A ship situation awareness data loss interpolation system and method based on Jili satellite

Abstract

This invention discloses a ship situational awareness data loss interpolation system and method based on the Geely satellite. The system includes a ship-side dynamic compression module, a satellite transmission and data reception module, a nonlinear logarithmic transformation module, a multi-scale adaptive decomposition module, a spectral filtering correlation module, and a nonlinear exponential transformation module. These modules work closely together to accurately reconstruct incomplete sparse time-series data caused by packet loss during Geely satellite transmission, thereby achieving complete ship situational awareness data loss interpolation based on Geely satellite. This invention has the following advantages: no retransmission is required, greatly saving satellite bandwidth and reducing latency; it overcomes the "peak clipping" effect and accurately restores nonlinear abrupt signal; it avoids trend drift and adapts to complex and fluctuating dynamic conditions; and it provides powerful filtering to suppress high-frequency noise interference under sparse samples.

Description

Technical Field

[0001] This invention relates to a ship situational awareness data loss interpolation system and method, specifically to a ship situational awareness data loss interpolation system and method based on the Geely satellite. Background Technology

[0002] As shipping systems evolve towards intelligence, ocean-going vessels need to transmit real-time information such as location, speed, and main engine operating status to shore-based centers for remote monitoring and intelligent decision-making. Currently, this data exchange primarily relies on communication networks such as the Geely satellite. However, due to the complex maritime environment, satellite communication often faces challenges such as limited bandwidth, high costs, and susceptibility to signal loss. To ensure data integrity, existing communication networks typically require ships to retransmit data upon detecting loss. When multiple ships simultaneously share limited satellite channels, this repeated retransmission severely strains already limited bandwidth, causing network congestion and significant delays in data received by shore, failing to meet the real-time requirements of situational awareness.

[0003] Existing network improvement methods mostly focus on increasing retransmission efficiency, neglecting the powerful computing capabilities of shore-based servers and failing to effectively utilize shore-based computing power to compensate for the limitations of satellite communication. To recover lost data without frequent retransmissions, the industry has also attempted to use conventional algorithms to estimate and fill in missing values. However, the operating conditions of ships at sea are extremely complex, and the resulting sensor data often experiences irregular and drastic fluctuations. Traditional models struggle to accurately adapt to these dynamic changes, easily leading to estimations that gradually deviate from reality. Furthermore, data collected by marine sensors often contains significant noise interference, making existing algorithms easily misled when historical data is insufficient. In addition, when ships encounter severe sea conditions or sudden equipment failures, relevant data often experiences disproportionately drastic changes. Existing conventional algorithms mostly handle only regular, gradual changes and cannot accurately calculate and reconstruct these complex emergencies, resulting in significant errors in the final data filling. Therefore, overcoming the problem of easily lost satellite communication data and accurately filling in ship status data under conditions of high interference and data mutation is a pressing technical challenge. Summary of the Invention

[0004] This invention provides a ship situational awareness data loss interpolation system and method based on the Geely satellite, using shore-based computing power to compensate for satellite bandwidth resources. The core of this invention is a nonlinear time-series reconstruction model that integrates multi-scale decomposition and spectral filtering. This model first uses multi-scale decomposition to capture dynamic trends. Then, in the frequency domain, the model uses spectral filtering to suppress noise interference under sparse samples. Furthermore, nonlinear spatial transformation converts abrupt signals into an additive form to handle multiplicative abrupt changes, thereby improving the model's ability to model sudden changes. Experiments on multiple public datasets and real ship data show that even in high packet loss environments, this model can still accurately reconstruct ship situational awareness data, outperforming existing mainstream models. This invention verifies the feasibility of compensating for satellite bandwidth resources with shore-based computing power, providing technical support for ship-shore collaborative sensing in next-generation intelligent shipping systems.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A ship situational awareness data loss interpolation system based on the Geely satellite includes a ship-side dynamic compression module, a satellite transmission and data reception module, a nonlinear logarithmic transform module, a multi-scale adaptive decomposition module, a spectral filtering correlation module, and a nonlinear exponential transform module, wherein:

[0007] The ship-end dynamic compression module is responsible for acquiring the original multi-dimensional situational awareness time-series data collected by ship sensors and calculating the local volatility of the data in real time. When the data is stable, it maintains a basic dead zone threshold to obtain a high compression ratio. When the data fluctuates violently, it adaptively expands the dead zone threshold, thereby performing dynamic dead zone compression on the original data and generating a sparse representation data sequence that removes redundancy but retains key mutation information.

