Power optical cable monitoring device and monitoring method

Abstract

The present application relates to the technical field of power communication monitoring, in particular to a kind of electric power cable monitoring device and monitoring method, comprising: collecting original optical signal containing external disturbance characteristic change in multiple predetermined positions of electric power cable, signal is demodulated and separated, obtain two kinds of characteristic vectors reflecting vibration mode and temperature change. After the alignment of the two types of vectors is associated in space-time, a multidimensional monitoring tensor is generated, a hidden Markov model describing the evolution of the physical state of the cable is constructed, the state transition probability is calculated through the model, and the abnormal state sequence matching the preset fault mode is identified. According to the abnormal sequence, the time axis evolution process of the monitoring tensor is traced back. The present application can effectively separate the vibration and temperature disturbance characteristics, avoid the mutual interference of different disturbance signals, make the abnormal state identification more consistent with the actual operation law of the cable, accurately determine the occurrence position and type of the initial disturbance, and improve the reliability and positioning accuracy of the cable abnormal monitoring.

Description

Technical Field

[0001] This invention relates to the field of power communication monitoring technology, and in particular to a power optical cable monitoring device and monitoring method. Background Technology

[0002] Conventional power fiber optic cable monitoring technologies typically acquire optical signals along the cable route using a single acquisition terminal. They then rely on basic signal demodulation methods to extract information related to external disturbances. Some technologies only perform feature analysis on single physical quantities such as vibration or temperature, relying on preset thresholds to determine abnormal cable conditions, thus achieving simple monitoring of the power fiber optic cable's operational status. These technologies can only perform overall analysis of the optical signal and cannot separate the signal characteristics caused by vibration and temperature. Furthermore, signals acquired from multiple locations are not matched and integrated in a spatiotemporal dimension, resulting in limitations in signal dimensionality.

[0003] Fixed threshold judgment methods cannot align with the dynamic changes in the physical state of optical cables, leading to misjudgments in abnormal state identification and an inability to match corresponding fault modes based on state evolution logic. During monitoring, it is impossible to trace the temporal evolution trajectory of abnormal signals, making it difficult to accurately determine the initial location of abnormal disturbances or distinguish their specific types, thus failing to meet the practical needs of refined monitoring of power optical cables. Therefore, it is necessary to independently demodulate and separate vibration and temperature characteristics of the original optical signal, perform spatiotemporal correlation processing on the separated features to form multidimensional monitoring data, calculate state transition probabilities based on state evolution models, identify abnormal state sequences, and trace back to locate the position and type of the initial disturbance. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology and to propose a power optical cable monitoring device and monitoring method.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for monitoring power optical cables, comprising: The original optical signal is acquired at multiple predetermined locations on the power optical cable, and the original optical signal contains characteristic changes caused by external disturbances; The original optical signal is subjected to feature demodulation and separation processing to obtain a first type of feature vector reflecting the vibration mode and a second type of feature vector reflecting the temperature change; The first type of feature vector and the second type of feature vector are spatiotemporally correlated and aligned to generate a fused multidimensional monitoring tensor. Based on the multidimensional monitoring tensor, a hidden Markov model describing the evolution of the physical state of the optical cable is constructed. The transition probabilities between different states are calculated using the hidden Markov model, and abnormal state sequences that match the preset fault mode are identified. Based on the abnormal state sequence, the evolution of the multidimensional monitoring tensor on the time axis is traced back to locate the position and type of the initial disturbance that caused the abnormal state sequence.

[0006] As a further aspect of the present invention, acquiring the original optical signal at multiple predetermined locations on the power optical cable includes: Multiple distributed fiber optic sensing and demodulation devices are deployed along the power optical cable; Configure each distributed fiber optic sensing demodulation device to collect backscattered light from the same power optical cable carrying the service when triggered by a preset time synchronization signal. Each distributed fiber optic sensing demodulation device converts the acquired analog optical signal into a digital optical intensity sequence; The digitized light intensity sequences collected by multiple distributed fiber optic sensing and demodulation devices are uploaded to the central processing unit via a communication network, where they are aggregated into the original light signal.

[0007] As a further aspect of the present invention, feature demodulation and separation processing is performed on the original optical signal to obtain a first type of feature vector reflecting the vibration mode and a second type of feature vector reflecting temperature change, including: The original optical signal is phase demodulated to extract its composite phase information in the time and frequency domains; The composite phase information is processed using a blind source separation algorithm to separate independent components from the mixed signal; Each isolated independent component is decomposed into different frequency bands by performing wavelet packet transform. From multiple frequency band signals that have undergone wavelet packet transform, time-frequency domain statistics characterizing the signal energy distribution, frequency center, and pulse characteristics are extracted. The time-frequency domain statistics are classified based on a pre-trained vibration and temperature feature classifier. Time-frequency domain statistics classified as vibration features are aggregated into the first type of feature vector, and time-frequency domain statistics classified as temperature features are aggregated into the second type of feature vector.

[0008] As a further aspect of the present invention, the first type of feature vector and the second type of feature vector are spatiotemporally correlated and aligned to generate a fused multidimensional monitoring tensor, including: Add a location code corresponding to the acquisition location and a time code corresponding to the acquisition timestamp to the first type of feature vector and the second type of feature vector, respectively; Establish a spatiotemporal correlation matrix to record the cross-correlation strength between the first type of feature vector and the second type of feature vector at different locations and time points; Based on the spatiotemporal correlation matrix, interpolation and alignment are performed on the first type of feature vectors and the second type of feature vectors from different locations but close in time, so that they are mapped to a unified spatiotemporal grid point; On a unified spatiotemporal grid point, the aligned first-type feature vector and the second-type feature vector are concatenated along the feature dimension. The spliced ​​high-dimensional vector sequence is rearranged according to time and spatial order to construct the multidimensional monitoring tensor.

