Laser ranging data processing method based on effective echo signal time sequence characteristics

Abstract

The invention discloses a laser ranging data processing method based on effective echo signal time sequence characteristics. The method comprises the following steps: firstly, preprocessing an original photon flight time sequence, calculating a residual error and a local space-time density feature, and constructing a feature vector; then, a mixed deep learning model composed of a one-dimensional convolutional neural network and a recurrent neural network is constructed and trained to learn waveform and sequence features of effective echoes; probability prediction and preliminary screening are carried out on data points by using the trained model; and finally, performing refined constraint on the candidate point set based on a satellite orbit dynamics principle, and eliminating outliers. According to the invention, through the fusion of deep learning and a physical model, the robustness, the automation degree and the precision of signal screening in complex scenes of strong noise, high dynamic, weak signals and the like are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of laser ranging technology, and in particular to a laser ranging data processing method based on the time series characteristics of effective echo signals. Background Technology

[0002] Laser ranging: Laser ranging is primarily based on the time-of-flight principle. A common method is pulsed laser ranging, which indirectly determines the distance by directly measuring the time interval between the transmitted and received beams. Depending on the target being observed, laser ranging systems are categorized into satellite laser ranging (SLR), space debris laser ranging (DLR), and lunar laser ranging (LLR), among others.

[0003] Satellite Laser Ranging (SLR) is a comprehensive technology integrating photoelectric detection, automatic control, electronic communication, and satellite orbit analysis. By recording the round-trip time of a laser pulse between the station and the satellite, SLR can accurately measure the distance between the Earth and the satellite. It plays an irreplaceable role in many fields, including determining the Earth's gravity field, measuring Earth's rotation parameters, monitoring crustal movement, precise satellite orbit determination, and establishing space geodetic benchmarks. To advance geodesy and geophysics research, SLR systems have evolved from the first to the third generation, undergoing a transformation from bulky, power-consuming, inaccurate, and highly manual-dependent systems to miniaturized, high-precision, and automated systems. With the continuous advancement of deep space exploration technology, the Global Geodetic Observation System (GGOS) has long demanded that SLR data have a standard point accuracy better than 1 mm, setting new and more stringent goals for technological development.

[0004] The emergence of high repetition rate (HRPR) SLR technology has made it possible to achieve millimeter-level ranging accuracy, laying the foundation for improved data quality by increasing the amount of echo data. However, the explosive growth in data volume has also placed a huge burden on data processing and storage units. More seriously, existing data processing methods have fundamental limitations in their methodology. Whether it is the standard point algorithm framework or various filtering techniques, their essence relies on statistical counting or distribution assumptions of echo photons, identifying effective signals by setting thresholds or based on the mean square error of the Gaussian distribution. The fundamental flaw of this method is that it simplifies a physical echo process with an inherent time-domain structure into a purely mathematical statistical problem. It focuses excessively on the "intensity" or "density" of the echo, while systematically ignoring the "time-domain distribution" characteristics of the SLR echo signal itself as a continuous event. This time-domain distribution is determined by the physical characteristics of the laser pulse, the detector response function, and atmospheric transmission effects, giving the effective echo a coherent and predictable pattern. Existing algorithms treat continuous waveforms containing rich physical information as discrete point sets for statistical analysis, thus losing key features that can effectively distinguish real signals from complex noise. Especially in low signal-to-noise ratio environments, the statistical characteristics of a weak but effective signal with a specific shape may be completely masked by background noise. Relying solely on statistical density can easily lead to misclassification as noise, resulting in a sharp decline in the robustness of the algorithm.

[0005] Furthermore, the inherent batch processing mode of existing workflows requires the collection of complete arc segment data before trend fitting, iterative filtering, and data compression, resulting in "long processing times and slow operation," which is completely unable to meet the real-time data processing requirements of high-repetition-rate SLR systems. This shallow mining of statistical features and reliance on human experience means that noise filtering, threshold setting, and data cropping all require manual intervention. This not only increases operational complexity but also introduces human error, reducing the accuracy and reliability of data processing, making it particularly passive when handling complex observation tasks.

[0006] In summary, existing SLR data processing technologies suffer from significant bottlenecks when dealing with high repetition rate systems, particularly in terms of robustness under low signal-to-noise ratios, real-time processing capabilities for massive amounts of data, and the degree of automation in the processing workflow. These limitations make it difficult to meet the ever-increasing demands for high precision, high real-time performance, and high reliability in data processing. Therefore, this invention proposes a laser ranging data processing method based on the time-series characteristics of effective echo signals. This method aims to overcome the shortcomings of existing technologies by mining and utilizing the temporal distribution characteristics of the echo signals to achieve automatic real-time data processing. This method is expected to significantly improve the accuracy, reliability, and real-time performance of SLR data processing, not only meeting system efficiency requirements but also unlocking cutting-edge applications requiring rapid response, such as real-time orbit determination and space debris collision early warning, thus shifting SLR's service capabilities from "post-event analysis" to "real-time service."

