High-end equipment state monitoring system based on optical fiber sensing network and fault early warning method thereof

Abstract

The application provides a high-end equipment state monitoring system based on an optical fiber sensing network and a fault early warning method thereof, and belongs to the technical field of equipment state monitoring and fault early warning. The method acquires time sequence physical field signals of multiple monitoring points on the structure of equipment through an optical fiber sensing network, constructs space-time graph data according to the spatial adjacency relationship of the monitoring points and a pre-labeled response transfer rate function; performs spatial feature aggregation and time state evolution on the space-time graph data, and extracts node transient response features of each monitoring point; determines a time sequence prediction influence index according to the change of bidirectional prediction errors between adjacent monitoring points, forms a directed correlation edge and combines it into an abnormal evolution candidate path; when the abnormal cumulative amount of the terminal monitoring point of the abnormal evolution candidate path exceeds a set threshold, early warning information containing a node sequence, a prediction dominant direction and a prediction leading time lag is output. Thus, reviewable equipment abnormal evolution candidate path early warning information is provided.

Description

Technical Field

[0001] This application relates to the field of high-end equipment condition monitoring technology, and more specifically, to a high-end equipment condition monitoring system based on fiber optic sensor networks and a fault early warning method thereof. Background Technology

[0002] When key structural components of high-end equipment operate under alternating loads and complex conditions, changes in their state, such as crack propagation and loosening of connections, may generate temporally related responses at multiple monitoring points. Fiber optic sensing technology can be used to acquire physical field signals from multiple monitoring points on the equipment structure. However, existing monitoring methods mostly focus on single-point threshold alarms or overall system anomaly detection, making it difficult to combine the spatial adjacency relationships of monitoring points to organize discrete abnormal responses into temporally correlated paths that are easy to verify.

[0003] Existing health monitoring methods typically identify equipment anomalies based on the signal amplitude of a single monitoring point or the statistical correlation between multiple monitoring points. However, the structural connectivity and response propagation differences between different monitoring points can affect the manifestation of abnormal responses at each monitoring point. Common operating conditions such as temperature changes and load fluctuations can also cause the signals of adjacent monitoring points to change simultaneously. Relying solely on single-point thresholds or statistical correlations without spatial structural constraints makes it difficult to distinguish between changes in common operating conditions and local anomaly responses with temporal chronology. It is also difficult to organize the abnormal states of multiple discrete monitoring points into easily verifiable path information, thus making it difficult to provide verifiable early warning information on candidate paths for equipment anomaly evolution. Summary of the Invention

[0004] This application provides a high-end equipment condition monitoring system and fault early warning method based on fiber optic sensor network, which can provide verifiable early warning information on candidate paths for abnormal equipment evolution.

[0005] In a first aspect, this application provides a method for early warning of faults in high-end equipment based on fiber optic sensor networks, comprising the following steps: The temporal physical field signals of multiple monitoring points on the equipment structure are acquired through fiber optic sensor networks. A spatial topology graph is constructed based on the spatial adjacency relationship of each monitoring point, and each edge of the spatial topology graph is weighted based on the pre-calibrated response transfer rate function between the monitoring points. The weighted spatial topology graph and the time-series physical field signal together constitute the spatiotemporal graph data. Spatial domain feature aggregation is performed on the spatiotemporal map data, and time domain state evolution processing is performed on the spatial aggregation feature sequence of each monitoring point to extract the fluctuation component that changes with time, forming the node transient response characteristics of each monitoring point. For every two adjacent monitoring points, based on the change in prediction error before and after adding the historical node transient response characteristics of one monitoring point to the prediction model of another monitoring point, the time-series prediction impact indicators in two directions are determined, the prediction advantage directions exceeding the preset impact threshold are selected to form directed correlation edges, and the directed correlation edges that are connected end to end and do not have duplicate nodes are combined into candidate paths for abnormal evolution. When the cumulative abnormality at the end monitoring point of the abnormal evolution candidate path exceeds a set threshold, an early warning message containing the abnormal evolution candidate path is issued.

[0006] Preferably, constructing a spatial topology map based on the spatial adjacency relationships of each monitoring point specifically includes: Obtain the physical spatial coordinates of each monitoring point on the equipment structure; Based on the physical space coordinates, the spatial adjacency relationship between any two monitoring points is determined. When there is a direct structural connection path between two points and there are no other monitoring points on the path, the two points are determined to be adjacent. A spatial topology graph is constructed using each monitoring point as a node and the connections between adjacent monitoring points as edges.

[0007] Preferably, assigning weights to each edge of the spatial topology graph based on the pre-calibrated response transfer rate function between the monitoring points specifically includes: For each edge in the spatial topology graph, a set of adjacent monitoring points are used. One monitoring point is used as the response reference point and the other monitoring point is used as the response observation point. The response signals of the two monitoring points under pulse excitation or step excitation are collected synchronously through the optical fiber sensing network. The auto-power spectral density of the response signal at the response reference point and the cross-power spectral density between the response signals at two monitoring points are determined. Based on the auto-power spectral density and the cross-power spectral density, the response transfer rate function within the characteristic frequency band of the equipment is determined, and the integral mean of the amplitude of the response transfer rate function within the characteristic frequency band of the equipment is used as the response transfer coefficient of the corresponding adjacent monitoring point pair. The response propagation coefficients corresponding to each edge in the spatial topology graph are normalized, and the normalized response propagation coefficients are used as the edge weights of the corresponding edges.

[0008] Preferably, the spatiotemporal graph data, which is composed of the weighted spatial topology graph and the temporal physical field signal, specifically includes: The time-series physical field signal is divided into multiple time frames according to a preset sampling interval, and each time frame contains the physical field signal value of each monitoring point at that moment; The physical field signal value of each time frame is assigned to each node of the spatial topology map according to the monitoring point location index, which serves as the node attribute of that time frame, forming a single-frame attribute map. The spatiotemporal graph data is constructed by arranging the single-frame attribute graphs of multiple consecutive time frames in chronological order, wherein each time frame corresponds to a graph snapshot, and there is a fixed time step between adjacent graph snapshots.

[0009] Preferably, the spatial domain feature aggregation is performed on the spatiotemporal map data, and the time domain state evolution processing is performed on the spatial aggregation feature sequence of each monitoring point to extract the fluctuation component that changes with time, forming the node transient response features of each monitoring point. Specifically, this includes: For each time frame in the spatiotemporal graph data, with each monitoring point as the target node, the physical field signal values ​​of its neighboring nodes are aggregated according to the edge weights of the spatial topology graph to obtain the spatial aggregation characteristics of each monitoring point in that time frame. The spatial aggregation features of each monitoring point in multiple consecutive time frames are arranged in chronological order to obtain the spatial aggregation feature sequence of each monitoring point. The spatial aggregation feature sequence of each monitoring point is smoothed to form a trend component that changes over time. The difference sequence between the spatial aggregation feature sequence and the trend component is extracted and the difference sequence is identified as the fluctuation component. The fluctuation components at each monitoring point are used to form the nodal transient response characteristics of the corresponding monitoring point.