[0009] As a further aspect of the present invention, based on the multidimensional monitoring tensor, a hidden Markov model describing the evolution of the physical state of the optical cable is constructed, including: Define a set of discrete hidden states, each hidden state corresponding to a physical state category of a power optical cable; Define an observation probability distribution associated with each hidden state, the observation probability distribution being used to describe the probability of generating a frame of data in the multidimensional monitoring tensor under a specific physical state; Using historical normal power fiber optic cable monitoring data, the parameters of the initial state probability, state transition probability matrix, and observation probability distribution of the Hidden Markov Model are trained unsupervised using the expectation-maximization algorithm. During the training process, state dwell time constraints are introduced to simulate the relatively stable physical state of the power optical cable. After training is complete, all parameters of the Hidden Markov Model are saved to obtain a Hidden Markov Model that can be used for state decoding.

[0010] As a further aspect of the present invention, the hidden Markov model is used to calculate the transition probabilities between different states and to identify an abnormal state sequence that matches a preset fault mode, including: The multidimensional monitoring tensor is input into the trained Hidden Markov Model in chronological order. The Viterbi algorithm is used to decode the hidden state sequence and simultaneously calculate the probability of being in each hidden state at each time step. Extract all state transition events between adjacent time steps from the decoded hidden state sequence; Based on the state transition probability matrix, calculate the ratio of the actual probability of each state transition event to the probability of the transition occurring in historical normal data; State transition events where the ratio is lower than or higher than a preset threshold are marked as abnormal transitions. Multiple consecutive abnormal transitions and their associated hidden states are combined into a candidate abnormal state sequence. The candidate abnormal state sequence is matched with the predefined fault mode state sequence in the knowledge base. Candidate abnormal state sequences whose similarity to any fault mode state sequence exceeds a matching threshold are determined as the final abnormal state sequence.

[0011] As a further aspect of the present invention, the method for constructing the predefined fault mode state sequence in the knowledge base includes: We collected a large number of historical power fiber optic cable fault cases, each case containing monitoring data of the entire process from normal state to fault state; For the monitoring data of each historical case, the original optical signal acquisition, feature demodulation and separation, spatiotemporal correlation alignment and hidden Markov model decoding are performed in sequence to generate the corresponding multidimensional monitoring tensor and decode the complete hidden state sequence. Domain experts annotate the hidden state sequence, marking out the key state subsequences that characterize the occurrence and development of the fault; Key state subsequences representing the same type of fault were extracted from different cases, and cluster analysis was performed to obtain the typical state evolution pattern of each type of fault. The typical state evolution pattern of each type of fault is encoded into a state symbol sequence and stored as a predefined fault mode state sequence in the knowledge base.

[0012] As a further aspect of the present invention, based on the abnormal state sequence, the evolution process of the multidimensional monitoring tensor on the time axis is traced back to locate the position and type of the initial perturbation that caused the abnormal state sequence, including: From the abnormal state sequence, the first time point that deviates from the normal state range is identified as the abnormal starting point; Extract all data within a time window centered on the anomaly initiation point from the multidimensional monitoring tensor; Analyze the changes in the energy distribution of the multidimensional monitoring tensor in the spatial dimension within the time window; One or more spatial locations near the anomaly initiation point where spatial energy suddenly increases or decreases are identified as suspected disturbance source locations. The type of the initial disturbance is determined by combining the fault mode type matched by the abnormal state sequence, as well as the performance of the first type of feature vector and the second type of feature vector near the suspected disturbance source location.

[0013] As a further aspect of the present invention, the method further includes a step of reconstructing and verifying the disturbance event: Based on the located suspected disturbance source and the determined type of the initial disturbance, the corresponding typical disturbance signal template is retrieved from the knowledge base; Using the typical disturbance signal template, matched filtering is performed in the corresponding spatiotemporal region of the multidimensional monitoring tensor to enhance and extract the complete disturbance signal waveform; The extracted disturbance signal waveform is parametrically fitted to obtain a set of key parameters describing the disturbance event, including amplitude, duration, and frequency components. The set of key parameters is input into a pre-trained perturbation verification model, which outputs a confidence score of the abnormal state sequence caused by the perturbation event. When the confidence score exceeds the verification threshold, the perturbation event is confirmed as the root cause of the abnormal state sequence, and a complete perturbation event report containing location, type, time, and key parameters is generated.

[0014] As a further aspect of the present invention, the present invention also includes a power optical cable monitoring device, the device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the power optical cable monitoring method described above.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: The original optical signals, which contain changes in external disturbance characteristics, are collected from multiple predetermined locations on the power optical cable. Feature demodulation and separation processing are performed to form a first type of feature vector reflecting the vibration mode and a second type of feature vector reflecting the temperature change. Vibration-related features and temperature-related features are presented independently, and the signal features of the two different causes will not be mixed with each other during the analysis process. The extraction boundaries of various disturbance features in the optical signal are clearer, and the signal changes caused by different external factors acting on the optical cable can be completely distinguished, thus refining the accuracy of signal feature analysis.

