An artificial intelligence-based edge device seismic data compression method and system

Abstract

The application relates to the technical field of seismic data processing, and discloses an edge device seismic data compression method and system based on artificial intelligence; the method comprises the following steps: collecting original one-dimensional seismic data; adapting a joint autoregressive hierarchical prior model from a two-dimensional image compression field to one-dimensional seismic data by the following modification: constructing a deep one-dimensional residual network as a main part of an analysis network and a synthesis network; replacing two-dimensional convolution operation in the joint autoregressive hierarchical prior model with one-dimensional convolution operation; replacing an up-sampling method in the joint autoregressive hierarchical prior model with one-dimensional sub-pixel convolution operation; designing a one-dimensional checkerboard context network to reconstruct a time sequence causal relationship; and performing end-to-end training on the modified joint autoregressive hierarchical prior model by optimizing a code rate and reconstruction distortion; the deep learning seismic data compression method has obvious advantages compared with traditional wavelet compression technology.

Description

Technical Field

[0001] This invention relates to the field of earthquake data processing technology, specifically to an artificial intelligence-based method and system for compressing earthquake data from edge devices. Background Technology

[0002] Effective compression of seismic data is crucial for fields such as seismic exploration, earthquake monitoring, disaster early warning, and earth science research. With the development of multi-dimensional, multi-volume, high-density, and high-sampling-rate seismic acquisition technologies, the volume of seismic data is experiencing explosive growth. This massive amount of data presents severe challenges to storage, transmission, processing, and analysis, especially in scenarios like edge computing where large amounts of data transmission are required, making the need for efficient compression technologies even more urgent.

[0003] The development of traditional seismic data compression technology has always revolved around how to achieve a balance between acceptable loss of fidelity (or complete losslessness) and the highest possible compression ratio, mainly including two technical paths: lossless compression and lossy compression.

[0004] Lossless compression technology, with its ability to perfectly reconstruct the core characteristics of original data, plays a crucial role in scenarios requiring absolute information fidelity, such as the permanent archiving of original data and certain post-processing tasks highly sensitive to minute changes in signals (e.g., accurate phase picking, reference data for full waveform inversion). These methods aim to reduce volume by eliminating statistical redundancy in the data; common techniques include entropy coding methods and the Lempel-Ziv algorithm. However, the inherent high information entropy and complex non-stationary and nonlinear characteristics of seismic data often render these classic lossless compression algorithms inadequate in practical applications. Their compression ratios typically struggle to exceed the 3:1 limit, a performance clearly insufficient to alleviate the data pressure demands of current massive seismic data reduction efforts, thus prompting researchers to focus more on lossy compression.

[0005] Lossy compression achieves higher compression ratios by discarding information that is insensitive or relatively minor to the human perception system. In seismic data processing, the key is to minimize the amount of data while ensuring that critical waveform information (such as amplitude, phase, and frequency) is not distorted. Transform domain coding is the mainstream strategy in traditional lossy compression. Early methods borrowed from image compression, such as Discrete Cosine Transform (DCT), while simple in principle, suffer from block artifacts when applied to continuous seismic trace data due to their inherent block processing method. Furthermore, their fixed transform basis struggles to effectively capture the complex time-frequency variations of seismic signals.

[0006] To overcome the limitations of Directed Transform (DCT), wavelet transform (WT), with its excellent time-frequency localization properties and multi-resolution analysis capabilities, has emerged and quickly become a hot topic in seismic data compression research. Researchers have developed various coding schemes suitable for seismic data based on wavelet theory, such as Embedded Zero Tree Wavelet (EZW) and Set Splitting into Hierarchical Tree (SPIHT) algorithms. These methods perform better in preserving the abrupt changes and edge information of seismic signals and possess progressive transmission capabilities. Although wavelet transform has made significant progress in seismic data compression, at high compression ratios, distortions such as ringing and smoothing effects can still mask weak but crucial geological anomaly signals. More importantly, any fixed, non-adaptive wavelet basis struggles to adapt well to seismic signals with drastically changing characteristics.

[0007] Traditional lossy compression methods face multiple technical challenges in seismic data processing. There is an inherent contradiction between signal fidelity and compression ratio; high compression ratios often lead to the loss of weak but crucial seismic information, thus affecting the accuracy of subsequent interpretation and inversion. Seismic data itself exhibits non-stationary and nonlinear characteristics, with significant variations in signal-to-noise ratio and complex and diverse frequency components. Traditional methods, which often rely on fixed transform bases or pre-defined models, struggle to adequately adapt to these complex characteristics. While some advanced algorithms, such as dictionary learning and certain wavelet coding methods, demonstrate good theoretical performance, their high computational complexity makes real-time processing difficult on resource-constrained edge devices. These methods typically require complex parameter adjustments for different geological conditions and exploration objectives, lacking necessary adaptive capabilities.

[0008] In recent years, deep learning has made groundbreaking progress in the field of data compression, especially in image and video compression, where its performance has surpassed traditional standards such as HEVC and VVC in certain metrics.

[0009] End-to-end learning compression frameworks based on Variational Autoencoders (VAEs) are currently the mainstream approach. These frameworks typically include an encoder neural network (analysis transform) that maps input data to a low-dimensional latent space representation; a quantization module that discretizes the latent representation; a decoder neural network (synthesis transform) that reconstructs the data from the quantized latent representation; and an entropy coding module that performs lossless compression on the quantized latent representation. The entire system is trained using a joint optimization rate (R) and distortion (D) objective function (L = R + λD).

[0010] The continuous improvement in rate-distortion performance of learning-based compression stems from the synergistic development of multiple technological dimensions. From the perspective of technological evolution, these improvements are mainly reflected in three interrelated research directions: optimization of transform network architecture, refinement of probabilistic modeling, and improvement of encoding and decoding efficiency.

[0011] In transform network design, the focus of existing technologies has gradually shifted from early local feature extraction to global dependency modeling. Theis et al. first introduced convolutional neural networks into end-to-end image compression, validating the potential of deep learning in nonlinear transform learning. However, the local receptive field of convolutional operations limits their ability to capture long-range correlations, a limitation that is particularly pronounced when processing data with complex spatial structures. To address this issue, Zou et al. and Koyuncu et al. proposed attention-based solutions. Their models seek a balance between computational complexity and modeling capability through windowed attention, while Contextformer explores the possibility of global context modeling. Although these methods have made progress in performance, their computational overhead and memory requirements remain important factors to consider in practical applications.