[0008] The satellite transmission and data reception module is responsible for sending the sparse representation data sequence to the shore-based center through the Geely satellite link. The shore-based receiver obtains the incomplete time sequence caused by packet loss in the satellite channel and constructs a corresponding binary mask sequence based on the successful reception status of the data, which is used to identify the missing positions of the data in the future.

[0009] The nonlinear logarithmic transformation module is responsible for receiving encoder input from the sequence data input and decoder input containing seasonal and trend terms, and transforming nonlinear mutation features into implicit additive space through logarithmic mapping.

[0010] The multi-scale adaptive decomposition module is responsible for the collaborative separation of seasonal and trend terms in the implicit reconstruction space of the data in the encoder and decoder of the deep reconstruction network.

[0011] The spectral filtering correlation module is responsible for suppressing the interference of sparse and noisy samples during autocorrelation modeling. It maps the input feature sequence to the frequency domain via fast Fourier transform. After calculating the frequency domain correlation, it sets a threshold based on the magnitude distribution of amplitude energy and generates a mask to accurately filter out low-amplitude high-frequency noise. Then, it restores the clean features to the time domain through inverse Fourier transform and performs time delay aggregation to complete global correlation extraction.

[0012] The nonlinear exponential transformation module is responsible for receiving the reconstruction results of seasonal and trend terms generated collaboratively by the deep reconstruction network decoder. Through inverse exponential transformation and sign restoration operations, the interpolated latent space sequence is accurately restored to the original physical scale of the ship data, and finally outputs complete ship situational awareness data without any missing data.

[0013] A method for interpolating lost ship situational awareness data using the aforementioned system includes the following steps:

[0014] Step 1: Acquire shipboard situational awareness data and perform dynamic dead zone compression:

[0015] Step 1.1: The shipborne edge computing node acquires the raw multi-dimensional situational awareness time-series data collected by the ship's sensors in real time;

[0016] Step 1.2: The ship-end dynamic compression module uses a dynamic dead zone compression strategy based on local volatility to process the original sensing data stream.

[0017] Step 1.3: The compressed sparse data stream is then transmitted to the shore via the Geely satellite link;

[0018] Step 2: Receiving Incomplete Data from the Shore and Constructing the Mask Sequence:

[0019] Step 2.1: The satellite transmission and data reception module receives compressed sensing data transmitted via the satellite link;

[0020] Step 2.2: Based on the received status, the shore-based computing center constructs a binary mask sequence with the same length as the original time series. Successfully received data is marked as Data that was not successfully received due to packet loss is marked as... .

[0021] Step 3: Spatial mapping based on the nonlinear logarithmic transformation module:

[0022] The masked sparse sequence is input into the front end of the ship situational awareness data loss interpolation system, and the original sequence is mapped from the multiplicative transformation space to the approximate additive space through the nonlinear logarithmic transformation module.

[0023] Step 4: Correlation feature extraction based on multi-scale decomposition and spectral filtering:

[0024] Step 4.1: Map the sequence The encoder and decoder of the input deep reconstruction network are used to collaboratively extract temporal features and perform interpolation reconstruction.

[0025] Step 4.2: Extract trends using the multi-scale adaptive decomposition module:

[0026] The multi-scale adaptive decomposition module adopts a dual-branch structure, extracting dynamic trends by introducing smoothing kernels of different scales, wherein:

[0027] The first branch uses a small-scale smooth kernel. Obtain a preliminary trend estimate:

[0028]

[0029] in, This is a feature for primary trend estimation. This indicates the average pooling operation. Indicates the boundary fill operation;

[0030] The second branch uses a large-scale smooth kernel. ( Perform two consecutive smoothing operations to extract deeper low-frequency trends:

[0031]

[0032] in, This is a transitional trend characteristic. It is a deep, low-frequency trend characteristic;

[0033] Subsequently, the two trends are concatenated along the feature dimension and then processed using learnable parameters. Generate fusion weights :

[0034]

[0035] in, For unnormalized fusion weights, This represents a concatenation operation along the feature dimension;

[0036] The weights are obtained after normalization using the Softmax function. And perform adaptive weighting to obtain the final trend term. and seasonal items :

[0037]

[0038] in, and Normalized weights The branch weight components correspond to the primary and deep trend characteristics;

