A ground station fault diagnosis system

Abstract

The application discloses a ground TT&C station fault diagnosis system and relates to the technical field of space TT&C.The application aims at the problems that the existing fault diagnosis of the ground TT&C station under the environment of low signal-to-noise ratio, multi-node concurrency and strong interference depends on fixed threshold, is difficult to depict the space-time correlation of faults and is highly sensitive to the quality of spectral data, etc., and the application is characterized in that: a plurality of monitoring branches are arranged before and after a plurality of key devices of a link, and multi-modal data such as spectrum, temperature, voltage and environmental parameters are collected, so that the diagnosis system no longer depends on single spectral information, and other observation quantities can be used to maintain the discrimination of the device state when the signal is seriously attenuated or interrupted for a short time; in cooperation with a data preprocessing module, the missing value compensation and targeted normalization or interval scaling are performed on the data from different sources, so that the characteristics of the same node at each time point are kept continuous and dimensionally unified, and the numerical stability of the training and reasoning process of the graph attention network is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of aerospace telemetry and control technology, and in particular to a ground telemetry and control station fault diagnosis system. Background Technology

[0002] As the core infrastructure for satellite communication and deep space exploration missions, ground control stations undertake functions such as receiving and transmitting telemetry, remote control, and data transmission signals for spacecraft. To ensure the reliability of the telemetry and control link, existing ground control stations are equipped with monitoring branches in the uplink and downlink, and combined with testing equipment such as spectrum analyzers, to monitor the radio frequency signals before and after key node equipment in real time, so as to assist in fault diagnosis and operation and maintenance decisions.

[0003] In actual operation, ground control stations operate in complex electromagnetic environments for extended periods. Conditions such as heavy rain, strong external interference signals, and aging equipment performance can all affect signal transmission, causing a significant decrease in the link signal-to-noise ratio. In particular, when multiple radio frequency devices and baseband devices work together, the status of each node has strong dynamism and coupling, requiring higher real-time and accuracy in fault diagnosis.

[0004] To address diagnostic scenarios in low signal-to-noise ratio environments, some solutions employ intelligent methods based on graph structures or data-driven approaches. For example, by constructing hierarchical directed graphs or multi-signal flow graph models, combined with node merging and effectiveness evaluation functions, the modeling scale of complex measurement and control systems can be simplified, and the fault search space can be reduced by using backtracking and forward reasoning. Other solutions integrate multi-source monitoring data such as spectral amplitude, equipment temperature, and environmental parameters, and use graph neural networks and other models to learn the dependencies between equipment nodes to distinguish between noise interference and real faults, improve the ability to identify concurrent faults, and reduce misjudgments caused by link attenuation.

[0005] However, existing fault diagnosis solutions still have certain limitations. On the one hand, many systems still rely on fixed thresholds or empirical rules for anomaly detection, making it difficult to adapt to the dynamic environment where interference intensity and noise levels change over time. This can easily lead to misjudging transient interference as equipment failure or masking early degradation characteristics of equipment. On the other hand, for telemetry and control links that work collaboratively with multiple nodes, faults may propagate between multiple levels of equipment. Some existing models only consider static topological relationships and cannot fully characterize the correlation between spectrum data and monitoring quantities in time and space, resulting in delays or even misjudgments in fault location. In addition, in scenarios with strong noise and low signal-to-noise ratio, the deterioration of spectrum data quality weakens the ability of traditional diagnostic algorithms to extract features, thereby affecting the continuity and reliability of ground telemetry and control station operation and maintenance. Summary of the Invention

[0006] In view of the aforementioned existing problems, the present invention is proposed.

[0007] This invention provides a ground tracking and control station fault diagnosis system to solve the problem that existing tracking and control stations rely on fixed thresholds for diagnosis in low signal-to-noise ratio and multi-node concurrent scenarios, making it difficult to accurately characterize fault propagation.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0009] This invention provides a ground-based telemetry and control station fault diagnosis system, which includes:

[0010] The monitoring branch is deployed at multiple nodes of the ground telemetry and control station link. Each node includes at least a low-noise amplifier, a down-converter, a power amplifier, and a baseband device. The monitoring branch is configured to collect multimodal data corresponding to the multiple nodes.

[0011] The graph attention network module, connected to the monitoring branch, is configured to: construct a device topology graph based on the device connection relationship of the ground control station link, wherein the nodes of the device topology graph represent ground control station devices, and the edges represent the link connections between devices; perform feature encoding on the nodes and / or edges based on the multimodal data, and dynamically allocate node weights through an attention mechanism to calculate the failure probability of each node;

[0012] The diagnostic module is connected to the graph attention network module and is configured to locate faulty nodes based on the fault probability.