[0012] The development of probabilistic modeling reflects a deepening understanding of the distribution characteristics of latent variables. The super-prior framework proposed by Ballé et al. laid the foundation for learned entropy models, the core idea of ​​which is to learn the prior distribution of latent variables through an auxiliary network. The effectiveness of this approach spurred the exploration of more complex probabilistic models. Cheng et al. introduced Gaussian mixture models to better fit the characteristics of non-Gaussian distributions, and He et al. further extended this to more general mixture distribution forms. However, increased model complexity often comes with increased parameter count and computational cost; how to achieve a balance between modeling accuracy and efficiency remains an open question.

[0013] Context modeling, as an important means of reducing bitrate, has evolved to reflect the continuous optimization of how spatial dependencies are utilized. Starting with the autoregressive method based on PixelCNN by Minnen et al., this field has undergone a significant shift from serial to parallel processing. The checkerboard pattern proposed by He et al. achieved partial parallelization, while Liu et al. and Koyuncu et al. further explored efficient capture of global context. These methods achieve varying degrees of trade-off between compression performance and decoding speed, but the optimal trade-off still depends on the specific application scenario.

[0014] From a practical perspective, variable bit rate technology and model lightweighting are key technologies driving the application of learning-based compression. Choi et al. and Cui et al. addressed the problem of single-model multi-bit rate from the perspectives of scaling factor learning and adaptive quantization, respectively, while Yang et al. further optimized this mechanism. Regarding model efficiency, Leguay et al. focused on lightweight architecture design, He et al. proposed a non-uniform grouping strategy, and Liu et al. and Agustsson et al. explored the possibility of hybrid architectures. These works demonstrate that practical needs are driving learning-based compression towards more refined engineering optimizations.

[0015] While deep learning paradigms have demonstrated superior performance in image and video data compression, their specialized application in seismic data processing is still in its early stages, and a universally accepted optimal technology paradigm has not yet been established. This is mainly due to a series of key factors: the scarcity of high-quality, large-scale seismic datasets limits model training effectiveness; the computational demands of complex deep architectures conflict with the hardware limitations of field monitoring equipment; and the lack of model interpretability affects the understanding of compression mechanisms. Therefore, the most critical challenge currently lies in achieving an effective balance between high compression ratios and the preservation of key seismological information (amplitude, phase, initial motion polarity, and spectral characteristics). Furthermore, with the rapid expansion of seismic monitoring networks and the continuous improvement of data sampling accuracy, traditional centralized data processing models are facing severe challenges. The introduction of edge computing technology provides an effective solution to this dilemma. By implementing compression processing directly at the data source, such as field seismometers and downhole sensors, it can fundamentally alleviate communication bandwidth pressure, reduce the storage burden on data centers, and significantly improve the response speed of earthquake early warning systems. This distributed processing architecture exhibits unique advantages in remote areas or emergency scenarios with limited communication, while also enhancing the disaster recovery capabilities and operational stability of the entire monitoring network under extreme conditions.

[0016] However, edge devices typically have limited computing power, small storage space, and strictly limited power consumption. This places extremely high demands on the efficiency, model size, and computational complexity of compression algorithms. Traditional computationally intensive compression algorithms or large deep learning models are difficult to deploy directly at the edge. Therefore, developing lightweight, efficient, low-power compression algorithms that maintain high compression performance and signal fidelity is a key technological bottleneck in edge seismic data processing.

[0017] In summary, existing seismic data compression methods, whether traditional or nascent deep learning approaches, all exhibit limitations to varying degrees when faced with ever-increasing data volumes, complex seismic data characteristics, and the demanding requirements of edge computing devices. There is an urgent need for a novel compression method that can balance high compression ratios, high signal fidelity, low computational complexity, and adapt to the diversity of seismic data and the limitations of edge devices. Summary of the Invention

[0018] To address the aforementioned technical problems, this invention provides a method and system for compressing seismic data from edge devices based on artificial intelligence.

[0019] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0020] In a first aspect, the present invention provides a method for compressing seismic data from edge devices based on artificial intelligence, comprising:

[0021] Acquire raw one-dimensional seismic data;

[0022] The joint autoregressive hierarchical prior model is adapted from the domain of two-dimensional image compression to one-dimensional seismic data through the following modifications:

[0023] (a) The modified joint autoregressive hierarchical prior model includes an analysis network, a super-prior network, a one-dimensional checkerboard context network, and a synthesis network; a deep one-dimensional residual network is constructed as the backbone of the analysis network and the synthesis network; the aim is to overcome the shortcomings of traditional shallow networks in capturing the long-range dependence of seismic signals and to solve the gradient vanishing problem in deep model training.

[0024] (b) Replace the two-dimensional convolution operation in the joint autoregressive hierarchical prior model with a one-dimensional convolution operation;

[0025] (c) Replace the upsampling method in the joint autoregressive hierarchical prior model with a one-dimensional subpixel convolution operation;

[0026] (d) The designed one-dimensional chessboard context network can reconstruct temporal causal relationships;

[0027] The modified joint autoregressive hierarchical prior model was trained end-to-end by optimizing the bitrate and reconstruction distortion, with the loss function... ,in, The bitrate for encoding the modified joint autoregressive hierarchical prior model. The mean square error between the reconstructed seismic data obtained after compression and restoration and the original seismic data;

[0028] Compression processing is performed in real time on an edge computing device to generate a compressed seismic data bitstream and transmit it over a wireless network.

[0029] In one embodiment, the analysis network employs a deep network consisting of multiple cascaded one-dimensional residual modules, which performs downsampling and feature extraction through one-dimensional convolution, and combines a generalized divisible normalization layer with the LeakyReLU activation function; the super-prior network employs an autoencoder architecture; the one-dimensional checkerboard context network employs a parallel autoregressive model consisting of one-dimensional causal mask convolutional layers; and the synthesis network employs a deep one-dimensional residual network symmetrical to the structure of the analysis network, which performs upsampling operations through one-dimensional sub-pixel convolution to reconstruct the signal.

[0030] In one embodiment, the modified joint autoregressive hierarchical prior model's operation includes compression and decompression phases;

[0031] During the compression stage, the analysis network transforms the input data into a first latent representation, and the super-prior network generates a second latent representation based on the first latent representation. Based on the quantized version of the second latent representation and combined with the autoregressive information provided by the one-dimensional checkerboard context network, the precise conditional probability model of the first latent representation is jointly determined, and the quantized versions of the two latent representations are entropy encoded accordingly to generate a compressed bitstream.