[0039] Step 4.3: Noise mitigation modeling using the spectral filtering correlation module:

[0040] The spectral filtering correlation module will input the Query( ) and Key ( Through Fast Fourier Transform Mapping to the frequency domain, frequency domain correlation is calculated using the Wiener-Khinchin theorem. :

[0041]

[0042] in, Indicates the complex conjugate operation;

[0043] Calculate its amplitude:

[0044]

[0045] And select the previous The amplitude of each frequency component is used as a threshold. ;

[0046] Construct the following binary mask function to filter out high-frequency noise:

[0047]

[0048] in, For constructing a frequency domain binary mask;

[0049] Filter the frequency domain correlation:

[0050]

[0051] Recover to the time domain using the inverse Fourier transform:

[0052]

[0053] in, This represents the inverse fast Fourier transform. The frequency domain correlation after filtering. To recover the correlation features to the time domain, a time-delay aggregation operation is performed to complete feature interpolation modeling;

[0054] Step 5: Nonlinear exponential inverse transform and output of the reconstruction result:

[0055] Step 5.1: After the decoder completes processing, it outputs the reconstructed seasonal components in the implicit space. Reconstructing components with trend items ;

[0056] Step 5.2: Using the nonlinear exponential transformation module, map the reconstructed features back to the physical scale of the original ship data.

[0057] set up For the dimensional projection matrix, the components are first merged and reconstructed to obtain:

[0058]

[0059] in, This is the initial reconstruction sequence in the latent space;

[0060] Subsequently, an inverse exponential transform, symmetric to the logarithmic transform, is performed:

[0061]

[0062] Final output This refers to the complete ship situational awareness time-series data after eliminating the impact of packet loss and accurately reconstructing the data.

[0063] Compared with the prior art, the present invention has the following advantages:

[0064] 1. No retransmission required, greatly saving satellite bandwidth and reducing latency. This invention, through a collaborative architecture of "ship-side dynamic compression + shore-side computing power reconstruction," tolerates the high packet loss rate during Geely satellite transmission, breaking through the network congestion bottleneck caused by frequent retransmissions in the traditional ARQ mechanism, and effectively ensuring the real-time performance of ship situational awareness data.

[0065] 2. Overcoming the "peak-shaving" effect and accurately restoring nonlinear abrupt signals. To address the proportional and drastic data abrupt changes caused by extreme sea conditions or equipment failures, this invention introduces nonlinear spatial logarithmic and exponential symmetric transformations, converting difficult-to-handle multiplicative abrupt changes into conventional additive calculations, significantly improving the model's ability to accurately interpolate sudden spike signals.

[0066] 3. Avoiding trend drift and adapting to complex and dynamic operating conditions. This invention employs a multi-scale adaptive decomposition strategy, using a dual-branch structure and smoothing processing at different scales to dynamically weight and fuse irregular ship fluctuation data, effectively avoiding trend drift errors caused by traditional fixed-window decomposition.

[0067] 4. Powerful filtering to suppress high-frequency noise interference under sparse samples. Addressing the high packet loss rate of Geely satellites and the inherent noise issues of maritime sensors, this invention performs threshold filtering based on amplitude energy in the frequency domain, accurately eliminating high-frequency noise and significantly improving the model's interpolation robustness in harsh communication environments. Attached Figure Description

[0068] Figure 1 A flowchart of the method for interpolating lost ship situational awareness data based on Geely satellite;

[0069] Figure 2 This is a diagram illustrating the principle of dynamic dead zone compression based on local volatility.

[0070] Figure 3 The overall architecture diagram of the nonlinear time-series reconstruction model for multi-scale decomposition and spectral filtering;

[0071] Figure 4 This is a structural diagram of the multi-scale adaptive decomposition module;

[0072] Figure 5 Here is a structural diagram of the spectral filtering correlation module;

[0073] Figure 6 A performance comparison chart of dynamic dead-zone compression methods under different data segments;

[0074] Figure 7 A comparison chart showing the interpolation effect for lost time-series data. Detailed Implementation

[0075] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.