[0016] After spatiotemporally aligning the first and second type of feature vectors, a multidimensional monitoring tensor is generated. Based on this tensor, a hidden Markov model describing the evolution of the physical state of the optical cable is constructed. The transition probabilities between different physical states are obtained through model computation. An abnormal state sequence matching the preset fault mode is selected. Based on the abnormal state sequence, the evolution process of the multidimensional monitoring tensor on the time axis is traced back to determine the location and type of the initial disturbance that caused the anomaly. The dynamic change process of the overall physical state of the optical cable can be completely depicted. The identification of abnormal states is consistent with the actual state change logic of the optical cable. The source tracing process of abnormal disturbances has a corresponding basis in spatiotemporal dimensions. The determination of the disturbance location and type can match the actual abnormal occurrence scenario of the optical cable. Attached Figure Description

[0017] Figure 1 This is a flowchart of a power optical cable monitoring method according to the present invention; Figure 2 A flowchart for acquiring raw optical signals at multiple predetermined locations on a power optical cable; Figure 3 A flowchart for generating multidimensional monitoring tensors for spatiotemporal correlation alignment. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] See Figure 1 This invention provides a method for monitoring power optical cables, the specific method including: By sensing changes in optical signals caused by external disturbances along the power optical cable, and analyzing these changes through signal processing and machine learning models, real-time monitoring and early fault warning of the cable's condition can be achieved. The method first acquires raw optical signals at multiple predetermined locations along the power optical cable. These raw signals contain characteristic changes caused by external disturbances. Subsequently, feature demodulation and separation processing is performed on the raw optical signals to obtain a first type of feature vector reflecting vibration modes and a second type of feature vector reflecting temperature changes. The two types of feature vectors are spatiotemporally correlated and aligned to generate a fused multidimensional monitoring tensor. Based on this tensor, a Hidden Markov Model (HMM) describing the evolution of the optical cable's physical state is constructed. The model is used to calculate state transition probabilities and identify abnormal state sequences matching preset fault modes. Finally, the temporal evolution of the multidimensional monitoring tensor is traced back based on the abnormal sequences to pinpoint the location and type of the initial disturbance causing the anomaly.

[0021] In one embodiment of the present invention, each distributed fiber optic sensing demodulation device is configured to collect backscattered light from the same power optical cable carrying the service, triggered by a preset time synchronization signal. Each distributed fiber optic sensing demodulation device converts the collected analog optical signal into a digitized optical intensity sequence, which is then transmitted to a central processing unit via a communication network. The central processing unit aggregates this sequence into a raw optical signal. Phase demodulation is performed on the raw optical signal to extract its composite phase information in the time and frequency domains. A blind source separation algorithm is applied to process the composite phase information, separating independent components from the mixed signal. Wavelet packet transform is performed on each separated independent component, decomposing it into different frequency bands. Time-frequency domain statistics characterizing the signal energy distribution, frequency center, and pulse characteristics are extracted from the transformed multi-frequency band signal. Based on a pre-trained vibration and temperature feature classifier, the time-frequency domain statistics are classified. Statistics classified as vibration features are aggregated into a first-class feature vector, and statistics classified as temperature features are aggregated into a second-class feature vector.

[0022] See Figure 2 A distributed fiber optic sensing demodulation device is deployed at predetermined intervals along the power optical cable. In practice, this interval can be 1 kilometer. Each distributed fiber optic sensing demodulation device integrates a laser source, a photodetector, and an analog-to-digital converter module. All distributed fiber optic sensing demodulation devices are configured to receive time synchronization pulses from the GPS timing module. When a pulse is received, each distributed fiber optic sensing demodulation device simultaneously injects a probe light pulse into the power optical cable and collects the backscattered light signal generated on the same power optical cable. Each distributed fiber optic sensing demodulation device converts the collected analog backscattered light signal into an electrical signal using a photodetector, and then converts it into a digitized light intensity sequence using an analog-to-digital converter module at a sampling rate of 100MHz. The length of the light intensity sequence matches the width of the probe light pulse and the acquisition time window. Each distributed fiber optic sensing demodulation device uploads the digitized light intensity sequence to the central processing unit located in the monitoring center through a dedicated power communication network. The central processing unit sorts and splices the light intensity sequences from different distributed fiber optic sensing demodulation devices according to timestamps and location identifiers to form a raw optical signal dataset covering the entire power optical cable.

[0023] In practical implementation, when the central processing unit demodulates the original optical signal, it uses coherent detection technology to extract the phase components of the optical signal, obtaining composite phase information containing time-domain phase fluctuations and frequency-domain phase spectra. A blind source separation algorithm based on independent component analysis is applied to process the composite phase information, separating independent vibration and temperature signal components from the mixed signal. The number of independent components corresponds to the number of potential disturbance sources along the power optical cable. Each separated independent component undergoes a three-level wavelet packet transform, decomposing it into eight different frequency bands. The frequency range of each band is determined by the wavelet basis function and the sampling rate. From the signals of each frequency band after wavelet packet transform, the signal energy, spectral centroid, and impulse factor within each time window are calculated, forming a set of time-frequency domain statistics describing the signal energy distribution, frequency center, and impulse characteristics.

[0024] In practice, the pre-trained vibration and temperature feature classifier is built based on support vector machines. The input of the classifier is time-frequency domain statistics, and the output is vibration feature labels or temperature feature labels. All time-frequency domain statistics that the classifier identifies as vibration features are aggregated in chronological order into a first-class feature vector, and the dimension of the first-class feature vector is equal to the number of statistics related to the vibration feature. All time-frequency domain statistics that the classifier identifies as temperature features are aggregated in chronological order into a second-class feature vector, and the dimension of the second-class feature vector is equal to the number of statistics related to the temperature feature.