[0032] In the decompression stage, the quantized version of the second latent representation is first decoded from the bitstream to reconstruct the initial probability model; then, a one-dimensional checkerboard context network is used to sequentially decode the quantized version of the first latent representation from the bitstream with the aid of autoregression; the synthesis network reconstructs the output data based on the decoded first latent representation.

[0033] In a second aspect, the present invention provides a compression system for implementing the method of any embodiment of the first aspect, comprising:

[0034] The data receiving module is responsible for reading the one-dimensional seismic data output by the seismometer in real time and processing it in segments according to a preset time window;

[0035] The preprocessing module performs data format conversion, normalization, and tensor reshaping of one-dimensional seismic data.

[0036] The compression inference module calls the optimized joint autoregressive hierarchical prior model to perform real-time compression of one-dimensional seismic data.

[0037] The data encapsulation module packages the compressed seismic data bitstream with metadata and transmits it to the backend server via the TCP / IP protocol.

[0038] Compared with the prior art, the beneficial technical effects of the present invention are:

[0039] The compression performance evaluation results based on rate-distortion analysis show that the deep learning seismic data compression method proposed in this invention exhibits significant advantages over traditional wavelet compression techniques. Tests on samples extracted from the STEAD dataset yielded the following results: Figures 13 to 15 Typical experimental examples show that the normalized mean square error (NMSE) of deep learning methods remains at a lower level within a compression ratio range of 10x to 90x. Quantitative analysis indicates that approximately 10x improvement in reconstruction accuracy (NMSE from 10x to 20x) is achieved in the low compression ratio region (10-20x). -2 The magnitude dropped to 10 -4 Up to 10 -3The compression ratio is on the order of magnitude (and maintains good performance stability as the compression ratio increases, even at a compression ratio of 90 times, it can still maintain acceptable reconstruction quality). This performance advantage directly translates into multiple positive effects in practical applications: it can accommodate more and longer-term observation data under the same storage space or transmission bandwidth constraints; it can significantly reduce storage and transmission costs under the same data quality requirements; and the high-fidelity reconstruction signal ensures the reliability and accuracy of subsequent seismic data processing, analysis, and inversion results, avoiding the impact of compression-introduced errors on the accuracy of geological understanding. Attached Figure Description

[0040] Figure 1 The diagram shows the overall network model of the existing joint autoregressive hierarchical prior model.

[0041] Figure 2 This is a schematic diagram of the overall process of the present invention.

[0042] Figure 3 This is a schematic diagram illustrating how the two-dimensional convolutional neural network is modified into a one-dimensional one-dimensional one according to the present invention.

[0043] Figure 4 This is a network model diagram of the analysis network and synthesis network of the combined autoregressive hierarchical prior model.

[0044] Figure 5 This is a network model diagram of the Hyperprior Analysis network and the Hyperprior Synthesis network, which combine autoregressive hierarchical prior models.

[0045] Figure 6 This is a network model diagram of the context network for the joint autoregressive hierarchical prior model.

[0046] Figure 7 This is a network model diagram of the improved Analysis Network and Synthesis Network of this invention.

[0047] Figure 8 This is a network model diagram of the improved Hyperprior Analysis and Hyperprior Synthesis networks of this invention.

[0048] Figure 9 This is a network model diagram of the one-dimensional chessboard-shaped Context Network proposed in this invention.

[0049] Figure 10 This invention presents a network model diagram of a deep residual network module with a stride (ResidualBlockWithStride). It proposes a residual mechanism, a LeaklyRelu activation function, and a generalized divisible normalization layer (GDN) working together to improve feature extraction capabilities.

[0050] Figure 11 This invention presents a network model diagram of a deep residual network module (ResidualBlock) without a stride, proposing a residual mechanism, LeaklyRelu activation function, and a generalized divisible normalization layer (GDN) to work together to improve feature extraction capabilities.

[0051] Figure 12 This is a schematic diagram of the compression system of the present invention.

[0052] Figure 13 This is a schematic diagram of the test results for Sample 1 of the STEAD earthquake dataset.

[0053] Figure 14 This is a schematic diagram of the test results for Sample 2 of the STEAD earthquake dataset.

[0054] Figure 15 This is a schematic diagram of the test results for Sample 3 of the STEAD earthquake dataset.

[0055] Figure 16 This is a flowchart of the training and optimization iteration process for the data compression method of the present invention.

[0056] Figure 17 This is a rate-distortion curve of the method of the present invention.

[0057] Figure 18 The image shows the effect of compressing sample 1 of the natural earthquake dataset using the compression method proposed in this invention.

[0058] Figure 19 The image shows the effect of compressing sample 1 of the natural earthquake dataset using the three-segment wavelet compression method.

[0059] Figure 20 The image shows the effect of compressing sample 2 of the natural earthquake dataset using the compression method proposed in this invention.

[0060] Figure 21 The image shows the effect of compressing sample 2 of the natural earthquake dataset using the three-segment wavelet compression method.

[0061] Figure 22 The image shows the effect of compressing sample 3 of the natural earthquake dataset using the compression method proposed in this invention.

[0062] Figure 23 The image shows the effect of compressing sample 3 of the natural earthquake dataset using the three-segment wavelet compression method. Detailed Implementation

[0063] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0064] Seismic data has several distinctive characteristics: First, it is massive in volume. For example, a medium-sized seismic monitoring array deploying 100 three-component geophones at a sampling rate of 1000Hz and 32-bit precision will generate approximately 100GB of raw data per day. Second, it has a complex data structure, containing both temporal and spatial information and exhibiting multi-scale characteristics. Third, signal and noise coupling is severe, with key information (such as seismic phase arrival time, amplitude changes, and spectral characteristics) often hidden in background noise. Fourth, in field environments, data acquisition equipment faces challenges such as limited power supply and communication bandwidth, especially in emergency scenarios such as disaster early warning, where real-time requirements are extremely high.

