An internet of things data compression transmission method based on semantic communication

Abstract

The application discloses a kind of based on semantic communication's thing networking data compression transmission method, it is related to data compression technical field, the method includes: step S1: obtaining original data block, record mutation mark in mutation data block;Step S2: construct semantic encoding model to output semantic feature vector and semantic code stream, reconstruct by decoding to obtain reconstructed data block;Step S3: calculate the residual amount of data block, calculate the quantization step of residual compensation code stream, to the cloud to obtain reconstructed data block;Step S4: using residual compensation amount and reconstructed data block obtains modified reconstruction block, completes data compression correction by splicing.The present application calculates residual compensation code stream to mutation data block, and adaptively adjusts feature dimension and quantization step, reconstructs data block according to time sequence splicing, avoids the distortion caused by traditional frequency domain truncation, has the advantages of guaranteeing thing networking terminal sequence reconstruction accuracy and the beneficial effects of significantly improving the correction accuracy of peak region.

Description

Technical Field

[0001] This invention relates to the field of data compression technology, and more specifically to an IoT data compression and transmission method based on semantic communication. Background Technology

[0002] In Internet of Things (IoT) applications, the physical signals (such as vibration, temperature, current, and pressure) collected by terminal devices are essentially time-series data that change continuously over time. Compared with traditional image or voice data, a significant characteristic of IoT data is the strong non-stationarity and abrupt changes in its statistical properties. For example, the vibration signal of industrial rotating machinery is approximately stationary during normal operation, but transient spikes appear when an impact fault occurs; the current in a smart grid experiences a step jump at the moment of load switching; and the strain response of a bridge structure exhibits local abrupt changes when a vehicle passes over it. These non-stationary components, such as spikes, steps, and edges, often contain crucial diagnostic or early warning information.

[0003] When data transitions from a stable segment to a region containing spikes or steps, the statically set basis functions or prediction coefficients no longer match the new data distribution. Specifically, in DCT compression, the energy of abrupt signal spreads to multiple high-frequency coefficients, while static compression strategies typically truncate these high-frequency coefficients, causing oscillations or ringing artifacts (Gibbs phenomenon) in the reconstructed signal near the abrupt change location. In predictive coding, static prediction coefficients cannot keep up with rapid signal changes, leading to a sharp increase in prediction residuals. The encoder, due to its fixed quantization range, saturates and truncates, resulting in step-like distortion in the reconstructed signal. These reconstruction errors exhibit self-persistence and diffusion characteristics. While data compression is typically processed independently in blocks or frames, predictive coding and other algorithms propagate state along the time axis. A residual coding error caused by a single abrupt change can affect the reconstruction of multiple subsequent samples, causing the error to spread continuously over time until it is partially reset at the next synchronization frame or the boundary of the coding block. In scenarios with continuous abrupt changes, the error may accumulate indefinitely, rendering the reconstructed data completely unusable over a long period. These defects mean that conventional compression methods cannot simultaneously guarantee reconstruction quality and high compression ratio when faced with the inherent non-stationary characteristics of IoT data. They may cause data omissions, false alarms, or even system instability due to artifacts or error propagation, and it is difficult to suppress reconstruction distortions introduced by sudden changes. Summary of the Invention

[0004] This invention provides an IoT data compression and transmission method based on semantic communication, which solves the problem that existing technologies are prone to signal distortion and data errors when compressing and transmitting IoT terminal data and reconstructing and repairing peak areas of IoT signals, making it difficult to guarantee the accuracy of the corrected signal while maintaining a high compression ratio.

[0005] This invention is achieved through the following technical solution: A semantic communication-based IoT data compression and transmission method, the method comprising: Step S1: Obtain the raw time-series data of the target signal of the target IoT terminal, collect the dynamic parameters of the target IoT terminal, divide the raw time-series data into multiple raw data blocks; perform mutation feature detection on the raw data blocks, mark the raw data blocks that meet the mutation feature detection as mutation data blocks, and record the mutation mark and position index of the mutation data blocks. Step S2: Construct a semantic coding model. Input all the original data blocks into the semantic coding model to output semantic feature vectors. Adjust the target dimension and quantization step size of the semantic feature vectors according to dynamic parameters and mutation markers to generate the semantic code stream corresponding to each original data block. Then, perform local decoding and reconstruction on the quantized semantic feature vectors to obtain the reconstructed data blocks. Step S3: Calculate the residual between the mutation data block and the reconstructed data block, and calculate the quantization step size of the residual compensation code stream based on the residual and dynamic parameters; transmit all semantic code streams and residual compensation code streams to the cloud server in the form of data packets, and decode and reconstruct all semantic code streams again on the cloud server to obtain the reconstructed data block used to splice the target signal. Step S4: Decode the residual compensation bitstream to obtain the residual compensation amount. Add the residual compensation amount to the data neighborhood of the reconstructed data block along the position index of the mutation mark to obtain the corrected reconstructed block. Concatenate all the corrected reconstructed blocks and the reconstructed data blocks without added residual compensation amount in time sequence. After processing the overlapping areas of the data with weighted smoothing, the compression correction of the original time sequence data is completed.

[0006] The process involves acquiring raw time-series data and dynamic parameters of the target IoT terminal, dividing the raw time-series data into multiple raw data blocks. Each raw data block undergoes mutation feature detection; raw data blocks meeting mutation feature criteria are labeled as mutation data blocks, and their mutation markers are recorded. A pre-trained semantic coding model is constructed, and all raw data blocks are input into the semantic coding model to extract corresponding semantic feature vectors. Based on the dynamic parameters and mutation markers, a semantic codestream corresponding to each raw data block is generated; simultaneously, the quantized semantic feature vectors are locally decoded and reconstructed to obtain reconstructed data blocks. The residual between the mutation data block and its corresponding reconstructed data block is calculated, and a residual compensation codestream is generated based on the residual and the dynamic parameters. The semantic codestream is decoded and reconstructed again on a cloud server to obtain reconstructed data blocks for splicing the target signal. The received residual compensation codestream is decoded to obtain the residual compensation amount, and according to the position index of the mutation marker, the residual compensation amount is superimposed onto a predetermined neighborhood of the corresponding reconstructed data block to obtain a corrected reconstructed block. All corrected reconstruction blocks and reconstructed data blocks without superimposed residual compensation are concatenated according to the original time sequence, thereby completing the compressed reconstruction and correction of abrupt change regions of the original time sequence data. This solves the problem that existing technologies are prone to signal distortion and data errors during the compressed transmission of IoT terminal data and the reconstruction and repair of peak regions of IoT signals, making it difficult to guarantee the accuracy of the corrected signal while maintaining a high compression ratio.