[0007] Existing technical solution one: The Normal point algorithm framework is the foundation of ILRS standardized data output. Currently, all SLR data uploaded to ILRS must be generated using full-rate data (frd) and normal point data (npt) according to the Normal point algorithm rules. Current data processing algorithms for SLR data must meet the requirements of the Normal point algorithm framework. As a standard data processing algorithm for SLR systems, its core lies in statistically counting echo photons and, combined with preset threshold conditions, filtering out what are considered valid signal points from a large amount of echo data, and then determining the distance between the satellite and the station through certain mathematical calculations. The algorithm framework mainly includes the following key steps: First, the predicted value is subtracted from the measured value to obtain the predicted residual. Outliers in the predicted residual are then removed using data identification and filtering methods. Second, a trend function is generated by polynomial fitting of the predicted residual. The fitted residual is obtained by subtracting the trend function from the predicted residual. Outliers outside 2.5 or 3 times the standard error of the fitted residual are removed using an iterative median filtering method. This process is repeated until the data accuracy no longer changes significantly. Finally, standard point intervals are divided according to the ILRS standard. Standard point data is generated within each interval using data compression technology. The standard point algorithm flow is as follows: Figure 1 As shown.

[0008] In 2012, to facilitate rapid conversion of observation targets by SLR and improve observation efficiency, ILRS redefined the time interval for standard point generation, as shown in Table 1. This modification allows high-repetition-rate SLR systems to generate standard points within time periods shorter than the standard point interval, facilitating rapid target switching by satellites.

[0009] Table 1. Standard point intervals corresponding to different track heights:

[0010] Target satellite orbital altitude Standard point interval length Representative satellite <500km 5s GRACE 500-800km 15s Sentinel-3 800km-2000km 30s Starlette, Stella 2000km-5000km 60s 5000km-15000km 120s LAGEOS 8000km-15000km 180s >15000km 300s Etalon, GNSS Lunar variable

[0011] Disadvantages of existing technology 1:

[0012] Lack of real-time capability and severe processing delays: The standard point algorithm operates in batch processing mode, requiring the acquisition of complete observation arc data before subsequent processing can begin. This delay is amplified when high-repetition-rate SLR systems generate massive data streams, resulting in long processing times and slow operations, failing to meet the rapid response requirements for real-time orbit determination.

[0013] Highly dependent on human intervention and with low automation: In the data preprocessing and key parameter adjustment stages, the algorithm relies on human experience. When the signal-to-noise ratio is low, "manual screen processing" is required to remove noise, which can easily lead to the loss of effective data, introduce subjective and human errors, and reduce the objectivity and consistency of data processing.

[0014] Over-reliance on statistical assumptions leads to poor robustness in low signal-to-noise ratio situations: The algorithm pre-screens the effective signal based on the statistical signal of the echo photons, simplifying the physical echo process into a mathematical statistical problem. When the signal is submerged by noise, the statistical features are covered up, making it difficult to distinguish the real signal, reducing accuracy, and easily causing misjudgment or loss, thus failing to meet the requirements of high-precision ranging.

[0015] The algorithm lacks universality and consistency is difficult to guarantee: the same algorithm produces different results when processing the same satellite data on different SLR systems. Adjustments need to be made to the performance parameters of each station system, which weakens its "standard" authority and affects data comparability and overall stability.

[0016] Existing Technology 2: In 1994, Ricklefs and Shelus et al. proposed the Poisson filtering algorithm based on the Poisson process of detector response for single-photon level LLR and SLR systems. The Poisson filtering algorithm uses a rectangular window to perform slope scanning on SLR data within a short time interval and performs histogram statistics on the scanning results. When the number of echo points counted within a unit time interval exceeds a threshold and conforms to the Poisson process, it is determined that there is valid echo data in that region. This method does not rely on high-precision prediction and trend functions, can quickly extract valid SLR echoes, and does not cause data loss. It is one of the representative high-precision SLR echo data processing methods currently available. Figure 2 The image shows the results of SLR data processing. The upper part shows the histogram method, and the lower part shows the Poisson filtering algorithm.

[0017] The specific steps of the Poisson algorithm are as follows:

[0018] Obtain the raw echo time series data recorded by the SLR system.

[0019] Define a sliding window: Set a time window of fixed width (i.e., a "rectangular window"). The width of this window is typically very short, matching the expected echo duration of a single laser pulse.

[0020] Sliding scan: This time window is slid forward along the entire observation time axis in fixed steps (e.g., a very small unit of time).

[0021] Counting: At each position within the window, count the number of echo photons contained within that time window.

[0022] Construct a histogram: Record the position of each window and its corresponding echo count value to form a statistical histogram with time as the horizontal axis and the number of echo points as the vertical axis.

[0023] Threshold check: Checks whether the echo count value for each time window on the histogram exceeds a preset threshold. This threshold is used to exclude purely random noise.

[0024] Poisson process test: For window regions where the count value exceeds the threshold, further examine whether the distribution of its echo counts conforms to the statistical characteristics of a Poisson process. This is a key step in distinguishing between valid signals and dense noise.

[0025] Marking valid signals: If the data within a time window simultaneously meets the conditions of "count value exceeds threshold" and "conforms to Poisson process", then valid SLR echo data is determined to exist within that time window.