[0010] Preferably, for every two adjacent monitoring points, based on the change in prediction error before and after incorporating the historical transient response characteristics of one monitoring point into the prediction model of another monitoring point, the time-series prediction impact indicators in both directions are determined, specifically including: For each pair of adjacent monitoring points, the transient response feature sequence of one monitoring point is taken as the first sequence, and the transient response feature sequence of the other monitoring point is taken as the second sequence. A first time-series prediction impact index of the first sequence on the second sequence and a second time-series prediction impact index of the second sequence on the first sequence are determined respectively. The direction corresponding to the one that exceeds the preset impact threshold and has the larger value among the first time-series prediction impact index and the second time-series prediction impact index is taken as the prediction advantage direction, and the time-series prediction impact index value corresponding to this direction is taken as the time-series prediction impact index of the adjacent monitoring point pair.

[0011] Preferably, the predicted advantage directions exceeding a preset influence threshold are selected to form directed correlation edges, and the directed correlation edges that are connected end to end and have no duplicate nodes are combined into candidate paths for abnormal evolution. Specifically, this includes: For each pair of adjacent monitoring points with the predicted advantage direction, the starting point of the predicted advantage direction is determined as the prediction leader node, and the ending point of the predicted advantage direction is determined as the prediction response node, forming a directed association edge from the prediction leader node to the prediction response node. The path search begins from the predicted leader node that has no incoming edges and proceeds along the directed associated edges. When the same node has multiple outgoing edges, path branches are formed respectively. When the next node has already appeared in the current path or the current node has no outgoing edges, stop searching the current path branch and determine the obtained continuous directed node sequence without duplicate nodes as the candidate path for abnormal evolution.

[0012] Preferably, when the cumulative anomaly amount corresponding to the monitoring point at the end of the abnormal evolution candidate path exceeds a set threshold, issuing a warning message containing the abnormal evolution candidate path specifically includes: Real-time tracking of the cumulative anomalies at the end monitoring points of each identified candidate path for abnormal evolution; When the cumulative amount of anomalies at any end monitoring point exceeds the set threshold, the complete node sequence of the corresponding candidate path for anomaly evolution, the predicted dominant direction of each directed associated edge, and the predicted lead time lag between adjacent nodes are extracted. The complete node sequence, the predicted advantage direction, the predicted lead time lag, and the early warning trigger signal are encapsulated into early warning information containing corresponding abnormal evolution candidate paths, and the early warning information is issued.

[0013] Preferably, the cumulative anomaly amount corresponding to the monitoring point at the end of the candidate path of anomaly evolution is determined in the following manner: Obtain the node transient response characteristics and preset normal baseline reference characteristics of the monitoring points at the end of the candidate path of the abnormal evolution; Calculate the deviation metric between the node transient response characteristics and the normal baseline reference characteristics; The deviation metric is accumulated within a preset time window to obtain the abnormal accumulation amount.

[0014] Secondly, this application provides a high-end equipment condition monitoring system based on a fiber optic sensor network, comprising: The acquisition module is used to acquire time-series physical field signals from multiple monitoring points on the equipment structure through a fiber optic sensor network. The processing module is used to construct a spatial topology map based on the spatial adjacency relationship of each monitoring point, and assign weights to each edge of the spatial topology map based on the pre-calibrated response transfer rate function between the monitoring points. The weighted spatial topology map and the time-series physical field signal together constitute the spatiotemporal map data. The processing module is also used to perform spatial domain feature aggregation on the spatiotemporal map data, perform time domain state evolution processing on the spatial aggregation feature sequence of each monitoring point, extract the fluctuation components that change with time, and form the node transient response features of each monitoring point. The processing module is also used to determine the time-series prediction impact index in two directions for every two adjacent monitoring points based on the change in prediction error before and after adding the historical node transient response characteristics of one monitoring point to the prediction model of another monitoring point, filter the prediction advantage direction that exceeds the preset impact threshold to form a directed correlation edge, and combine the directed correlation edges that are connected end to end and do not have duplicate nodes into anomaly evolution candidate paths. The execution module is used to issue a warning message containing the abnormal evolution candidate path when the abnormal accumulation corresponding to the monitoring point at the end of the abnormal evolution candidate path exceeds a set threshold.

[0015] Thirdly, this application provides a computer device, which includes a memory and a processor. The memory stores code, and the processor is configured to acquire the code and execute the above-described high-end equipment fault early warning method based on fiber optic sensor networks.

[0016] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-mentioned high-end equipment fault early warning method based on fiber optic sensor networks.

[0017] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: This application constructs a spatial topology map based on the spatial adjacency relationship of monitoring points on the equipment structure, and uses the response transfer rate function of adjacent monitoring points under calibration conditions to determine edge weights, so that the spatial aggregation of monitoring data is constrained by structural connection relationship and response coupling difference. The spatial aggregation feature sequence is decomposed into trend component and fluctuation component, and the fluctuation component is used to form the node transient response characteristics, which can reduce the impact of slow common changes on short-term temporal relationship analysis. By comparing the time series prediction influence indicators of two directions of adjacent monitoring points, directed associated edges with predictive leading relationship can be screened, and abnormal evolution candidate paths are formed under the conditions of non-repeating nodes and cycle stopping. When the abnormal accumulation of the monitoring point at the end of the abnormal evolution candidate path exceeds the set threshold, the node sequence, the predicted dominant direction and the predicted leading time lag are output, thereby providing verifiable equipment abnormal evolution candidate path early warning information. Attached Figure Description

[0018] Figure 1 This is an exemplary flowchart of a high-end equipment fault early warning method based on fiber optic sensor networks, according to some embodiments of this application. Figure 2 This is a schematic diagram illustrating the process of constructing a spatial topology map according to some embodiments of this application; Figure 3 This is a flowchart illustrating the process of determining time-series forecast impact indicators according to some embodiments of this application; Figure 4 This is a schematic diagram of the structure of a high-end equipment condition monitoring system based on an optical fiber sensor network, according to some embodiments of this application. Figure 5 This is a schematic diagram of the structure of a computer device that implements a high-end equipment fault early warning method based on fiber optic sensor networks, according to some embodiments of this application. Detailed Implementation

[0019] To better understand the technical solution of this application, the technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] refer to Figure 1 The figure is an exemplary flowchart of a high-end equipment fault early warning method based on an optical fiber sensor network, according to some embodiments of this application. The high-end equipment fault early warning method based on an optical fiber sensor network mainly includes the following steps: In step 101, the time-series physical field signals of multiple monitoring points on the equipment structure are acquired through an optical fiber sensor network.