[0013] As a preferred embodiment of the ground telemetry and control station fault diagnosis system of the present invention, the multimodal data includes spectrum data, equipment temperature data, voltage data and environmental parameter data, wherein the environmental parameter data includes humidity data and interference index data.

[0014] As a preferred embodiment of the ground control station fault diagnosis system of the present invention, it further includes a data preprocessing module connected to the monitoring branch, wherein the data preprocessing module is configured to perform normalization processing and missing value compensation on the multimodal data.

[0015] As a preferred embodiment of the ground tracking and control station fault diagnosis system of the present invention, the graph attention network module is further configured as follows:

[0016] The graph attention network is trained based on historical fault case data to optimize the allocation of node weights. The historical fault case data includes fault data acquired in a low signal-to-noise ratio environment and / or low signal-to-noise ratio synthetic data obtained through link simulation.

[0017] As a preferred embodiment of the ground telemetry and control station fault diagnosis system of the present invention, the diagnosis module is further configured as follows:

[0018] The failure probability of each node is compared with the alarm threshold corresponding to the node. When the failure probability exceeds the corresponding alarm threshold, a failure alarm is triggered. The alarm threshold is a dynamic threshold that is adaptively adjusted based on historical error data.

[0019] As a preferred embodiment of the ground tracking and control station fault diagnosis system described in this invention, the graph attention network module is further configured as follows:

[0020] The dependencies between nodes in the device topology graph are modeled by spatiotemporal sequence learning to capture the propagation path of faults in the device topology graph.

[0021] As a preferred embodiment of the ground telemetry and control station fault diagnosis system of the present invention, the graph attention network module operates under a multi-task learning framework and is configured to simultaneously perform fault classification tasks and fault propagation path prediction tasks based on a shared feature extraction layer. The fault classification task outputs the fault status of each node, and the fault propagation path prediction task outputs the propagation sequence or propagation probability of the fault on the equipment topology graph.

[0022] As a preferred embodiment of the ground telemetry and control station fault diagnosis system of the present invention, it further includes an output module connected to the diagnosis module and / or the graph attention network module. The output module is configured to visualize the location result of the fault node and the corresponding fault propagation path on the display interface, and to highlight the fault node and its adjacent nodes on the equipment topology map.

[0023] As a preferred embodiment of the ground telemetry and control station fault diagnosis system of the present invention, the interference index data is calculated based on the spectrum data, and the interference index data at least represents one or more of the following: spectrum energy density, in-band and out-of-band energy ratio and / or illegal signal occupancy duration within the monitoring frequency band.

[0024] As a preferred embodiment of the ground tracking and control station fault diagnosis system of the present invention, the graph attention network module and the diagnosis module are deployed in the server or edge computing node of the ground tracking and control station, implemented by the processor executing program code stored in the memory, and interacting with the monitoring branch through a local area network or serial bus.

[0025] The beneficial effects of this invention are as follows: Addressing the problems of existing fault diagnosis methods for ground-based telemetry and control stations in low signal-to-noise ratio, multi-node concurrency, and strong interference environments—such as reliance on fixed thresholds, difficulty in depicting the spatiotemporal correlation of faults, and high sensitivity to spectral data quality—this invention deploys monitoring branches before and after multiple key devices in the link and collects multimodal data such as spectrum, temperature, voltage, and environmental parameters. This allows the diagnostic system to move beyond relying solely on single spectral information, maintaining device status assessment even with severe signal attenuation or short-term interruptions using other observations. Furthermore, the data preprocessing module performs missing value compensation and targeted normalization or interval scaling on data from different sources, ensuring the continuity and uniformity of features of the same node at various time points. This significantly improves the numerical stability of the graph attention network's training and inference processes, thereby enhancing its sensitivity to minute abnormal changes. At the model level, the topology graph constructed based on the physical connection relationships of devices embeds the real link structure. The graph attention network, based on this, adaptively allocates neighbor weights through spatial attention, and combines time windows and temporal attention to achieve spatiotemporal sequence joint modeling. This enables automatic extraction of fault propagation paths and key influencing nodes from multi-node historical sequences, compared to static topology or single-time-division methods. The system is better suited for handling concurrent and cascading fault scenarios involving multiple devices. By training with historical fault data including low signal-to-noise ratio scenarios and synthetic data generated from link simulation, the model's discrimination boundary under strong noise conditions is specifically strengthened, resulting in better robustness to spectrum quality degradation. Regarding alarm strategies, a dynamic threshold mechanism based on historical error statistics is introduced, enabling the system to automatically tighten or loosen alarm thresholds according to changes in environmental noise, interference levels, and equipment aging, reducing false alarms caused by instantaneous interference and missed alarms caused by long-term slow degradation. Simultaneously, a multi-task learning framework for fault state discrimination and fault propagation path prediction is adopted for unified spatiotemporal feature representation, allowing both tasks to share features meaningful to the physical propagation process. This improves computational efficiency and enhances the consistency and interpretability of fault location results and propagation paths. Combined with a visualization output module, fault nodes and their propagation links are intuitively highlighted on the device topology map, enabling maintenance personnel to more quickly locate the root cause device and the potential affected area. Without significantly increasing hardware overhead, this helps improve the accuracy of fault diagnosis, alarm reliability, and maintenance continuity of ground control stations in complex electromagnetic environments. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0027] Figure 1 This is a schematic diagram of the ground control station fault diagnosis system in the embodiment. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0029] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0030] For example, the terms “first” and “second” used in this application are only used to distinguish and describe similar objects, to differentiate the first object from another object, and are not used to describe a specific order or sequence, nor should they be interpreted as indicating or implying relative importance.