[0026] Result generation: Extract the echo data within the time window that is determined to be a valid signal, and use it as the final filtering result.

[0027] End: The algorithm ends when the sliding window has scanned the entire observation data sequence.

[0028] Disadvantages of existing technology 2:

[0029] Lacking real-time performance and suffering from severe processing delays: Similar to the standard point algorithm, the Poisson filtering algorithm requires a complete sliding scan and histogram statistics of the entire observed arc segment to determine the valid echo. This inherent process makes it impossible for the algorithm to output data in real-time or near real-time, which has a fundamental limitation in modern applications that require rapid response.

[0030] Strong dependence on ideal statistical models: The core assumption of the Poisson algorithm is that the effective echo strictly follows the Poisson distribution. However, in actual SLR systems, factors such as detector effects, multiphoton echoes, or atmospheric turbulence can cause the actual echo distribution to deviate from the ideal distribution. The model assumptions are seriously inconsistent with the actual distribution, and the accuracy of the algorithm is significantly reduced.

[0031] The Poisson algorithm suffers from sensitivity to parameter selection and difficulty in tuning: its performance is highly dependent on the sliding window width and the counting threshold. A window that is too wide or too narrow, or a threshold that is too high or too low, will affect the results. The optimal parameter value needs to be repeatedly adjusted according to the observation conditions. It lacks adaptive ability and affects the degree of automation and robustness.

[0032] Insufficient robustness to complex noise backgrounds: The algorithm is good at filtering out random uniform background noise, but it may be misjudged as a valid signal due to sudden or structured noise because the noise accumulates in a short time and meets the judgment conditions, resulting in data pollution.

[0033] Existing technology three: The image processing method aims to map the raw SLR residual data into a binary image, and combined with image processing techniques, achieve noise removal and automatic signal extraction by analyzing the density distribution and linear shape features of the data. The specific steps are as follows:

[0034] Data is mapped to a binary image: the residual (O - C) between the original observations and the theoretical values, as well as the data at the main wave moment, are amplified by an appropriate factor and rounded, and then mapped to a binary image matrix. Simultaneously, a sequence matrix is ​​generated to record the original data position corresponding to each pixel, facilitating subsequent signal recovery.

[0035] Region labeling and coarse denoising: Based on the density characteristics of the observed data, connected components are labeled in the binary image, and the area of ​​each connected component (measured in pixels) is calculated. Given that the signal density is higher than the noise density, an area threshold is set to separate large areas (possibly signal) from small areas (preliminarily identified as noise).

[0036] Linearity analysis: Calculate the centroid coordinates of each connected region, and identify signal regions that are approximately linearly distributed by analyzing the consistency of slopes between regions (such as performing slope difference histogram statistics) and eliminating noise regions that deviate from the linearity characteristics.

[0037] Signal region recovery: For regions initially identified as noise, calculate the centroid angle between them and adjacent signal regions. If the angle is close to 0° or 180° (meeting the linearity criterion), the region is reclassified as a signal region.

[0038] Inverse mapping signal recovery: Based on the non-zero pixel positions in the denoised binary image, the original data is located using a sequence matrix, and the corresponding ranging signal points are extracted, thus completing signal extraction. The mapped binary image is shown below. Figure 3 As shown.

[0039] The disadvantages of existing technology three:

[0040] The method is highly dependent on parameter settings: its effectiveness depends heavily on the setting of multiple parameters such as data mapping magnification factor, region separation area threshold, connectivity mode, and linearity judgment threshold, which need to be adjusted according to different ranging data.

[0041] Insufficient versatility and poor adaptability to different data: Parameter settings are closely related to echo rate. High echo rate conventional ranging and low echo rate diffuse reflection ranging require different parameter combinations, resulting in poor universality.

[0042] Parameter selection relies on experience, which affects the degree of automation: the threshold value used in this method is determined by the "range type and experience", and weak signal data with high noise needs to be "carefully adjusted". The method is not fully automated and its application depends on the operator's experience.

[0043] Processing only at a single scale: This method "only extracts signals at one scale", which does not make full use of multi-scale information and has room for improvement in terms of noise filtering effect and universality.

[0044] Information loss may occur during the data mapping process: When the original data is mapped to a binary image, although the magnification factor can reduce the probability, multiple data points may still be mapped to the same pixel, resulting in the loss or confusion of some information.

[0045] Existing Technology 4: Deep learning technology, as an emerging technology in the field of SLR / DLR data processing in recent years, mainly utilizes the Long Short-Term Memory (LSTM) network of recurrent neural networks in deep learning. By capturing and preserving long-term dependencies in time-series data, it processes noisy, missing, or irregular sequence data to achieve noise suppression and signal extraction, such as... Figure 4 As shown, the specific process is as follows:

[0046] 1. Data labeling: Overlap the original observation data with the target data selected by professional observers. The overlapping part is labeled as signal data, and the rest is labeled as noise data.

[0047] 2. Data preprocessing: Divide the data into fixed-length (e.g., 300) sequence data using a sliding window, and perform unit vector normalization to speed up model training and improve robustness.