[0021] It should be noted that the fiber optic sensing network in this application refers to a distributed array of optical fibers or fiber grating sensors deployed along the surface or interior of the equipment structure, which acquires strain or vibration information at each measuring point by demodulating optical signals. The method of this application can be applied to fault early warning of high-end equipment, such as: the main shaft and gearbox of large fans, the housing and pipeline of industrial pump and compressor units, the mill stand and transmission shaft system, the stator and bearing housing of large motors, and the pressure-bearing housing of towers and reactors.

[0022] In some embodiments, the acquisition of time-series physical field signals of multiple monitoring points on the equipment structure through an optical fiber sensing network can be achieved in the following manner: a single-mode optical fiber with multiple fiber Bragg gratings etched on it is pasted along a preset path on or inside the equipment structure, with each grating position serving as a monitoring point; the optical fiber is connected to a fiber Bragg grating demodulator, and a broadband light source built into the demodulator is injected into the optical fiber through a coupler. Each grating reflects back narrowband light with a center wavelength corresponding to its own grating pitch. When the structure experiences strain or vibration, the grating pitch changes, and the reflected wavelength drifts accordingly. The demodulator continuously acquires the real-time wavelength of each grating at a preset sampling frequency and converts the wavelength drift into a strain value according to a pre-calibrated wavelength-strain sensitivity coefficient, thereby obtaining a strain signal sequence of each monitoring point over time, which is the time-series physical field signal of that monitoring point.

[0023] It should also be noted that the sampling frequency in this application is typically 100 Hz to 2 kHz, and can be selected according to the dynamic response characteristics of the monitored equipment. The wavelength-strain sensitivity coefficient in this application can be pre-calibrated in the following way: the fiber segment containing the grating is pasted on the surface of a beam of equal strength, and a known displacement or load is applied to the free end of the beam through a micrometer screw or weights to generate a uniform standard strain on the beam surface. The real-time center wavelength of each grating is recorded simultaneously to obtain the wavelength drift under different strain levels. The wavelength drift and strain value are linearly regressed, and the regression slope is the wavelength-strain sensitivity coefficient of this batch of fiber optic gratings.

[0024] In step 102, a spatial topology graph is constructed based on the spatial adjacency relationship of each monitoring point, and each edge of the spatial topology graph is weighted based on the pre-calibrated response transfer rate function between the monitoring points. The weighted spatial topology graph and the time-series physical field signal together constitute the spatiotemporal graph data.

[0025] In some embodiments, reference Figure 2 As shown in the figure, this is a schematic flowchart of constructing a spatial topology map according to some embodiments of this application. In this embodiment, the spatial topology map can be constructed based on the spatial adjacency relationship of each monitoring point using the following steps: In step 1021, the physical spatial coordinates of each monitoring point on the equipment structure are obtained; In step 1022, based on the physical space coordinates, the spatial adjacency relationship between any two monitoring points is determined. When there is a direct structural connection path between two points and there are no other monitoring points on the path, the two points are determined to be adjacent. In step 1023, a spatial topology graph is constructed using each monitoring point as a node and the connection relationship between adjacent monitoring points as edges.

[0026] In this application, spatial adjacency refers to the adjacency determination relationship between two monitoring points where there is a direct structural connection path between them in the equipment structure that is not interrupted by other monitoring points.

[0027] In practice, firstly, after the fiber optic sensor network is deployed, based on the three-dimensional design model of the equipment structure or precise on-site measurement data, the three-dimensional spatial coordinates of each monitoring point on the surface or inside the equipment structure are determined one by one. These three-dimensional spatial coordinates are the physical spatial coordinates of the monitoring point. Secondly, for the set of all monitoring points, the geometric and structural relationships between two monitoring points are checked one by one. During the check, the physical spatial coordinates of the two monitoring points are used as reference endpoints, and path tracing is performed along the continuous load-bearing path on the surface or inside the equipment structure. If a connecting path that meets the above-mentioned adjacent determination conditions can be found, the pair of monitoring points are marked as adjacent. If it is necessary to cross structural interfaces, non-load-bearing accessories, or if the path inevitably passes through... If the physical spatial coordinates of the third monitoring point are not considered adjacent, after all pairs are checked, all monitoring point pairs marked as adjacent are recorded as a set of monitoring point pairs with spatial adjacency. Then, each monitoring point is abstracted as a single node in the graph, and all monitoring points form a node set. All monitoring point pairs with spatial adjacency recorded in the previous step are read one by one, and an undirected connection edge is created between the two nodes of each pair of monitoring points. All connection edges form an edge set. The node set and the edge set together form a spatial topology graph. Finally, the spatial topology graph constructed with monitoring points as nodes and the connection relationship between adjacent monitoring points as edges is used as a data model to represent the spatial layout and adjacency relationship of monitoring points.

[0028] In some embodiments, assigning weights to the edges of the spatial topology graph based on a pre-calibrated response transfer rate function between the monitoring points can be achieved using the following steps: For each edge in the spatial topology graph, a set of adjacent monitoring points are used. One monitoring point is used as the response reference point and the other monitoring point is used as the response observation point. The response signals of the two monitoring points under pulse excitation or step excitation are collected synchronously through the optical fiber sensing network. The auto-power spectral density of the response signal at the response reference point and the cross-power spectral density between the response signals at two monitoring points are determined. Based on the auto-power spectral density and the cross-power spectral density, the response transfer rate function within the characteristic frequency band of the equipment is determined, and the integral mean of the amplitude of the response transfer rate function within the characteristic frequency band of the equipment is used as the response transfer coefficient of the corresponding adjacent monitoring point pair. The response propagation coefficients corresponding to each edge in the spatial topology graph are normalized, and the normalized response propagation coefficients are used as the edge weights of the corresponding edges.

[0029] The response transfer rate function between adjacent monitoring points is determined by the self-power spectral density of the response reference point and the cross-power spectral density between the response signals of the two monitoring points. It characterizes the amplitude variation of the response observation point relative to the response reference point within a specified excitation mode and equipment characteristic frequency band. The amplitude of the response transfer rate function is averaged and normalized over the frequency band to form side weights, enabling spatial aggregation processing to distinguish the strength of response coupling between different adjacent monitoring points under calibration conditions.

[0030] In practice, firstly, for each pair of adjacent monitoring points corresponding to each edge in the spatial topology diagram, one monitoring point is used as the response reference point and the other as the response observation point. The response signals of the two points under pulse excitation or step excitation are synchronously collected through the fiber optic sensor network. On the matrix in the neighborhood of the structure where the response reference point is located, avoiding the fiber optic grating sensor body, a force hammer with a force sensor is used to apply pulse excitation to the surface of the structure, or a step excitation is applied by rapidly releasing the preload, so that the local structure where the response reference point is located generates a broadband transient response. At the same time as applying the excitation, the center wavelength of the response reference point and the response observation point is synchronously acquired in parallel across all channels using a fiber optic grating demodulator. The acquired raw data is a sequence of wavelengths changing over time. Based on the pre-calibrated wavelength-strain sensitivity coefficient, the wavelength sequence of the response reference point and the wavelength sequence of the response observation point are converted into their respective strain response time histories. These two strain response time histories are called the response signal of the response reference point and the response signal of the response observation point, respectively.