[0065] In modern earthquake monitoring systems, real-time data transmission is of paramount importance. The suddenness and destructiveness of earthquakes necessitate rapid response capabilities from monitoring systems, especially in critical scenarios such as earthquake early warning, aftershock monitoring, and post-earthquake emergency rescue. Even a few seconds of data delay can have serious consequences. However, many earthquakes occur in remote mountainous areas and uninhabited regions lacking 4G or 5G signal coverage, relying solely on BeiDou satellite communication for data transmission. This transmission rate is only 228 bytes per minute, equivalent to a maximum of approximately 328KB of data transmitted per day. Given that a single earthquake monitoring instrument generates 1GB of raw data daily, there is a massive gap of up to 3000 times between transmission demand and available bandwidth. Under such stringent bandwidth constraints, real-time transmission is impossible without extremely efficient data compression. Therefore, developing earthquake data compression technologies suitable for extreme bandwidth environments has become a core technological bottleneck in the construction of earthquake monitoring systems. Traditional seismic data compression methods (such as wavelet transform, Huffman coding, and the SPIHT algorithm) face multiple challenges: limited compression efficiency, typically making it difficult to simultaneously achieve high compression ratios (>10:1) and high signal fidelity; high computational resource requirements, making it difficult to process high-sampling-rate data streams in real time; insufficient specialization, failing to optimize for the unique time-frequency characteristics of seismic data; and fixed compression strategies, making it difficult to dynamically adjust the compression ratio based on data importance. These limitations severely restrict the long-term deployment capability, real-time response capability, and large-scale array observation capability of seismic monitoring systems.

[0066] Artificial intelligence-based compression algorithms offer revolutionary advantages over traditional methods. First, through deep learning, the algorithm can automatically identify and preserve key patterns and features in seismic signals, providing higher reconstruction quality at the same compression rate while maintaining the accuracy of key phase identification. Second, end-to-end trained neural network models can capture the high-dimensional nonlinear characteristics and spatiotemporal correlations of seismic data, extracting the optimal representation, surpassing the expressive power of traditional transform domain methods. Third, attention-based intelligent compression strategies can adaptively allocate different bit resources to different signal components, prioritizing the preservation of key information such as phases and spectral peaks, thus optimizing compression efficiency. Fourth, neural networks can learn data distribution characteristics, solving the joint probability distribution problem in lossless compression and approaching the lower bound of information entropy. Fifth, through model quantization, knowledge distillation, and inference acceleration techniques, complex models can be compressed to a computationally acceptable scale for edge devices, achieving low-power, real-time processing.

[0067] Deploying AI compression algorithms to edge devices has unique value: it reduces the burden on data sources, fundamentally solving the bottlenecks in big data transmission and storage; it reduces dependence on communication networks, enhancing the robustness of the monitoring system in harsh environments and under communication interruption conditions; and it enables "intelligent sensing" and "intelligent recording," allowing the system to dynamically adjust sampling rates and compression strategies based on the importance of earthquake events, thus optimizing resource allocation.

[0068] In practical applications, the technical solution of this invention demonstrates significant advantages in low-bandwidth transmission environments. In disaster early warning in remote areas, this technology enables the early warning system to respond quickly under narrowband network conditions through efficient data compression, gaining valuable early warning time for areas with poor communication conditions. In marine resource exploration, it overcomes the bandwidth cost constraints of satellite communication, enabling the real-time transmission of large-scale seismic exploration data and significantly reducing operational costs. In field scientific research, it makes the deployment of long-term, high-density observation networks in network-limited environments such as deserts and polar regions a reality, providing high-quality real-time observation data transmission capabilities for fields such as seismology and geophysics.

[0069] Overall, this invention innovatively solves key technical bottlenecks in the seismic data acquisition process by combining cutting-edge artificial intelligence technology and edge computing architecture, providing a new technical path for the efficient transmission of seismic data, promoting technological progress in fields such as earthquake monitoring, disaster early warning, and resource exploration, and has significant scientific value and broad application prospects.

[0070] Traditional seismic data compression methods suffer from significant technical bottlenecks. Classical algorithms such as wavelet transform and DCT are limited by fixed basis functions, resulting in compression ratios of only 3:1 to 5:1, which is insufficient to meet the needs of large-scale monitoring. Existing deep learning compression technologies are mainly geared towards image processing, lacking specific designs for one-dimensional seismic time-series data, and are computationally complex, making them unsuitable for edge devices. This invention redesigns the network architecture, transforming the hierarchical prior mechanism originally designed for two-dimensional image spatial correlation into prior modeling suitable for one-dimensional time-series data, significantly improving compression efficiency. Simultaneously, a lightweight model design enables real-time operation on edge platforms such as Jetson Orin Nano, completing seismic data compression locally. The compressed data is then transmitted via 4G, WiFi, or BeiDou short message service, fundamentally alleviating data transmission and storage pressures.

[0071] This invention, based on a joint autoregressive and hierarchical prior model, innovatively adapts it from the field of image compression to one-dimensional seismic data compression, proposing a deep learning-based seismic data compression method and system specifically designed for edge devices. This system achieves efficient end-to-end compression by jointly optimizing bitrate and distortion, aiming for high compression ratios, high reconstruction quality, and efficient operation on edge devices.

[0072] 1. Original Joint Autoregressive Hierarchical Prior Model Structure

[0073] The joint autoregressive and hierarchical priors model is a probabilistic generative model architecture that integrates autoregressive modeling mechanisms with hierarchical prior distributions (see network model diagram). Figure 1 This model decomposes complex high-dimensional data into conditionally dependent sequential representations and uses an autoregressive mechanism to capture the temporal dependencies and structured patterns within the data. Simultaneously, the model employs a hierarchical prior distribution to constrain latent variables at different levels, thereby effectively mitigating overfitting and improving generalization performance while maintaining the model's expressive power.

[0074] Figure 1The system architecture of this technical solution is demonstrated, forming a complete encoder-decoder. During encoding, the input image is first transformed into a compact latent representation space by an Analysis Network; this latent representation is then further abstracted and compressed by a Hyperprior Analysis Network. To achieve efficient entropy coding, a Hyperprior Synthesis Network and a Context Network jointly construct an accurate and content-adaptive probability distribution for the latent representation. The Entropy Bottleneck module, based on this probability model, quantizes the latent representation and performs lossless entropy coding on the quantized representation. In the decoding path, each module performs its corresponding inverse operation, and the Synthesis Network completes the final image reconstruction.