[0076] This invention provides a ship situational awareness data loss interpolation system based on the Geely satellite. The system includes a ship-side dynamic compression module, a satellite transmission and data reception module, a nonlinear logarithmic transformation module, a multi-scale adaptive decomposition module, a spectral filtering correlation module, and a nonlinear exponential transformation module. These modules work closely together to accurately reconstruct incomplete sparse time-series data caused by packet loss during Geely satellite transmission, thereby achieving complete ship situational awareness data loss interpolation based on the Geely satellite. Wherein:

[0077] The ship-end dynamic compression module is responsible for acquiring the raw multi-dimensional situational awareness time-series data collected by the ship's sensors and calculating the local volatility of the data in real time. When the data is stable, a basic dead zone threshold is maintained to obtain a high compression ratio. When the data fluctuates drastically, the dead zone threshold is adaptively expanded, thereby performing dynamic dead zone compression on the raw data to generate a sparse representation data sequence that removes redundancy but retains key mutation information.

[0078] The satellite transmission and data reception module is responsible for transmitting the sparse representation data sequence to the shore-based center via the Gillette satellite link. The shore-based receiver acquires the incomplete time sequence caused by packet loss in the satellite channel and constructs a corresponding binary mask sequence based on the successful data reception status, which is used to identify the missing data locations later.

[0079] The nonlinear logarithmic transformation module, as the front-end processing stage of the model, is responsible for receiving the encoder input from the sequence data input and the decoder input containing seasonal and trend terms. This module transforms the nonlinear mutation features into an implicit additive space through logarithmic mapping, facilitating the stable learning of subsequent deep networks.

[0080] The multi-scale adaptive decomposition module is responsible for the collaborative separation of seasonal and trend terms in the implicit reconstruction space of the encoder and decoder of the deep reconstruction network. This module employs a dual-branch parallel structure, internally introducing smoothing kernels of different scales to perform single and multiple consecutive smoothing processes on the input sequence to expand the receptive field. Subsequently, learnable fusion weights are used to adaptively weight the two smoothed features, accurately extracting local details and global low-frequency trends, avoiding trend drift caused by a single fixed window.

[0081] The spectral filtering correlation module is responsible for suppressing interference from sparse and noisy samples during autocorrelation modeling. This module maps the input feature sequence to the frequency domain via a Fast Fourier Transform (FFT). After calculating the frequency domain correlation, it sets a threshold based on the magnitude distribution of amplitude energy and generates a mask to accurately filter out low-amplitude high-frequency noise. Subsequently, an Inverse Fourier Transform (IFT) is used to restore the clean features to the time domain, and time-delay aggregation is performed to complete global correlation extraction.

[0082] The nonlinear exponential transformation module, as the output backend of the time-series reconstruction model, is responsible for receiving the reconstruction results of seasonal and trend terms generated collaboratively by the deep reconstruction network decoder. Through inverse exponential transformation and sign restoration operations, it accurately restores the interpolated latent space sequence to the original physical scale of the ship data, and finally outputs complete ship situational awareness data without any missing data.

[0083] In this invention, a multi-scale adaptive decomposition module is arranged in the multi-layer network structure of the encoder and decoder to collaboratively separate and extract complex dynamic trends and seasonal features. For example... Figure 4 As shown, this module receives sequence input and employs a dual-branch parallel structure: its left path performs a single smoothing process (small-scale kernel) on the sequence to capture local short-term fluctuations; its right path performs a first smoothing process (large-scale kernel) and a second smoothing process (large-scale kernel) consecutively to extract global deep low-frequency trends. Subsequently, the smoothing results from both paths enter the Softmax weighted fusion node for adaptive weight allocation, thereby generating a drift-resistant trend term output. The seasonal term output is then obtained by separating the input residuals.

[0084] In this invention, spectral filtering correlation modules are alternately arranged in the network to suppress interference from sparse and noisy samples during autocorrelation modeling. For example... Figure 5 As shown, this module first retrieves the Query( Sequence input and Key ( The input sequence is mapped to the frequency domain using a Fast Fourier Transform (FFT) and frequency domain correlation calculation is performed. Then, a crucial amplitude threshold filtering stage is initiated, where a threshold mask is set to precisely remove high-frequency noise interference. The frequency domain features after noise reduction filtering are then fully recovered to the time domain using an Inverse Fast Fourier Transform (IFFT), and finally compared with the corresponding Value(s). The sequence inputs are merged, a delay aggregation operation is performed, and a highly robust global correlation feature is output.