[0031] Secondly, the response signals at the response reference point and the response observation point are processed to eliminate linear trend terms, and a Hanning window is applied to suppress spectral leakage. Then, cross-power spectral density analysis is performed to calculate the auto-power spectral density function of the response signal at the response reference point, and the cross-power spectral density function of the response signal at the response reference point and the response signal at the response observation point. In the frequency domain, the amplitude of the cross-power spectral density function is compared with the amplitude of the auto-power spectral density function at each frequency point to obtain a set of amplitude ratio curves that vary with frequency. This curve is the response transferability function, which characterizes the relationship between the response amplitude of the response observation point and the response reference point at different frequencies under a specified excitation mode. Next, based on the structural dynamics model or modal test results of the monitored equipment, the distribution range of its low-order natural frequencies is obtained, and the frequency range covering the first few main modes is taken as the characteristic frequency band of the equipment. Within the frequency range of the characteristic frequency band, the amplitude curve of the response transferability function is integrated in the frequency domain and divided by the bandwidth of the characteristic frequency band of the equipment to obtain the integral mean. This integral mean is the response transfer coefficient of the adjacent monitoring point pair.

[0032] Then, obtain the response propagation coefficient corresponding to each edge in the spatial topology graph, and then gather all the response propagation coefficients of all edges together, find the maximum and minimum values, and use the min-max normalization method to subtract the minimum value from each response propagation coefficient, and then divide by the difference between the maximum and minimum values, so as to linearly map each response propagation coefficient to the interval between 0 and 1, and obtain the normalized response propagation coefficient; then assign the normalized response propagation coefficient corresponding to each edge in the spatial topology graph as the edge weight of the edge.

[0033] In some embodiments, the spatiotemporal graph data, which is composed of the weighted spatial topology graph and the temporal physical field signal, can be achieved by the following steps: The time-series physical field signal is divided into multiple time frames according to a preset sampling interval, and each time frame contains the physical field signal value of each monitoring point at that moment; The physical field signal value of each time frame is assigned to each node of the spatial topology map according to the monitoring point location index, which serves as the node attribute of that time frame, forming a single-frame attribute map. The spatiotemporal graph data is constructed by arranging the single-frame attribute graphs of multiple consecutive time frames in chronological order, wherein each time frame corresponds to a graph snapshot, and there is a fixed time step between adjacent graph snapshots.

[0034] The spatiotemporal map data in this application is a complete multi-frame snapshot sequence used to carry the evolution of the physical field signals of the monitoring points on the spatial topology map over time.

[0035] In specific implementation, firstly, a preset sampling interval is set, which is an integer multiple of the sampling period of the fiber optic sensor network. The sampling period is the time interval between two adjacent sampling actions when the fiber optic grating demodulator performs full-channel synchronous acquisition. Then, taking the moment when the fiber optic sensor network starts acquisition as the starting time reference, starting from the starting time reference, a time point is taken at every preset sampling interval as the identifier time of the time frame. At the identifier time of each time frame, the single sample value corresponding to the identifier time on the time axis is obtained from the time-series physical field signal of each monitoring point, which is used as the physical field signal value of each monitoring point in that time frame. The physical field signal values ​​of all monitoring points in that time frame are combined together to form the set of physical field signal values ​​of the monitoring points in that time frame. The data set containing the physical field signal values ​​of each monitoring point at the identifier time, with each identifier time as the time frame unit, is used as the data content of the corresponding time frame.

[0036] Secondly, for each time frame, the physical field signal values ​​of each monitoring point within that time frame are read. The monitoring point location index refers to the unique numerical number assigned to each monitoring point during the deployment phase of the fiber optic sensor network. This number corresponds one-to-one with the physical spatial coordinates of the monitoring point on the equipment structure. When constructing the spatial topology map, the unique numerical number of each monitoring point is bound to the nodes in the map, establishing a direct mapping relationship from the monitoring point location index to the nodes of the spatial topology map. Based on this mapping relationship, the physical field signal value of each monitoring point in the current time frame is assigned to the attribute field of the node in the spatial topology map that is bound to the unique numerical number of that monitoring point. The physical field signal value is stored in the node attribute field as a floating-point number. After the assignment is completed, each node in the spatial topology map retains the original adjacency relationship information and edge weights, and adds the physical field signal value of that time frame as a node attribute, thus obtaining a spatial topology map structure with node attributes.

[0037] Then, the single-frame attribute maps corresponding to multiple time frames generated sequentially are arranged into an ordered sequence. The number of consecutive time frames is determined according to the observation duration requirements of the anomaly evolution process in actual monitoring. The product of the number of frames and the preset sampling interval covers the observation time window of the target anomaly evolution process. Each single-frame attribute map in the sequence is a snapshot, and the time interval between two adjacent snapshots is equal to the preset sampling interval, which serves as a fixed time step. The snapshot sequence formed by the single-frame attribute maps of consecutive time frames in chronological order records the changes in the physical field signal values ​​of each monitoring point at discrete sampling times, thus constituting spatiotemporal map data.

[0038] In step 103, spatial domain feature aggregation is performed on the spatiotemporal map data, and time domain state evolution processing is performed on the spatial aggregation feature sequence of each monitoring point to extract the fluctuation components that change with time and form the node transient response features of each monitoring point.

[0039] In some embodiments, spatial domain feature aggregation is performed on the spatiotemporal map data, and time domain state evolution processing is performed on the spatial aggregation feature sequence of each monitoring point to extract the fluctuation component that changes with time, forming the nodal transient response characteristics of each monitoring point. This can be achieved by the following steps: For each time frame in the spatiotemporal graph data, with each monitoring point as the target node, the physical field signal values ​​of its neighboring nodes are aggregated according to the edge weights of the spatial topology graph to obtain the spatial aggregation characteristics of each monitoring point in that time frame. The spatial aggregation features of each monitoring point in multiple consecutive time frames are arranged in chronological order to obtain the spatial aggregation feature sequence of each monitoring point. The spatial aggregation feature sequence of each monitoring point is smoothed to form a trend component that changes over time. The difference sequence between the spatial aggregation feature sequence and the trend component is extracted and the difference sequence is identified as the fluctuation component. The fluctuation components at each monitoring point are used to form the nodal transient response characteristics of the corresponding monitoring point.

[0040] Spatial aggregation features are used to characterize the comprehensive response of a single monitoring point within a single time frame, combining the physical field signal information of its spatial neighborhood. The spatial aggregation feature sequence records the process of how the spatial aggregation features of a single monitoring point change over time. The spatial aggregation feature sequence is smoothed to form a trend component, and the difference sequence between the spatial aggregation feature sequence and the trend component forms a fluctuation component. This fluctuation component characterizes the rapid response change of the corresponding monitoring point relative to the slowly changing portion, thus forming the node's transient response features.