[0007] Furthermore, a semantic coding model is constructed using a temporal convolutional network, the contents of which include: Input layer: configured to receive raw data blocks, perform adaptive normalization on each channel of the raw data blocks, set an initial convolution kernel for the raw data blocks, pass the raw data blocks to the convolutional backbone network, and set the number of output channels of the input layer to be equal to the number of internal channels of the convolutional residual block; Convolutional backbone network: includes multiple convolutional residual blocks, each residual block includes: dilated causal convolutional layer, activation function layer and dropout layer. The number of stacked convolutional residual blocks and the dilation rate exponent of the dilated causal convolutional layer are selected according to the energy consumption level of the target IoT terminal. Feature aggregation layer: connected to the output of the convolutional backbone network, it uses global average pooling to compress the time dimension into a one-dimensional feature vector; Fully connected projection layer: linearly maps the one-dimensional feature vector to the target semantic dimension and outputs a semantic feature vector.

[0008] Furthermore, an initial convolution stride and a downsampling convolution stride are set for the initial convolution kernel, and a window limit is set for the time window length of the original data block; when the time window length of the original data block does not exceed the window limit, the input layer uses the initial convolution stride; when the time window length of the original data block exceeds the window limit, the input layer uses the downsampling convolution stride.

[0009] Furthermore, the process of selecting the stacking number of convolutional residual blocks and the dilation rate sequence of the dilated causal convolutional layer based on the energy consumption level of the target IoT terminal is set as follows: Establish a mapping function between the energy consumption level of the target IoT terminal and the temporal convolutional network. Based on the configuration factor of the target IoT terminal according to the mapping function, set a configuration threshold for the configuration factor and preset a dilation rate index sequence with values ​​arranged from low to high. Preset the maximum and minimum values ​​of the stacking number of convolutional residual blocks. When the configuration factor is higher than the configuration threshold, the number of convolutional residual blocks stacked is set to the maximum value, and all inflation rate indices in the inflation rate index sequence are used. When the configuration factor is lower than or equal to the configuration threshold, the number of convolutional residual blocks stacked is set to the minimum value, and the inflation rate index in the inflation rate index sequence is the lowest among the top 50% of the values.

[0010] Furthermore, the process of establishing the mapping function includes: Obtain the terminal power value of the target IoT terminal and represent it as β, and preset the minimum operating power value βmin and the maximum reference power value βmax; set the computing resources allocated to the target IoT terminal in the temporal convolutional network and represent them as R, and set the upper limit value Rmax and the lower limit value Rmin of the computing resources. The mapping function is then expressed as: , The configuration factor is denoted as λ, and satisfies: .

[0011] Furthermore, the process of generating the semantic code stream includes: The dynamic parameters include the channel quality score, terminal power value, and task accuracy index of the target IoT terminal; the benchmark dimension value representing the semantic dimension benchmark is extracted from the semantic feature vector; let the semantic dimension of each original data block be the target dimension value, let the mutation marker be m, and m∈{0,1}; When m=1, it indicates that the current original data block is a mutation data block. The target dimension value of the current original data block is set to equal the baseline dimension value, and the quantization step size of the current original data block is calculated using the channel quality score and the task accuracy index. When m=0, it indicates that the current raw data block is a non-mutation data block. The target dimension value of the current raw data block is calculated using the channel quality score, terminal power value and task accuracy index, and the quantization step size is determined using the target dimension value. The semantic feature vector is compressed to the target dimension using a linear projection matrix. Then, each component of the semantic feature vector is numerically discretized according to the quantization step size to obtain a discrete symbol sequence. The discrete symbols in the discrete symbol sequence are entropy encoded to generate a semantic code stream.

[0012] Furthermore, the process of calculating the quantization step size of the current raw data block using the channel quality score and the task precision index includes: Let Qi be the quantization step size of the current raw data block, Qb be the initial value of the quantization step size, Qmax be the reference maximum quantization step size of the raw time series data, γ be the channel quality score of the current raw data block, and η be the task precision index. The formula for calculating the initial quantization step size when m=1 is: , Where η0 represents the reference value of the mission accuracy index, and δ represents the minimum positive number to prevent zeroing; Let the quantization step size of the current original data block when m=1 be expressed as: , Where k1 represents the channel correction coefficient.

[0013] Furthermore, the process of calculating the target dimension value of the current raw data block using the channel quality score, terminal power value, and task accuracy index includes: Let the baseline dimension value be denoted as D0, the target dimension value as Di, the channel quality score of the current raw data block as γ, the task accuracy index as η, and the terminal power value as β. The formula for calculating the target dimension value when m=0 is expressed as: , Where η0 represents the reference value of the mission accuracy index, and δ represents the minimum positive number to prevent zero.

[0014] Furthermore, the process of determining the quantization step size using the target dimension value includes: Let the quantization step size of the current raw data block be denoted as Qi; set the reference minimum quantization step size for the raw time series data as Qmin; set the reference maximum and minimum dimensions for the target dimension value, denoted as Dmax and Dmin respectively; set the step size adjustment exponent μ. The formula for calculating the quantization step size when m=0 is: .

[0015] Furthermore, the calculation process of the residual compensation bitstream includes: Let γ represent the channel quality score of the current raw data block, η represent the task accuracy index, and β represent the terminal power value; set the residual compensation coefficient k2, and let Qre represent the quantization step size of the residual compensation bitstream. The formula for calculating the quantization step size of the residual compensated bitstream is: , Where T represents the maximum absolute value of the residual segment in the current mutation data block, Tre represents the normalized reference value of the residual, and δ represents the minimum positive number to prevent zero.

[0016] Compared with the prior art, the present invention has the following advantages and beneficial effects: 1. By detecting mutation features and indexing locations, non-stationary components are explicitly separated. With the help of dynamic parameters and mutation markers, the feature dimension and quantization step size are adaptively adjusted. This can preserve more semantic information in the mutation region while maintaining a high overall compression ratio, avoiding distortion caused by traditional frequency domain truncation and significantly improving the correction accuracy of the peak region. 2. Calculate the residual amount for the mutated data block and generate an independent residual compensation code stream. The residual compensation amount obtained by decoding is added to the neighborhood of the corresponding reconstructed data block only along the position index. The error is strictly limited to the local area of ​​the mutation and cannot propagate backward, thus avoiding the problem of error persistence and diffusion and ensuring the stability of long-term sequence reconstruction of IoT terminals. 3. A semantic coding model is adopted to replace the fixed frequency domain transformation. At the same time, by accurately superimposing the residual compensation amount to the neighborhood of the mutation location, and by using weighted smoothing to process the overlapping areas of all reconstructed data blocks in time sequence, the Gibbs artifacts caused by mutation reconstruction are effectively suppressed, the discontinuity of block boundaries is eliminated, and the robustness of the IoT system is improved. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0019] Examples, such as Figure 1 As shown, this embodiment is an IoT data compression and transmission method based on semantic communication, which includes: Step S1: Obtain the raw time-series data of the target signal of the target IoT terminal, collect the dynamic parameters of the target IoT terminal, divide the raw time-series data into multiple raw data blocks; perform mutation feature detection on the raw data blocks, mark the raw data blocks that meet the mutation feature detection as mutation data blocks, and record the mutation mark and position index of the mutation data blocks. Step S2: Construct a semantic coding model. Input all the original data blocks into the semantic coding model to output semantic feature vectors. Adjust the target dimension and quantization step size of the semantic feature vectors according to dynamic parameters and mutation markers to generate the semantic code stream corresponding to each original data block. Then, perform local decoding and reconstruction on the quantized semantic feature vectors to obtain the reconstructed data blocks. Step S3: Calculate the residual between the mutation data block and the reconstructed data block, and calculate the quantization step size of the residual compensation code stream based on the residual and dynamic parameters; transmit all semantic code streams and residual compensation code streams to the cloud server in the form of data packets, and decode and reconstruct all semantic code streams again on the cloud server to obtain the reconstructed data block used to splice the target signal. Step S4: Decode the residual compensation bitstream to obtain the residual compensation amount. Add the residual compensation amount to the data neighborhood of the reconstructed data block along the position index of the mutation mark to obtain the corrected reconstructed block. Concatenate all the corrected reconstructed blocks and the reconstructed data blocks without added residual compensation amount in time sequence. After processing the overlapping areas of the data with weighted smoothing, the compression correction of the original time sequence data is completed.