[0031] This application proposes a ground telemetry and control station fault diagnosis system, combined with... Figure 1 As shown, the system includes:

[0032] The monitoring branch is deployed at multiple nodes in the ground telemetry and control station link. Each node includes at least a low-noise amplifier, a down-converter, a power amplifier, and a baseband device. The monitoring branch is configured to collect multimodal data corresponding to multiple nodes.

[0033] The graph attention network module, connected to the monitoring branch, is configured to: construct a device topology graph based on the device connection relationship of the ground telemetry and control station links, where the nodes of the device topology graph represent ground telemetry and control station devices, and the edges represent the link connections between devices; perform feature encoding on nodes and / or edges based on multimodal data, and dynamically allocate node weights through an attention mechanism to calculate the failure probability of each node;

[0034] The diagnostic module, connected to the graph attention network module, is configured to locate faulty nodes based on fault probability.

[0035] In one embodiment, the multimodal data includes spectrum data, device temperature data, voltage data, and environmental parameter data, wherein the environmental parameter data includes humidity data and interference index data.

[0036] In one embodiment, a data preprocessing module is also included, connected to the monitoring branch, the data preprocessing module being configured to perform normalization processing and missing value compensation on the multimodal data;

[0037] The normalization and missing value compensation are performed as follows:

[0038] Step a: When the monitoring branch writes the spectrum data, equipment temperature data, voltage data and environmental parameter data of each node into the cache sequence, the data preprocessing module traverses each time series, marks invalid sampled values ​​and breakpoints on the time axis, forms a set of missing locations, and distinguishes them into isolated missing points and continuous missing intervals, providing a location basis for subsequent interpolation and statistical compensation.

[0039] In this embodiment, invalid sampled values ​​can be understood as observations whose output from the monitoring branch exceeds the device's range, fails to resolve, or significantly deviates from the physical reasonable range of such sensing quantities. Breakpoints on the time axis refer to situations where the difference between two adjacent sampling timestamps is significantly greater than the sampling period. After identifying the above invalid samples or breakpoints, the data preprocessing module uniformly includes the corresponding time positions into the missing position set, and distinguishes between isolated missing points and continuous missing intervals based on whether there are valid observations between adjacent missing positions. Cases where there are valid observations before and after and the missing value only appears in a single or very few time indices are considered isolated missing points, while cases where multiple adjacent time indices are continuously missing or appear at the beginning or end of the sequence and cannot be sandwiched by observations before and after are considered continuous missing intervals. This division method facilitates subsequent processing using interpolation compensation and statistical compensation methods respectively, taking into account both time continuity and statistical stability from an engineering perspective.

[0040] Step b: For isolated missing points where valid observations exist both before and after, the data preprocessing module uses linear interpolation to compensate for the current time point based on the observations from the previous and next valid time points on the time axis. The calculation method is as follows:

[0041] ,

[0042] in, Indicates the node index Modal feature index Time Index The observed values ​​after linear interpolation compensation, Indicates the index of the same node Same modality feature index Index of the previous valid time The original observations at that location, Indicates the index of the same node Same modality feature index Index at the next valid time The original observations at that location, This indicates the time index that currently needs compensation. This represents the time index of the most recent valid observation before the current time. This represents the time index of the most recent valid observation after the current time. This indicates the device node index in the ground control station link. Represents a multimodal feature index;

[0043] This interpolation step maintains the continuity of the device observation sequence on the timeline, avoiding the introduction of abrupt changes due to isolated missing data.