[0085] In this invention, such as Figure 3 As shown, time series decomposition is interspersed within the network structure, used to perform basic decomposition and reorganization operations on intermediate latent variables between the extraction of different features. This operation is used to perform basic decomposition and reorganization operations on intermediate latent variables between the extraction of different features (such as spectral filtering features and multi-scale decomposition features). Specifically, at the beginning of each layer of the encoder, time series decomposition is used to initially split the input features into seasonal and trend components, which are then fed into the subsequent spectral filtering correlation module and multi-scale adaptive decomposition module for deep modeling, thereby achieving multi-dimensional feature decoupling of complex ship perception data.

[0086] In this invention, the adaptive decomposition module is followed by a feedforward fully connected layer, which is used to perform nonlinear spatial mapping and feature dimension transformation on the extracted temporal features.

[0087] In this invention, after the encoder completes multi-layer feature extraction, the key features are extracted. Matrix and The matrix is ​​then passed to the spectral filtering correlation module on the decoder side for cross-end interaction. Finally, the decoder integrates all features to generate and output the accurately reconstructed complete sequence data.

[0088] This invention also provides a method for interpolating lost ship situational awareness data using the above-mentioned modules, such as... Figure 1 As shown, the method includes the following steps:

[0089] Step 1: Acquire shipboard situational awareness data and perform dynamic dead-zone compression. The shipboard edge computing node acquires the raw multi-dimensional situational awareness time-series data collected by the ship's sensors in real time. To adapt to the bandwidth limitations of the Geely satellite, the shipboard dynamic compression module adopts a dynamic dead-zone compression strategy based on local volatility to process the raw sensing data stream.

[0090] Figure 2 This diagram illustrates the principle of dynamic dead zone compression based on local volatility, and compares the differences between static and dynamic dead zone compression.

[0091] Specifically, the amount of content in the sliding window is calculated in real time. Initial sensing data Standard deviation :

[0092]

[0093] in, This represents the mean of the data within that interval.

[0094] Subsequently, the volatility coefficients of adjacent intervals were calculated. :

[0095]

[0096] in, The standard deviation of the data within the current sliding window. This is the standard deviation of the data within the previous adjacent sliding window. Based on this fluctuation coefficient, the dead zone threshold for the compression process is determined. Perform adaptive updates:

[0097]

[0098] in, The dead zone threshold updated at the current moment. This is the dead zone threshold of the previous moment.

[0099] This updated rule makes it possible to operate in a stable sailing state. Maintaining a small threshold to increase the compression ratio, while in the event of sudden changes The threshold is quickly relaxed to preserve key mutation information, thus adapting to signal fluctuations. The compressed sparse data stream is then transmitted to the shore via the Gilead satellite link.

[0100] Step 2: Receiving incomplete data on the shore and constructing the mask sequence.

[0101] The satellite transmission and data reception module receives compressed sensing data transmitted via the satellite link. Due to packet loss in satellite communication, the received data is an incomplete sequence with missing values.

[0102] Based on the received state, the onshore computing center constructs a binary mask sequence with the same length as the original time series. Successfully received data is marked as... Data that was not successfully received due to packet loss is marked as... .

[0103] Step 3: Spatial mapping based on the nonlinear logarithmic transformation module.

[0104] The masked sparse sequence is input into the front end of the nonlinear time series reconstruction model, and the original sequence is mapped from the multiplicative transformation space to the approximate additive space through the nonlinear logarithmic transformation module.

[0105] Let the input multivariate time series be... Perform the following logarithmic transformation on its elements:

[0106]

[0107] in, The sequence after mapping to additive space, Represents a symbolic function. This indicates element-wise multiplication.

[0108] This operation compresses the numerical dynamic range while preserving the original symbolic information, converting abrupt signals into an additive form that is easy for deep networks to process.

[0109] Step 4: Correlation feature extraction based on multi-scale decomposition and spectral filtering.

[0110] The mapped sequence The encoder and decoder of the input deep reconstruction network are used to collaboratively extract temporal features and perform interpolation reconstruction.

[0111] This step includes using a multi-scale adaptive decomposition module to extract trends and using a spectral filtering correlation module for noise-resistant modeling.