[0020] In step S1, acquiring the raw time-series data of the target IoT terminal signal means directly collecting or receiving the uncompressed raw sampling sequence output from the target IoT terminal, such as vibration sensors in industrial rotating machinery, current monitoring devices in smart grids, strain gauges in bridge structures, etc., through wired or wireless communication interfaces. This sequence represents physical quantity measurements at discrete time points, such as vibration acceleration, temperature, current amplitude, pressure, etc., and has the time-series characteristics of continuous time sampling. Acquisition methods include, but are not limited to: active reporting by the terminal, cloud-based command retrieval, and edge gateway cache reading. This raw time-series data is the input basis for all subsequent compression, transmission, and reconstruction processes. Collecting the dynamic parameters of the target IoT terminal means acquiring auxiliary information characterizing the terminal's time-varying or configuration-related aspects under the current operating environment, which is used for subsequent adaptive adjustment of the compression strategy. In specific implementations, the dynamic parameters may include load parameters, channel quality parameters, and equipment operating condition parameters of the terminal. Cutting the continuous long sequence of raw time-series data into several non-overlapping or partially overlapping sub-series units according to a preset time window length or number of sample points, each unit is called a raw data block. In practice, the partitioning methods can include fixed-length partitioning and event-driven partitioning. Fixed-length partitioning means that every N sampling points constitute a data block, where N is determined based on the signal sampling rate and typical stable time periods. Event-driven partitioning means that the block boundaries are dynamically adjusted based on the preliminary pre-triggering conditions for mutation detection, so that the mutation point is located within the block rather than at the block boundary.

[0021] For each original data block obtained from the data segmentation, a preset mutation discrimination criterion is used to evaluate the local stationarity of the time series data within the data block block-by-block or sliding evaluation to identify whether it contains abrupt components that meet physical or statistical definitions. These abrupt components include, but are not limited to, impact spikes, step jumps, edge inflection points, or transient amplitude transitions. In specific implementations, the mutation discrimination criterion can use conventional discrimination methods, such as setting thresholds for the time-domain amplitude change rate or signal energy value of the data block, identifying data blocks that reach the threshold as containing abrupt components. For original data blocks that are determined to contain abrupt components, they are marked as abrupt data blocks and assigned a machine-readable mutation marker; and the position index corresponding to the abrupt data block in the entire original time series data is recorded simultaneously. The position index includes at least one of the following: data block number, relative time offset within the block, global sampling point sequence number, or timestamp, used for precise addressing in subsequent steps for locating the abrupt region, residual compensation, and weighted smoothing processing.

[0022] In step S2, the semantic encoding model is used to map the original time-series data blocks to a low-dimensional semantic feature space. In specific implementations, the semantic encoding model can be a variational autoencoder, a neural network model, or a state-space model. The variational autoencoder has been successfully applied in time-series semantic compression and is suitable for resource-constrained IoT terminal devices. All the original data blocks obtained in step S1 are input one by one into the semantic encoding model to calculate and extract the semantic feature vector corresponding to each data block. This semantic feature vector is a compact representation of the original data at the semantic level, and its initial dimension is determined by the model structure. The target dimension and quantization step size of the semantic feature vector are dynamically adjusted. When a mutation marker indicates that the current data block is a mutated data block, the target dimension is increased and the quantization step size is decreased to preserve the detailed information of the mutation. When a mutation marker indicates that the data block is stationary, the target dimension is decreased and the quantization step size is increased to achieve higher compression efficiency. The semantic feature vector is truncated or projected according to the adjusted target dimension (e.g., retaining the first few principal components), and the truncated vector is scalar-quantized or vector-quantized according to the adjusted quantization step size to generate the semantic code stream corresponding to the original data block. The semantic bitstream is the core compressed data that is ultimately transmitted to the cloud server. The quantized semantic feature vector is then input into the local decoder of the semantic coding model to perform local decoding and reconstruction, resulting in a reconstructed data block corresponding to the original data block. This reconstructed data block is used in subsequent step S3 to calculate the residual (i.e., the difference between the abruptly generated data block and the reconstructed data block) in order to generate a residual compensation bitstream. Local decoding and reconstruction are performed only on the terminal side, without increasing the burden on the cloud server. Its main function is to ensure that the residual calculation is based on the reconstructed signal consistent with the cloud-side decoding, guaranteeing the accuracy of the compensation.

[0023] In step S3, the cloud server receives data packets from the IoT terminal. Each data packet encapsulates a semantic code stream corresponding to the same original data block and an existing residual compensation code stream. The cloud server first parses the data packets and extracts all semantic code streams; then, it calls a semantic decoding model that matches the semantic encoding model on the terminal side, independently performing decoding and reconstruction operations on each semantic code stream to generate preliminary reconstructed data blocks corresponding to each original data block. The reconstructed data blocks on the cloud server have the same physical meaning and are obtained in the same way as the reconstructed data blocks on the local end in step S2. They are compressed and transmitted in the form of semantic code streams, which meets the low-load channel requirements of IoT terminals. These preliminary reconstructed data blocks have not yet incorporated residual compensation information; their function is to provide the basic outline or main components of the signal as the basic material for subsequent timing splicing. The reconstructed data block obtained at this time does not include residual compensation. Therefore, it is suitable for the basic layer reconstruction of non-mutation regions and mutation regions where residual compensation will be superimposed. This achieves the separation of terminal encoding calculation and cloud decoding reconstruction, allowing the terminal to focus on lightweight semantic feature extraction and local residual calculation, while the cloud is responsible for complete signal reconstruction and subsequent fusion processing. This is suitable for IoT terminals with limited computing power and cloud collaborative architecture with abundant resources.