[0041] In specific implementation, for each time frame in the spatiotemporal graph data, taking each monitoring point as the target node, and aggregating the physical field signal values ​​of the neighboring nodes of the corresponding target node according to the edge weights of the spatial topology graph, the spatial aggregation features of each monitoring point in that time frame can be obtained in the following way: For each graph snapshot in the spatiotemporal graph data, each monitoring point in the spatial topology graph is taken as the target node in turn. All adjacent nodes directly connected to the target node are obtained from the spatial topology graph, and the edge weights on each connecting edge are read. The physical field signal value of the target node itself is regarded as a self-loop connection contribution with a weight of 1, and is normalized and weighted averaged together with the physical field signal values ​​of each adjacent node according to their respective edge weights. The physical field signal value of the target node is added to the weighted physical field signal value of each adjacent node, and then divided by the sum of the self-loop weight of the target node (1) and the edge weights of all adjacent edges. The resulting value is used as the spatial aggregation feature of the target node in the current time frame. The normalized weighted average operation is performed on all monitoring points in each graph snapshot to obtain the spatial aggregation features of each monitoring point in each time frame.

[0042] In addition, in specific implementation, the spatial aggregation features of each monitoring point on multiple consecutive time frames can be arranged in chronological order to obtain the spatial aggregation feature sequence of each monitoring point. This can be achieved in the following way: For each monitoring point, starting from the first snapshot of the spatiotemporal map data, the spatial aggregation feature values ​​of the monitoring point in each consecutive time frame are extracted sequentially. The number of consecutive time frames is determined according to the preset multiple of the moving average window width in the subsequent time series evolution processing, usually three to five times the moving average window width, to ensure sufficient sample size when extracting trend components. Then, according to the chronological order of the snapshots, these spatial aggregation feature values ​​are arranged into a one-dimensional array. The position index of each element in the array corresponds to the time series number of the snapshot in the spatiotemporal map data. This array is used as the spatial aggregation feature sequence of the monitoring point.

[0043] In addition, in practical implementation, the spatial aggregated feature sequences of each monitoring point can be smoothed to form a trend component that changes over time. The difference sequence between the spatial aggregated feature sequence and the trend component can be determined as the fluctuation component. This can be achieved as follows: For the spatial aggregated feature sequence of each monitoring point, a moving average method is used for smoothing. The window width of the moving average is determined based on the shortest evolution time of the monitored equipment target anomaly and a preset sampling interval. The neighborhood mean of the spatial aggregated feature sequence is taken point by point using the window width to form a trend component reflecting slow response changes. Then, the trend component at the corresponding time is subtracted from the original spatial aggregated feature sequence point by point to form a fluctuation component reflecting relatively fast response changes. The fluctuation components of each monitoring point are arranged in chronological order to form the node transient response feature sequence of the corresponding monitoring point. The node transient response feature sequence is used to determine the subsequent time-series prediction impact indicators.

[0044] The spatially aggregated feature sequence is decomposed into trend and fluctuation components to separate the slowly changing parts from the relatively fast changing parts in the monitoring point response. Subsequent time-series prediction impact analysis uses the node transient response features formed by the fluctuation components to reduce the impact of slow, shared changes on the short-term temporal sequence of adjacent monitoring points. These node transient response features represent the rapid response changes of discrete monitoring points under the current sampling and spatial aggregation conditions, and are used to form predictive leader relationships that require further verification.

[0045] In step 104, for every two adjacent monitoring points, based on the change in prediction error before and after adding the historical node transient response characteristics of one monitoring point to the prediction model of another monitoring point, the time-series prediction influence indicators in two directions are determined, the prediction advantage directions exceeding the preset influence threshold are selected to form directed correlation edges, and the directed correlation edges that are connected end to end and have no duplicate nodes are combined into candidate paths for abnormal evolution.

[0046] In some embodiments, reference Figure 3 As shown in the figure, this is a flowchart illustrating the determination of time-series prediction impact indicators according to some embodiments of this application. In this embodiment, for every two adjacent monitoring points, the determination of time-series prediction impact indicators in two directions can be achieved by the following steps, based on the change in prediction error before and after adding the historical transient response characteristics of one monitoring point to the prediction model of another monitoring point: In step 1041, for each group of adjacent monitoring points, the node transient response feature sequence of one monitoring point is taken as the first sequence, and the node transient response feature sequence of the other monitoring point is taken as the second sequence. In step 1042, a first time-series prediction impact index of the first sequence on the second sequence and a second time-series prediction impact index of the second sequence on the first sequence are determined respectively. In step 1043, the direction corresponding to the one that exceeds the preset influence threshold and has the larger value among the first time-series prediction influence index and the second time-series prediction influence index is taken as the prediction advantage direction, and the time-series prediction influence index value corresponding to this direction is taken as the time-series prediction influence index of the adjacent monitoring point pair.

[0047] The first and second time-series prediction influence indices represent the degree of reduction in prediction error caused by introducing the historical transient response characteristics of another monitoring point when predicting the current transient response characteristics of a monitoring point. The two time-series prediction influence indices are used to compare the predictive leadership relationship of transient response changes between adjacent monitoring points. Directions exceeding a preset influence threshold and having larger values ​​are considered the predictive advantage directions for forming directed correlation edges.

[0048] In practical implementation, a baseline prediction model and an extended prediction model are constructed for the first and second sequences, respectively. When the second sequence is the predicted object, the baseline prediction model uses its own historical values ​​to predict the current value of the second sequence, while the extended prediction model uses the historical values ​​of both the first and second sequences to predict the current value of the second sequence. Both models use data from the same continuous time period and the same candidate order range, and select the prediction order according to the Akaike information content criterion. The first baseline prediction error and the first extended prediction error are determined based on the residuals of the baseline prediction model and the extended prediction model, respectively. The degree to which the first baseline prediction error is reduced relative to the first extended prediction error is determined as the first time series prediction impact index. The predicted relationships of the first and second sequences are swapped, and the second time series prediction impact index is determined in the same way.

[0049] Furthermore, in specific implementation, the first and second time-series predicted impact indicators are compared with preset impact thresholds. The preset impact thresholds are determined based on monitoring data collected by the equipment under normal operating conditions with no confirmed faults. The time-series predicted impact indicators of each adjacent monitoring point pair under normal operating conditions are sorted, and the time-series predicted impact indicator value corresponding to the 95th percentile is taken as the preset impact threshold. If the first time-series predicted impact indicator exceeds the preset impact threshold and is greater than the second time-series predicted impact indicator, the direction from the monitoring point corresponding to the first sequence to the monitoring point corresponding to the second sequence is determined as the predicted dominant direction. If the second time-series predicted impact indicator exceeds the preset impact threshold and is greater than the first time-series predicted impact indicator, the direction from the monitoring point corresponding to the second sequence to the monitoring point corresponding to the first sequence is determined as the predicted dominant direction. If neither of the time-series predicted impact indicators exceeds the preset impact threshold, then the adjacent monitoring point pair is not determined as a predicted dominant direction.

