresponding to the low-value data remained. The statistical summary data is retained in local storage for transmission in step S7. All data marked as high-value data... and its corresponding window index and data type index Combining high-value data sets The data is stored in the local memory of the node device. The GPS trajectory data consists of three components (…). ) and subcarrier state data ( This data does not participate in the above scoring and labeling process; its window aggregation value or complete window vector is required data to be transmitted. Spectral feature mean vector. As mandatory data, it is not included in the high / low value selection. The multi-source data fusion and importance scoring process is as follows: Figure 2 As shown.

[0039] In step S5, an adaptive dictionary learning sparse coding method is used to further sparsely represent and compress the selected high-value data, the mandatory GPS aggregation values, the mandatory subcarrier state vectors, and the mandatory spectral feature mean vectors.

[0040] For each time window Construct the data vector to be encoded ,in It is a fixed dimension. The components are arranged in the following order: the first 5 positions correspond to to The five scalar data types, among which and Enter the aggregated value in this window based on the high / low value status. (High value) or fixed normalized mean (Low value) Always fill in the aggregate value of this window ; 6th to 6th The position is the mean vector of the spectral features that must be transmitted. ;at last Each position represents a mandatory complete subcarrier state vector. .in That is, the first of the aforementioned fixed normalization parameters The mean of the components, because the first Component index at each position With data type index The values ​​are the same ( , Since the mean vector of spectral features and the subcarrier state vector always exist, all windows are sparsely encoded, and the vector dimension is fixed.

[0041] Let the data vector to be encoded be Before sparse coding, the data vector to be encoded... The components are preprocessed by normalization: using preset fixed normalization parameters, that is, subtracting the corresponding fixed mean from each component. Then divide by the corresponding fixed standard deviation ,in This is for component indexing. The calibration method for fixed normalization parameters is as follows: take the data collected before system deployment. Data from historical monitoring periods, Values This value is based on the following considerations: when Greater than or equal to At that time, according to the central limit theorem, the error in estimating the sample mean is less than 1 / 3. Stable normalized parameter estimates can be obtained; if Less than If the variance of the normalized parameter estimate is large, it may affect the performance of subsequent sparse coding. If the historical data collected before system deployment is insufficient... If the cycle is 1, the system default value will be used. , During the calibration phase, fill the first five positions with the window aggregate values ​​of the corresponding data types. (That is, assuming all types are high value), 6th to 7th Fill in the mean vector of spectral features at each position. ,at last Fill the positions with the subcarrier state vector. This forms the calibration vector to be encoded; for all Each calibration vector in a historical monitoring cycle is encoded according to its components. Calculate the arithmetic mean as Calculate the standard deviation as If the calculated standard deviation Then set it to After calibration, the system deployment tool writes the fixed normalization parameters into the local memory of the node-side devices and platform-side devices; if the platform-side devices do not have these parameters pre-configured, they are distributed through the configuration interface during system deployment. These fixed normalization parameters are stored in the local memory of the node-side devices and platform-side devices and remain unchanged during system operation. The fixed normalized mean value filled in the first 5 positions for low-value types in step S5... That is, the first Fixed mean of each component The data after normalization preprocessing has unified dimensions, and the regularization parameter of sparse coding is... Values Since the fixed normalization parameters do not change with the monitoring period, the atoms learned by the overcomplete dictionary in each monitoring period are always located in the same normalization space, ensuring the effectiveness of dictionary learning.

[0042] Complete dictionary Past data saved from the local memory of the node device using an adaptive dictionary learning method The data is learned online from the data vector to be encoded (i.e., the normalized vector) for each monitoring period, where... Values to Integers between [values]. During the initial phase of initial deployment of node-side devices or system reset (i.e., dictionary cold start phase), these do not yet exist in local memory. Historical data vectors for each monitoring period. At this stage, a complete dictionary is used. Initial atoms are generated directly using a random initialization method, specifically: each atom The components are from the mean value Standard deviation is Random sampling is performed on each atom from a standard normal distribution. Norm normalization. During the dictionary cold start phase, the node device uses the initialized overcomplete dictionary to perform the sparse coding operation in step S5. Before the start of the first monitoring cycle, the node device transmits this initial cold start dictionary to the platform device via the dictionary synchronization channel. The platform device receives and stores this dictionary for decoding. When the node device accumulates... After the historical data vectors for each monitoring period are obtained, the complete dictionary learning update process begins according to the steps described in S5. The periodic triggering conditions were executed normally.

[0043] Due to the total number of moments in each monitoring cycle There may be differences, resulting in aggregation matrices for different monitoring periods. number of rows Different. When constructing the historical data vector set, the following alignment method is used: Take all In each monitoring cycle median of values As a uniform number of windows. For Value greater than The monitoring cycle, from which random selections are made. A data vector for each time window; for Value less than The monitoring cycle is used to linearly interpolate and extend the data vector in the time dimension to... Each point. After the above alignment operation, all The number of data vectors in each monitoring period is consistent, and they can be concatenated into a historical data vector set for dictionary learning. When the total number of historical data vectors is... At that time, a historical data matrix was formed. ,in This represents the total number of historical data vectors.

[0044] Complete dictionary Include Atoms Total number of atoms The value of is related to the dimension of the vector to be encoded. Linkage, to satisfy the overcomplete conditions ( ), specifically defined as ,in This is for floor function operations. In a typical deployment scenario (total number of subcarriers)... ,at this time )Down, , The number of atoms increases linearly with dimension, always satisfying In the total number of subcarriers In larger extreme scenarios, this formula... limit The upper limit does not exceed At the same time, through ensure At least for This ensures that the overcompleteness condition is strictly met under all deployment conditions. . For the first Atom vectors, For atomic indexes.