[0024] In step S4, based on the mutation location index indicated by the mutation marker, a neighborhood window is determined in the reconstructed data block, centered at that location and extending to both sides of the time axis by a preset length or an adaptively calculated length. The residual compensation amount obtained after decoding the residual compensation bitstream is then superimposed point-by-point onto each data sample in the reconstructed data block that falls within the neighborhood window, according to the correspondence of the sampling points. The new data sequence generated after superposition constitutes the corrected reconstructed block. The length of the neighborhood window is adaptively determined based on the mutation type (e.g., spike-type mutation or step-type mutation) and at least one of the dynamic parameters to ensure that the residual compensation only acts on the critical local area affected by the mutation, avoiding additional distortion or bitrate overhead introduced by global superposition. The corrected reconstructed block is a local signal segment formed by superimposing the residual compensation amount on the basic reconstructed data after local or cloud decoding at the location index corresponding to the mutated data block, thus correcting the mutation distortion. The original data block for which no mutation was detected is a signal segment that is only semantically decoded and reconstructed without additional residual compensation. These two types of data blocks are joined together according to their chronological order in the original time series data, i.e., the time index or positional order retained when dividing the original data blocks in step S1, to form a continuous reconstructed time series. To prevent edge jumps introduced by independent block processing or residual compensation, In this embodiment, a certain length of overlap interval is reserved at the splicing boundary of two adjacent reconstruction blocks. This overlap interval is typically taken from several sampling points at the front and / or rear of each reconstruction block, and its length is determined according to dynamic parameters or preset rules, for example, 5% to 10% of the block length or equal to the signal autocorrelation length. The two adjacent reconstruction blocks include those between a modified reconstruction block and a normal reconstruction block, between two modified reconstruction blocks, and between two normal reconstruction blocks. Within this overlap region, the two adjacent blocks provide their respective reconstruction values ​​for the original signal in the same time interval. Within the overlap region, the reconstruction values ​​from the two adjacent reconstruction blocks are weighted and fused. As a feasible implementation example, a linear weighting method can be used in specific applications, which is enumerated as follows: Let the value of the k-th sampling point of the i-th reconstruction block within the overlap interval be xi(k), and the value of the corresponding sampling point of the (i+1)-th reconstruction block be xi+1(k), then the fused reconstruction value... Where w(k) represents a monotonically decreasing weighting function decreasing from 1 to 0. All overlapping and non-overlapping regions after weighted smoothing are concatenated sequentially in time to obtain the final high-fidelity reconstructed signal after compressed transmission of the original time-series data. Distortion in abrupt regions of this reconstructed signal is significantly suppressed, while the overall compression ratio is maintained.

[0025] Furthermore, as a feasible implementation method, a semantic coding model is constructed using a temporal convolutional network, the content of which includes: Input layer: configured to receive raw data blocks, perform adaptive normalization on each channel of the raw data blocks, set an initial convolution kernel for the raw data blocks, pass the raw data blocks to the convolutional backbone network, and set the number of output channels of the input layer to be equal to the number of internal channels of the convolutional residual block; Convolutional backbone network: includes multiple convolutional residual blocks, each residual block includes: dilated causal convolutional layer, activation function layer and dropout layer. The number of stacked convolutional residual blocks and the dilation rate exponent of the dilated causal convolutional layer are selected according to the energy consumption level of the target IoT terminal. Feature aggregation layer: connected to the output of the convolutional backbone network, it uses global average pooling to compress the time dimension into a one-dimensional feature vector; Fully connected projection layer: linearly maps the one-dimensional feature vector to the target semantic dimension and outputs a semantic feature vector.

[0026] The input layer is configured to receive the raw data block obtained in step S1. This raw data block is a fixed-length time series containing one or more physical channels, such as simultaneously acquiring multi-dimensional signals like vibration, current, and temperature. The input layer does not change the time order or sampling interval of the data block, directly feeding it into the subsequent network for feature extraction, ensuring the integrity of the signal's temporal structure. The adaptive normalization refers to: performing a linear transformation on all sample points within the current data block based on the real-time statistics of that channel, so that the normalized data has zero mean and unit variance; or dynamically adjusting based on the historical statistical parameters of the terminal device. After the raw data block is normalized, the input layer sets a set of initial convolutional kernels for it. The initial convolutional kernels are one-dimensional convolutional filters, with a length less than the time length of the data block, sliding along the time axis to perform convolution operations on the input signal. After the initial convolution processing, the number of channels in the output feature map of the input layer is set to be equal to the number of internal channels in each convolutional residual block in the convolutional backbone network. The number of internal channels refers to the output feature dimension (i.e., the number of convolutional kernels) of the dilated causal convolutional layer in the residual block. By forcing the number of channels to be equal, the output features of the input layer can be directly fed into the first convolutional residual block without additional dimensionality transformation.

[0027] The dilated causal convolutional layer employs a causal convolutional structure, ensuring that the output at the current moment depends only on the input at the current and historical moments, without introducing future information, thus meeting the real-time data stream processing requirements of IoT terminals. Simultaneously, an inflation factor (inflation rate) is introduced, allowing the convolutional kernel to sample along the time dimension with exponential strides without increasing the number of parameters, thereby exponentially expanding the receptive field corresponding to each output neuron. The activation function layer, connected after the dilated causal convolutional layer, introduces a nonlinear transformation, improving the network's ability to fit complex mutation patterns. The dropout layer randomly sets the output of some neurons to zero with a certain probability during training to prevent overfitting. The input of the feature aggregation layer is connected to the output of the convolutional backbone network. The feature aggregation layer uses global average pooling to average and compress the entire feature map along the time dimension, summing the time series values ​​of each channel into a scalar, thereby transforming the feature map into a one-dimensional feature vector with a length equal to the number of channels. The fully connected projection layer receives the one-dimensional feature vector output by the feature aggregation layer as input and maps it from its original length to a preset target semantic dimension through a fully connected linear transformation.

[0028] Furthermore, as a feasible implementation, an initial convolution stride and a downsampling convolution stride are set for the initial convolution kernel, and a window limit is set for the time window length of the original data block; when the time window length of the original data block does not exceed the window limit, the input layer uses the initial convolution stride; when the time window length of the original data block exceeds the window limit, the input layer uses the downsampling convolution stride.