[0050] In some embodiments, the following steps can be used to select predicted advantage directions that exceed a preset influence threshold to form directed association edges, and combine the head-to-tail directed association edges that are connected and have no duplicate nodes to form anomaly evolution candidate paths: For each pair of adjacent monitoring points with the predicted advantage direction, the starting point of the predicted advantage direction is determined as the prediction leader node, and the ending point of the predicted advantage direction is determined as the prediction response node, forming a directed association edge from the prediction leader node to the prediction response node. The path search begins from the predicted leader node that has no incoming edges and proceeds along the directed associated edges. When the same node has multiple outgoing edges, path branches are formed respectively. When the next node has already appeared in the current path or the current node has no outgoing edges, stop searching the current path branch and determine the obtained continuous directed node sequence without duplicate nodes as the candidate path for abnormal evolution.

[0051] Anomaly evolution candidate path refers to a continuous sequence of nodes formed by directed associated edges with predictive leading relationships under spatial adjacency constraints, which does not contain duplicate nodes. It is used to represent the anomaly evolution order of discrete monitoring points that needs to be further verified in conjunction with equipment structure and operating conditions.

[0052] In practice, the predicted dominant direction and corresponding time-series predicted influence index of each adjacent monitoring point pair in the spatial topology map are read one by one. The adjacent monitoring point pairs with the predicted dominant direction are retained as the set of monitoring point pairs to be used to construct directed association edges. Since the predicted dominant direction used when forming the directed association edge has already been filtered by a preset influence threshold, the same threshold is not used to repeatedly filter the time-series predicted influence index when constructing the directed association edge.

[0053] For each pair of adjacent monitoring points to be constructed with directed association edges, the predicted advantage direction formed by the adjacent monitoring points in the time series prediction impact index determination stage is read. The starting point of the predicted advantage direction is marked as the prediction leader node, and the ending point of the predicted advantage direction is marked as the prediction response node. A directed association edge from the prediction leader node to the prediction response node is formed in the spatial topology graph. The corresponding time series prediction impact index value and the predicted advantage direction are stored in the attribute field of the directed association edge.

[0054] For all directed edges, count the incoming and outgoing edges of each node. Start the path search along the directed edges from the predicted leader node with no incoming edges, and record the nodes already visited in the current path. When the same node has multiple outgoing edges, copy the current node sequence and form path branches along different outgoing edges. When the next node already exists in the current path, stop expanding along that directed edge to avoid duplicate searches caused by directed cycles. When the current node has no outgoing edges, identify the continuous directed node sequence without duplicate nodes as a candidate path for abnormal evolution. For directed edges that cannot be visited from nodes without incoming edges, select the unvisited node as the starting point and perform the same search, and mark the node sequence with cyclic relationships as a path to be reviewed.

[0055] In step 105, when the cumulative abnormality corresponding to the end monitoring point of the abnormal evolution candidate path exceeds a set threshold, an early warning message containing the abnormal evolution candidate path is issued.

[0056] In some embodiments, when the cumulative abnormality corresponding to the end monitoring point of the abnormal evolution candidate path exceeds a set threshold, issuing an early warning message containing the abnormal evolution candidate path can be achieved through the following steps: Real-time tracking of the cumulative anomalies at the end monitoring points of each identified candidate path for abnormal evolution; When the cumulative amount of anomalies at any end monitoring point exceeds the set threshold, the complete node sequence of the corresponding candidate path for anomaly evolution, the predicted dominant direction of each directed associated edge, and the predicted lead time lag between adjacent nodes are extracted. The complete node sequence, the predicted advantage direction, the predicted lead time lag, and the early warning trigger signal are encapsulated into early warning information containing corresponding abnormal evolution candidate paths, and the early warning information is issued.

[0057] The cumulative abnormality is used to reflect the cumulative result of the degree of deviation of the end monitoring point from the normal state over a continuous period of time; the prediction leading time lag is determined by the model lag step number that forms the corresponding time series prediction influence index and the fixed time step of the spatiotemporal map data, and is used to represent the time interval corresponding to when the transient response characteristics of the historical node of the prediction leading node have a large incremental prediction effect on the transient response characteristics of the current node of the prediction response node.

[0058] In specific implementation, when the accumulated anomaly at any end monitoring point exceeds a set threshold, starting from the end monitoring point exceeding the threshold, the system traces back segment by segment along the opposite direction of the directed associated edges to the starting node of the corresponding candidate path for anomaly evolution. The monitoring point numbers traversed are recorded sequentially, and the recording order is reversed to form a complete node sequence from the starting node to the end node. Simultaneously, the predicted dominant direction stored on each directed associated edge is read segment by segment, and the candidate lag step number with the greatest effect on reducing prediction error in the extended prediction model is read. The product of the candidate lag step number and the fixed time step of the spatiotemporal map data is determined as the prediction leading lag between corresponding adjacent nodes. The complete node sequence, each segment of predicted dominant direction, and the prediction leading lag together form the candidate path information for anomaly evolution. The temporal resolution of the prediction leading lag is not less than the fixed time step of the spatiotemporal map data, and the spatial resolution of the candidate path for anomaly evolution is jointly limited by the monitoring point deployment location and spatial adjacency relationship. The set threshold is determined based on historical data of the accumulated anomaly at the corresponding monitoring points under normal equipment operating conditions.

[0059] In specific implementation, the complete node sequence, predicted advantageous directions of each segment, predicted leading time lag, and early warning trigger signal can be encapsulated and output together in the following way: The complete node sequence, predicted advantageous directions of each segment, predicted leading time lag between adjacent nodes, the current cumulative abnormality value of the end monitoring point, and the timestamp of the time exceeding the threshold are organized into a structured diagnostic information data block to generate an early warning trigger signal. This early warning trigger signal includes an early warning level identifier and the time of early warning occurrence. The early warning level identifier is determined based on the extent to which the cumulative abnormality exceeds a set threshold. The early warning trigger signal is used as the message header, and the structured diagnostic information data block is used as the message body, combined into a complete early warning output message body. This early warning output message body is sent to the display terminal and event storage database of the equipment monitoring system through a data communication interface. The early warning output message body, consisting of the early warning trigger signal as the message header and the diagnostic information data block containing the complete node sequence, predicted advantageous directions of each segment, and predicted leading time lag as the message body, is the final early warning information containing candidate paths for abnormal evolution.

[0060] Preferably, in some embodiments, the cumulative anomaly amount corresponding to the monitoring point at the end of the candidate path of anomaly evolution can be achieved by the following steps: Obtain the node transient response characteristics and preset normal baseline reference characteristics of the monitoring points at the end of the candidate path of the abnormal evolution; Calculate the deviation metric between the node transient response characteristics and the normal baseline reference characteristics; The deviation metric is accumulated within a preset time window to obtain the abnormal accumulation amount.