[0045] The goal of sparse coding is to find a sparse coefficient vector. , making Approximating the normalized data vector to be encoded in the least squares sense ,at the same time Minimize the number of non-zero elements: ; in It is a sparse coefficient vector. express Norm, express Norm, For regularization parameters, Values .

[0046] The above optimization problem is solved using an iterative threshold shrinkage algorithm. The initial value for the iteration... Let it be a vector of all zeros, that is The iterative shrinkage threshold algorithm's first step The update rule for the next iteration is: ;in Index for iteration count, For the first The sparse coefficient vector after the nth iteration. For step size parameters, For an overcomplete dictionary matrix The transpose of the matrix, This is a soft-thresholding shrinkage operator. To ensure the convergence of the iterative shrinkage thresholding algorithm, the step size parameter... The value of must satisfy ,in For matrix The largest eigenvalue. In embodiments of the present invention, an overcomplete dictionary... Each atom is normalized (i.e. , ), making In actual implementation, take ,in The calculation is repeated after each dictionary update. The termination condition for the iterative shrinking threshold algorithm is: if... If convergence is achieved, the iteration is stopped; otherwise, if the relative change between the sparse coefficient vectors of two adjacent iterations is less than a preset threshold, the iteration is stopped. Stop iteration when, i.e. ,in Values ; or when the number of iterations Reaching the preset maximum number of iterations Stop iteration when, where Values The iteration stops when any one of the three termination conditions is met. (Soft threshold shrinkage operator) Acting on vectors Each component: ;in The vector after soft threshold shrinkage is the first The values ​​of each component, The threshold parameter for the soft threshold shrinkage operator. This is a sign function; its value is [value] when the independent variable is greater than zero. When less than zero, the value is taken as When it equals zero, it takes the value of , For vectors The Each component.

[0047] Complete dictionary The dictionary is learned online from historical monitoring data using an adaptive dictionary learning method. The dictionary learning process alternates between sparse encoding and dictionary update steps. The complete update process of dictionary learning occurs every [period missing]. Each monitoring cycle is executed once, among which Values The cumulative data rate of dictionary updates is related to the number of time windows in step S2. Directly related. When When taking the smaller value A larger dictionary allows for more training samples to be obtained through dictionary updates, but the computational cost of a single update also increases accordingly; when When taking a larger value Smaller size reduces the computational cost of dictionary updates, but may result in insufficient sample diversity. Is Typical value to The dictionary was determined under certain conditions, balancing the sufficiency of samples for dictionary updates with the reasonableness of computational overhead. During each monitoring cycle, the node-side device uses the currently available overcomplete dictionary when performing sparse coding in step S5. Encode the code without triggering a dictionary update step. When the accumulation reaches [a certain value]... After one monitoring cycle, the node-end device uses this A complete dictionary update process is performed on all accumulated data vectors to be encoded (after normalization) within each monitoring period. Furthermore, if the currently used dictionary is incomplete... In continuous The average reconstruction error within a monitoring period exceeds a preset threshold. ( Values If so, the dictionary update will be triggered ahead of time, without waiting for the dictionary to be full. One cycle. The average reconstruction error is defined as: ;in To assess sample size, For the first The normalized vector of each evaluation sample. For the first A sparse coefficient vector of each evaluation sample.

[0048] The maximum cumulative number of samples in the data matrix to be encoded used for dictionary updates is: When the accumulated number of data vectors to be encoded exceeds At that time, random sampling was used to uniformly sample data from the accumulated data. One sample is used for dictionary updates. If The number of data vectors to be encoded accumulated within a monitoring period is less than ( Values If the dictionary update is not executed in that round, it will be skipped and will wait for more data to accumulate in the next update cycle before being executed. The value is This value is relative to the atomic dimension. In other words, far greater than This satisfies the basic requirement in dictionary learning that the number of samples should be much greater than the atomic dimension. If The possible values ​​will be expanded in future implementations. It should be adjusted accordingly. That is, each atom corresponds to at least One training sample. Skip to update. After that, even if the sample size is insufficient It also forces a dictionary update, using all existing samples for the update.

[0049] In the dictionary update step, the current sparse coefficient matrix is ​​fixed. Update the dictionary Each column. Let the column to be updated be the... The column atoms are Define the set of sample indices that use this atom. ,in For use of the A set of sample indices for atoms. For sample index, For the first The sample at the th The sparsity coefficient on each atom. If If empty, then the atom Reinitialize to a random unit vector (each component starts from the mean of 1 / 2). Standard deviation is After sampling from the standard normal distribution, (norm normalization), and skip the singular value decomposition update step for that atom. If If the result is not empty, then define the residual matrix: ;in To remove the first The residual matrix after contributions from each atom The matrix consists of all normalized samples to be encoded. The total number of data vectors to be encoded. Each column corresponds to a normalized data vector to be encoded. , , For atomic indexing, , For the first Atom vectors, For the first The coefficient vector of each atom. The updating of column atoms is achieved through the following: In the index set Singular value decomposition is performed on a restricted matrix: ;in for Only retain Index the submatrix corresponding to the column. It is a left singular matrix. It is a singular value matrix. This is a right singular matrix. The updated atoms. Pick The first column, corresponding sparsity coefficients Pick The first column, in which For the updated number Atom vectors, For the updated number A vector of coefficients of atoms. This represents the first diagonal element of the singular value matrix. Due to the left singular matrix of the singular value decomposition... Each column is a unit vector, therefore the updated atoms Automatic satisfaction The normalization properties of each atom are naturally preserved after the dictionary is updated, without the need for additional normalization operations.