[0029] In IoT time-series data compression scenarios, the length of the original data blocks is not fixed. When dividing the data into blocks in step S1, a variable-length block strategy may be adopted based on mutation markers, dynamic parameters, or transmission protocol requirements. If a uniform fixed convolution stride is used for all data blocks, it may lead to: short blocks using large strides, resulting in over-compression of temporal information and loss of key mutation details; and long blocks using small strides, resulting in excessively long feature maps with an excessively long temporal dimension, increasing the computational burden and memory consumption of the subsequent convolutional backbone network, which is detrimental to low-power terminal deployment. Two different convolution strides are set for the initial convolutional kernel of the input layer: an initial convolution stride and a downsampling convolution stride, and the stride used is dynamically selected based on the comparison between the time window length of the original data block and the preset window limit. When the data block length is short and does not exceed the window limit, the initial convolution stride is used, usually set to 1, to retain the complete temporal resolution; when the data block length is long and exceeds the window limit, a downsampling convolution stride is used, usually an integer greater than 1, such as 2 or 4, to downsample the input signal in the time dimension while maintaining computational efficiency. The window limit can be set based on rules of thumb or historical data. Regardless of the stride, the feature map obtained after the initial convolution will be fed into a convolutional backbone network with a fixed structure. To ensure the backbone network's consistent processing of the time dimension, an adaptive interpolation layer can be added after the input layer or before the backbone network to unify the time length of feature maps with different output lengths to a preset range, or the dilation parameter of the backbone network can be dynamically adjusted with the input length.

[0030] Furthermore, as a feasible implementation, the process of selecting the stacking number of convolutional residual blocks and the dilation rate sequence of the dilated causal convolutional layer based on the energy consumption level of the target IoT terminal is set as follows: Establish a mapping function between the energy consumption level of the target IoT terminal and the temporal convolutional network. Based on the configuration factor of the target IoT terminal according to the mapping function, set a configuration threshold for the configuration factor and preset a dilation rate index sequence with values ​​arranged from low to high. Preset the maximum and minimum values ​​of the stacking number of convolutional residual blocks. When the configuration factor is higher than the configuration threshold, the number of convolutional residual blocks stacked is set to the maximum value, and all inflation rate indices in the inflation rate index sequence are used. When the configuration factor is lower than or equal to the configuration threshold, the number of convolutional residual blocks stacked is set to the minimum value, and the inflation rate index in the inflation rate index sequence is the lowest among the top 50% of the values.

[0031] The mapping function refers to the mathematical correspondence rule for calculating the configuration factor based on the energy consumption level of the target IoT terminal, such as low, medium, or high energy consumption levels, or energy constraint indicators such as remaining battery percentage and power supply method. This mapping function can be a linear function, a piecewise function, or a lookup table mapping. The configuration factor is used to quantitatively characterize the terminal's comprehensive configuration capabilities in terms of computing power, storage, and energy. A higher value indicates that the terminal can support the computational needs of larger-scale or more complex temporal convolutional networks. The configuration threshold is a preset numerical limit used to divide the configuration factor into two or more intervals. The specific value of the configuration threshold can be preset according to experimental or application requirements. By comparing the currently calculated configuration factor with this configuration threshold, it is determined whether the terminal belongs to "high configuration capability" or "low configuration capability," thereby determining the subsequent network structure parameter selection strategy. The dilation rate index sequence is a pre-designed array of indices arranged in ascending order (i.e., from low to high). Each index value in this sequence corresponds to the dilation rate index of the dilated causal convolutional layer of a convolutional residual block in the temporal convolutional network. The inflation rate exponent is typically a non-negative integer, and the actual inflation rate is equal to the base (usually 2) raised to the power of this exponent. For example, a preset sequence of [0,1,2,3,4] indicates that the inflation rates of each layer are 1, 2, 4, 8, and 16, respectively. The length of this sequence determines the maximum number of dilated causal convolutional layers that can be configured in the network. The stacking quantity refers to the total number of convolutional residual blocks contained in the network. This invention pre-sets an allowed upper limit (maximum value) and an allowed lower limit (minimum value) for this stacking quantity, constituting an adjustable range for network depth. The maximum value corresponds to the deepest network structure when the terminal has sufficient energy, and the minimum value corresponds to the shallowest network structure when energy is limited. The specific values ​​of the maximum and minimum values ​​are jointly determined based on the design capacity of the temporal convolutional network, the semantic feature complexity of the target task, and the computing power of a typical IoT terminal. The inflation rate exponents that are in the first 50% of the sequence are selected from the sequence, arranged from the lowest values, and the number of selected exponents does not exceed half the total length of the sequence. As examples of odd and even numbers of values, if the preset inflation rate index sequence is 0, 1, 2, 3, 4, 5, 6, and the length of the inflation rate index sequence is 7, then the three lowest values ​​in the first 50% are taken, i.e., rounded down, corresponding to indices of 0, 1, and 2. If the sequence is 0, 1, 2, 3, and the length of the inflation rate index sequence is 4, then the two lowest values ​​in the first 50% are taken, i.e., 0 and 1.

[0032] Furthermore, as a feasible implementation method, the process of establishing the mapping function includes: Obtain the terminal power value of the target IoT terminal and represent it as β, and preset the minimum operating power value βmin and the maximum reference power value βmax; set the computing resources allocated to the target IoT terminal in the temporal convolutional network and represent them as R, and set the upper limit value Rmax and the lower limit value Rmin of the computing resources. The mapping function is then expressed as: , The configuration factor is denoted as λ, and satisfies: .

[0033] The computing resource R represents the available computing power or power budget for the semantic coding model, which in specific applications can correspond to CPU utilization, millions of operations per second, inference time, or power consumption quota. The maximum reference energy value βmax is usually taken as the nominal value when the terminal is fully charged. In the formula, β-βmin represents the increment of the current energy value beyond the minimum working energy value, indicating the terminal's current excess energy. Only when β>βmin is there additional energy available to allocate more computing resources; if β=βmin, the difference is 0. The βmax-βmin part represents the variable range of energy, used to normalize the absolute difference of energy to the [0,1] interval. The ratio represents the normalized power ratio, ranging from 0 to 1. A ratio of 0 indicates that power is at its minimum (β = βmin), while a ratio of 1 indicates that power has reached its maximum reference value (β ≥ βmax). This ratio linearly reflects the sufficiency of power. Rmax - Rmin represents the dynamic adjustment range of allocable computing resources; this difference determines the amplification factor of the power ratio on the final resource R. The factor is added to the lower limit of computing resources Rmin, representing the additional resource amount added based on the power ratio on top of the minimum resource limit. This ensures that even when power is at its minimum, the basic resource of Rmin can be allocated; when power is sufficient, the resource reaches the upper limit Rmax. This mapping function ensures that the more abundant the power, the more computing resources the model can use, thereby automatically adjusting the network size and achieving an adaptive balance between energy consumption and compression performance.

[0034] Furthermore, as a feasible implementation method, the semantic code stream generation process includes: The dynamic parameters include the channel quality score, terminal power value, and task accuracy index of the target IoT terminal; the benchmark dimension value representing the semantic dimension benchmark is extracted from the semantic feature vector; let the semantic dimension of each original data block be the target dimension value, let the mutation marker be m, and m∈{0,1}; When m=1, it indicates that the current original data block is a mutation data block. The target dimension value of the current original data block is set to equal the baseline dimension value, and the quantization step size of the current original data block is calculated using the channel quality score and the task accuracy index. When m=0, it indicates that the current raw data block is a non-mutation data block. The target dimension value of the current raw data block is calculated using the channel quality score, terminal power value and task accuracy index, and the quantization step size is determined using the target dimension value. The semantic feature vector is compressed to the target dimension using a linear projection matrix. Then, each component of the semantic feature vector is numerically discretized according to the quantization step size to obtain a discrete symbol sequence. The discrete symbols in the discrete symbol sequence are entropy encoded to generate a semantic code stream.