[0044] Specifically, before performing linear interpolation, the sampling period can be set according to the monitoring requirements of the ground telemetry and control station for the link status and the sampling rate of the monitoring branch hardware. It is usually selected in the range of milliseconds to seconds to ensure that the interpolated observation sequence can cover the changes in the state before and after the fault without excessively increasing the amount of computation. After obtaining the interpolation result, the data preprocessing module can refer to the engineering upper and lower bounds of the modal feature to check the range of the compensation value. When the compensation value exceeds the physical allowable range, it is truncated within the upper and lower bounds to avoid interpolation amplification errors caused by abnormal neighborhood values. For features that change slowly and are not sensitive to instantaneous fluctuations, in this embodiment, isolated missing points can be simply replaced by the effective observation value of the previous or next moment as a degenerate implementation. Its input and output form is consistent with linear interpolation, which can reduce the computational overhead in deployment scenarios with limited resources or higher real-time requirements.

[0045] Step c: For consecutive missing intervals appearing at the start or end of the sequence, or exceeding a preset threshold in length, due to a lack of sufficient temporal proximity information, the data preprocessing module compensates based on observation statistics from similar devices at the same time; this compensation is written as:

[0046] ,

[0047] in, Indicates the node index Modal feature index Time Index The observed values ​​are after mean compensation by similar equipment. Indicates time index First Nodes of the same type in modal feature index The original observations below, Indicating in modal feature index The number of nodes of the same type participating in the mean calculation. Indicates the node index in a set of nodes of the same type;

[0048] In actual deployment, nodes of the same type can be filtered according to device model, operating frequency band or functional position in the topology diagram to reduce the impact of differences in device characteristics on the compensation results.

[0049] Furthermore, nodes of the same type can be divided into several groups based on metadata such as device type and operating frequency band. Nodes within each group are comparable in terms of hardware structure and operating conditions. In actual configuration, a member set can be maintained for each type of node, and only nodes that are currently in operation and have valid observations are selected from this set to participate in mean compensation, so as to ensure that the statistical results reflect the typical level of healthy devices. When the number of nodes of the same type under a certain time index is too small or there is an extreme case where all nodes of this type are missing, the data preprocessing module can degenerate to use statistical results over a wider range, such as using the mean of all available nodes in the same link or the same node in historical time for compensation, prioritizing that the compensation results are within the reasonable range of engineering, thereby improving the compensability of continuous missing intervals without introducing other new modules.

[0050] Step d: After missing value compensation, for features with approximately unimodal distributions such as temperature, voltage, and some spectral statistics, the data preprocessing module uses a standardization method based on historical statistics to centralize these features numerically to near zero mean and unit magnitude, facilitating stable gradient maintenance for the graph attention network during training and inference. This process is described as follows:

[0051] ,

[0052] in, Indicates the node index Modal feature index Time Index Standardized observations This represents the observations obtained after missing value compensation. Represents the modal feature index in the training dataset. Historical average, Represents the modal feature index in the training dataset. Historical standard deviation This represents a stable term that prevents numerical oscillations when the denominator is zero and suppresses extreme variances; its value ranges from... to between;

[0053] and It can be obtained through historical sample statistics during the offline stage and updated at fixed time intervals, so that the normalization parameters slowly follow environmental changes;

[0054] For example, historical sample statistics can be calculated based on multimodal data collected in a recent period using a sliding time window. The window length can be set to several minutes to several hours, depending on the measurement and control task cycle and the rate of environmental change. When updating at fixed time intervals, the mean and standard deviation can be updated in batches when the system load is low to reduce the impact on online inference. For individual feature dimensions, when the statistically obtained standard deviation is close to zero or the number of effective samples is significantly insufficient, the data preprocessing module can treat the feature as an approximate constant and directly use the current mean as the normalized output, or set a minimum standard deviation lower limit for the dimension to avoid numerical amplification caused by dividing by a very small number during the normalization process. These processing methods do not change the feature dimensions received by the subsequent graph attention network, but only improve the numerical stability.