[0061] The preset normal baseline reference features are formed by statistically analyzing the node transient response features of the corresponding monitoring point in multiple time frames when the equipment is in a confirmed fault-free normal operating state. It is used as a comparison benchmark to measure the degree of deviation of the current node transient response features of the terminal monitoring point from the normal state. The deviation metric is used to quantify the degree of difference between the current node transient response features of the terminal monitoring point and the normal baseline reference features.

[0062] In specific implementation, after forming an abnormal evolution candidate path, the transient response characteristics of the corresponding end monitoring point in the current time frame are read, and the normal baseline reference characteristics of the end monitoring point are read from a preset baseline database. The normal baseline reference characteristics are formed by statistically analyzing the transient response characteristics of the monitoring point in multiple time frames under normal operating conditions with confirmed no faults. The absolute value of the difference between the current node transient response characteristics and the normal baseline reference characteristics is calculated, and the absolute value of the difference is used as the deviation metric value of the current time frame. At the arrival of each time frame, the current deviation metric value is added to the cumulative calculation within a preset time window, and historical deviation metric values ​​exceeding the preset time window are removed to form the abnormal cumulative amount of the end monitoring point at the current moment.

[0063] In summary, this application spatially organizes monitoring data based on the spatial adjacency of monitoring points and the response transfer rate function under calibration conditions. It uses fluctuation components to form transient response characteristics of nodes and utilizes time-series predictive influence indicators between adjacent monitoring points to form directed correlation edges with predictive leading relationships. During the path search process, it forms candidate paths for abnormal evolution through non-repeating nodes and cyclic stopping conditions. Then, it triggers early warnings by combining the abnormal accumulation of terminal monitoring points relative to normal baseline reference characteristics. This provides verifiable early warning information for candidate paths of abnormal equipment evolution.

[0064] On the other hand, in some embodiments, this application provides a high-end equipment condition monitoring system based on fiber optic sensor networks, with reference to... Figure 4 The figure is a schematic diagram of the structure of a high-end equipment condition monitoring system based on an optical fiber sensor network according to some embodiments of this application. The high-end equipment condition monitoring system 400 based on an optical fiber sensor network includes: an acquisition module 401, a processing module 402, and an execution module 403, which are described below: Acquisition module 401, in this application, is mainly used to acquire time-series physical field signals of multiple monitoring points on the equipment structure through an optical fiber sensor network; Processing module 402, in this application, is used to construct a spatial topology map based on the spatial adjacency relationship of each monitoring point, and to assign weights to each edge of the spatial topology map based on the pre-calibrated response transfer rate function between the monitoring points. The weighted spatial topology map and the time-series physical field signal together constitute spatiotemporal map data. In specific implementation, the processing module 402 in this application is also used to perform spatial domain feature aggregation on the spatiotemporal map data, perform time domain state evolution processing on the spatial aggregation feature sequence of each monitoring point, extract the fluctuation component that changes with time, and form the node transient response features of each monitoring point. In addition, the processing module 402 in this application is also used to determine the time-series prediction influence index in two directions for each pair of adjacent monitoring points, based on the change in prediction error before and after adding the historical node transient response characteristics of one monitoring point to the prediction model of another monitoring point, filter the prediction advantage direction that exceeds the preset influence threshold to form a directed correlation edge, and combine the directed correlation edges that are connected end to end and do not have duplicate nodes into anomaly evolution candidate paths. The execution module 403 in this application is mainly used to issue a warning message containing the abnormal evolution candidate path when the abnormal accumulation corresponding to the end monitoring point of the abnormal evolution candidate path exceeds a set threshold.

[0065] In addition, this application also provides a computer device, which includes a memory and a processor. The memory stores code, and the processor is configured to acquire the code and execute the above-described high-end equipment fault early warning method based on fiber optic sensor networks.

[0066] In some embodiments, reference Figure 5 This figure is a schematic diagram of the structure of a computer device implementing a high-end equipment fault early warning method based on an optical fiber sensor network, according to some embodiments of this application. The high-end equipment fault early warning method based on an optical fiber sensor network in the above embodiments can... Figure 5 The computer device shown is used to implement this, and the computer device 500 includes at least one processor 501, a communication bus 502, a memory 503, and at least one communication interface 504.

[0067] Processor 501 can be a general-purpose central processing unit (CPU) or an application-specific integrated circuit (ASIC).

[0068] The communication bus 502 can be used to transmit information between the aforementioned components.

[0069] Memory 503 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital versatile optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. Memory 503 may exist independently and be connected to processor 501 via communication bus 502. Memory 503 may also be integrated with processor 501.

[0070] The memory 503 stores program code for executing the scheme of this application, and its execution is controlled by the processor 501. The processor 501 executes the program code stored in the memory 503. The program code may include one or more software modules. The high-end equipment fault early warning method based on fiber optic sensor network in the above embodiment can be implemented by the processor 501 and one or more software modules in the program code in the memory 503.

[0071] Communication interface 504 uses any transceiver-like device to communicate with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0072] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single-CPU) processor or a multi-core (multi-CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0073] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.

[0074] In addition, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned high-end equipment fault early warning method based on fiber optic sensor networks.

[0075] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0076] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for early warning of faults in high-end equipment based on fiber optic sensor networks, characterized in that, include: The temporal physical field signals of multiple monitoring points on the equipment structure are acquired through fiber optic sensor networks. A spatial topology graph is constructed based on the spatial adjacency relationship of each monitoring point, and each edge of the spatial topology graph is weighted based on the pre-calibrated response transfer rate function between the monitoring points. The weighted spatial topology graph and the time-series physical field signal together constitute the spatiotemporal graph data. Spatial domain feature aggregation is performed on the spatiotemporal map data, and time domain state evolution processing is performed on the spatial aggregation feature sequence of each monitoring point to extract the fluctuation component that changes with time, forming the node transient response characteristics of each monitoring point. For every two adjacent monitoring points, based on the change in prediction error before and after adding the historical node transient response characteristics of one monitoring point to the prediction model of another monitoring point, the time-series prediction impact indicators in two directions are determined, the prediction advantage directions exceeding the preset impact threshold are selected to form directed correlation edges, and the directed correlation edges that are connected end to end and do not have duplicate nodes are combined into candidate paths for abnormal evolution. When the cumulative abnormality at the end monitoring point of the abnormal evolution candidate path exceeds a set threshold, an early warning message containing the abnormal evolution candidate path is issued.

2. The method as described in claim 1, characterized in that, Constructing a spatial topology map based on the spatial adjacency relationships of each monitoring point specifically includes: Obtain the physical spatial coordinates of each monitoring point on the equipment structure; Based on the physical space coordinates, the spatial adjacency relationship between any two monitoring points is determined. When there is a direct structural connection path between two points and there are no other monitoring points on the path, the two points are determined to be adjacent. A spatial topology graph is constructed using each monitoring point as a node and the connections between adjacent monitoring points as edges.