[0050] The termination condition for the alternating iteration of the sparse coding step and the dictionary update step is: when the change in the dictionary after two consecutive complete alternating iterations is less than a preset threshold. Stop at time, that is ,in Values , The first iteration before the last alternation update One atomic vector; or when the number of alternating iterations reaches the preset maximum number. Stop at time Values The alternating iteration stops when either of the two termination conditions is met.

[0051] For each data vector to be encoded, the sparse coding process described above yields a corresponding sparse coefficient vector. .because The vast majority of elements are zero; in actual storage and transmission, only the non-zero coefficients and their corresponding atom indices need to be stored. Let... The number of non-zero elements is ,in The number of non-zero elements. express The number of non-zero elements. The amount of data stored and transmitted from... The number of values ​​decreased to A number. This is the sparse coefficient matrix of all samples used in the current iteration step. The sparse coefficient matrix is ​​stored in the local memory of the node device, and the two are numerically equivalent after iterative convergence. The sparse coefficient vector is the sparse coefficient vector of all data vectors to be encoded. Forming a sparse coefficient matrix ,in The total number of data vectors to be encoded. For the first A sparse coefficient vector of a data vector to be encoded. Sparse coefficient matrix. It is stored in the local memory of the node device.

[0052] The node device immediately increments the local dictionary version number after the dictionary update is completed. The new dictionary version is sent via a dictionary synchronization channel; after successfully receiving and verifying the new dictionary, the platform device sets its local dictionary version number to the received version number. The adaptive dictionary learning sparse coding process is as follows: Figure 3 As shown.

[0053] In step S6, an asymmetric autoencoder is introduced as a supplementary compression encoding mechanism. Encoding compression is performed on the node-side device, and decoding and restoration are performed on the platform-side device.

[0054] Asymmetric autoencoders consist of encoder networks and decoder network Composition. Encoder network Deployed at the node level, its network structure is relatively lightweight and computationally complex. Decoder network The network structure of devices deployed on the platform is relatively complex, resulting in high computational complexity. Encoder network It consists of three fully connected layers, with the number of neurons in each layer being as follows: , , Encoder network In the three fully connected layers, the first and second layers use the Modified Linear Unit (ReLU) activation function, which is defined as... The third layer (output layer) does not use an activation function and directly outputs a fixed dimension. A real-valued vector. Decoder network. It consists of three fully connected layers, with the number of neurons in each layer being as follows: , , Decoder network In the three fully connected layers, the first and second layers use the ReLU activation function, while the third layer (output layer) does not use an activation function and directly outputs the reconstructed data vector. Due to the training set Each training sample in the dataset is a sparse coefficient vector. Therefore, the input data dimension Equal to the dimension of the sparse coefficient vector ,Right now Encoder network The number of neurons in the first layer is set to The number of neurons in the output layer is fixed. Decoder network The number of neurons in the first layer is set to The number of neurons in the output layer is . The maximum dimension of the preset latent space is set to a value of [value]. ; .

[0055] Set up encoder network The parameters are decoder network The parameters are For the input data vector The encoder maps it to a fixed dimension. Hidden space: ;in The complete latent space vector output by the encoder. The largest dimension of the hidden space. For the input data dimensions, Indicated by The encoder network with parameters for the input The mapping result.

[0056] Hidden space dimension determined by the node-end device based on the current monitoring cycle ( ),from Before the middle cut Each component serves as the latent space coding vector for actual transmission. : ;in for The One portion, This is the component index. The platform-side device receives the latent space encoded vector. Then, fill it with zeros to the dimension. The zero-padded vector is obtained. Then the decoder network restores it to the reconstructed data: ;in The reconstructed data vector output by the decoder. Indicated by The decoder network with parameters encodes the latent space vector after padding with zeros. The mapping result.

[0057] The asymmetric autoencoder is trained using a large amount of historical monitoring data as the training set. Training set Each training sample The sparse coefficient vector is obtained by collecting data from the node device during past monitoring cycles and then performing sparse coding in step S5. The training set is stored in the local memory of the platform device and the asymmetric autoencoder is trained by the platform device in an offline state. The total number of training samples, For the first There are 10 training samples. The training objective is to minimize the reconstruction error of all training samples. ; During training, the encoder network and decoder network The parameters are jointly optimized using the backpropagation algorithm and the stochastic gradient descent optimizer. In each iteration, a cut length is randomly selected. (exist to (Uniformly randomized values ​​between) will be used to generate the complete latent space vector output by the encoder. The To the Each component is set to zero to obtain the latent space vector used for training. The data is then input into the decoder for reconstruction; the reconstruction error in the training target is based on the latent space vector after random truncation and zero-padding. The computation is performed. The decoder trained in this way can effectively reconstruct latent space vectors padded with zeros of arbitrary truncation lengths. During actual transmission, the node-side device performs the truncation operation, and the platform-side device performs the zero-padding operation, consistent with the training process.

[0058] In the initial stage (cold start stage) before the asymmetric autoencoder has completed training, the node device outputs the sparse coefficient vector from step S5. The data is transmitted directly to the platform device in a compact representation (non-zero coefficients and their corresponding atomic indices) to build the initial training set. After receiving the compact representation, the platform-side device fills the non-zero coefficients into a length of [length missing] according to the atomic index. The complete training samples are obtained from the all-zero vector. Platform-side devices use... Complete the first training of the asymmetric autoencoder and connect the encoder network. Optimization parameters Synchronize to node-side devices. The cold start phase continues. One monitoring cycle, among which Values After the cold start phase, data from subsequent monitoring periods are processed using the trained encoder network. Compressed transmission is then performed. Subsequently, in each monitoring cycle, the node device randomly selects from the sparse coefficient vector output in step S5. Each vector, in a compact representation, is appended to the data and transmitted to the platform device for training set updates. Values to Integers between [a certain range].