[0048] 3. Constructing the LSTM network model: A binary classification network model structure based on the LSTM network is adopted. The preprocessed data first enters a unidirectional single-layer LSTM layer with 300 neurons (64-dimensional hidden layer parameters) to capture long-term dependencies, and then the output of the last neuron of the LSTM is fed into the linear layer.

[0049] 4. Loss Function Optimization: Addressing the low signal-to-noise ratio and significant imbalance between positive and negative samples in laser ranging data, the Focal loss function replaces the traditional binary cross-entropy loss function. Attenuation factors and focal length parameters are introduced to reduce the weight of easily classified samples, allowing the model to focus more on weak signal samples that are difficult to classify, thus improving performance under imbalanced conditions.

[0050] 5. Probability Calculation and Output: The output of the linear layer enters the Sigmoid activation layer, and the probability value of each data point belonging to the signal data is obtained through the activation function.

[0051] 6. Threshold-based classification: Based on preset thresholds, the data is classified into noise or signal data to achieve real-time identification of space debris laser ranging signals.

[0052] The disadvantages of existing technology four:

[0053] The model architecture is simplistic and lacks the ability to extract local waveform features. This technique only learns the macroscopic time-series trend of effective echo signals, lacks a local feature extraction module, and struggles to capture microscopic waveform features. When faced with "ghost waves" or complex noise, its ability to finely identify waveforms is limited.

[0054] The feature engineering is simple, but it does not incorporate physical prior local spatiotemporal density features. This technique only normalizes the residual sequence, making it difficult to effectively identify weak echo signals at low signal-to-noise ratios.

[0055] The technique lacks time-series verification based on physical or geometric constraints. It directly uses threshold filtering to determine data after outputting probabilities, lacking physical or geometric constraints. This results in the final ranging data containing points that do not conform to the actual orbital motion, affecting orbit determination accuracy.

[0056] The training time is long and the real-time performance is poor. The algorithm model training takes more than 40 hours, which affects the model update speed.

[0057] The technology suffers from low automation and poor flexibility. Training data construction relies on manual signal selection, making effective data acquisition difficult. It lacks flexibility when facing new scenarios or requiring rapid model updates, and the high cost of data acquisition limits its large-scale application. Summary of the Invention

[0058] The technical solution of this invention to solve the above-mentioned technical problems is to provide a laser ranging data processing method based on the time series characteristics of effective echo signals, comprising the following steps:

[0059] S1. Data preprocessing: Obtain the raw photon flight time series data collected by the laser ranging system, calculate its residual sequence relative to the theoretical prediction value, and calculate the local spatiotemporal density features of each data point based on the residual sequence, and construct a feature vector containing residual and density features as model input.

[0060] S2. Model Construction and Training: Construct a hybrid deep learning model that can simultaneously extract the microscopic waveform features and macroscopic temporal correlation features of the signal. Train the model using a dataset labeled with effective echo labels so that it learns the characteristics and dynamic laws of the effective echo signal in the time domain.

[0061] The hybrid deep learning model is a combination architecture of a one-dimensional convolutional neural network and a recurrent neural network; wherein, the one-dimensional convolutional neural network is used to extract the local waveform structure features of the echo signal, and the recurrent neural network is a long short-term memory network or a gated recurrent unit, used to capture the long-range dependence and temporal dynamics of the echo signal sequence.

[0062] S3. Signal Filtering: Input the preprocessed data to be processed into the trained hybrid deep learning model to obtain the probability prediction value of each data point belonging to the valid echo; perform preliminary filtering based on the probability prediction value to obtain a set of candidate valid echo points;

[0063] S4. Post-processing: Based on the principles of satellite orbit dynamics, construct a geometric or physical constraint model, refine the candidate valid echo point set, remove outliers that do not meet the constraints, and output the final valid echo data.

[0064] Furthermore, in step S1, the local spatiotemporal density feature is obtained by calculating the number of neighboring points of each data point within a preset temporal radius and residual radius neighborhood, which is used to characterize the clustering of the signal in the spatiotemporal domain.

[0065] Furthermore, in step S2, Focal Loss is used as the loss function during model training, as follows:

[0066] ,

[0067] in, This represents the model's predicted probability for the true class. (Parameters) Used to balance the proportional weights of positive and negative samples The focusing parameter is used to reduce the weight of easily classified samples, so that the model can focus on training weak signal samples that are difficult to distinguish.

[0068] Furthermore, in step S3, the preliminary screening uses a low preset probability threshold to obtain a set of candidate valid echo points with high recall.

[0069] Furthermore, in step S4, the refined processing based on the principle of satellite orbital dynamics specifically involves: using a random sampling consensus algorithm to fit the candidate valid echo point set with a multinomial model, and identifying points with fitting residuals greater than a set threshold as outliers and removing them.

[0070] Furthermore, the method is specifically applied to ground target observation scenarios; in step S2, the hybrid deep learning model focuses on learning the waveform fingerprint features of the system impulse response function through a one-dimensional convolutional neural network in order to distinguish between real ground target echoes and interference signals generated by detector saturation.