[0075] The structure and function of each module are described in detail below:

[0076] like Figure 4 As shown, the core transformation unit consists of an analysis network and a synthesis network. Figure 4 The left side is the analysis network, which is responsible for receiving image data and performing a nonlinear downsampling transformation on the input data y through a deep structure consisting of multiple 2D convolutional layers and a generalized normalization layer (GDN), thereby generating the first-level latent representation Latent y. Figure 4 The right side is the synthesis network, which adopts a mirror structure. It performs an inverse upsampling transformation on the latent representations ẑ and ŷ generated by the advanced prior analysis network and the entropy bottleneck module through a series of two-dimensional transposed convolutional layers and inverse generalized divisive normalization layers (IGDN) to reconstruct a high-quality image.

[0077] To achieve efficient entropy coding, the statistical distribution of the latent representation y must be accurately modeled. This technical solution employs a hybrid probability model that combines hierarchical and autoregressive approaches.

[0078] The model first utilizes a hierarchical prior structure to capture the global relevance of the latent representations. (Refer to...) Figure 5 , Figure 5The left side shows the Hyperprior Analysis network, which receives Latent y and further represents it as a hyperprior latent representation Latent z. This network consists of a series of straddle convolutions with LeakyReLU linear units. Figure 5 On the right is the Hyperprior Synthesis network, which takes the quantized version ẑ of Latent z as input and decodes it through a mirror transposed convolutional structure to generate the first set of parameters for parameterizing the global probability distribution of y.

[0079] Building upon this, to further utilize the local spatial redundancy present in the latent representation, the system also includes a context network, constructed as follows: Figure 6 As shown, this network introduces an efficient spatial autoregressive model, namely the checkerboard context model. This model divides the spatial locations in y into two disjoint sets, like the black and white squares on a checkerboard. This design requires only two masked convolution operations to complete the conditional probability modeling of the entire feature map: the first step uses a class A mask to predict the probability of elements in the first set (yellow checkerboard); the second step uses a class B mask to predict the probability of elements in the second set (yellow checkerboard) given the elements in the first set. Compared to the traditional point-by-point raster scan autoregressive model, this two-step prediction mechanism has stronger parallel processing capabilities, significantly reduces the computational overhead required for context modeling, and achieves a good balance between performance and efficiency.

[0080] Finally, the two sets of probability distribution parameters generated by the super-prior synthesis network and the context network are fed together into the entropy bottleneck module, which is implemented by the quantizer and entropy encoder. This module is responsible for performing quantization and entropy coding. It utilizes this highly accurate hybrid probability model to losslessly compress the discretized latent representations ŷ and ẑ into the final binary bitstream, thereby achieving an extremely high compression ratio.

[0081] At the technical implementation level, this model structure constructs a progressive modeling framework from global to local by introducing a multi-level hierarchy of latent variables. The upper-level prior distribution controls the modeling of global semantic information, while the lower-level conditional distribution is responsible for the generation of fine-grained features. This joint modeling strategy not only effectively handles long-range dependencies in the data but also enhances the model's learning ability for complex data distributions through hierarchical inductive bias, demonstrating excellent performance in tasks such as sequence generation, image modeling, and multimodal learning.

[0082] The network design of this invention draws on this basic architecture and makes key improvements for application scenarios on one-dimensional seismic data.

[0083] 2. Adaptation of the Joint Autoregressive Hierarchical Prior Model Structure to Seismic Data

[0084] Seismic data compression technology has long faced a fundamental contradiction between compression performance and computational efficiency. While traditional linear transformation methods such as wavelet transform and DCT are computationally simple, their fixed mathematical basis functions struggle to adaptively learn the complex non-stationary characteristics of seismic data, resulting in compression ratios consistently hovering at a low level of 3:1 to 5:1. This limitation stems from the highly time-varying and sudden nature of seismic signals, which fixed basis functions cannot effectively capture.

[0085] With the development of deep learning technology, learning-based compression methods have made breakthrough progress in the image domain, among which the joint autoregressive hierarchical prior model represents the current state-of-the-art technology. However, these techniques are entirely based on the spatial correlation design of two-dimensional images, and their network architecture, feature extraction mechanisms, and context modeling methods cannot be directly applied to one-dimensional seismic time-series data. This technological difference between domains has become a key bottleneck restricting the improvement of seismic data compression performance.

[0086] The core innovation of this invention lies in identifying the universality of image compression technology in its compression principles and overcoming existing bottlenecks through systematic cross-domain technology transfer. This invention deeply analyzes the mathematically fundamental similarity between the spatial correlation of image data and the temporal correlation of seismic data, discovering that both can learn their inherent statistical patterns through deep neural networks. Based on this insight, this invention proposes an end-to-end, data-driven closed-loop system development and optimization for machine learning, see [link to relevant documentation]. Figure 2 .

[0087] At the network architecture level, this invention employs, as follows: Figure 1 The overall system framework is shown, but all core computational modules have been fundamentally redesigned. Unlike existing technologies that simply process two-dimensional spatial data, this invention dynamically replaces all two-dimensional convolutional operations in the original model with one-dimensional convolutional operations, see [link to relevant documentation]. Figure 3 This replacement is not a simple reduction in dimensionality, but rather a redefinition of the modeling method for temporal correlation by introducing deep residual learning, improving the upsampling mechanism, and reconstructing the autoregressive model, enabling it to effectively learn the physical characteristics and causal relationships of seismic wave propagation.

[0088] A fundamental improvement of this invention lies in abandoning the original shallow two-dimensional convolutional network design and instead employing a deep residual architecture specifically designed for one-dimensional time-series data to construct the analysis network and synthesis network. Existing shallow structures are insufficient to capture the complex non-stationary characteristics and long-range dependencies prevalent in seismic signals—such as the correlations between data points geographically distant on the time axis caused by reflections from deep geological structures. The deep residual network used in this invention (its structure is described in [reference needed]) Figure 7 , 10 11) Through its skip connection mechanism, it allows for the construction of unprecedented network depth, thereby establishing a hierarchical feature extraction process that can capture short-term seismic wavelet details while effectively modeling long-period trends and dependencies. This structure ensures that even in extremely deep networks, accurate phase and temporal information from the lower layers can be losslessly propagated to higher layers, effectively solving the gradient vanishing problem in deep network training. Furthermore, the unique combination of LeakyReLU and GDN activation functions within the residual module is a specific optimization for the statistical characteristics of seismic data. GDN effectively normalizes neuron responses, reducing statistical redundancy of features, thus generating a more compact and probabilistically modelable latent representation.