[0059] After training, the encoder network Optimization parameters The data is transmitted from the platform device to the node device via a wireless link. After receiving the data, the node device transmits it to the local encoder network. The parameters are updated to The parameter synchronization operation is performed once after each training session. The platform device waits for an acknowledgment signal from the node device after parameter transmission is complete. If this occurs within the preset timeout period... If no confirmation signal is received, the parameters will be retransmitted. Values Seconds. The node device receives the complete encoder network parameters. Next, the received data undergoes integrity verification using a cyclic redundancy check algorithm. If the verification passes, the node device writes the parameters into the local encoder network. In the model file, immediately increment the local encoder version number. Then it returns an acknowledgment signal to the platform device. If the verification fails, an acknowledgment signal will be returned. Request retransmission. After receiving the acknowledgment signal, the platform device parses the status code. If the status code is... Then, the process is completed synchronously, and the platform device will increment the local encoder version number. If the status code is or If this occurs, it will trigger a retransmission of parameters, with a maximum of [number] retransmissions. Second-rate, After all retransmissions fail, a synchronization error log is recorded. If the platform device retransmits parameters due to not receiving an acknowledgment signal after a timeout, the node device recognizes it as the same parameters that have already been processed and only returns an acknowledgment signal again. The local encoder version number will no longer be incremented repeatedly. The parameter synchronization process is consistent with the complete dictionary in step S5. The synchronization processes are independent of each other, each using different transmission channels or different time slots within the same channel.

[0060] training set A sliding window mechanism is used for periodic updates. The platform-side device maintains a capacity of... The training sample queue, in which Values After each monitoring cycle of data transmission is completed by the node device, the uncompressed sparse coefficient vector randomly selected within that cycle is... Training samples are added to the queue in a first-in, first-out (FIFO) manner. When the number of samples in the queue reaches... Afterwards, for each new sample added, the oldest sample in the queue is removed. The asymmetric autoencoder is retrained every [period]. It is triggered once per monitoring cycle, among which Values During retraining, all samples in the current queue are used; after retraining, the encoder network is... Optimization parameters The data is resynchronized to the node devices. This incremental update mechanism enables the asymmetric self-encoder to continuously adapt to changes in the statistical characteristics of the current monitoring data.

[0061] In the actual transmission process, the sparse coefficient vector output in step S5 First, input the encoder network. Encoding and compression are performed to obtain the complete latent space vector. Then, based on the current hidden space dimension Before truncation Each component serves as the latent space coding vector for actual transmission. The node-end device will Transmitted to the platform device. The platform device receives... Then, fill it with zeros. Wei Dedao By decoder network Reconstruct Combined with the overcomplete dictionary synchronized in step S5 Restore the original data A complete dictionary After each dictionary learning update is completed, the node device immediately transmits the data to the platform device via the dictionary synchronization channel. Encoder network parameter synchronization and overcomplete dictionary synchronization share the same wireless transmission channel. When both synchronization operations are pending execution simultaneously, overcomplete dictionary synchronization is performed first, followed by encoder network parameter synchronization after dictionary synchronization is complete. The synchronization operation is initiated by the node device. When the monitoring task is idle, the node device actively queries the platform device for dictionaries or parameters to be synchronized. The platform device returns the corresponding data according to the priority of the synchronization queue.

[0062] Latent space dimension of asymmetric autoencoders Dynamically adjusted based on actual bandwidth conditions. Hidden space dimension. With the dynamic selection of threshold in step S4 The values ​​are related. The larger the value, The smaller the value, the better. Specifically: ; in The maximum dimension of the implicit space is preset. The minimum dimension of the implicit space is preset. Values The latent space encoding vectors corresponding to all data vectors to be encoded. Forming the latent space encoding matrix ,in For the first The latent space encoding vector of a data vector to be encoded. Latent space encoding matrix. It is stored in the local memory of the node device.

[0063] Hidden Space Dimension The dynamic adjustment enables the node-end device to adaptively adjust the compression ratio of the autoencoder according to the current channel state: when the available bandwidth is sufficient, Smaller Larger encoders retain more original information, resulting in higher reconstruction accuracy; however, when available bandwidth is limited, Larger A smaller channel size allows the encoder to achieve a higher compression ratio, but the reconstruction accuracy decreases accordingly. This adjustment mechanism forms an adaptive closed loop of "channel state - compression ratio - reconstruction accuracy" within step S6.

[0064] In step S7, the latent space encoding matrix corresponding to the high-value data selected in step S4 is... And a statistical summary of the low-value data stored in step S4 (including only...) and The average value of data types marked as low value within the corresponding time window and variance ), and carry out unified organization and transmission.

[0065] For high-value data and data that must be transmitted, directly use the latent space encoding vector corresponding to each window output in step S6. As a transmission payload. For low-value data (only) or (Data types marked as low value) transmit their statistical summaries, including the mean of that data type within the corresponding time window. and variance : ; in For the first Within the first time window Variance of various data types The first one defined in step S1 At the [time]th moment The scalar representative value of the monitoring data type. The range of values ​​is The three components of GPS trajectory data ( Subcarrier status data ( Neither the spectral feature mean vector nor the spectral feature mean vector transmits statistical summaries; their information is already contained in the window latent space encoded vector.