[0071] Furthermore, the method is specifically applied to low-orbit satellite observation scenarios; in step S2, the hybrid deep learning model focuses on learning the dynamic laws of rapidly changing nonlinear residual curves through recurrent neural networks in order to achieve continuous locking of signals across the entire observation arc.

[0072] Furthermore, the method is specifically applied to medium- and high-orbit satellite observation scenarios; in step S1, the local spatiotemporal density features are given higher weights when constructing feature vectors, and in step S2, the hybrid deep learning model maintains trajectory prediction capability when the signal is intermittently interrupted through the memory mechanism of the recurrent neural network, thereby achieving rapid signal relocking after the interruption.

[0073] The technical solution of this invention fundamentally improves the intelligence level and overall performance of satellite laser ranging echo signal screening by constructing an intelligent processing framework that integrates data-driven and physical prior knowledge. Its beneficial effects are mainly reflected in the following aspects: First, by utilizing the powerful feature learning capabilities of a hybrid deep learning model, it effectively overcomes the failure problems of traditional methods in extreme scenarios such as strong noise and saturation interference (e.g., "ghost waves" in ground target observation), high dynamic nonlinear residuals (e.g., slope flipping when low-orbit satellites pass overhead), and extremely low signal-to-noise ratio and intermittent signal interruptions (e.g., in medium- and high-orbit satellite observation), achieving highly robust signal extraction. Second, through end-to-end point-by-point classification and real-time inference, this method significantly reduces the reliance on manual intervention and empirical parameter tuning, greatly improving the automation and real-time performance of the data processing flow. Finally, by combining a post-processing mechanism based on orbital dynamic constraints, the physical rationality of the output signal is ensured, thereby achieving significant improvements in several key indicators such as system delay calibration accuracy, full-arc data utilization, and continuous weak signal locking capability, providing reliable technical support for achieving high-precision, fully automated, and all-weather satellite laser ranging observation. Attached Figure Description

[0074] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0075] Figure 1 This provides a framework for the standard point algorithm process.

[0076] Figure 2 The results of SLR data processing using different algorithms and the histogram method Poisson filtering algorithm;

[0077] Figure 3 This is the mapped binary image;

[0078] Figure 4 This refers to a deep learning SLR / DLR data processing algorithm based on an LSTM network model.

[0079] Figure 5 This is a diagram of the 1D-CNN-LSTM model architecture used in this invention;

[0080] Figure 6 This is a flowchart of the laser ranging data processing method based on the time series characteristics of effective echo signals according to the present invention.

[0081] Figure 7 This is a residual map of the original observations from the ground-target SLR system.

[0082] Figure 8 A probability distribution map predicted by the deep learning model of the ground-target SLR system;

[0083] Figure 9 Figure 1. Effective echo distribution and Gaussian fitting results after screening of ground-target SLR systems;

[0084] Figure 10 A comparative diagram showing the effects of residual processing during the overhead phase of low-Earth orbit satellites;

[0085] Figure 11 This is a schematic diagram of weak signal interruption recovery based on LSTM memory mechanism. Detailed Implementation

[0086] This invention proposes a laser ranging data processing method based on the time series characteristics of effective echo signals. It aims to significantly improve the robustness, automation, and accuracy of signal selection in complex scenarios such as strong noise, high dynamics, and weak signals by integrating deep learning with physical models.

[0087] The laser ranging data processing method based on the time series characteristics of effective echo signals proposed in this invention will be described below in specific embodiments:

[0088] In the technical solution of this embodiment, such as Figure 6 As shown, a laser ranging data processing method based on the time series characteristics of effective echo signals includes the following steps:

[0089] S1. Data preprocessing: Obtain the raw photon flight time series data collected by the laser ranging system, calculate its residual sequence relative to the theoretical prediction value, and calculate the local spatiotemporal density features of each data point based on the residual sequence, and construct a feature vector containing residual and density features as model input.

[0090] S2. Model Construction and Training: Construct a hybrid deep learning model capable of simultaneously extracting microscopic waveform features and macroscopic temporal correlation features of a signal, such as... Figure 5 As shown, the model is trained using a dataset labeled with valid echoes, enabling it to learn the characteristics and dynamic patterns of valid echo signals in the time domain.

[0091] The hybrid deep learning model is a combination architecture of a one-dimensional convolutional neural network and a recurrent neural network; wherein, the one-dimensional convolutional neural network is used to extract the local waveform structure features of the echo signal, and the recurrent neural network is a long short-term memory network or a gated recurrent unit, used to capture the long-range dependence and temporal dynamics of the echo signal sequence.

[0092] S3. Signal Filtering: Input the preprocessed data to be processed into the trained hybrid deep learning model to obtain the probability prediction value of each data point belonging to the valid echo; perform preliminary filtering based on the probability prediction value to obtain a set of candidate valid echo points;

[0093] S4. Post-processing: Based on the principles of satellite orbit dynamics, construct a geometric or physical constraint model, refine the candidate valid echo point set, remove outliers that do not meet the constraints, and output the final valid echo data.