[0089] Figure 7 The left side shows the improved analysis network of this invention. Figure 7 The left side shows the improved synthetic network of this invention. Figure 8 The left side shows the improved prior analysis network of this invention. Figure 8 The right side shows the improved advanced prior synthesis network of this invention.

[0090] In the upsampling stage of signal reconstruction, this invention also performs a key optimization. In existing technologies, the inherent checkerboard artifacts of two-dimensional transposed convolution during upsampling cause unacceptable distortion to seismic waveforms requiring high fidelity. This invention completely replaces this technique with one-dimensional subpixel convolution in both the Synthesis Network and Hyperprior Synthesis Network (see [link to Synthesis Network]). Figure 7 and Figure 8This technique improves resolution by parametrically rearranging the feature channels to the time dimension. This process avoids introducing additional convolution kernels, making it a purer upsampling method with less prior bias. Its advantage lies in generating smoother, more natural waveforms while preserving the spectral content and temporal accuracy of the original signal to the greatest extent possible, which is crucial for subsequent seismic data interpretation. This invention replaces the upsampling method in the joint autoregressive hierarchical prior model with a one-dimensional sub-pixel convolution operation, achieving smoother and more accurate seismic waveform reconstruction.

[0091] More challenging technological breakthroughs lie in the one-dimensional transformation of context modeling and the reconstruction of probabilistic model context modeling methods. Existing context networks (such as...) Figure 6 The model shown employs a two-dimensional checkerboard pattern to capture spatial neighborhood relationships. However, this model's assumptions are physically invalid for time series with a strict chronological order. To address this fundamental flaw, this invention designs a context network that strictly adheres to temporal causal relationships (see...). Figure 9 This network replaces the two-dimensional spatial mask with a one-dimensional causal mask, forcing that when predicting the probability distribution at any time point t, the information it relies on can only come from historical time points prior to t. This design is not a simple dimensional modification, but a necessary condition to ensure that the model learns a true conditional probability distribution. Its technical benefits are significant: a physically consistent probability model can more accurately predict the statistical distribution of data, allowing the entropy bottleneck module implemented by the quantizer and entropy encoder to utilize a more optimized encoding table, representing the same information with far fewer bits than the non-causal model, approaching the theoretical performance limit of lossless compression systems.

[0092] In summary, this invention enhances the depth and breadth of feature representation through deep residual structures, improves the fidelity of signal reconstruction through one-dimensional subpixel convolution, and ensures the physical authenticity and accuracy of probabilistic modeling through a one-dimensional causal autoregressive model. This successfully and systematically transforms a two-dimensional image compression framework into a high-performance compression system specifically designed for one-dimensional seismic data. These interconnected improvements work synergistically to achieve unprecedented compression efficiency for one-dimensional seismic data.

[0093] The network architecture proposed in this invention performs end-to-end training through rate-distortion optimization, such as... Figure 2 As shown. The loss function is... .in, This represents the overall loss function, which is the objective to be minimized during network training; Represents the encoding bitrate, which measures the number of bits required for compressed data, calculated in bits per pixel; This indicates reconstruction distortion, used to quantify the quality difference between the original and reconstructed data, and is measured using mean squared error or perceptual loss. In actual training, The choice of parameters has a significant impact on the final performance and needs to be adjusted according to the specific application scenario. When the value is smaller, the system tends to achieve a higher compression ratio but at the cost of reduced reconstruction quality; conversely, it prioritizes reconstruction accuracy.

[0094] To address the computational resource constraints of edge deployments, this invention optimizes computational efficiency at every stage of network design. By employing lightweight techniques such as depthwise separable convolution and quantized perceptual training, complex deep neural networks are compressed into efficient models suitable for edge devices, significantly reducing computational complexity and memory usage while maintaining compression performance.

[0095] This systematic technological innovation not only solved the performance bottleneck of seismic data compression, but more importantly, it provided a successful example for cross-domain technology transfer, proving that by deeply understanding the essential characteristics of different data modalities, advanced technologies can be effectively transferred between different application fields.

[0096] The advantage of this architecture lies in its integrity and adaptability. Compared to traditional methods, the system of this invention can automatically adjust the compression strategy according to the actual distribution of the data, rather than relying on preset transformations or dictionaries. At the same time, end-to-end training ensures coordinated optimization among the modules, avoiding suboptimal solutions that may result from step-by-step optimization.

[0097] 3. Training methods

[0098] The proposed model employs an end-to-end training strategy for optimization, which jointly updates all parameters of the analysis network, synthesis network, super-prior network, and context adaptation module. The overall optimization objective is to minimize a composite loss function designed to balance rate-distortion performance.

[0099] Loss function: The entire model (including the analysis network, synthesis network, super-prior network, and context-dependent parameter prediction part) undergoes end-to-end joint optimization. Rate distortion loss:

[0100] ;

[0101] in, This represents the overall loss function, which is the objective to be minimized during network training; Represents the encoding bitrate, which measures the number of bits required for compressed data, calculated in bits per pixel; This indicates reconstruction distortion and is used to quantify the quality difference between the original data and the reconstructed data. It is calculated using the mean square error method.

[0102] Optimizer: To minimize the rate-distortion loss function described above, this invention employs a gradient-based optimization algorithm. Specifically, this invention uses the Adam (Adaptive Moment Estimation) optimizer for model training.

[0103] 4. Algorithm deployment for edge devices

[0104] This invention constructs a complete real-time seismic data compression and transmission system to verify the practicality of the above technical solutions. The system deploys a trained deep learning compression model to edge computing devices, achieving on-site compression and efficient transmission of seismic data. Figure 12 As shown, the entire system adopts a distributed architecture design and achieves end-to-end optimized processing from data acquisition to remote storage through edge computing technology.

[0105] 4.1 System Architecture Design

[0106] This system adopts an edge-cloud collaborative distributed architecture. Core components include: edge computing devices using the NVIDIA Jetson Orin Nano platform, integrating 1024 CUDA cores and 32 Tensor Cores, providing 40 TOPS of AI inference computing power with power consumption controlled within the range of 7W to 15W, meeting the low-power requirements for field deployment. The data acquisition end is equipped with a three-component digital seismometer, a built-in 32-bit ADC, supporting a maximum sampling frequency of 1000Hz, and achieving high-bandwidth data transmission via a USB 3.0 interface. For network connectivity, the system supports both Gigabit Ethernet and 4G / 5G wireless modules, ensuring network connectivity in different deployment environments.