[0025] In practice, the aggregation of feature vectors can be achieved through weighted summation, for example, the first type of feature vectors... element It is obtained by weighting multiple time-frequency domain statistics belonging to vibration characteristics, and the calculation formula is as follows:

[0026] in: Indicates the first Within the first time window Time-frequency domain statistics of a vibration characteristic These are the weighting coefficients for the corresponding statistics. This represents the total number of vibration characteristic statistics; similarly, the elements of the second type of characteristic vector are weighted and aggregated from temperature characteristic statistics in the same way.

[0027] It is understandable that the acquisition and digitization of the original optical signal relies on time-synchronized distributed fiber optic sensing demodulation equipment to ensure the temporal consistency of data from multiple locations. It is also understandable that the combination of phase demodulation, blind source separation, and wavelet packet transform can effectively distinguish signal changes caused by vibration and temperature, providing a clear foundation for subsequent feature extraction. Optionally, the sampling rate of the analog-to-digital conversion module can be adjusted according to the length of the optical cable. If the total length of the optical cable exceeds 50 kilometers, the sampling rate can be increased to 200MHz to improve spatial resolution. Optionally, the number of wavelet packet transform layers can be expanded to four to increase the fineness of frequency band division and adapt to more complex disturbance signal characteristics.

[0028] In one embodiment of the present invention, see [reference] Figure 3 To address this, location codes for the first and second class feature vectors and time codes for the corresponding acquisition timestamps are added, respectively. A spatiotemporal correlation matrix is ​​established to record the cross-correlation strength between the two types of feature vectors at different locations and time points. Based on the spatiotemporal correlation matrix, interpolation and alignment are performed on the two types of feature vectors from different locations but with similar times, mapping them to a unified spatiotemporal grid point. At the unified grid point, the aligned two types of feature vectors are concatenated along the feature dimension, and the concatenated high-dimensional vector sequence is rearranged in temporal and spatial order to construct a multidimensional monitoring tensor. A set of discrete hidden states is defined to correspond to different physical state categories of the power optical cable, and an observation probability distribution associated with each hidden state is defined to describe the probability of generating a frame of data in the multidimensional monitoring tensor under a specific state. Using historical normal monitoring data, the expectation-maximization algorithm is used to perform unsupervised training on the initial state probability, state transition probability matrix, and observation probability distribution parameters of the Hidden Markov Model. State dwell time constraints are introduced during training to simulate the stability of the physical state of the optical cable. After training, the model parameters are saved to obtain a Hidden Markov Model that can be used for state decoding.

[0029] Location and time codes are added to the first and second type feature vectors, respectively. The location code uses the distance value based on the starting point of the power fiber optic cable, and the time code uses a Unix timestamp accurate to milliseconds. A spatiotemporal correlation matrix is ​​established, where rows correspond to different acquisition locations, columns correspond to different time points, and matrix elements are the cross-correlation coefficients of the first and second type feature vectors at adjacent time points at the same location. The cross-correlation coefficients reflect the spatiotemporal correlation strength between the two types of features. Based on the spatiotemporal correlation matrix, the first and second type feature vectors from different locations with a time difference less than a preset threshold are time-aligned using a linear interpolation method, mapping all vectors to a unified time grid with fixed intervals. Simultaneously, in the spatial dimension, discrete locations along the fiber optic cable are mapped to spatial grid points with constant spacing. On a unified spatiotemporal grid, the aligned first-class and second-class feature vectors are concatenated along the feature dimension. The concatenated high-dimensional vector contains all the information of vibration features, temperature features, position coordinates and timestamps. The concatenated high-dimensional vector sequence is arranged in chronological order, and the vectors at different positions at the same time point are arranged in spatial order to construct a three-dimensional multidimensional monitoring tensor. The three dimensions of the multidimensional monitoring tensor correspond to time, space and features, respectively.

[0030] A set of discrete hidden states is defined, the number of which is determined based on the possible physical state categories of the power fiber optic cable, including normal operation, slight vibration interference, continuous temperature rise, mechanical damage, and fiber breakage. Each hidden state corresponds to one physical state category. An observation probability distribution associated with each hidden state is defined, using a multivariate Gaussian distribution to describe the probability of generating a frame of data in a multidimensional monitoring tensor under a specific physical state. A frame of data in the multidimensional monitoring tensor is the set of feature vectors of all spatial locations on a time slice. Using historical normal power fiber optic cable monitoring data, the expectation-maximization algorithm is used to perform unsupervised training on the initial state probability, state transition probability matrix, and mean vector and covariance matrix of the observation probability distribution of the Hidden Markov Model. The historical normal data consists of 30 consecutive days of fault-free monitoring records. During training, a state dwell time constraint is introduced, limiting the duration of the same hidden state to no less than the minimum dwell time, set at 10 sampling intervals, to simulate the relatively stable characteristics of the actual physical state of the power fiber optic cable. After training, all parameters of the Hidden Markov Model (HMM) are saved, including the initial state probability vector, the state transition probability matrix, and the mean vector and covariance matrix of the multivariate Gaussian distribution corresponding to each state, resulting in an HMM that can be used for state decoding. In some embodiments, the elements of the spatiotemporal correlation matrix are calculated from the cosine similarity between the first-class feature vector and the second-class feature vector at the same time step, using the following formula:

[0031] in: Indicates the first First type feature vector at each location With the Second type feature vector at each position Cross-correlation coefficients between them Represents the vector dot product. This represents the vector magnitude. In some embodiments, the time interval between uniform spatiotemporal grid points is set to 1 second, and the spatial interval is set to 100 meters, to meet the conventional spatiotemporal resolution requirements of power fiber optic cable monitoring.