[0055] Step e: For features with physical upper and lower bounds or engineering-specified ranges, such as humidity, energy ratio within and outside the frequency band, and interference index, the data preprocessing module uses interval scaling to map these features to a unified interval, reducing scale differences between different units; this processing is written as:

[0056] ,

[0057] in, Indicates the node index Modal feature index Time Index The observed values ​​after interval scaling, This represents the observations after missing value compensation. Represents the modal feature index in the training dataset. The historical minimum or preset lower bound, Represents the modal feature index in the training dataset. The historical maximum value or preset upper limit;

[0058] In implementation, a normalization or range scaling strategy can be bound to different modalities through configuration files. A more suitable scaling range can be used for features with obvious skewness, so as to avoid the data being over-compressed to a narrow numerical range.

[0059] Similarly, the upper and lower bounds used for interval scaling can preferentially adopt the empirical minimum and maximum values ​​from historical data, or they can be slightly relaxed within the physical allowable range according to engineering safety margins to avoid occasional extreme values ​​hitting the boundary for a long time. When calculating the scaling results, the data preprocessing module can internally limit the current observation value to not be less than the predetermined lower bound and not greater than the predetermined upper bound. For observation values ​​exceeding this range, the upper and lower bounds are used as substitutes to ensure that the scaled values ​​always fall within the target interval. For some interference exponents or energy ratios with obvious long-tail distributions, this embodiment can also first perform a gentle truncation on the original dimensions before performing interval scaling to reduce the impact of extreme noise samples on the overall numerical range and subsequent feature learning, while still maintaining the output as a dimensionless value within the same unified interval.

[0060] Step f: After the above compensation and normalization are completed, the data preprocessing module splices the normalization results of each modality under the same node and the same time index in a fixed order to generate node-level feature vectors and writes them into the input buffer of the graph attention network module. In this way, each node in the device topology graph corresponds to a multi-dimensional feature that has completed missing compensation and normalization processing at each time index, which is used for subsequent node feature encoding and fault probability calculation.

[0061] Specifically, this paper constructs a complete data cleaning chain from three dimensions: time axis, inter-node statistics, and feature scale. For missing values, it distinguishes between isolated missing values ​​and long-term missing values, and uses interpolation of nearby time points and compensation by the mean of similar devices to restore the monitoring sequence to a continuous form. At the same time, it uses the statistical characteristics of observations of similar nodes to suppress the bias of data distribution caused by severe anomalies of individual devices. The normalization part classifies and processes statistical features of different modalities. One category is based on unimodal distribution quantities such as temperature and voltage, and its variation magnitude is controlled by standardization to reduce numerical instability during training. The other category is based on quantities with physical boundaries such as humidity and interference index, and the mapping range is unified by interval scaling to reduce the scale difference between different units. Finally, multimodal features are uniformly expressed in terms of missing modes and numerical scales, which is beneficial for graph attention networks to use a unified threshold to judge the degree of fault in low signal-to-noise ratio scenarios and maintain relatively stable output performance in fault diagnosis and fault propagation path modeling tasks.

[0062] Optionally, after the splicing of features of each modality is completed, the feature dimension of each node on a single time index can be automatically determined according to the actual number of sensor channels connected and the number of statistics calculated for each channel. Under typical configuration, the feature dimension of each node is in the range of tens to hundreds of dimensions, which can cover information such as spectrum, temperature, voltage and environmental parameters, without putting unbearable pressure on the storage and computing resources of the subsequent graph attention network. When a modality is completely missing on a certain time index due to equipment maintenance or temporary shutdown, the data preprocessing module can fill the corresponding dimension according to the default engineering safety value or the stable value in the recent period of the modality, and mark it internally for use in subsequent analysis, thereby ensuring that the feature vector structure is fixed and the length is consistent, which makes it easy to keep the model structure unchanged in the training set and inference stages.

[0063] In one embodiment, the graph attention network module is further configured as follows:

[0064] The graph attention network is trained based on historical fault case data to optimize the allocation of node weights. The historical fault case data includes fault data acquired in a low signal-to-noise ratio environment and / or low signal-to-noise ratio synthetic data obtained through link simulation.

[0065] In one embodiment, the diagnostic module is further configured as follows:

[0066] The failure probability of each node is compared with the alarm threshold corresponding to the node. When the failure probability exceeds the corresponding alarm threshold, a failure alarm is triggered. The alarm threshold is a dynamic threshold that is adaptively adjusted based on historical error data.