[0066] All data to be transmitted is organized into a transmission data vector according to time order and data type order. ,in To transmit data vectors, The total dimension of the data to be transmitted. For the actual constructed transmission data vector The dimension, this value is the same as the one predicted in step S4. They can be different. exist Confirm after construction is complete. The specific meaning varies depending on the load type: when the load type is hour, For transmitting data vectors The number of elements; when the load type is hour, For a compact representation of sparse coefficients, the total number of non-zero coefficient values ​​and atomic indices (i.e., (Total number of non-zero coefficients). In the packet header, Adopt a unified Bit field encoding. During normal transmission (payload type is...) ), The construction method is: indexed by time window. from arrive Arrangement: Within each time window, first arrange the latent space encoding vector corresponding to that window. (This vector, output from step S6, corresponds to the vector to be encoded, consisting of all scalar types within the window, the three GPS components, the mean vector of spectral features, and the subcarrier state vector.) Then, the low-value data types within the window (only...) are arranged... or The mean of those marked as low value and variance Index by data type Arrange from smallest to largest, and for each low-value data type, first calculate the mean and then the variance.

[0067] During the normal transmission phase, the node device first encapsulates and sends the main data packet (payload type). Then encapsulate and send auxiliary training data packets (payload type) The time window number field for the auxiliary training data package takes the value of The length of the list of non-zero coefficients in the window is Bit.

[0068] During the cold start transmission phase or the training sample additional transmission phase (payload type is...) ), Without quantization, the data payload is directly packaged as follows: each sparse coefficient is compactly represented by a vector of sparse coefficients. The non-zero coefficient values ​​and their corresponding atomic indices are composed of; the node-end device assigns each non-zero coefficient value to... Encoded in single-precision IEEE 754 floating-point format, each atom index is... The data packet header is encoded in bit-unsigned integer format, and all compact representations are concatenated into a byte stream in sample order; the minimum and maximum quantization range fields in the data packet header are padded with... The quantization bit field is filled with .

[0069] The adaptive coding module adjusts the signal-to-noise ratio of the current channel. and available bandwidth Dynamically select the number of quantization bits and coding rate Signal-to-noise ratio Defined as signal power With noise power The ratio: ;in For signal-to-noise ratio, For signal power, Noise power. Signal power. The signal strength is obtained by the signal strength detection module of the node device by taking the root mean square value of the signal strength at each frequency point within the monitoring period. ,in For the first At the [time]th moment Signal strength values ​​at each frequency sampling point. Noise power. Measurements are obtained by the node-end equipment during a preset silent period before the start of the monitoring cycle. During the silent period, the node-end equipment does not emit any signals and only receives ambient noise. Take the root mean square value of the signal strength at all frequency points within this time period. The length of the silent period is... , Values The signal power is measured in seconds during a quiet period before the start of each monitoring cycle. Calculated after the monitoring period ends and before data packet encapsulation begins, using all data collected and stored in local memory during the monitoring period. Signal strength data at each time point Calculate the root mean square value. Noise power. Measurements were taken during the quiet period before the start of the monitoring cycle, and the results are stored in local memory for use in the current monitoring cycle. Therefore, the signal-to-noise ratio... The calculation is completed after the monitoring period ends and before the quantization operation is executed, with a quantization bit depth. According to this The value is determined and used for the quantization encoding of data collected in the current monitoring cycle. If the measurement fails during a silent period, the value from the last successful measurement is used. The value is used as an alternative, and a warning message indicating that the noise measurement failed is logged in the local log.

[0070] Quantization bits With signal-to-noise ratio The relationship is: ; in For the number of quantization bits, This is a preset upper limit for the number of quantization bits. This is a preset lower limit for the number of quantization bits. Indicated by Logarithmic operations with base 0. Values , Values . The value is based on the following: In a typical low-altitude wireless signal monitoring scenario, the dynamic range of signal strength is approximately... dB Bit quantization ( (each quantization level) is sufficient to cover the dynamic range and ensure that... Quantization accuracy in dB. The value is determined based on the following: when the signal-to-noise ratio is extremely low ( ), quantization bits reduced to Bit( (Each quantization level) can guarantee basic data identifiability while avoiding the transmission of too many useless bits under poor channel conditions.

[0071] Encoding rate With available bandwidth and total amount of data The relationship is: ; in For encoding rate, ,when When indicates no redundant coding, when This indicates the introduction of forward error correction coding redundancy. This is the current transmission time slot length. If Then define It skips the data packet encapsulation and transmission steps, waiting for the next monitoring cycle before transmitting. When the payload type is When, define It does not introduce forward error correction redundancy. Forward error correction coding uses Reed-Solomon coding, with each symbol having a size of [missing information]. A bit is a byte of data, where each symbol corresponds to one byte. When the payload type is... When the quantized integer sequence is concatenated into a continuous bit stream in sequence, if it is insufficient... End-of-bit bits are padded with zeros to the byte boundary; the platform determines the quantization bit depth during decoding. and number of elements Re-segment the bitstream. Let the total number of bits in the quantized bitstream be... The number of bytes after padding with zeros to the byte boundary is This number of bytes is used as the number of raw data symbols. When the payload type is At this time, the data payload itself is a byte stream, and its number of bytes is the same as the number of original data symbols; no quantization segmentation is performed. The total number of symbols in Reed-Solomon coding is... The number of redundant check symbols is the total number of symbols minus the number of original data symbols. When The number of time redundancy check symbols is zero when When introduced Redundancy. Number of elements required in the data packet header. (Load type) The time represents the total number of values ​​in the compact representation, the number of padding bits, and the number of original data symbols, for use by the platform for decoding.