[0112] The multi-scale adaptive decomposition module employs a dual-branch structure, extracting dynamic trends by introducing smoothing kernels of different scales. The first branch uses a small-scale smoothing kernel. Obtain a preliminary trend estimate:

[0113]

[0114] in, This is a feature for primary trend estimation. This indicates the average pooling operation. This indicates a boundary fill operation. The second branch uses a large-scale smoothing kernel. ( Perform two consecutive smoothing operations to extract deeper low-frequency trends:

[0115]

[0116] in, This is a transitional trend characteristic. This represents a deep, low-frequency trend feature. Subsequently, the two trends are concatenated along the feature dimension and then processed using learnable parameters. Generate fusion weights :

[0117]

[0118] in, For unnormalized fusion weights, This represents the concatenation operation along the feature dimension. The weights are obtained after normalization using the Softmax function. And perform adaptive weighting to obtain the final trend term. and seasonal items :

[0119]

[0120] in, and Normalized weights The branch weight components correspond to the primary and deep trend characteristics.

[0121] The spectral filtering correlation module will input the Query( ) and Key ( Through Fast Fourier Transform Mapping to the frequency domain, frequency domain correlation is calculated using the Wiener-Khinchin theorem. :

[0122]

[0123] in, This represents the complex conjugate operation. Calculate its magnitude:

[0124]

[0125] And select the previous The amplitude of each frequency component is used as a threshold. Construct the following binary mask function to filter out high-frequency noise:

[0126]

[0127] in, This is the constructed frequency domain binary mask. Frequency domain correlation is then filtered.

[0128]

[0129] Recover to the time domain using the inverse Fourier transform:

[0130]

[0131] in, This represents the inverse fast Fourier transform. The frequency domain correlation after filtering. To recover the correlation features in the time domain, a time-delay aggregation operation is performed to complete the feature interpolation modeling.

[0132] Step 5: Nonlinear exponential inverse transform and output of the reconstruction result.

[0133] After the decoder completes processing, it outputs the reconstructed seasonal components in the implicit space. Reconstructing components with trend items At this point, the reconstructed features are mapped back to the physical scale of the original ship data using a nonlinear exponential transformation module. Let... For the dimensional projection matrix, the components are first merged and reconstructed to obtain:

[0134]

[0135] in, This is the initial reconstructed sequence in the latent space. Subsequently, an inverse exponential transform, symmetric to the logarithmic transform, is performed:

[0136]

[0137] Final output This refers to the complete ship situational awareness time-series data after eliminating the impact of packet loss and accurately reconstructing the data.

[0138] Figure 6Taking data segments 1 and 4 as examples, the performance differences between the two dead-zone compression strategies are compared. In data segment 1, the traditional static dead-zone compression method achieves a compression ratio of 1.52 by applying a fixed threshold. However, this strong truncation operation leads to significant information loss, with the root mean square error (RMSE) of the reconstructed sequence reaching 0.028. In contrast, the dynamic dead-zone compression method adopted in this invention can adaptively adjust the threshold range according to the local fluctuation amplitude of the data. When the data fluctuates drastically, the algorithm automatically widens the compression boundary to preserve the key transient structure, resulting in a significant reduction in the RMSE to 0.003, almost approaching the zero error level, while the compression ratio remains within the effective range of 1.19. Data segment 4 shows a consistent trend, with dynamic compression significantly reducing the RMSE from 0.020 of the static method to 0.001. The above results show that the dynamic dead zone compression mechanism based on local volatility adjustment can effectively identify sudden changes in ship operation data, avoid sacrificing important transient features to improve the compression rate, and thus achieve reliable data compression while ensuring extremely high reconstruction accuracy. This helps to alleviate the pressure on satellite communication bandwidth from the data source.

[0139] Figure 7 The performance of the time-series reconstruction model under extreme communication conditions was further examined. A severe scenario with a packet loss rate as high as 40% was selected, and the analysis focused on the torque parameters of the ship's shafting system. Shafting torque itself exhibits typical nonlinear abrupt changes. Within the time step range of 300 to 400, affected by simulated extreme sea conditions or equipment status changes, the actual torque record shows a sharp drop followed by a rapid rebound (shown by the black solid line in the figure). The iTransformer model on the right (blue dashed line) shows significant lag in fitting this range and fails to accurately capture the lowest point of torque, exhibiting a typical peak reduction phenomenon. The SpecDformer model proposed in this invention on the left (red dashed line) closely follows the actual black curve, accurately reproducing the deep V-shaped transient change process. This advantage is mainly due to the nonlinear spatial transformation module included in the model, which implicitly reconstructs multiplicative disturbances in logarithmic additive space, thus maintaining the ability to characterize drastic fluctuations even under high packet loss conditions. This comparison demonstrates that SpecDformer exhibits strong robustness in the face of strong nonlinear signals caused by extreme packet loss and complex sea conditions. It overcomes the shortcomings of traditional additive decomposition methods, which tend to lose signal peak and valley details, and provides a reliable data interpolation method for real-time and accurate perception of the shore-side situation.