[0067] In one embodiment, the graph attention network module is further configured as follows:

[0068] The dependencies between nodes in the device topology graph are modeled by spatiotemporal sequence learning, which is used to capture the propagation path of faults in the device topology graph;

[0069] The steps for learning and modeling the dependencies between nodes in a device topology graph using spatiotemporal sequence learning are as follows:

[0070] Step 1: After the data preprocessing module outputs the multimodal normalized features, record the nodes. In time index The feature vector is The graph attention network module constructs the temporal input using a time window approach; for each current time index... (satisfy ), select a length of The set of physical sampling times corresponding to the sliding window is denoted as:

[0071] ,

[0072] in, Indicated by time index This is the physical time set of each sampling point within the timing window at the end. Indicates the current time index. Indicates the length of the timing window (in terms of the number of sampling points), subscript The tag that represents a window. This represents the time interval between two adjacent sampling points;

[0073] In implementation, window Inner and Node The corresponding input sequence is This is used for subsequent spatiotemporal joint modeling;

[0074] In this embodiment, the length of the time window can be configured according to the typical time scale from the occurrence of the fault to its significant impact on the link indicators. When the fault evolves rapidly, a shorter window can be selected to improve the temporal resolution. When the fault evolves slowly, the window length can be appropriately increased to cover the complete change process. The sampling time interval is consistent with the sampling period of the monitoring branch, thereby ensuring that the time index within the window corresponds one-to-one with the physical time. For cases where the time index at the beginning of the sequence is less than the window length, the graph attention network module can construct a complete window by repeating the feature vector of the earliest moment or filling it with statistically obtained steady-state features at the beginning of the window, so that each time index corresponds to a fixed-length input sequence, avoiding changes in the network structure between different time indices due to sequence shortage.

[0075] Step 2, indexing at each time point Above, based on the link connection relationship in the topology diagram of the ground telemetry and control station equipment, nodes are... Building a Neighbor Set And perform graph attention updates; spatial attention scores are written as:

[0076] ,

[0077] in, Indicates time index The upper neighbor node Point to target node Spatial attention midpoint score, Trainable vectors representing spatial attention mechanisms. Top bid This represents the vector transpose operation. This represents a trainable weight matrix that performs a linear transformation on the node features. Indicates time index upper node The input feature vector, Indicates time index Upper neighbor node The input feature vector, symbol This represents a vector concatenation operation. This represents the index of the target node in the device topology graph. Represents nodes Connected neighbor node index, Indicates a time index;

[0078] To obtain the intermediate score Then, for the nodes Normalization is performed on all neighbors to obtain spatial attention weights:

[0079] ,

[0080] in, Indicates time index Upper neighbor node For nodes Spatial attention normalization weights, This represents an activation function that performs a non-linear transformation on the intermediate scores of attention (e.g., a linear rectified function with a leakage coefficient). Indicates a node The neighbor node index when traversing all neighbors. This indicates the relationship between nodes in the device topology diagram. The set of neighbors consisting of all nodes that have a link connection;

[0081] node In time index The spatial aggregation representation on can be written as:

[0082] ,

[0083] in, Indicates time index Nodes after merging neighbor information The spatial representation vector, This represents a non-linear activation function (e.g., an exponential linear unit) that acts on each dimension of the vector.

[0084] Through the aforementioned spatial attention mechanism, the network adaptively allocates neighbor weights based on link connectivity at each time segment, and transmits the features of fault-related nodes to the spatial representation of the target node in the topology graph according to their weights. middle;

[0085] Specifically, the device connection relationships used to construct the neighbor set can be directly derived from the existing link configuration data of the ground telemetry and control station. When generating the topology graph, only device pairs with actual signal transmission paths in the engineering are written as adjacent nodes into the neighbor set to avoid introducing unnecessary fictitious edges. In typical site configurations, the number of neighbors of a single node is usually one to several. The complexity of spatial attention calculation increases linearly with the number of neighbors, which is convenient for deployment on actual servers or edge nodes. When a node has only a single neighbor or no effective neighbors in the topology graph, the graph attention network module can degenerate into applying a linear transformation only to the node's own characteristics or treating it as a self-loop neighbor for aggregation. While keeping the input and output interfaces unchanged, it still provides a spatial representation, thereby preventing the node from being unable to participate in subsequent spatiotemporal modeling due to temporary link adjustments or the offline status of some devices.