[0072] Before quantization (load type only) First, count the transmitted data vectors. All elements in the whole Maximum value in each dimension and minimum value ,in For the element index of the transmitted data vector, , for The One element. Set the quantization range to... .for Each element in The quantized integer is obtained by uniform quantization according to the following formula. : ; when At that time, all element quantization values ​​are Quantification range and use Single-precision IEEE 754 floating-point format encoding, used as data payload (payload type) (start) Bits are transmitted along with data packets. Quantization bit depth. use Bit unsigned integer encoding, following closely behind and after.

[0073] Transmit data vector Each element is quantized according to the quantization method described above (load type only). This yields a quantized integer sequence. This integer sequence is then converted into a byte stream using the bitstream packing method described above, and then processed according to the encoding rate. Perform forward error correction coding to generate an encoded data packet containing redundancy check information.

[0074] The final data packets are encapsulated according to a preset frame format. The frame header is fixed as follows. Bits, containing Synchronization word of bits (fixed value) )and The version number of the bit. The data type identifier is extended to... Bit ( The total number of data types defined in step S1, in this invention ,Right now Bits), encoded using network byte order (big-endian), with each data type corresponding to one. Bit sub-identifiers, indexed by data type from arrive The order of the numbers is encoded as follows: This indicates the existence of this type of data. This indicates that it does not exist. The timestamp is... Bit unsigned integer, representing a number of bits. Year moon day hour point The number of milliseconds since the second (Coordinated Universal Time). Number of time windows. use Bit unsigned integer encoding; when the payload type is In this case, the field represents the total number of samples in a compact representation. Latent space dimension use Bit-based unsigned integer encoding. Payload type uses... Bit encoding: This represents the latent space encoding vector. This represents a compact representation of sparse coefficients. The dictionary version number and encoder version number are respectively... Bit unsigned integer encoding, initial version number is Increases with each update Data selection bitmap is only available for load types. It exists at time, and its length is Bits. The list of non-zero coefficients for a window is only available if the load type... It exists at any time, and its length is the total number of samples multiplied by 1. Bits. Number of elements. The number of original data symbols adopted Bit encoding, padding bits are used Bit encoding. The frame check sequence is... Bit Cyclic Redundancy Check (CRC) code, the CRC calculation covers all bytes from the frame header to the data payload, and the generator polynomial adopts the CRC-16-CCITT standard polynomial. The initial value is After encapsulation, the node device sends the data packet to the platform device via a wireless link.

[0075] After receiving the data packet, the platform device sequentially performs frame synchronization, forward error correction decoding, bit stream segmentation, and dequantization (payload type only). decoder network The process includes inverse operations such as reconstruction, sparse coefficient restoration, and denormalization. Upon receiving a data packet, the platform device first checks the dictionary version number and encoder version number. If the local version number is less than the version number in the data packet, the data packet is temporarily stored and decoded only after the corresponding version is synchronized.

[0076] During dequantization, the platform-side device extracts data from the data payload header. , and Then, perform dequantization as follows to convert the quantized integer. Restored to the reconstructed floating-point value : ; in For the dequantized first The reconstructed value of each element. After dequantization, the platform device calculates the value based on the number of time windows in the packet header. Hidden Space Dimension And payload type, for the dequantized transmission data vector Perform the segmentation.

[0077] When the load type is At that time, the platform-side devices are indexed according to the time window. from arrive Extract a length of within each time window Latent space encoding vector Then select the bitmap based on the data. and The number of low-value data types is determined by corresponding bits, and the mean and variance of these low-value data types are extracted sequentially. For each window, a latent space encoding vector is extracted. The platform-side device fills it with zeros. Wei Dedao Input decoder network Obtain the reconstructed sparse coefficient vector Combined with the overcomplete dictionary synchronized in step S5 Reconstruct the normalized data vector to be encoded for this window. Then, using preset fixed normalization parameters, the normalization is... Perform inverse normalization operation, by component implement The actual reconstructed value of the data vector to be encoded is obtained. Finally, based on the order of the vectors to be encoded defined in step S5, from... The process sequentially extracts the aggregated value of high-value types, the aggregated value of the three GPS components, the mean vector of spectral features, and the subcarrier state vector. For low-value data types, the platform-side devices directly use the extracted mean. and variance replace The fixed mean of the first 5 positions .

[0078] When the load type is At that time, the platform-side device extracts data from each sample according to the list of non-zero coefficients in the window, in sample order. Each non-zero coefficient is assigned a number of values ​​(coefficient value first, then atom index). The non-zero coefficients are then filled into a length of [length missing] according to their atom indices. From the all-zero vector, the sparse coefficient vector reconstruction value of the sample is obtained. It is used to build the initial training set or update the training set.

[0079] At this point, the platform-side equipment has fully recovered the main characteristic information of the monitoring data for all time windows within the monitoring period, including the mean vector of spectral features, GPS trajectory aggregation values, subcarrier state vectors, and scalar aggregation values ​​and statistical summaries of high and low value types. The asymmetric autoencoder and data encapsulation and transmission process is as follows: Figure 4 As shown.