[0032] It is understandable that by adding location and time encoding and establishing a spatiotemporal correlation matrix, the dependence of vibration and temperature characteristics along the optical cable direction and time axis can be effectively captured. It is also understandable that using a Hidden Markov Model to model the optical cable state evolution can utilize sequence probability characteristics to characterize the transition patterns between states, providing a theoretical basis for subsequent anomaly identification. Optionally, location encoding can use latitude and longitude values ​​in a geographic coordinate system, suitable for non-linear power optical cable paths; optionally, the observation probability distribution can be a Gaussian mixture model to better fit the complex distribution pattern of the multidimensional monitoring tensor data.

[0033] In one embodiment of the present invention, a multidimensional monitoring tensor is input sequentially into a trained Hidden Markov Model (HMM), and the Viterbi algorithm is used to decode the hidden state sequence, calculating the probability of being in each hidden state at each time step. All adjacent state transition events are extracted from the decoded state sequence, and the ratio of the actual probability of each transition event to its probability in historical normal data is calculated based on the state transition probability matrix. Transition events with ratios lower than or higher than a preset threshold are marked as anomalous transitions. Multiple consecutive anomalous transitions and associated hidden states are combined into a candidate anomalous state sequence. The candidate sequence is then subjected to pattern matching with predefined fault mode state sequences in a knowledge base, and the candidate sequence whose similarity to any fault mode sequence exceeds a matching threshold is determined as the final anomalous state sequence.

[0034] The multidimensional monitoring tensor is input sequentially into the trained Hidden Markov Model (HMM). The temporal dimension of the multidimensional monitoring tensor contains continuous monitoring frames, with each frame corresponding to spatial-feature data at a sampling time. The Viterbi algorithm is used to decode the optimal hidden state sequence. The Viterbi algorithm, based on dynamic programming, finds the hidden state path most likely to generate the observation sequence. During decoding, the posterior probability of the model being in each hidden state at each time step is calculated and recorded simultaneously. The posterior probability reflects the likelihood of the model being in a certain state given the observed data. From the decoded hidden state sequence, all state transition events between adjacent time steps are extracted. Each state transition event is represented by a state pair consisting of the previous state and the current state. Based on the state transition probability matrix of the HMM, the theoretical probability of each state transition event in the model is queried, i.e., the value of the corresponding element in the state transition probability matrix. The actual frequency of this state transition event in the historical normal dataset is obtained as its empirical probability in the historical normal dataset. The ratio of the actual probability of each state transition event to the historical empirical probability is calculated. A ratio greater than 1 indicates a transition frequency higher than the historical normal, while a ratio less than 1 indicates a frequency lower than the historical normal. State transition events with a ratio below a preset lower threshold of 0.5 or above a preset upper threshold of 2.0 are marked as abnormal transitions. Abnormal transitions mean that the state change does not conform to the normal behavior pattern. Multiple abnormal transitions that occur consecutively in time and their associated intermediate hidden states are combined in chronological order to form a candidate abnormal state sequence. The length of the candidate sequence must contain at least two abnormal transitions.

[0035] Candidate anomalous state sequences are matched against predefined fault mode state sequences in the knowledge base. The matching process calculates the edit distance between the candidate sequence and each fault mode sequence. The edit distance is defined as the minimum number of state replacement, insertion, or deletion operations required to make the two sequences equal. Candidate anomalous state sequences whose normalized edit distance similarity with any fault mode state sequence exceeds a matching threshold of 0.8 are determined as the final anomalous state sequences. The formula for calculating the normalized edit distance similarity is:

[0036] in: To edit distance, The maximum value of the lengths of the two sequences. Similarity scores are assigned. See Table 1 for example data comparing state transition probabilities and anomaly detection: Table 1: Probability of State Transition Events and Anomaly Judgment Table State transition (from beginning to end) Model transition probability Historical experience probability probability ratio Is it abnormal? Normal → Slight vibration 0.15 0.14 1.07 no Slight vibration → mechanical damage 0.02 0.001 20.0 yes Normal → Continuous temperature rise 0.08 0.09 0.89 no Continuous temperature rise → Fiber breakage fault 0.05 0.01 5.0 yes In some embodiments, the actual probability of a state transition event can be directly adopted as the normalized value of the forward-backward probability product calculated during Viterbi decoding to more accurately reflect the transition probability under the current observation sequence. In some embodiments, the screening of candidate abnormal state sequences can be constrained by length, retaining only sequences containing at least three consecutive abnormal transitions to reduce false alarms from transient disturbances. It is understood that by comparing the actual probability of state transition with the historical baseline probability, the degree of abnormality in state evolution can be quantified, effectively distinguishing random fluctuations from real fault precursors. It is also understood that using edit distance for pattern matching can flexibly address individual differences in the fault development process and improve the robustness of anomaly identification. Optionally, the threshold for the probability ratio can be dynamically adjusted according to the monitoring sensitivity requirements; for high-security scenarios, the lower limit can be set to 0.3 and the upper limit to 3.0. Optionally, the similarity matching threshold can be set with differentiated values ​​for different types of fault modes, for example, 0.75 for mechanical damage faults and 0.9 for fiber breakage faults.