[0107] 4.2 Software System Implementation

[0108] The software architecture is built on the NVIDIA JetPack SDK, integrating core components such as CUDA Toolkit, cuDNN, and TensorRT. Model deployment follows an optimized path from PyTorch to ONNX and then to TensorRT, fully utilizing hardware acceleration capabilities. The system mainly consists of four functional modules: a data receiving module responsible for real-time reading of MiniSEED or binary format data output from seismometers and processing it in segments according to preset time windows; a preprocessing module performing data format conversion, normalization, and tensor reshaping; a compression inference module calling the optimized neural network model to perform real-time compression calculations; and a data encapsulation module packaging the compressed bitstream and metadata, transmitting it to the backend server via TCP / IP protocol.

[0109] 4.3 Workflow and Performance Characteristics

[0110] The system operation process is as follows: continuously acquired seismic waveform data is segmented into time windows ranging from 1 to 5 seconds, then standardized in format by a preprocessing module. This data is then input into a trained compression network for real-time inference calculations. The compressed data, along with metadata such as timestamps and station information, is encapsulated and transmitted. The backend server receives the data, stores it, and calls the corresponding decompression model to reconstruct the original waveform when needed.

[0111] 4.3 Conclusion

[0112] This embodiment verifies the feasibility of the technical solution of the present invention in actual earthquake monitoring scenarios, realizes a complete closed loop of real-time compression processing at the edge and remote data reconstruction, and provides an effective technical solution for bandwidth optimization and cost control of earthquake data acquisition networks.

[0113] like Figure 12 As shown, the seismic data acquisition module collects seismic data and sends the raw seismic data to the Jetson Orin Nano. The Jetson Orin Nano then sends the compressed bitstream to the server for decompression and storage. The transmission methods are divided into two categories: BeiDou short message service and WiFi, 4G, Ethernet, etc.

[0114] 5. Verification Indicators

[0115] Validation Plan: To verify the performance of the method proposed in this invention in practical applications, tests will be conducted in real-world industrial scenarios such as seismic exploration or vibration monitoring. Real seismic or vibration data will be collected for compression and reconstruction experiments. The main evaluation criteria will be the reconstruction quality of the model under different compression rates, with metrics including mean square error (MSE), normalized mean square error (NMSE), normalized root mean square error (NRMSE), and signal-to-noise ratio (SNR).

[0116] Mean squared error (MSE) is a commonly used metric to measure the difference between predicted and actual values. It is the average of the squared residuals (the squares of the differences) between the actual and model-predicted values, defined as follows:

[0117] .

[0118] The total number of sampling points. For the raw seismic data in the first The values ​​of each sampling point For the compressed and reconstructed seismic data in the first The normalized mean square error (NMSE) is a normalized form of MSE relative to the energy of the original data, reflecting the proportion of error in the overall data. It is defined as follows:

[0119] .

[0120] Normalized root mean square error (NRMSE) is the result of the root mean square error (RMSE) relative to a normalized benchmark of the original data. The following uses the maximum value of the original data. As a normalization benchmark, it is defined as follows:

[0121] .

[0122] The maximum value of the original seismic data will be used. At the same time, the computational latency of the compression and decompression process will be rigorously tested to ensure that it meets the real-time or near-real-time application requirements of edge devices, and to verify the adaptability of the algorithm to different working conditions and different types of seismic / vibration signals.

[0123] Figure 13 This is sample 1 from the STEAD seismic dataset. The compression ratio is 67x, and the NMSE is 0.0041.

[0124] Figure 14 This is sample 2 from the STEAD seismic dataset. The compression ratio is 10x, and the NMSE is 0.0009.

[0125] Figure 15 This is sample 3 from the STEAD seismic dataset. The compression ratio is 16x, and the NMSE is 0.0019.

[0126] To verify the practicality and effectiveness of the AI-based edge device seismic data compression method proposed in this invention, a complete implementation and verification experiment is described in detail. This experiment, from data preparation and system deployment to performance evaluation, comprehensively demonstrates the feasibility of this technical solution in real-world application scenarios.

[0127] Figure 16 The diagram illustrates the complete workflow of AI-based data compression. This workflow includes data cleaning and standardization in the preprocessing stage, expansion and transformation in the data augmentation stage, and the construction of labeled data and ontology for the training dataset. The core data compression neural network adopts an encoder-feature extraction-decoder architecture, achieving model-based processing through a training optimization module. After data acquisition and preprocessing are completed in the real-time data acquisition and preprocessing stage, the system undergoes actual deployment testing, performance verification and evaluation, and continuous optimization of the network based on feedback results, forming a complete closed-loop optimization system.

[0128] (1) Experimental dataset

[0129] This experiment uses the Stanford Earthquake Dataset (STEAD) as the basic data source. The STEAD dataset contains more than one million seismic waveform records, covering seismic events in different geographical regions around the world, and has good representativeness and diversity.

[0130] Training Dataset Construction: 8,000,000 high-quality seismic records were selected from the STEAD dataset as the model training data. All training data underwent a standardized preprocessing process, including linear detrending to eliminate long-period drift, mean removal to eliminate DC components, tapering to reduce boundary effects, and 0.5Hz high-pass filtering to remove low-frequency noise interference. To enhance the model's generalization ability, the training set also included data samples under different signal-to-noise ratios, ranging from -20dB to 40dB.

[0131] Test Dataset Construction: A test set containing 35,000 seismic records was independently constructed, employing a stratified sampling strategy to ensure a balanced distribution under different signal-to-noise ratio (SNR) conditions. Specifically, 5,000 records were selected for each SNR interval: -20dB to -10dB, -10dB to 0dB, 0dB to 10dB, 10dB to 20dB, 20dB to 30dB, 30dB to 40dB, and above 40dB, ensuring the comprehensiveness and objectivity of the algorithm performance evaluation.

[0132] (2) System deployment environment

[0133] Hardware platform configuration: The NVIDIA Jetson Orin Nano 8GB development kit was selected as the edge computing deployment platform. This platform is equipped with an Ampere architecture GPU (1024 CUDA cores, 32 Tensor Cores) and a 6-core ARM Cortex-A78AE CPU, providing 40 TOPS of AI inference performance, with power consumption controlled within the range of 7W to 15W, meeting the low power consumption requirements of edge deployment.