[0080] Example 1: In Example 1, the node device of the low-altitude wireless signal monitoring system is a hexacopter UAV equipped with a spectrum analysis module and a GPS receiver module. While conducting low-altitude inspection flights along a railway line, this UAV continuously collects spectrum data, GPS trajectory data, signal strength time-series data, and subcarrier status data. Monitoring cycle. Values That is, drones in Data was collected at that time. Group monitoring data. Total number of data types. Values These correspond to six types of data: spectrum data, GPS longitude, GPS latitude, GPS altitude, signal strength time-series data, and subcarrier status data. Number of time windows. Values Each time window contains One original data point. The highest order of the Chebyshev polynomial. Values That is, the spectrogram data at each time step is compressed to The Chebyshev approximation coefficients, and the window spectral feature mean vector dimension are... Weighting coefficients for importance scores , , The values ​​are respectively , , The values ​​in this group were obtained through a grid search method in the most recent The optimization was performed based on historical data from several monitoring periods, with the objective of maximizing the Pearson correlation coefficient between the reconstructed data from the platform and the original data. (The total number of atoms in the complete dictionary is also mentioned.) Values Total number of subcarriers Values Dimension of the vector to be encoded The maximum dimension of the latent space of an asymmetric autoencoder. Values The smallest dimension of the hidden space Values Perform all operations of Example 1 in the order of steps S1 to S7, ensuring the positioning accuracy of the UAV by the platform device (positioning error less than...). (meters) and signal recognition accuracy (recognition accuracy greater than) Under the premise that the actual amount of data transmitted is approximately Number of values ​​(per time window) Each vector is a latent space encoded vector, with dimensions... The average is approximately ,total (Number of values). The total amount of raw monitoring data includes a complete spectrum data matrix. ( Subcarrier state data matrix ( ) and initial data matrix ( ), totaling approximately Tens of thousands of values. The compression ratio (the ratio of the total amount of original data to the amount of data transmitted) is approximately... far higher .

[0081] Example 2: Based on Example 1, Example 2 further specifies the initialization method for the overcomplete dictionary. The overcomplete dictionary in Example 1... Random initialization is used, meaning each atom of the dictionary... Each component is generated by random sampling from a standard normal distribution. The overcomplete dictionary in Example 2... An initialization method based on clustering of historical data vectors to be encoded is adopted.

[0082] Specifically, before step S5 is executed, the node-side device reads the most recent data from its local memory. The historical data vector to be encoded for each monitoring period, where Values . Values The basis is: determined by autocorrelation analysis of historical monitoring data—when Less than The data diversity included in the time dictionary initialization is insufficient. Greater than The statistical characteristics of outdated data deviate significantly from the current monitoring scenario. To achieve a compromise, each historical data vector to be encoded is normalized to obtain a sample set. K-means clustering is then performed on the sample set, with the number of cluster centers set to [value missing]. If the number of valid samples Then you will get directly There are cluster centers. If the number of valid samples... and Then first K-means clustering was performed to determine the number of cluster centers, resulting in... The first cluster center; then for the second... Cluster centers ( ),generate Each of the perturbation replicas has an average value superimposed on the central replica. Standard deviation is Gaussian random perturbation, a total of 10 atoms; if the total number exceeds 100 atoms Then truncate from the end to If the number of valid samples is zero, then revert to the random initialization method of Example 1. Use the cluster centers or the perturbed cluster centers as the overcomplete dictionary. The initial atoms, i.e. ,in For the first The initial value of each atom, For the first Cluster initialization is performed only during the cold start phase of the overcomplete dictionary, providing an initial value close to the data distribution for dictionary learning. After cluster initialization, the dictionary update process described in step S5 (alternating iterations of sparse coding and dictionary updates) is performed in subsequent monitoring cycles according to... The periodic triggering conditions are executed normally, and the dictionary atoms are continuously optimized to adapt to the latest statistical characteristics of the current monitoring data. The clustering initialization method and the subsequent dictionary update process work together—the former provides a good starting point, and the latter enables continuous adaptive optimization.

[0083] The clustering initialization method ensures that the initial atoms of the overcomplete dictionary are roughly distributed in the main clustering regions of the actual data vector to be encoded, accelerating the convergence of the alternating iterations between sparse coding and dictionary updates during subsequent dictionary learning. Compared with the random initialization method in Example 1, the dictionary learning in Example 2 reduces the number of iterations required to achieve the same reconstruction accuracy by approximately [number missing]. .

[0084] Example 3: Based on Example 1, Example 3 increases the latent space dimension of the asymmetric autoencoder. The dynamic adjustment mechanism was further defined. The hidden space dimension in Example 1... Based on the dynamically selected threshold in step S4 The value is adjusted in real time. Values , Values .

[0085] When available bandwidth When there is ample supply, The value is relatively small. For example, when hour, The latent space has a large dimension, and the encoder network... Retain more original data information. When available bandwidth When nervous, The value is relatively large. For example, when hour, The latent space has a small dimension, and the encoder network... Achieve higher compression ratios. Hidden space dimension. The dynamic adjustment enables the node-end device to adaptively adjust the compression ratio of the autoencoder according to the current channel state: when the available bandwidth is sufficient, Smaller Larger encoders retain more original information, resulting in higher reconstruction accuracy; however, when available bandwidth is limited, Larger A smaller encoder allows for a higher compression ratio, but this results in a corresponding decrease in reconstruction accuracy. This adaptive adjustment mechanism enables the entire compressed transmission system to continuously track changes in channel conditions.