3. The method as described in claim 2, characterized in that, Assigning weights to each edge of the spatial topology graph based on the pre-calibrated response transfer rate function between the monitoring points specifically includes: For each edge in the spatial topology graph, a set of adjacent monitoring points are used. One monitoring point is used as the response reference point and the other monitoring point is used as the response observation point. The response signals of the two monitoring points under pulse excitation or step excitation are collected synchronously through the optical fiber sensing network. The auto-power spectral density of the response signal at the response reference point and the cross-power spectral density between the response signals at two monitoring points are determined. Based on the auto-power spectral density and the cross-power spectral density, the response transfer rate function within the characteristic frequency band of the equipment is determined, and the integral mean of the amplitude of the response transfer rate function within the characteristic frequency band of the equipment is used as the response transfer coefficient of the corresponding adjacent monitoring point pair. The response propagation coefficients corresponding to each edge in the spatial topology graph are normalized, and the normalized response propagation coefficients are used as the edge weights of the corresponding edges.

4. The method as described in claim 3, characterized in that, The spatiotemporal graph data, which is composed of the weighted spatial topology graph and the temporal physical field signal, specifically includes: The time-series physical field signal is divided into multiple time frames according to a preset sampling interval, and each time frame contains the physical field signal value of each monitoring point at that moment; The physical field signal value of each time frame is assigned to each node of the spatial topology map according to the monitoring point location index, which serves as the node attribute of that time frame, forming a single-frame attribute map. The spatiotemporal graph data is constructed by arranging the single-frame attribute graphs of multiple consecutive time frames in chronological order, wherein each time frame corresponds to a graph snapshot, and there is a fixed time step between adjacent graph snapshots.

5. The method as described in claim 1, characterized in that, Spatial domain feature aggregation is performed on the spatiotemporal map data, and time domain state evolution processing is performed on the spatial aggregation feature sequences of each monitoring point to extract the fluctuation components that change with time, forming the nodal transient response features of each monitoring point. Specifically, this includes: For each time frame in the spatiotemporal graph data, with each monitoring point as the target node, the physical field signal values ​​of its neighboring nodes are aggregated according to the edge weights of the spatial topology graph to obtain the spatial aggregation characteristics of each monitoring point in that time frame. The spatial aggregation features of each monitoring point in multiple consecutive time frames are arranged in chronological order to obtain the spatial aggregation feature sequence of each monitoring point. The spatial aggregation feature sequence of each monitoring point is smoothed to form a trend component that changes over time. The difference sequence between the spatial aggregation feature sequence and the trend component is extracted and the difference sequence is identified as the fluctuation component. The fluctuation components at each monitoring point are used to form the nodal transient response characteristics of the corresponding monitoring point.

6. The method as described in claim 1, characterized in that, For every two adjacent monitoring points, based on the change in prediction error before and after incorporating the historical transient response characteristics of one monitoring point into the prediction model of the other, the specific time-series prediction impact indicators in both directions are determined, including: For each pair of adjacent monitoring points, the transient response feature sequence of one monitoring point is taken as the first sequence, and the transient response feature sequence of the other monitoring point is taken as the second sequence. A first time-series prediction impact index of the first sequence on the second sequence and a second time-series prediction impact index of the second sequence on the first sequence are determined respectively. The direction corresponding to the one that exceeds the preset impact threshold and has the larger value among the first time-series prediction impact index and the second time-series prediction impact index is taken as the prediction advantage direction, and the time-series prediction impact index value corresponding to this direction is taken as the time-series prediction impact index of the adjacent monitoring point pair.

7. The method as described in claim 1, characterized in that, Directed edges are formed by filtering predicted advantageous directions that exceed a preset influence threshold, and these connected edges without duplicate nodes are combined into candidate paths for abnormal evolution. Specifically, this includes: For each pair of adjacent monitoring points with the predicted advantage direction, the starting point of the predicted advantage direction is determined as the prediction leader node, and the ending point of the predicted advantage direction is determined as the prediction response node, forming a directed association edge from the prediction leader node to the prediction response node. The path search begins from the predicted leader node that has no incoming edges and proceeds along the directed associated edges. When the same node has multiple outgoing edges, path branches are formed respectively. When the next node has already appeared in the current path or the current node has no outgoing edges, stop searching the current path branch and determine the obtained continuous directed node sequence without duplicate nodes as the candidate path for abnormal evolution.

8. The method as described in claim 1, characterized in that, When the cumulative anomaly amount corresponding to the monitoring point at the end of the candidate path of the abnormal evolution exceeds a set threshold, an early warning message containing the candidate path of the abnormal evolution is issued, specifically including: Real-time tracking of the cumulative anomalies at the end monitoring points of each identified candidate path for abnormal evolution; When the cumulative amount of anomalies at any end monitoring point exceeds the set threshold, the complete node sequence of the corresponding candidate path for anomaly evolution, the predicted dominant direction of each directed associated edge, and the predicted lead time lag between adjacent nodes are extracted. The complete node sequence, the predicted advantage direction, the predicted lead time lag, and the early warning trigger signal are encapsulated into early warning information containing corresponding abnormal evolution candidate paths, and the early warning information is issued.

9. The method as described in claim 1, characterized in that, The cumulative anomaly amount corresponding to the monitoring point at the end of the candidate path for anomaly evolution is determined in the following way: Obtain the node transient response characteristics and preset normal baseline reference characteristics of the monitoring points at the end of the candidate path of the abnormal evolution; Calculate the deviation metric between the node transient response characteristics and the normal baseline reference characteristics; The deviation metric is accumulated within a preset time window to obtain the abnormal accumulation amount.

10. A high-end equipment condition monitoring system based on fiber optic sensor networks, which uses the method described in any one of claims 1 to 9 for fault early warning, characterized in that, The system includes: The acquisition module is used to acquire time-series physical field signals from multiple monitoring points on the equipment structure through a fiber optic sensor network. The processing module is used to construct a spatial topology map based on the spatial adjacency relationship of each monitoring point, and assign weights to each edge of the spatial topology map based on the pre-calibrated response transfer rate function between the monitoring points. The weighted spatial topology map and the time-series physical field signal together constitute the spatiotemporal map data. The processing module is also used to perform spatial domain feature aggregation on the spatiotemporal map data, perform time domain state evolution processing on the spatial aggregation feature sequence of each monitoring point, extract the fluctuation components that change with time, and form the node transient response features of each monitoring point. The processing module is also used to determine the time-series prediction impact index in two directions for every two adjacent monitoring points based on the change in prediction error before and after adding the historical node transient response characteristics of one monitoring point to the prediction model of another monitoring point, filter the prediction advantage direction that exceeds the preset impact threshold to form a directed correlation edge, and combine the directed correlation edges that are connected end to end and do not have duplicate nodes into anomaly evolution candidate paths. The execution module is used to issue a warning message containing the abnormal evolution candidate path when the abnormal accumulation corresponding to the monitoring point at the end of the abnormal evolution candidate path exceeds a set threshold.

Images