[0086] Step 3, in obtaining the window Spatial representation of each time index within Subsequently, the graph attention network module uses a self-attention approach to model the temporal dependencies within the same node in the time dimension; this can be achieved by first performing a linear mapping on each time point in the sequence:

[0087] , , ,

[0088] in, Indicates time index node Time query vector, Indicates time index node Time key vector, Indicates time index node The time value vector, This represents the trainable weight matrix that maps the spatial representation to the temporal query vector. This represents a trainable weight matrix that maps the spatial representation to a time key vector. This represents a trainable weight matrix that maps a spatial representation to a time-valued vector. Indicates time index node Spatial aggregation representation, Indicates time index node Spatial aggregation representation, It means falling in the window Historical time index within;

[0089] Based on the above mapping, node In time index The time attention weights can be written as:

[0090] ,

[0091] in, Indicates indexing by time Historical time index within the window when predicting the target. For nodes Time attention normalization weights This represents the inner product similarity between the query vector and the key vector. The feature dimension representing the temporal attention space is used for scaling the inner product result. This represents the time index when iterating through all historical time indices within the window; the exponential function is used for this purpose. Used to map similarity to non-negative weights;

[0092] After obtaining the temporal attention weights, the node In time index The temporal aggregation is represented as:

[0093] ,

[0094] in, This indicates the node after comprehensively considering the influence of various historical moments within the window. In time index The spatiotemporal joint representation vector, Indicates time index Time Index Attention weights Indicates time index The time value vector at that location;

[0095] Through this time attention step, the network can automatically mine the historical moments that have the greatest impact on the current state between different time indices, such as key changes within a few sampling periods before and after the failure, thereby forming a model of the failure evolution process in the time dimension.

[0096] Furthermore, the linear mapping matrix used to generate the time query vector, time key vector, and time value vector can be configured as a square matrix or a matrix with appropriate dimensionality reduction according to the dimension of the node space representation. Its output dimension is the feature dimension of the time attention space. In engineering, this dimension is usually set to between tens and hundreds to achieve a balance between expressive power and computational cost. In order to avoid the time attention weights being too concentrated or too even over long sequences, the graph attention network module can constrain the time attention distribution by adjusting the feature dimension and the specific form of the nonlinear activation function during the training phase. This makes it easier for the network to give significantly higher weights to key sampling moments before and after the fault than to other moments, thereby enhancing the ability to capture key signals of fault evolution.

[0097] Step 4, the graph attention network module will connect the nodes In time index spatiotemporal joint representation The subsequent fault probability estimation header and fault propagation path prediction header are fed into the multi-task learning framework, which outputs the node fault state and its propagation sequence or propagation probability on the device topology graph; because Simultaneously, by integrating the spatial dependencies between topological neighbors and the evolutionary characteristics within the time window, the hop-by-hop propagation process of faults on the link will be reflected as a joint pattern of spatial attention weights and temporal attention weights, thereby achieving explicit modeling of fault propagation paths.

[0098] Specifically, the graph attention network module is divided into three parts: windowed temporal construction, spatial graph attention, and temporal attention. The temporal window segments continuous observations using length parameters and sampling intervals, so that each time index is associated with a fixed-length historical segment, facilitating the observation of the transition from a normal state to an abnormal state in subsequent steps. The spatial graph attention part utilizes the link structure in the device topology graph to adaptively allocate the weights of each neighboring node in feature aggregation, so that the propagation of the fault signal on the topology is encoded into the intermediate representation vector. The temporal attention part establishes a weighted summation relationship on the historical sequence of the same node, highlighting the key moments that have a significant impact on the current state and weakening the interference of long-term stable intervals. By linking the attention mechanisms of the spatial and temporal dimensions, the final spatiotemporal representation of the node records both the propagation path on the link and the evolution trajectory in time, providing a feature basis with practical physical meaning for subsequent fault probability calculation and fault propagation path prediction.

[0099] Optionally, when training the above spatiotemporal representation using multiple tasks, the node fault markers given in the historical operation and maintenance records and the typical propagation paths obtained from the fault investigation process can be used as supervision signals to construct target outputs for fault classification and propagation path prediction. In actual data, when only the fault status is provided at certain times and complete propagation path information is lacking, the loss can be calculated only for the fault classification branch, and the loss of the samples corresponding to the propagation path prediction branch can be set to zero. This allows for continuous optimization of the shared feature extraction layer without discarding samples with partial labels. When the training set is completely lacking in propagation path labels, the multi-task structure can naturally degenerate into a single-task network containing only the fault classification branch. In this case, the spatiotemporal representation is still used to calculate the fault probability of each node, and the calculation module of the propagation path prediction branch can remain idle without affecting the normal operation of the model in fault diagnosis.