[0037] In one embodiment of the present invention, a large number of historical power fiber optic cable fault cases are collected. Each case contains monitoring data of the entire process from normal operation to fault. For each case, a multidimensional monitoring tensor is generated using the same process as described above, and the hidden state sequence is decoded. Domain experts annotate the hidden state sequences, marking key state subsequences characterizing the occurrence and development of the fault. Key subsequences of similar faults are extracted from different cases and subjected to cluster analysis to obtain typical state evolution patterns for each type of fault. The typical evolution patterns of each type of fault are encoded as state symbol sequences and stored as predefined fault mode state sequences in a knowledge base.

[0038] A large number of historical power fiber optic cable fault cases were collected, including five years of operation records and laboratory simulated fault tests of the same power fiber optic cable. Each case contained multi-channel monitoring data of the entire process from normal state to fault state, covering backscattered light intensity, phase fluctuation, and time-frequency domain characteristics. For the monitoring data of each historical case, the raw optical signal acquisition, feature demodulation and separation, spatiotemporal correlation alignment, and hidden Markov model decoding were performed sequentially to generate a corresponding multidimensional monitoring tensor and decode a complete hidden state sequence. Each symbol in the hidden state sequence represents the physical state category of the power fiber optic cable at a sampling time. An expert group composed of more than three power fiber optic cable operation and maintenance experts with more than ten years of experience independently labeled the hidden state sequence of each historical case. Based on the corresponding field logs and fault reports, the experts marked the key state subsequences representing the occurrence and development of the fault in the sequence. The starting point of the key state subsequence was the time point when the abnormal state first appeared, and the ending point was the time point when the fault was confirmed. The expert group cross-validated all independent labeling results. When there were disagreements, consensus was reached through discussion, forming a set of consistently labeled key state subsequences.

[0039] Key state subsequences representing similar faults are extracted from different historical cases. For example, all subsequences of mechanical damage caused by external construction are grouped into one category, and all subsequences of abnormal temperature rise caused by insulation aging are grouped into another. Cluster analysis is performed on the key state subsequences of each fault category. The clustering algorithm uses a distance metric based on dynamic time warping, which can handle the differences in fault development speed in different cases. Through cluster analysis, the subsequences of each fault category are divided into several clusters, each cluster representing a typical evolution pattern of that fault category. The central sequence of each cluster is selected as the representative of that pattern. The typical state evolution pattern of each fault category is encoded into a fixed-length state symbol sequence according to the state symbol system of a Hidden Markov Model. Each position in the state symbol sequence corresponds to a specific state, which is stored as a predefined fault pattern state sequence in the knowledge base. During storage, metadata such as fault type, typical duration, and confidence level are associated. See Table 2, which shows the pattern encoding of mechanical damage faults: Table 2: Failure Mode State Sequence Table for Mechanical Damage Fault Mode Number State symbol sequence Typical duration (minutes) Data source: number of cases M-001 N→V1→V2→D1→F 12 8 M-002 N→V1→D1→F 8 5 M-003 N→V2→D2→F 18 3 Note: N represents the normal state, V1 / V2 represents different levels of vibration, D1 / D2 represents different types of mechanical damage, and F represents fiber breakage fault.

[0040] In some embodiments, cluster analysis may employ hierarchical clustering methods, controlling the granularity of clusters by setting distance thresholds to ensure high consistency in sequence morphology within each cluster. In some embodiments, wildcard symbols may be introduced into the encoding of fault mode state sequences to represent multiple possible states at a given location, adapting to the uncertainty in the fault evolution process. It is understood that combining expert annotation with cluster analysis can extract common patterns from diverse historical fault data, forming a standardized fault mode library. It is also understood that using dynamic time warping to address time series alignment issues can effectively capture the temporal differences in fault development across different cases. Optionally, during expert annotation, visualization tools can be used to display the correspondence between hidden state sequences and original monitoring data, assisting in determining the boundaries of key subsequences. Optionally, the knowledge base can store multiple representations of fault mode state sequences, including original state sequences and smoothed simplified sequences, to accommodate matching requirements of varying precision.

[0041] In one embodiment of the present invention, the first time point deviating from the normal state range in the abnormal state sequence is identified as the abnormal starting point. All data within a time window centered on this point are extracted from the multidimensional monitoring tensor. The changes in the energy distribution of the tensor in the spatial dimension within the window are analyzed, and one or more locations near the abnormal starting point where the spatial energy suddenly increases or decreases are located as suspected disturbance source locations. Combining the fault mode type matched by the abnormal sequence and the performance of the first and second type feature vectors near the suspected source, the initial disturbance type is comprehensively determined. Based on the located suspected location and the determined disturbance type, a corresponding typical disturbance signal template is called from the knowledge base, and matched filtering is performed in the corresponding spatiotemporal region of the multidimensional monitoring tensor to enhance and extract the complete disturbance signal waveform. The extracted waveform is parametrically fitted to obtain a set of key parameters describing the disturbance event, such as amplitude, duration, and frequency components. The key parameters are input into a pre-trained disturbance verification model. The model outputs a confidence score of the abnormal sequence caused by the disturbance event. When the score exceeds the verification threshold, the disturbance event is confirmed as the root cause, and a complete disturbance event report containing location, type, time, and key parameters is generated.