[0035] The channel quality score reflects the current wireless transmission channel conditions; a higher value indicates a better channel, allowing for the transmission of more data streams, while a lower value requires more aggressive compression. The terminal power value reflects remaining energy and is used for non-mutational block adjustment to avoid excessive power consumption during low battery periods. The task accuracy index characterizes the sensitivity of downstream tasks to reconstruction accuracy. In specific implementations, the channel quality score can be any common channel parameter of the target IoT terminal, or a weighted sum of several channel parameters; this embodiment does not impose a unique limitation. Channel parameters used in specific applications may include signal-to-noise ratio, signal received power, channel quality indicator, packet loss rate, channel throughput estimate, etc. Similarly, the task accuracy index can be any downstream signal parameter of the target IoT terminal, or a weighted sum of several downstream signal parameters; this embodiment does not impose a unique limitation. Downstream signal parameters used in specific applications may include false alarm rate tolerance, mean square error tolerance, task response delay, residual tolerance, SLA data quality level, etc. When m=1, the mutation block carries critical diagnostic information, and dimensionality reduction should not sacrifice fidelity. Terminal power consumption is not considered because the number of mutation blocks is typically small, and even low-power terminals should prioritize transmission quality. The quantization step size is adjusted based on channel conditions and task accuracy requirements. When m=0, the number of non-mutation blocks is large, allowing for flexible dimensionality reduction based on available resources to save bit rate and transmission power.

[0036] Furthermore, as a feasible implementation, the process of calculating the quantization step size of the current raw data block using the channel quality score and the task accuracy index includes: Let Qi be the quantization step size of the current raw data block, Qb be the initial value of the quantization step size, Qmax be the reference maximum quantization step size of the raw time series data, γ be the channel quality score of the current raw data block, and η be the task precision index. The formula for calculating the initial quantization step size when m=1 is: , Where η0 represents the reference value of the mission accuracy index, and δ represents the minimum positive number to prevent zeroing; Let the quantization step size of the current original data block when m=1 be expressed as: , where k1 represents the channel correction coefficient.

[0037] The initial value of the quantization step Qb is the basis for determining whether and how to reduce the maximum quantization step after comparing the task precision exponent η with the preset reference benchmark η0. When the task precision requirement is lower than the benchmark, no refinement is performed; when it is higher than the benchmark, the step size linearly decreases inversely with η. The reference maximum step size Qmax represents the maximum allowed quantization interval (corresponding to the coarsest quantization), which is preset by the system. The reference benchmark value η0 of the task precision exponent is a preset constant, for example, set to 0.8 or 1.0. In the calculation formula of the initial value of the quantization step, for the part, when the task precision exponent η is higher, the denominator η + δ is larger, and the ratio is smaller; conversely, when η is lower, the ratio is larger. The ratio represents the relative relationship between the expected precision and the actual precision. If the actual precision η has reached or exceeded the reference benchmark η0, then the ratio ≤ 1; if the actual precision is lower than the reference benchmark, then the ratio > 1. If the fractional part is greater than 1, the min value is taken as 1; if the fractional part is less than 1, the min value is taken as the fractional value itself. η ≥ η0 means that when the task requires the reconstruction precision not to be lower than the reference benchmark, a fine quantization strategy inversely proportional to the precision exponent is enabled. The higher the precision, the finer the step size, and more information is retained. η < η0 means that when the precision requirement is lower than the reference threshold, at this time the task precision requirement is not high, and the system believes that there is no need to provide more than basic quantization precision for the current mutation block, and directly uses the maximum step size to save bitstream resources. Because the basic step size is already the largest, even if the subsequent channel quality is better, it will not be finer than Qmax. The channel correction coefficient k1 represents a preset positive real number, which can be set by the empirical rule; usually set to 0 < k1 < 1, and is used to control the correction intensity of the channel quality on the step size. The larger k1 is, the stronger the reduction effect of the channel quality on the step size.

[0038] In the quantization step calculation formula, the part represents the correction factor. When the channel quality is the worst, the quantization step is equal to Qb, that is, no reduction is made to the basic step size. When the channel is very poor, even if fine quantization is adopted, a large number of quantization symbols may not be correctly recovered due to bit errors, so the code rate should not be increased, and a coarser basic step size should be maintained. When the channel quality improves, the step size decreases. The channel condition allows more information to be transmitted, and a finer quantization step can be allocated to the mutation block to retain more semantic details and improve the reconstruction fidelity. To ensure Qi > 0, it must be satisfied that That is, k1 < (1 / γmax). Since γ is usually normalized to [0,1], k1 needs to be set to < 1 in specific implementations. In practical applications, as an example, k1 can take values ​​such as 0.5 and 0.8. Under the premise of ensuring the key information of the mutation block, the quantization granularity is dynamically optimized by utilizing channel quality information: the better the channel, the finer the quantization; the worse the channel, the coarser the quantization.

[0039] Furthermore, as a feasible implementation method, the process of calculating the target dimension value of the current raw data block using the channel quality score, terminal power value, and task accuracy index includes: Let the baseline dimension value be denoted as D0, the target dimension value as Di, the channel quality score of the current raw data block as γ, the task accuracy index as η, and the terminal power value as β. The formula for calculating the target dimension value when m=0 is expressed as: , Where η0 represents the reference value of the mission accuracy index, and δ represents the minimum positive number to prevent zero.

[0040] In the calculation formula As the numerator of the fraction, when γ,β∈[0,1], the numerator ∈[1,2]. The better the channel and the more abundant the power, the larger the numerator, the higher the target dimension, allowing for the transmission of more semantic information. Resources are significantly improved only when both are good; if one of them is 0, the numerator degenerates to 1, and the dimension is determined only by the denominator. The constant 1 ensures that even under the worst resource conditions, the numerator is not 0, and the dimension will not decrease to 0. In the formula... As the denominator, the higher the task precision requirement, the larger η is, the larger the denominator is, and thus the smaller Di is. Bitrate resources are preferentially allocated to mutation blocks, which have a fixed baseline dimension and finer quantization. Non-mutation blocks sacrifice dimension for overall compression efficiency. When the channel is good, power is sufficient, and task precision requirements are low, the numerator is large and the denominator is small, with Di greater than D0; when the channel is poor, power is low, and task precision requirements are high, the numerator is close to 1 and the denominator is large, with Di less than D0. The reference baseline quantity η0 is a preset constant, for example, set to 0.8 or 1.0, used to normalize η. The zero-prevention minimum positive number δ is used to avoid a denominator of zero. As a specific example, the specific value of the zero-prevention minimum positive number δ can be set to 10. -6 .