[0134] Software Environment Setup: The system is built on NVIDIA JetPack SDK 5.1.2 and runs on Ubuntu 20.04LTS. The development environment includes CUDA Toolkit 11.4, cuDNN 8.6.0, TensorRT 8.5.1, and PyTorch 1.13.0 optimized for the ARM64 architecture. To improve inference efficiency, TensorRT is used to quantize and optimize the trained PyTorch model, converting the FP32 precision model to INT8 precision, significantly improving inference speed while maintaining compression quality.

[0135] Model deployment process: First, the PyTorch trained model is converted to the ONNX intermediate format. Then, TensorRT is used for graph optimization and accuracy calibration, finally generating an inference engine optimized for the Jetson platform. The deployed compressed model includes a complete encoding network (analysis network, super-prior network, context network) and an entropy encoder, supporting real-time seismic data compression processing.

[0136] (3) Expected performance and effects

[0137] Compression performance: such as Figure 17 As shown, this embodiment is expected to enable efficient compression of real seismic data on Jetson Orin Nano. Figure 17 The rate-distortion curves shown demonstrate that the deep learning-based seismic data compression method comprehensively outperforms wavelet methods under any compression ratio. Experimental results show that the proposed compression method performs excellently on the Jetson OrinNano embedded platform. For example, for natural earthquake data, it can achieve a compression ratio of over 10 times while maintaining a low normalized mean square error (NMSE). Figure 18 As shown, the compression ratio of sample 1 is 16.12, and the NMSE is 0.0046; Figure 20 The compression ratio of sample 2 is 10.22, and the NMSE is 0.0093; while Figure 22 The middle sample 3 achieved a compression ratio of 114.37 and an NMSE of only 0.0059, fully ensuring the accurate reconstruction of key phase information (P-wave, S-wave arrival time, amplitude, etc.). In contrast, the traditional three-segment wavelet compression method (implementation method see appendix) performed significantly worse under the same conditions, such as... Figure 19 , Figure 21 , Figure 23 As shown, their compression ratios are only 6.22, 7.22 and 10.67, respectively, while the corresponding NMSEs are as high as 0.028, 0.037 and 0.015, respectively, which further verifies the significant advantages of the proposed method in the field of seismic data compression.

[0138] Processing speed: If processed in 1-second data segments, the model's encoding latency is expected to be controlled within tens to two hundred milliseconds, fully meeting the requirements of real-time processing. The Jetson Orin Nano's GPU can process the computation of multiple data channels or time segments in parallel.

[0139] Power consumption: The Jetson Orin Nano typically consumes between 7W and 12W, making it suitable for long-term field monitoring sites powered by solar or batteries.

[0140] This embodiment demonstrates that the proposed AI-based edge device seismic data compression method and system are practically feasible, can be effectively integrated into existing seismic monitoring systems, and bring significant performance improvements and economic benefits. By deploying an optimized deep learning model on a modern edge AI platform such as Jetson Orin Nano, efficient and intelligent seismic data compression can be successfully achieved at the forefront of data acquisition.

[0141] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0142] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0143] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for compressing seismic data from edge devices based on artificial intelligence, characterized in that, include: Acquire raw one-dimensional seismic data; The joint autoregressive hierarchical prior model is adapted from the domain of two-dimensional image compression to one-dimensional seismic data through the following modifications: (a) The modified joint autoregressive hierarchical prior model includes an analysis network, a super-prior network, a one-dimensional checkerboard context network, and a synthesis network; a deep one-dimensional residual network is constructed as the backbone of the analysis network and the synthesis network. (b) Replace the two-dimensional convolution operation in the joint autoregressive hierarchical prior model with a one-dimensional residual network operation; (c) Replace the upsampling method in the joint autoregressive hierarchical prior model with a one-dimensional subpixel convolution operation; (d) The designed one-dimensional chessboard context network can reconstruct temporal causal relationships; The modified joint autoregressive hierarchical prior model was trained end-to-end by optimizing the bitrate and reconstruction distortion, with the loss function... ,in, The bitrate for encoding the modified joint autoregressive hierarchical prior model. The mean square error between the reconstructed seismic data obtained after compression and restoration and the original seismic data; Compression processing is performed in real time on an edge computing device to generate a compressed seismic data bitstream and transmit it over a wireless network.

2. The method for compressing seismic data from edge devices based on artificial intelligence according to claim 1, characterized in that, The analysis network employs a deep network consisting of multiple cascaded one-dimensional residual modules, which performs downsampling and feature extraction through one-dimensional convolution, combined with a generalized divisible normalization layer and the LeakyReLU activation function; the super-prior network adopts an autoencoder architecture; the one-dimensional checkerboard context network employs a parallel autoregressive model consisting of one-dimensional causal mask convolutional layers; the synthesis network employs a deep one-dimensional residual network symmetrical to the analysis network structure, which performs upsampling operations through one-dimensional sub-pixel convolution to reconstruct the signal.

3. The method for compressing seismic data from edge devices based on artificial intelligence according to claim 1, characterized in that, The modified joint autoregressive hierarchical prior model's working process includes compression and decompression stages; During the compression stage, the analysis network transforms the input seismic data into a first latent representation, and the super-prior network generates a second latent representation based on the first latent representation. Based on the quantized version of the second latent representation and combined with the autoregressive information provided by the one-dimensional checkerboard context network, the precise conditional probability model of the first latent representation is jointly determined, and the quantized versions of the two latent representations are entropy encoded accordingly to generate a compressed bitstream. In the decompression stage, the quantized version of the second latent representation is first decoded from the compressed bitstream to reconstruct the initial probability model; Subsequently, a one-dimensional checkerboard context network is used to sequentially decode the quantized version of the first latent representation from the bitstream with the aid of autoregression; the synthetic network reconstructs the seismic data based on the decoded first latent representation.

4. A compression system for implementing the method according to any one of claims 1 to 3, characterized in that, include: The data receiving module is responsible for reading the one-dimensional seismic data output by the seismometer in real time and processing it in segments according to a preset time window; The preprocessing module performs data format conversion, normalization, and tensor reshaping of one-dimensional seismic data. The compression inference module calls the optimized joint autoregressive hierarchical prior model to perform real-time compression of one-dimensional seismic data. The data encapsulation module packages the compressed seismic data bitstream with metadata and transmits it to the backend server via the TCP / IP protocol.

Images