[0042] From the sequence of abnormal states, the time point corresponding to the first hidden state deviating from the normal state range is identified as the anomaly initiation point. The normal state range is determined by the distribution of hidden states in historical normal monitoring data; states exceeding the distribution boundary are considered deviations. In the multidimensional monitoring tensor, all data within a set time window centered on the anomaly initiation point and extending before and after it are extracted. The time window length is set to 60 seconds, covering the evolution process before and after the anomaly initiation. The changes in the energy distribution of the multidimensional monitoring tensor in the spatial dimension within the time window are analyzed. Spatial energy is calculated as the norm sum of the squares of the vibration feature vector and the temperature feature vector at each spatial location, reflecting the overall signal strength at that location. One or more spatial locations near the anomaly initiation point where spatial energy experiences a sudden increase or a sharp decrease are identified as suspected disturbance source locations. A sudden increase is defined as energy exceeding three standard deviations of the average energy of the previous time window, and a sharp decrease is defined as energy falling below three standard deviations of the average energy. By combining the fault mode type matched by the abnormal state sequence, and the performance of the first and second type feature vectors near the suspected disturbance source location, the type of initial disturbance is comprehensively judged. For example, if the fault mode is mechanical damage and the vibration characteristics are significantly increased, it is determined to be an external force impact disturbance.

[0043] Based on the located suspected disturbance source and the determined initial disturbance type, a typical disturbance signal template is retrieved from the knowledge base. This template is derived from standard signal waveforms extracted from historical cases of similar faults. Using this template, matched filtering is performed within the corresponding spatiotemporal region of the multidimensional monitoring tensor. Matched filtering is achieved by convolving the template with local data to enhance and extract the complete disturbance signal waveform. The extracted disturbance signal waveform is then parametrically fitted using a damped sine function to obtain a set of key parameters describing the disturbance event, including amplitude, duration, attenuation coefficient, and dominant frequency. This set of key parameters is input into a pre-trained disturbance verification model, which is a multilayer perceptron-based classifier. The model outputs a confidence score (range 0 to 1) indicating that the disturbance event leads to an abnormal state sequence. When the confidence score exceeds the verification threshold of 0.85, the disturbance event is confirmed as the root cause of the abnormal state sequence, and a complete disturbance event report containing location, type, time, and key parameters is generated. The report records the precise time of the disturbance, the distance from the optical cable starting point, the disturbance type, and details of the fitted parameters.

[0044] In some embodiments, the spatial energy calculation may use only the norm of the first type of eigenvector to examine the impact of vibration disturbances alone, which is suitable for scenarios that exclude temperature change interference. In some embodiments, the template for matched filtering may be selected with different waveform shapes according to the type of disturbance, such as a short pulse template for impact disturbances and a long period template for continuous disturbances. It can be understood that by tracing back the spatiotemporal energy changes of the multidimensional monitoring tensor, the physical disturbance source can be deduced from the state anomaly, achieving end-to-end fault tracing. It can also be understood that by combining parametric fitting with a verification model, the causal relationship between disturbances and state anomalies can be quantitatively evaluated, improving the reliability of the location results. Optionally, the time window length can be dynamically adjusted according to the fault mode, with 30 seconds for rapid faults and 120 seconds for slow, gradual faults. Optionally, the threshold of the disturbance verification model can be set according to the fault level, with 0.9 for high-risk faults and 0.8 for general faults.

[0045] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for monitoring power optical cables, characterized in that, include: The original optical signal is acquired at multiple predetermined locations on the power optical cable, and the original optical signal contains characteristic changes caused by external disturbances; The original optical signal is subjected to feature demodulation and separation processing to obtain a first type of feature vector reflecting the vibration mode and a second type of feature vector reflecting the temperature change; The first type of feature vector and the second type of feature vector are spatiotemporally correlated and aligned to generate a fused multidimensional monitoring tensor. Based on the multidimensional monitoring tensor, a hidden Markov model describing the evolution of the physical state of the optical cable is constructed. The transition probabilities between different states are calculated using the hidden Markov model, and abnormal state sequences that match the preset fault mode are identified. Based on the abnormal state sequence, the evolution of the multidimensional monitoring tensor on the time axis is traced back to locate the position and type of the initial disturbance that caused the abnormal state sequence.

2. The power optical cable monitoring method as described in claim 1, characterized in that, The acquisition of raw optical signals at multiple predetermined locations on the power optical cable includes: Multiple distributed fiber optic sensing and demodulation devices are deployed along the power optical cable; Configure each distributed fiber optic sensing demodulation device to collect backscattered light from the same power optical cable carrying the service when triggered by a preset time synchronization signal. Each distributed fiber optic sensing demodulation device converts the acquired analog optical signal into a digital optical intensity sequence; The digitized light intensity sequences collected by multiple distributed fiber optic sensing and demodulation devices are uploaded to the central processing unit via a communication network, where they are aggregated into the original light signal.

3. The power optical cable monitoring method as described in claim 1, characterized in that, The original optical signal is subjected to feature demodulation and separation processing to obtain a first type of feature vector reflecting the vibration mode and a second type of feature vector reflecting temperature change, including: The original optical signal is phase demodulated to extract its composite phase information in the time and frequency domains; The composite phase information is processed using a blind source separation algorithm to separate independent components from the mixed signal; Each isolated independent component is decomposed into different frequency bands by performing wavelet packet transform. From multiple frequency band signals that have undergone wavelet packet transform, time-frequency domain statistics characterizing the signal energy distribution, frequency center, and pulse characteristics are extracted. The time-frequency domain statistics are classified based on a pre-trained vibration and temperature feature classifier. Time-frequency domain statistics classified as vibration features are aggregated into the first type of feature vector, and time-frequency domain statistics classified as temperature features are aggregated into the second type of feature vector.