Claims

1. A ship situational awareness data loss interpolation system based on Geely satellite, characterized in that... The system includes a ship-end dynamic compression module, a satellite transmission and data reception module, a nonlinear logarithmic transformation module, a multi-scale adaptive decomposition module, a spectral filtering correlation module, and a nonlinear exponential transformation module, wherein: The ship-end dynamic compression module is responsible for acquiring the original multi-dimensional situational awareness time-series data collected by ship sensors and calculating the local volatility of the data in real time. When the data is stable, it maintains a basic dead zone threshold to obtain a high compression ratio. When the data fluctuates violently, it adaptively expands the dead zone threshold, thereby performing dynamic dead zone compression on the original data and generating a sparse representation data sequence that removes redundancy but retains key mutation information. The satellite transmission and data reception module is responsible for sending the sparse representation data sequence to the shore-based center through the Geely satellite link. The shore-based receiver obtains the incomplete time sequence caused by packet loss in the satellite channel and constructs a corresponding binary mask sequence based on the successful reception status of the data, which is used to identify the missing positions of the data in the future. The nonlinear logarithmic transformation module is responsible for receiving encoder input from the sequence data input and decoder input containing seasonal and trend terms, and transforming nonlinear mutation features into implicit additive space through logarithmic mapping. The multi-scale adaptive decomposition module is responsible for the collaborative separation of seasonal and trend terms in the implicit reconstruction space of the data in the encoder and decoder of the deep reconstruction network. The spectral filtering correlation module is responsible for suppressing the interference of sparse and noisy samples during autocorrelation modeling. It maps the input feature sequence to the frequency domain via fast Fourier transform. After calculating the frequency domain correlation, it sets a threshold based on the magnitude distribution of amplitude energy and generates a mask to accurately filter out low-amplitude high-frequency noise. Then, it restores the clean features to the time domain through inverse Fourier transform and performs time delay aggregation to complete global correlation extraction. The nonlinear exponential transformation module is responsible for receiving the reconstruction results of seasonal and trend terms generated collaboratively by the deep reconstruction network decoder. Through inverse exponential transformation and sign restoration operations, the interpolated latent space sequence is accurately restored to the original physical scale of the ship data, and finally outputs complete ship situational awareness data without any missing data.

2. The ship situational awareness data loss interpolation system based on Geely satellite as described in claim 1, characterized in that... The multi-scale adaptive decomposition module is arranged in the multi-layer network structure of the encoder and decoder. The module receives the sequence input and adopts a dual-branch parallel structure: its left path performs a smoothing process on the sequence to capture local short-term fluctuations; its right path performs a smoothing process and a second smoothing process to extract the global deep low-frequency trend. Subsequently, the smoothed results from both paths are fed into the Softmax weighted fusion node for adaptive weight allocation, thereby generating a trend term output that resists drift. The seasonal term output is then obtained by separating the input residuals.

3. The ship situational awareness data loss interpolation system based on Geely satellite as described in claim 1, characterized in that... At the beginning of each layer of the encoder, the input features are initially divided into seasonal and trend components through time series decomposition. These components are then fed into the subsequent spectral filtering correlation module and multi-scale adaptive decomposition module for in-depth modeling, thereby achieving multi-dimensional feature decoupling of complex ship perception data.

4. The ship situational awareness data loss interpolation system based on Geely satellite as described in claim 1, characterized in that... The adaptive decomposition module is followed by a feedforward fully connected layer, which is used to perform nonlinear spatial mapping and feature dimension transformation on the extracted temporal features.