[0041] The channel quality score γ and the terminal power value β are embedded in the numerator in the form of a product. This design allows the semantic dimension to monotonically increase with available resources. When channel conditions are good and the terminal has sufficient power, the numerator increases, the dimension increases, allowing for the transmission of richer semantic information, thereby improving the reconstruction quality of non-abrupt regions. When the channel deteriorates or power is insufficient, the numerator approaches 1, the dimension automatically shrinks, and the bitstream length is reduced to save transmission power consumption and bandwidth. This design enables the compression strategy to match the physical layer resource status in real time, avoiding resource waste or transmission failure caused by a fixed dimension.

[0042] Furthermore, as a feasible implementation method, the process of determining the quantization step size using the target dimension value includes: Let the quantization step size of the current raw data block be denoted as Qi; set the reference minimum quantization step size for the raw time series data as Qmin; set the reference maximum and minimum dimensions for the target dimension value, denoted as Dmax and Dmin respectively; set the step size adjustment exponent μ. The formula for calculating the quantization step size when m=0 is: .

[0043] When Di = D0, the fractional part The ratio is 0; when Di = Dmin, the ratio is 1; when Di is between the two, the ratio is in the (0,1) interval. The fractional part represents the degree of compression of the semantic dimension relative to the baseline dimension. The larger the ratio, the more severe the dimension compression, that is, Di is closer to Dmin, meaning that the non-mutation block is considered less important and can withstand coarser quantization. By adjusting μ, the sensitivity of the quantization step size to dimension compression can be flexibly controlled to adapt to the fidelity requirements of non-mutation regions in different application scenarios. The step size adjustment exponent μ is used for non-linear adjustment. When μ = 1, it has a linear relationship, and the step size increases proportionally with the dimension compression ratio. When μ > 1, the power function is convex downward. When the ratio is small, The step size is smaller, meaning it increases slowly, maintaining finer quantization even with slight dimensionality compression, until the step size increases rapidly when the ratio approaches 1. When 0 < μ < 1, the power function is convex, and the step size increases rapidly with the ratio. As a special extreme case, when μ = 0... When the partial value equals 1, Qi = Qmax is always the coarsest quantization. When the compression ratio is 0, Qi = Qmin, giving the finest quantization; when the compression ratio is 1, Qi = Qmax, giving the coarsest quantization. The intermediate ratio is linearly mapped according to the exponentially adjusted value, so that the step size changes continuously within the range of [Qmin, Qmax]. Through the exponent μ, the system can adjust the response speed of the step size to dimensionality changes for different business needs. For scenarios that need to retain the details of non-mutation regions, using μ > 1 can maintain finer quantization even when the dimensionality decreases slightly; for scenarios that only focus on mutation information, using μ < 1 can quickly switch to coarse quantization to maximize the compression ratio. When Di = D0, that is, when resources are sufficient and dimensionality is not reduced, the finest quantization is automatically used to ensure that non-mutation regions can also obtain high-quality reconstruction; when Di = Dmin, that is, under extreme compression, the coarsest quantization is used to save bitrate to the extreme.

[0044] Furthermore, as a feasible implementation method, the calculation process of the residual compensation bitstream includes: Let γ represent the channel quality score of the current raw data block, η represent the task accuracy index, and β represent the terminal power value; set the residual compensation coefficient k2, and let Qre represent the quantization step size of the residual compensation bitstream. The formula for calculating the quantization step size of the residual compensated bitstream is: , Where T represents the maximum absolute value of the residual segment in the current mutation data block, Tre represents the normalized reference value of the residual, and δ represents the minimum positive number to prevent zero.

[0045] The residual normalization reference value represents a preset constant used to make T dimensionless, for example, taking 1 / 10 of the expected residual amplitude or the signal dynamic range. The quantization step size of the residual compensation bitstream is used to quantize the reconstructed residual signal of the abrupt change block. The smaller the step size, the finer the residual compensation and the higher the reconstruction accuracy of the abrupt change region; the larger the step size, the coarser the compensation and the lower the bit rate. The residual compensation coefficient is a positive real number used to scale the quantization granularity of the residual compensation as a whole, and can be preset according to the system's dependence on residual correction combined with empirical rules. The channel quality score, task accuracy index, and terminal power value are all normalized values. The residual compensation coefficient k2 is used as a fixed gain to determine the overall tendency of residual compensation. If the system wants to prioritize saving bit rate, a smaller k2 can be selected; if high fidelity is required, a larger k2 can be selected. The higher the task accuracy requirement, the larger the numerator, leading to an increase in Qre. When η, γ, and β are all large, the target IoT terminal has high task requirements, good channel, and sufficient power. In this case, the large product leads to a small Qre, thus requiring fine compensation. If any one of these factors is very small, such as extremely poor channel, insensitive task, or extremely low power, the product approaches 0, the denominator approaches δ, and the step size is extremely large. In this case, fine compensation is practically abandoned. As a feasible specific application of this embodiment, the task accuracy index represents the residual tolerance value, and the channel quality score represents the channel quality indicator. The denominator term... This represents dynamic adaptive processing of residuals. A larger T indicates a larger residual amplitude, requiring a finer quantization step size to avoid saturation distortion. Therefore, increasing the logarithmic term increases the denominator and decreases the step size. The logarithmic function is used to compress the large dynamic range and avoid bitrate explosion caused by excessively small step sizes. The logarithmic formula avoids the problem of bitrate explosion due to excessively small step sizes caused by a single extremely large residual point; simultaneously, it automatically passivates at small residuals, maintaining robustness to noise. When T is very small, close to 0, ln(1+0)=0, and only δ remains in the denominator. At this point, the step size is extremely large, essentially achieving no compensation processing. The zero-prevention minimal positive number δ is used to ensure that the denominator is always positive. The semantic bitstream has undergone basic reconstruction of the mutation blocks. Residual compensation, as an enhancement part, has its step size controlled by the calculation formula in this implementation method, realizing layered compression of the basic layer and the on-demand enhancement layer, effectively suppressing error propagation that easily occurs during data compression and transmission from IoT terminals. The inverse proportional structure organically combines resource conditions, task requirements, and the actual size of the residuals, providing amplitude adaptation for the logarithmic term and implementing bottleneck constraints for the multiplicative term.