4. The power optical cable monitoring method as described in claim 2, characterized in that, The first type of feature vector and the second type of feature vector are spatiotemporally correlated and aligned to generate a fused multidimensional monitoring tensor, including: Add a location code corresponding to the acquisition location and a time code corresponding to the acquisition timestamp to the first type of feature vector and the second type of feature vector, respectively; Establish a spatiotemporal correlation matrix to record the cross-correlation strength between the first type of feature vector and the second type of feature vector at different locations and time points; Based on the spatiotemporal correlation matrix, interpolation and alignment are performed on the first type of feature vectors and the second type of feature vectors from different locations but close in time, so that they are mapped to a unified spatiotemporal grid point; On a unified spatiotemporal grid point, the aligned first-type feature vector and the second-type feature vector are concatenated along the feature dimension. The spliced ​​high-dimensional vector sequence is rearranged according to time and spatial order to construct the multidimensional monitoring tensor.

5. The power optical cable monitoring method as described in claim 4, characterized in that, Based on the aforementioned multidimensional monitoring tensor, a hidden Markov model describing the evolution of the physical state of the optical cable is constructed, including: Define a set of discrete hidden states, each hidden state corresponding to a physical state category of a power optical cable; Define an observation probability distribution associated with each hidden state, the observation probability distribution being used to describe the probability of generating a frame of data in the multidimensional monitoring tensor under a specific physical state; Using historical normal power fiber optic cable monitoring data, the parameters of the initial state probability, state transition probability matrix, and observation probability distribution of the Hidden Markov Model are trained unsupervised using the expectation-maximization algorithm. During the training process, state dwell time constraints are introduced to simulate the relatively stable physical state of the power optical cable. After training is complete, all parameters of the Hidden Markov Model are saved to obtain a Hidden Markov Model that can be used for state decoding.

6. The power optical cable monitoring method as described in claim 5, characterized in that, The hidden Markov model is used to calculate the transition probabilities between different states and to identify abnormal state sequences that match a preset fault mode, including: The multidimensional monitoring tensor is input into the trained Hidden Markov Model in chronological order. The Viterbi algorithm is used to decode the hidden state sequence and simultaneously calculate the probability of being in each hidden state at each time step. Extract all state transition events between adjacent time steps from the decoded hidden state sequence; Based on the state transition probability matrix, calculate the ratio of the actual probability of each state transition event to the probability of the transition occurring in historical normal data; State transition events where the ratio is lower than or higher than a preset threshold are marked as abnormal transitions. Multiple consecutive abnormal transitions and their associated hidden states are combined into a candidate abnormal state sequence. The candidate abnormal state sequence is matched with the predefined fault mode state sequence in the knowledge base. Candidate abnormal state sequences whose similarity to any fault mode state sequence exceeds a matching threshold are determined as the final abnormal state sequence.

7. The power optical cable monitoring method as described in claim 6, characterized in that, The methods for constructing the predefined fault mode state sequences in the knowledge base include: We collected a large number of historical power fiber optic cable fault cases, each case containing monitoring data of the entire process from normal state to fault state; For the monitoring data of each historical case, the original optical signal acquisition, feature demodulation and separation, spatiotemporal correlation alignment and hidden Markov model decoding are performed in sequence to generate the corresponding multidimensional monitoring tensor and decode the complete hidden state sequence. Domain experts annotate the hidden state sequence, marking out the key state subsequences that characterize the occurrence and development of the fault; Key state subsequences representing the same type of fault were extracted from different cases, and cluster analysis was performed to obtain the typical state evolution pattern of each type of fault. The typical state evolution pattern of each type of fault is encoded into a state symbol sequence and stored as a predefined fault mode state sequence in the knowledge base.

8. The power optical cable monitoring method as described in claim 6, characterized in that, Based on the abnormal state sequence, the evolution of the multidimensional monitoring tensor on the time axis is traced back to locate the position and type of the initial perturbation that caused the abnormal state sequence, including: From the abnormal state sequence, the first time point that deviates from the normal state range is identified as the abnormal starting point; Extract all data within a time window centered on the anomaly initiation point from the multidimensional monitoring tensor; Analyze the changes in the energy distribution of the multidimensional monitoring tensor in the spatial dimension within the time window; One or more spatial locations near the anomaly initiation point where spatial energy suddenly increases or decreases are identified as suspected disturbance source locations. The type of the initial disturbance is determined by combining the fault mode type matched by the abnormal state sequence, as well as the performance of the first type of feature vector and the second type of feature vector near the suspected disturbance source location.

9. The power optical cable monitoring method as described in claim 8, characterized in that, The method also includes steps for reconstructing and verifying the perturbation event: Based on the located suspected disturbance source and the determined type of the initial disturbance, the corresponding typical disturbance signal template is retrieved from the knowledge base; Using the typical disturbance signal template, matched filtering is performed in the corresponding spatiotemporal region of the multidimensional monitoring tensor to enhance and extract the complete disturbance signal waveform; The extracted disturbance signal waveform is parametrically fitted to obtain a set of key parameters describing the disturbance event, including amplitude, duration, and frequency components. The set of key parameters is input into a pre-trained perturbation verification model, which outputs a confidence score of the abnormal state sequence caused by the perturbation event. When the confidence score exceeds the verification threshold, the perturbation event is confirmed as the root cause of the abnormal state sequence, and a complete perturbation event report containing location, type, time, and key parameters is generated.

10. A power fiber optic cable monitoring device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the power optical cable monitoring method as described in any one of claims 1 to 9.

Images