Claims

1. A wireless data compression and transmission method based on data characteristics, characterized in that, Includes the following steps: Step S1: Collect various types of raw monitoring data at multiple times, and construct an initial data matrix based on the raw monitoring data. The raw monitoring data includes spectrum data, signal strength time series data, and subcarrier state data. The subcarrier state data is organized into a subcarrier state data matrix. Step S2: The segmented aggregation approximation method is used to perform time-domain dimensionality reduction on each column of the initial data matrix to obtain the aggregation matrix. The fluctuation amount between adjacent time windows is extracted based on the aggregation matrix to obtain the fluctuation feature matrix. Step S3: The Chebyshev polynomial approximation method is used to perform frequency domain nonlinear feature modeling on the spectrum data to obtain the spectrum feature matrix; Step S4: The aggregation matrix, fluctuation feature matrix and spectrum feature matrix are fused. The importance score of the average signal strength and signal strength time series data in the spectrum graph within each time window in the aggregation matrix is ​​calculated to obtain the comprehensive importance score. Based on the comprehensive importance score and the available bandwidth of the current wireless link, the data to be transmitted is selected, and the remaining data is discarded. Step S5: Adaptive dictionary learning sparse coding method is used to perform sparse representation and compression coding on the aggregated values ​​corresponding to the transmitted data, the GPS trajectory aggregated values ​​of the aggregated matrix, the subcarrier state vector of the subcarrier state data matrix, and the mean vector of the spectrum features of the spectrum feature matrix to obtain a sparse coefficient vector. Step S6: The sparse coefficient vector is compressed and encoded using an asymmetric autoencoder to output a latent space encoded vector, and the latent space encoded vector is decoded and restored. Step S7: Organize the latent space coding vector corresponding to the transmitted data and the statistical summary of the discarded data into a transmitted data vector, and encapsulate the transmitted data vector into a data packet after adaptive encoding.

2. The wireless data compression and transmission method based on data features according to claim 1, characterized in that, In step S2, the time series of each column of data in the initial data matrix is ​​divided into multiple time windows of equal length. The arithmetic mean of all data points in each time window is calculated as the aggregate value of that time window. The aggregate values ​​of all time windows constitute an aggregate matrix. The difference between the aggregate values ​​of adjacent time windows is calculated as the fluctuation amount. All fluctuation amounts constitute a fluctuation feature matrix.

3. The wireless data compression and transmission method based on data features according to claim 1, characterized in that, In step S3, the frequency values ​​of the spectrum data are linearly mapped to a normalized interval. The spectrum data is then approximated using Chebyshev polynomials on the normalized interval. The Chebyshev approximation coefficients are calculated using discrete Chebyshev transform. The Chebyshev approximation coefficients at all times are arranged in chronological order to form a spectrum feature matrix.

4. The wireless data compression and transmission method based on data features according to claim 1, characterized in that, In step S4, the average signal strength of the spectrum and the absolute value of the fluctuation of the signal strength time series data within each time window are calculated as the fluctuation amplitude contribution. The standard deviation of the aggregated value of the signal strength time series data across all time windows is calculated as the statistical significance contribution. The average absolute value of the correlation coefficient between the signal strength time series data and the spectral features within each time window is calculated as the spectral correlation contribution. The fluctuation amplitude contribution, statistical significance contribution, and spectral correlation contribution are normalized and then weighted and summed to obtain a comprehensive importance score.

5. The wireless data compression and transmission method based on data features according to claim 4, characterized in that, In step S4, the method for selecting transmission data based on the comprehensive importance score and the available bandwidth of the current wireless link is as follows: a dynamic selection threshold is set, which is adaptively adjusted according to the ratio of the available bandwidth of the current wireless link to the total amount of data to be transmitted; the average signal strength of the spectrum and the signal strength time series data with a comprehensive importance score greater than or equal to the dynamic selection threshold are used as transmission data, and the average signal strength of the spectrum and the signal strength time series data with a comprehensive importance score less than the dynamic selection threshold are used as discarded data.

6. The wireless data compression and transmission method based on data features according to claim 1, characterized in that, In step S5, the adaptive dictionary learning sparse coding method includes: constructing an overcomplete dictionary containing multiple atoms, each atom having the same dimension as the fixed dimension of the data vector to be encoded; solving the sparse coding objective function through an iterative shrinking threshold algorithm to obtain a sparse coefficient vector; learning the overcomplete dictionary online from historical monitoring data by alternately executing the sparse coding step and the dictionary update step; and updating each column of atoms in the overcomplete dictionary through singular value decomposition in the dictionary update step.

7. The wireless data compression and transmission method based on data features according to claim 6, characterized in that, In step S5, the data vector to be encoded is constructed as follows: for each time window, the aggregated value, GPS trajectory aggregated value, spectral feature mean vector and subcarrier state vector corresponding to the transmitted data are arranged in a fixed order and uniformly filled into the fixed dimension.

8. The wireless data compression and transmission method based on data features according to claim 1, characterized in that, In step S6, the training method for the asymmetric autoencoder is as follows: during the cold start phase, the sparse coefficient vector is directly transmitted to the decoder network in a compact representation to construct the initial training set. The decoder network completes the first training of the asymmetric autoencoder using the initial training set and synchronizes the optimized parameters of the trained encoder network to the encoder network. In subsequent monitoring periods, the encoder network uses the trained encoder network to encode and compress the sparse coefficient vector.

9. The wireless data compression and transmission method based on data features according to claim 1, characterized in that, In step S6, the first few components are extracted from the complete latent space vector output by the encoder network based on the latent space dimension determined by the current monitoring period as the latent space coding vector for actual transmission; the latent space dimension is dynamically adjusted according to the value of the dynamic selection threshold in step S4, and the larger the dynamic selection threshold, the smaller the latent space dimension.

10. The wireless data compression and transmission method based on data features according to claim 1, characterized in that, In step S7, the statistical summary of discarded data includes the mean and variance of discarded data within the corresponding time window; adaptive coding includes dynamically selecting the quantization bit depth based on the current channel signal-to-noise ratio and dynamically selecting the coding rate based on the available bandwidth and the total amount of data in the transmitted data vector, with the coding rate determining the number of redundant check symbols in the forward error correction coding; the data packet is encapsulated according to a preset frame format, which includes a frame header, data type identifier, timestamp, number of time windows, latent space dimension, payload type, dictionary version number, encoder version number, data selection bitmap, number of elements, number of padding bits, number of original data symbols, data payload, and frame check sequence.

Images