[0094] Furthermore, in step S1, the local spatiotemporal density feature is obtained by calculating the number of neighboring points of each data point within a preset temporal radius and residual radius neighborhood, which is used to characterize the clustering of the signal in the spatiotemporal domain.

[0095] Furthermore, in step S2, Focal Loss is used as the loss function during model training, as follows:

[0096] ,

[0097] in, This represents the model's predicted probability for the true class. (Parameters) Used to balance the proportional weights of positive and negative samples The focusing parameter is used to reduce the weight of easily classified samples, so that the model can focus on training weak signal samples that are difficult to distinguish.

[0098] Furthermore, in step S3, the preliminary screening uses a low preset probability threshold to obtain a set of candidate valid echo points with high recall.

[0099] Furthermore, in step S4, the refined processing based on the principle of satellite orbital dynamics specifically involves: using a random sampling consensus algorithm to fit the candidate valid echo point set with a multinomial model, and identifying points with fitting residuals greater than a set threshold as outliers and removing them.

[0100] Furthermore, the method is specifically applied to ground target observation scenarios; in step S2, the hybrid deep learning model focuses on learning the waveform fingerprint features of the system impulse response function through a one-dimensional convolutional neural network in order to distinguish between real ground target echoes and interference signals generated by detector saturation.

[0101] Furthermore, the method is specifically applied to low-orbit satellite observation scenarios; in step S2, the hybrid deep learning model focuses on learning the dynamic laws of rapidly changing nonlinear residual curves through recurrent neural networks in order to achieve continuous locking of signals across the entire observation arc.

[0102] Furthermore, the method is specifically applied to medium- and high-orbit satellite observation scenarios; in step S1, the local spatiotemporal density features are given higher weights when constructing feature vectors, and in step S2, the hybrid deep learning model maintains trajectory prediction capability when the signal is intermittently interrupted through the memory mechanism of the recurrent neural network, thereby achieving rapid signal relocking after the interruption.

[0103] Implementation Method 1: Ground Target Observation Scenario (System Delay Calibration under Short-Range Strong Noise);

[0104] This implementation method is specifically applied to ground target calibration tasks in laser ranging systems. Ground targets are typically very close to the measuring station (usually within tens of meters to several kilometers), resulting in extremely strong laser echo signals, often mixed with strong atmospheric backscatter noise and electronic self-excitation noise caused by detector saturation. Traditional thresholding or centroid methods struggle to accurately separate the true main peak from the saturated nonlinear waveform, often requiring frequent manual adjustments of the distance gate and attenuator, severely impacting the automation and accuracy of system time delay calibration.

[0105] In this scenario, the deep learning model of this invention (especially the 1D-CNN layer) plays a dominant role. Instead of relying on a simple intensity threshold, it effectively distinguishes between genuine target echoes and spurious "ghost waves" caused by detector saturation by learning the unique fingerprint characteristics of the system's Impulse Response Function (IRF) (such as rising edge slope, full width at half maximum, and other waveform morphology). Experiments show that this method can accurately extract the main wave peak in strong background noise and saturation interference during ground target observation, enabling precise calculation of the system's one-way transmission and reception delays.

[0106] Software architecture diagrams or modular functional block diagrams, such as Figure 6 As shown;

[0107] Ground target data acquisition module: controls the laser ranging system to observe ground targets at known distances and acquire raw echo time series containing strong noise and saturation effects.

[0108] Benchmarking module: Select some historical ground target data with stable waveforms and moderate signal-to-noise ratio, and use standard point algorithm or waveform decomposition algorithm to perform benchmarking to establish the morphological standard of "true echo".

[0109] Training data construction module: The labeled data is augmented (e.g., simulating different degrees of saturation distortion) and organized into deep learning training samples containing features of [normalized residual, local density].

[0110] Deep learning model module: Constructs a network structure with 1D-CNN as the main component and LSTM as the auxiliary component. CNN is used to identify the unique waveform microstructure of ground target echoes, while LSTM is used to maintain the locking stability of the time series.

[0111] Real-time filtering module: Deploy the trained model to FPGA or edge computing unit to perform point-by-point discrimination of the real-time collected ground target echo photons and output valid signal identifiers.

[0112] Time delay calculation module: Calculates the measurement distance using only the filtered effective echo time, and inverts the two-way loop delay value of the system by combining the known theoretical distance of the ground target.

[0113] Calibration data output module: Generates system calibration reports that conform to ILRS specifications, including system delay values, root mean square error (RMS), and other indicators.