[0046] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A semantic communication based IoT data compression transmission method, characterized in that, The method includes: Step S1: Obtain the raw time-series data of the target signal of the target IoT terminal, collect the dynamic parameters of the target IoT terminal, divide the raw time-series data into multiple raw data blocks; perform mutation feature detection on the raw data blocks, mark the raw data blocks that meet the mutation feature detection as mutation data blocks, and record the mutation mark and position index of the mutation data blocks. Step S2: Construct a semantic coding model. Input all the original data blocks into the semantic coding model to output semantic feature vectors. Adjust the target dimension and quantization step size of the semantic feature vectors according to dynamic parameters and mutation markers to generate the semantic code stream corresponding to each original data block. Then, perform local decoding and reconstruction on the quantized semantic feature vectors to obtain the reconstructed data blocks. Step S3: Calculate the residual between the mutation data block and the reconstructed data block, and calculate the quantization step size of the residual compensation code stream based on the residual and dynamic parameters; transmit all semantic code streams and residual compensation code streams to the cloud server in the form of data packets, and decode and reconstruct all semantic code streams again on the cloud server to obtain the reconstructed data block used to splice the target signal. Step S4: Decode the residual compensation bitstream to obtain the residual compensation amount, add the residual compensation amount to the data neighborhood of the reconstructed data block along the position index of the mutation mark to obtain the corrected reconstructed block, and concatenate all the corrected reconstructed blocks with the reconstructed data blocks without added residual compensation amount in time sequence. After processing the overlapping area of ​​the data with weighted smoothing, the compression correction of the original time sequence data is completed. A semantic coding model is constructed using temporal convolutional networks, and its contents include: Input layer: configured to receive raw data blocks, perform adaptive normalization on each channel of the raw data blocks, set an initial convolution kernel for the raw data blocks, pass the raw data blocks to the convolutional backbone network, and set the number of output channels of the input layer to be equal to the number of internal channels of the convolutional residual block; Convolutional backbone network: includes multiple convolutional residual blocks, each residual block includes: dilated causal convolutional layer, activation function layer and dropout layer. The number of stacked convolutional residual blocks and the dilation rate exponent of the dilated causal convolutional layer are selected according to the energy consumption level of the target IoT terminal. Feature aggregation layer: connected to the output of the convolutional backbone network, it uses global average pooling to compress the time dimension into a one-dimensional feature vector; Fully connected projection layer: linearly maps the one-dimensional feature vector to the target semantic dimension and outputs a semantic feature vector.

2. The data compression transmission method based on semantic communication for Internet of Things according to claim 1, characterized in that, An initial convolution stride and a downsampling convolution stride are set for the initial convolution kernel, and a window limit is set for the time window length of the original data block. When the time window length of the original data block does not exceed the window limit, the input layer uses the initial convolution stride; when the time window length of the original data block exceeds the window limit, the input layer uses the downsampling convolution stride. 3.The method of claim 1, wherein, The process of selecting the stacking number of convolutional residual blocks and the dilation rate sequence of the dilated causal convolutional layer based on the energy consumption level of the target IoT terminal is set as follows: Establish a mapping function between the energy consumption level of the target IoT terminal and the temporal convolutional network. Based on the configuration factor of the target IoT terminal according to the mapping function, set a configuration threshold for the configuration factor and preset a dilation rate index sequence with values ​​arranged from low to high. Preset the maximum and minimum values ​​of the stacking number of convolutional residual blocks. When the configuration factor is higher than the configuration threshold, the number of convolutional residual blocks stacked is set to the maximum value, and all inflation rate indices in the inflation rate index sequence are used. When the configuration factor is lower than or equal to the configuration threshold, the number of convolutional residual blocks stacked is set to the minimum value, and the inflation rate index in the inflation rate index sequence is the lowest among the top 50% of the values.

4. The data compression transmission method based on semantic communication for Internet of Things according to claim 3, characterized in that, The process of establishing the mapping function includes: Obtain the terminal power value of the target IoT terminal and represent it as β, and preset the minimum operating power value βmin and the maximum reference power value βmax; set the computing resources allocated to the target IoT terminal in the temporal convolutional network and represent them as R, and set the upper limit value Rmax and the lower limit value Rmin of the computing resources. The mapping function is then expressed as: , The configuration factor is denoted as λ, and satisfies: .

5. The IoT data compression and transmission method based on semantic communication according to claim 1, characterized in that, The process of generating the semantic code stream includes: The dynamic parameters include the channel quality score, terminal power value, and task accuracy index of the target IoT terminal; the benchmark dimension value representing the semantic dimension benchmark is extracted from the semantic feature vector; let the semantic dimension of each original data block be the target dimension value, let the mutation marker be m, and m∈{0,1}; When m=1, it indicates that the current original data block is a mutation data block. The target dimension value of the current original data block is set to equal the baseline dimension value, and the quantization step size of the current original data block is calculated using the channel quality score and the task accuracy index. When m=0, it indicates that the current raw data block is a non-mutation data block. The target dimension value of the current raw data block is calculated using the channel quality score, terminal power value and task accuracy index, and the quantization step size is determined using the target dimension value. The semantic feature vector is compressed to the target dimension using a linear projection matrix. Then, each component of the semantic feature vector is numerically discretized according to the quantization step size to obtain a discrete symbol sequence. The discrete symbols in the discrete symbol sequence are entropy encoded to generate a semantic code stream.

6. The IoT data compression and transmission method based on semantic communication according to claim 5, characterized in that, The process of calculating the quantization step size of the current raw data block using the channel quality score and the task accuracy index includes: Let Qi be the quantization step size of the current raw data block, Qb be the initial value of the quantization step size, Qmax be the reference maximum quantization step size of the raw time series data, γ be the channel quality score of the current raw data block, and η be the task precision index. The formula for calculating the initial quantization step size when m=1 is: , Where η0 represents the reference value of the mission accuracy index, and δ represents the minimum positive number to prevent zeroing; Let the quantization step size of the current original data block when m=1 be expressed as: , Where k1 represents the channel correction coefficient.

7. The IoT data compression and transmission method based on semantic communication according to claim 5, characterized in that, The process of calculating the target dimension value of the current raw data block using channel quality score, terminal power value, and mission accuracy index includes: Let the baseline dimension value be denoted as D0, the target dimension value as Di, the channel quality score of the current raw data block as γ, the task accuracy index as η, and the terminal power value as β. The formula for calculating the target dimension value when m=0 is expressed as: , Where η0 represents the reference value of the mission accuracy index, and δ represents the minimum positive number to prevent zero.

8. The IoT data compression and transmission method based on semantic communication according to claim 7, characterized in that, The process of determining the quantization step size using the target dimension value includes: Let the quantization step size of the current raw data block be denoted as Qi; set the reference minimum quantization step size for the raw time series data as Qmin; set the reference maximum and minimum dimensions for the target dimension value, denoted as Dmax and Dmin respectively; set the step size adjustment exponent μ. The formula for calculating the quantization step size when m=0 is: .

9. The IoT data compression and transmission method based on semantic communication according to claim 1, characterized in that, The calculation process of the residual compensation bitstream includes: Let γ represent the channel quality score of the current raw data block, η represent the task accuracy index, and β represent the terminal power value; set the residual compensation coefficient k2, and let Qre represent the quantization step size of the residual compensation bitstream. The formula for calculating the quantization step size of the residual compensated bitstream is: , Where T represents the maximum absolute value of the residual segment in the current mutation data block, Tre represents the normalized reference value of the residual, and δ represents the minimum positive number to prevent zero.

Images