[0100] In one embodiment, the graph attention network module operates under a multi-task learning framework and is configured to simultaneously perform a fault classification task and a fault propagation path prediction task on the basis of a shared feature extraction layer. The fault classification task outputs the fault state of each node, and the fault propagation path prediction task outputs the propagation sequence or propagation probability of the fault on the device topology graph.

[0101] In one embodiment, it further includes an output module connected to the diagnostic module and / or the graph attention network module. The output module is configured to visualize the location results of the fault node and the corresponding fault propagation path on the display interface, and to highlight the fault node and its adjacent nodes on the device topology map.

[0102] In one embodiment, the interference index data is calculated based on the spectrum data, and the interference index data at least characterizes one or more of the following: spectral energy density, in-band and out-of-band energy ratio, and / or duration of illegal signal occupation within the monitoring frequency band.

[0103] In one embodiment, the graph attention network module and the diagnostic module are deployed in the server or edge computing node of the ground telemetry and control station. They are implemented by the processor executing program code stored in memory and interact with the monitoring branch through a local area network or serial bus.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0105] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A ground-based telemetry and control station fault diagnosis system, characterized in that, include: The monitoring branch is deployed at multiple nodes of the ground telemetry and control station link. Each node includes at least a low-noise amplifier, a down-converter, a power amplifier, and a baseband device. The monitoring branch is configured to collect multimodal data corresponding to the multiple nodes. The graph attention network module, connected to the monitoring branch, is configured to: construct a device topology graph based on the device connection relationship of the ground control station link, wherein the nodes of the device topology graph represent ground control station devices, and the edges represent the link connections between devices; perform feature encoding on the nodes and / or edges based on the multimodal data, and dynamically allocate node weights through an attention mechanism to calculate the failure probability of each node; The diagnostic module is connected to the graph attention network module and is configured to locate faulty nodes based on the fault probability. The graph attention network module is also configured to: The dependencies between nodes in the device topology graph are modeled by spatiotemporal sequence learning to capture the propagation path of faults in the device topology graph; The graph attention network module operates within a multi-task learning framework and is configured to simultaneously perform fault classification and fault propagation path prediction tasks based on a shared feature extraction layer. The fault classification task outputs the fault status of each node, and the fault propagation path prediction task outputs the propagation sequence or probability of the fault on the device topology graph.

2. The ground telemetry and control station fault diagnosis system as described in claim 1, characterized in that, The multimodal data includes spectrum data, device temperature data, voltage data, and environmental parameter data, wherein the environmental parameter data includes humidity data and interference index data.

3. The ground telemetry and control station fault diagnosis system as described in claim 2, characterized in that, It also includes a data preprocessing module connected to the monitoring branch, the data preprocessing module being configured to perform normalization processing and missing value compensation on the multimodal data.

4. The ground telemetry and control station fault diagnosis system as described in claim 3, characterized in that, The graph attention network module is further configured as follows: The graph attention network is trained based on historical fault case data to optimize the allocation of node weights. The historical fault case data includes fault data acquired in a low signal-to-noise ratio environment and / or low signal-to-noise ratio synthetic data obtained through link simulation.

5. A ground-based telemetry and control station fault diagnosis system as described in claim 4, characterized in that, The diagnostic module is further configured as follows: The failure probability of each node is compared with the alarm threshold corresponding to the node. When the failure probability exceeds the corresponding alarm threshold, a failure alarm is triggered. The alarm threshold is a dynamic threshold that is adaptively adjusted based on historical error data.

6. A ground-based telemetry and control station fault diagnosis system as described in claim 5, characterized in that, It also includes an output module connected to the diagnostic module and / or the graph attention network module. The output module is configured to visualize the location result of the fault node and the corresponding fault propagation path on the display interface, and to highlight the fault node and its adjacent nodes on the device topology map.

7. A ground-based telemetry and control station fault diagnosis system as described in claim 2, characterized in that, The interference index data is calculated based on the spectrum data, and the interference index data at least represents one or more of the following: spectral energy density within the monitoring frequency band, in-band and out-of-band energy ratio, and / or duration of illegal signal occupation.

8. A ground-based telemetry and control station fault diagnosis system as described in any one of claims 1 to 7, characterized in that, The graph attention network module and the diagnostic module are deployed in the server or edge computing node of the ground telemetry and control station. They are implemented by the processor executing the program code stored in the memory and interact with the monitoring branch through the local area network or serial bus.

Images