[0114] The specific implementation process is as follows:

[0115] First, the laser ranging system is switched to ground target observation mode to record the raw photon flight time series. The system reads the observation data and calculates the observation residual (OC) using the known ground target distance. For example... Figure 7 As shown, due to the extremely strong ground target echo causing detector saturation, the original residual image typically includes a dense band of real signal, a "ghost wave" band with a fixed time difference (electronic afterpulse), and a large amount of random background noise, in addition to the actual signal band. In this case, a simple density clustering algorithm can easily misclassify the high-density "ghost wave" as a valid signal. The preprocessed data tensor is fed into a trained deep learning model. The 1D-CNN layer in the model acts as the core feature extractor, scanning the local waveform microstructure of each data point. Based on the learned impulse response function (IRF) fingerprint features, the model distinguishes between the steep rising edge of the real echo and the trailing shape of the "ghost wave." Figure 8 As shown, the model outputs a predicted probability score for each data point. The predicted probability of the true echo point is significantly higher than the threshold (e.g., ...). Those that are not marked in dark are marked in light; conversely, deceptive "ghost waves" and background noise are marked in light because their waveform characteristics do not match and their prediction probability is extremely low.

[0116] The system performs real-time coarse screening based on probability thresholds, eliminating low-confidence points and retaining only the set of valid echoes with high probability. Subsequently, the filtered data is used for Gaussian waveform fitting or centroid calculation. For example... Figure 9 As shown, the data after removing "ghost wave" interference exhibits a normal distribution. The system extracts the center time of this distribution as the observation flight time. Combined with the launch time Theoretical distance from ground target According to the formula The final system delay value is calculated, and the calibration accuracy RMS index is output.

[0117] It has the following beneficial effects:

[0118] First, it significantly improves robustness in low signal-to-noise ratio scenarios. The deep learning model directly learns the local shape and temporal correlation of the effective echo, no longer relying on Poisson statistics or density thresholds, thus enabling it to capture real signals in noisy backgrounds and improve the effective echo recognition rate.

[0119] Second, it enables real-time automatic filtering. The model is a point-by-point classification structure, which can directly output valid echo labels during the observation process, avoiding the batch processing delay of the standard point algorithm and meeting the real-time requirements of space target calibration.

[0120] Third, it reduces reliance on human experience. Once the model is trained, its parameters are entirely data-driven, eliminating the need for manual settings of window widths, thresholds, or trend functions, thus making the screening process more stable and consistent.

[0121] Fourth, improve the accuracy of time delay calibration. Because the effective echo extraction is more reliable, the delay estimation is no longer affected by noise points, thereby improving the accuracy of time delay calibration in the transceiver separation system.

[0122] Implementation Method 2: Low Earth Orbit Satellite Observation Scenario (High Dynamic Nonlinear Residual Processing);

[0123] This implementation method is applied to observation missions using low Earth orbit (LEO) satellites (such as Starlette, Swarm, and GRACE). LEO satellites move at extremely high angular velocities, causing the OC (observation-calculation) residual curve to be non-linear, exhibiting significant nonlinear characteristics. Especially during the satellite's overhead transit phase, the residual slope can drastically reverse within a short period (e.g., from negative to positive). Traditional "polynomial fitting + statistical filtering" methods typically assume that the data follows a linear or low-order polynomial distribution over a short time. Therefore, at the overhead transit moment, the fitting model is prone to failure, leading to the removal of valid points and the loss of data in critical segments. The hardware and software architecture of this implementation method is basically the same as that of implementation method 1, the difference being in the construction of the training data and the emphasis of the model:

[0124] 1. Differentiated Data Construction: When constructing the training set, a large number of historical LEO satellite data samples containing nonlinear residual trends (such as “V” or “S” shaped residual curves) were deliberately added, and a wider time window was reserved in the preprocessing stage to accommodate rapidly changing residual trajectories.

[0125] 2. Model Emphasis: Unlike the ground target scenario, which focuses on CNN for waveform fingerprint extraction, this implementation uses LSTM (Long Short-Term Memory) as the dominant technology. Leveraging LSTM's dynamic modeling capabilities for time series, the model no longer attempts to "fit" a line, but instead learns the dynamics of residual changes. Even when the residual slope changes drastically during the overshoot phase, LSTM can accurately predict the "bending" path of subsequent effective echoes based on the preceding trajectory.

[0126] Beneficial effects: The residual processing effect during the overpass phase of low-Earth orbit satellites is comparable to... Figure 10 As shown, fully automated processing across the entire arc segment is achieved, resolving the "overshoot failure" problem. The model learns the dynamic continuity of the satellite orbit rather than its simple geometry, thus effectively capturing rapidly changing residual curves without requiring manual segmentation or intervention on overshoot data. This significantly improves the data utilization rate of low-Earth orbit satellites. It also avoids the erroneous deletion of valid points due to fitting errors, ensuring high-quality data output across the entire observation arc from arrival at orbit (AOS) to departure at orbit (LOS).

[0127] Implementation Method 3: Medium-to-High Orbit Satellite Observation Scenario (Long-Range Weak Signal); This implementation method is applied to observation missions of medium-to-high orbit (MEO / HEO) satellites (such as LAGEOS, Etalon, Galileo). These satellites are extremely far away (up to 20,000 kilometers or more), and the echo signals are at the single-photon level, with extremely low signal-to-noise ratios. Furthermore, the signals are often intermittently interrupted due to cloud cover, atmospheric turbulence, or telescope tracking jitter. Existing Poisson filtering heavily relies on statistical density per unit time. When the signal is interrupted or becomes extremely sparse, the algorithm often fails to generate effective density peaks, leading to missed detections or the inability to quickly re-lock onto the signal after an interruption.