5. A method for interpolating lost ship situational awareness data using the system described in any one of claims 1-4, characterized in that... The method includes the following steps: Step 1: Acquire shipboard situational awareness data and perform dynamic dead zone compression; Step 2: Receiving Incomplete Data from the Shore and Constructing the Mask Sequence: Step 2.1: The satellite transmission and data reception module receives compressed sensing data transmitted via the satellite link; Step 2.2: Based on the received status, the shore-based computing center constructs a binary mask sequence with the same length as the original time series. Successfully received data is marked as Data that was not successfully received due to packet loss is marked as... ; Step 3: Spatial mapping based on the nonlinear logarithmic transformation module: The masked sparse sequence is input into the front end of the ship situational awareness data loss interpolation system, and the original sequence is mapped from the multiplicative transformation space to the approximate additive space through the nonlinear logarithmic transformation module. Step 4: Correlation feature extraction based on multi-scale decomposition and spectral filtering: Step 4.1: Map the sequence The encoder and decoder of the input deep reconstruction network are used to collaboratively extract temporal features and perform interpolation reconstruction. Step 4.2: Extract trends using the multi-scale adaptive decomposition module: The multi-scale adaptive decomposition module adopts a dual-branch structure, extracting dynamic trends by introducing smoothing kernels of different scales, wherein: The first branch uses a small-scale smooth kernel. Obtain a preliminary trend estimate: in, This is a feature for primary trend estimation. This indicates the average pooling operation. Indicates the boundary fill operation; The second branch uses a large-scale smooth kernel. Perform two consecutive smoothing operations to extract deeper low-frequency trends: in, This is a transitional trend characteristic. It is a deep low-frequency trend characteristic. ; Subsequently, the two trends are concatenated along the feature dimension and then processed using learnable parameters. Generate fusion weights : in, For unnormalized fusion weights, This represents a concatenation operation along the feature dimension; The weights are obtained after normalization using the Softmax function. And perform adaptive weighting to obtain the final trend term. and seasonal items : in, and Normalized weights The branch weight components correspond to the primary and deep trend characteristics; Step 4.3: Noise mitigation modeling using the spectral filtering correlation module: The spectral filtering correlation module will input Sequence and The sequence is subjected to Fast Fourier Transform Mapping to the frequency domain, frequency domain correlation is calculated using the Wiener-Khinchin theorem. : in, Indicates the complex conjugate operation; Calculate its amplitude: And select the previous The amplitude of each frequency component is used as a threshold. ; Construct the following binary mask function to filter out high-frequency noise: in, For constructing a frequency domain binary mask; Filter the frequency domain correlation: Recover to the time domain using the inverse Fourier transform: in, This represents the inverse fast Fourier transform. The frequency domain correlation after filtering. To recover the correlation features to the time domain, a time-delay aggregation operation is performed to complete feature interpolation modeling; Step 5: Nonlinear exponential inverse transform and output of the reconstruction result: Step 5.1: After the decoder completes processing, it outputs the reconstructed seasonal components in the implicit space. Reconstructing components with trend items ; Step 5.2: Using the nonlinear exponential transformation module, the reconstructed features are mapped back to the physical scale of the original ship data.

6. The ship situational awareness data loss interpolation method based on Geely satellite according to claim 5, characterized in that... The specific steps of step 1 are as follows: Step 1.1: The shipborne edge computing node acquires the raw multi-dimensional situational awareness time-series data collected by the ship's sensors in real time; Step 1.2: The ship-end dynamic compression module uses a dynamic dead zone compression strategy based on local volatility to process the original sensing data stream. Step 1.3: The compressed sparse data stream is then sent to the shore via the Geely satellite link.

7. The ship situational awareness data loss interpolation method based on Geely satellite according to claim 6, characterized in that... The specific steps of step 1.2 are as follows: Real-time calculation of the content volume of the sliding window Initial sensing data Standard deviation : in, This represents the mean of the data within that interval. Subsequently, the volatility coefficients of adjacent intervals were calculated. : in, The standard deviation of the data within the current sliding window. This represents the standard deviation of the data within the previous adjacent sliding window. Based on this fluctuation coefficient, the dead zone threshold for the compression process is determined. Perform adaptive updates: in, The dead zone threshold updated at the current moment. This is the dead zone threshold of the previous moment.

8. The ship situational awareness data loss interpolation method based on Geely satellite according to claim 5, characterized in that... In step 3, let the input multivariate time series be... Perform the following logarithmic transformation on its elements: in, The sequence after mapping to additive space, Represents a symbolic function. This indicates element-wise multiplication.

9. The ship situational awareness data loss interpolation method based on Geely satellite according to claim 5, characterized in that... The specific steps of step 5.2 are as follows: set up For the dimensional projection matrix, the components are first merged and reconstructed to obtain: in, This is the initial reconstruction sequence in the latent space; Subsequently, an inverse exponential transform, symmetric to the logarithmic transform, is performed: Final output This refers to the complete ship situational awareness time-series data after eliminating the impact of packet loss and accurately reconstructing the data.

Images