[0128] The core improvement of this implementation lies in the weight adjustment and inference strategy of feature engineering: 1) For sparse signals, this implementation assigns higher weights to local spatio-temporal density in the input features. Even weak signals exhibit a fundamental statistical difference in their spatio-temporal clustering compared to random noise with a Poisson distribution. The model can keenly capture this weak clustering through high-dimensional feature mapping. 2) Utilizing the gating mechanism of LSTM (forget gate and input gate), the model establishes a signal trajectory pipeline spanning a long time. When cloud cover causes signal interruption (e.g., an interruption of 5-10 seconds), LSTM can maintain the predicted state in the interrupted region based on the trajectory trend before the interruption and the orbital dynamics constraints. Once the signal reappears, the model can immediately associate it with the historical trajectory, rather than treating it as noise. A schematic diagram of weak signal interruption recovery using the LSTM memory mechanism is shown below. Figure 11 As shown.

[0129] Beneficial effects: Significantly improved recall under weak signal conditions. Compared to Poisson filtering, which mainly focuses on the density of the "current window," this method utilizes deep learning to mine the non-statistical features of weak signals. Experiments show that in LAGEOS satellite processing, the effective echo extraction rate is improved by approximately 15%. It possesses signal locking capability resistant to interruptions. Thanks to the long-term memory capability of LSTM, the system can make logical connections across signal interruption areas, enabling continuous orbit data output even under intermittent and harsh observation conditions. This is crucial for generating long arc segment normal points.

[0130] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A laser ranging data processing method based on the time series characteristics of effective echo signals, characterized in that, Includes the following steps: S1. Data preprocessing: Obtain the raw photon flight time series data collected by the laser ranging system, calculate its residual sequence relative to the theoretical prediction value, and calculate the local spatiotemporal density features of each data point based on the residual sequence, and construct a feature vector containing residual and density features as model input. S2. Model Construction and Training: Construct a hybrid deep learning model that can simultaneously extract the microscopic waveform features and macroscopic temporal correlation features of the signal. Train the model using a dataset labeled with effective echo labels so that it learns the characteristics and dynamic laws of the effective echo signal in the time domain. The hybrid deep learning model is a combination architecture of a one-dimensional convolutional neural network and a recurrent neural network; wherein, the one-dimensional convolutional neural network is used to extract the local waveform structure features of the echo signal, and the recurrent neural network is a long short-term memory network or a gated recurrent unit, used to capture the long-range dependence and temporal dynamics of the echo signal sequence. S3. Signal filtering: Input the preprocessed data to be processed into the trained hybrid deep learning model to obtain the probability prediction value of each data point belonging to the valid echo. Based on the predicted probability values, a preliminary screening is performed to obtain a set of candidate valid echo points; S4. Post-processing: Based on the principles of satellite orbit dynamics, construct a geometric or physical constraint model, refine the candidate valid echo point set, remove outliers that do not meet the constraints, and output the final valid echo data.

2. The method according to claim 1, characterized in that, In step S1, the local spatiotemporal density feature is obtained by calculating the number of neighboring points of each data point in the preset temporal radius and residual radius neighborhood, which is used to characterize the clustering of the signal in the spatiotemporal domain.

3. The method according to claim 1, characterized in that, In step S2, Focal Loss is used as the loss function during model training, as follows: ,in, This represents the model's predicted probability for the true class. (Parameters) Used to balance the proportional weights of positive and negative samples The focusing parameter is used to reduce the weight of easily classified samples, so that the model can focus on training weak signal samples that are difficult to distinguish.

4. The method according to claim 1, characterized in that, In step S3, the preliminary screening uses a low preset probability threshold to obtain a set of candidate valid echo points with high recall.

5. The method according to claim 1, characterized in that, In step S4, the refined processing based on the principle of satellite orbit dynamics specifically involves: using a random sampling consensus algorithm to fit the candidate valid echo point set with a multinomial model, and identifying points with fitting residuals greater than a set threshold as outliers and removing them.

6. The method according to any one of claims 1 to 5, characterized in that, The method is specifically applied to ground target observation scenarios; in step S2, the hybrid deep learning model focuses on learning the waveform fingerprint features of the system impulse response function through a one-dimensional convolutional neural network in order to distinguish between real ground target echoes and interference signals generated by detector saturation.

7. The method according to any one of claims 1 to 5, characterized in that, The method is specifically applied to low-orbit satellite observation scenarios; in step S2, the hybrid deep learning model focuses on learning the dynamic laws of rapidly changing nonlinear residual curves through recurrent neural networks in order to achieve continuous locking of signals across the entire observation arc.

8. The method according to any one of claims 1 to 5, characterized in that, The method is specifically applied to medium- and high-orbit satellite observation scenarios. In step S1, the local spatiotemporal density features are given higher weights when constructing feature vectors. In step S2, the hybrid deep learning model maintains trajectory prediction capability when the signal is intermittently interrupted through the memory mechanism of the recurrent neural network, thereby achieving rapid signal relocking after the interruption.

Images