Optical access network PON port fault prediction and operation and maintenance support method and system

Abstract

The invention provides an optical access network PON port fault prediction and operation and maintenance support method and system, and the method comprises the steps: collecting real-time multi-index sequences of optical power, bit error rate, temperature, frame loss rate, business load and the like, automatically extracting a long-term trend based on variational mode decomposition, constructing a residual sequence, generating a state transition map through employing a Markov transition field algorithm, and carrying out the fault prediction and operation and maintenance support of the PON port. And future operation state distribution prediction is realized through the LSTM neural network. And after an abnormal threshold is dynamically adjusted in combination with quantile regression, a time sequence prediction result and a threshold response are fused to output a comprehensive fault probability score, and an early warning signal is triggered through exponential smoothing and continuity check, so that the progressive fault detection accuracy and advancement can be improved, and an efficient and steady prediction and early warning basis is provided for intelligent operation and maintenance of the PON port.

Description

Technical Field

[0001] This invention relates to the field of dynamic monitoring and fault prediction technology for PON port operation status, and in particular to a method and system for fault prediction and operation and maintenance support of PON port in optical access network. Background Technology

[0002] In the current field of intelligent operation and maintenance and fault prediction for power optical access networks (PON), anomaly detection and operation and maintenance management solutions based on real-time monitoring data have been widely adopted. Mainstream systems typically collect key operating parameters of PON ports, such as optical power, bit error rate, temperature, frame loss rate, and service load, and combine this with simple statistical analysis or single-threshold-based judgment methods to detect and warn of abnormal port states. Furthermore, some high-end solutions have introduced time series analysis and machine learning methods, using historical feature extraction, trend analysis, and deep learning models to achieve preliminary automated identification and prediction of port faults and performance degradation, laying the foundation for improving the efficiency and accuracy of network operation and maintenance.

[0003] However, the existing technologies for PON port fault monitoring and prediction still have the following prominent shortcomings: (1) Fixed thresholds are difficult to adapt to the dynamic operating environment of PON ports and cannot be adjusted in a timely manner according to changes in actual services, environment or resources. False alarms and missed detections occur frequently, resulting in low reliability of the early warning system. (2) Existing methods rely on manual experience to adjust parameters for long-term drift and periodic disturbances of port operating baseline fluctuations, and cannot achieve automatic and robust adjustment of abnormal boundaries; (3) When complex trends such as gradual degradation and population aging occur, the original technology is slow to respond when the signal noise is large or the abnormal characteristics are not obvious, and cannot achieve early and sensitive risk intervention. (4) Although deep learning time series modeling has predictive capabilities, it is difficult to effectively combine with real-time monitoring strategies if it lacks dynamic threshold adaptive compensation, making actual deployment and implementation difficult. (5) As the scale of optical access network expands and the density of access terminals increases, relying solely on hardware upgrades or single rule optimizations is insufficient to support the higher requirements of intelligent operation and maintenance and refined fault prediction. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method and system for predicting and supporting the operation and maintenance of PON ports in optical access networks.

[0005] The technical solution of this invention is implemented as follows: A method for fault prediction and operation and maintenance support of PON port in optical access network, comprising: S1: Collect PON port runtime sequence data, including optical power, bit error rate, temperature, frame loss rate, and service load indicators, to form a set of raw observation sequences.

[0006] S2: Perform variational mode decomposition (VMD) on the original observation sequence, and separate the low-frequency dominant component as a dynamic baseline based on the preset number of decomposition layers and penalty factor.

[0007] S3: Calculate the residual sequence between the current observation and the dynamic baseline, and convert the residual sequence into a two-dimensional state transition map based on the Markov Transition Field (MTF) algorithm, wherein the state division adopts the equal-width bin method to set the discrete state.

[0008] S4: Input the state transition map into the Long Short-Term Memory (LSTM) network model, and use the neural network to train and predict the state transition probability distribution for multiple future time steps.

[0009] S5: Construct a quantile regression model based on historical residual sequences, use linear programming to estimate quantile thresholds, and dynamically adjust quantile levels according to the variance change rate Δσ² within the sliding window.

[0010] S6: Combine the abnormal state probability P output by LSTM with the adaptive threshold Th to calculate the comprehensive fault probability score.

[0011] S7: Perform exponential smoothing on the fault probability score sequence, use a smoothing factor to generate a stationary prediction curve, and trigger a progressive fault warning when the predicted value exceeds the threshold Th for multiple consecutive time steps.

[0012] The present invention also provides a PON port fault prediction and operation and maintenance support system for optical access networks, which uses the above-mentioned PON port fault prediction and operation and maintenance support method to perform PON port fault prediction and operation and maintenance support for optical access networks.

[0013] The present invention provides a method and system for PON port fault prediction and operation and maintenance support in optical access networks, which has the following beneficial effects: (1) This invention significantly improves the sensitivity and stability of anomaly detection by constructing a dual discrimination mechanism of "dynamic baseline + adaptive threshold". It uses variational mode decomposition (VMD) to extract low-frequency dominant components from multi-dimensional time series data as a dynamic baseline reflecting the normal operation trend of the port, effectively removing non-fault disturbances caused by business cycle fluctuations and environmental changes, and avoiding misjudging normal fluctuations as anomalies. On this basis, a dynamic threshold adjustment strategy based on quantile regression is introduced, which adaptively adjusts the warning boundary according to the historical distribution characteristics of the residual sequence and the real-time fluctuation intensity. During the stable operation period of the system, a high confidence level (such as 98% quantile) is used to suppress false alarms, and when the fluctuation intensifies, it is automatically downgraded to a lower quantile (such as 90%) to capture potential signs of degradation in advance. This mechanism overcomes the dependence of traditional methods on a single threshold and realizes the synergistic optimization of detection sensitivity and robustness. It is especially suitable for identifying progressive performance degradation caused by group aging in high-density access environments. (2) This invention innovatively integrates Markov Transition Field (MTF) and Long Short-Term Memory Network (LSTM) to construct a state transition map prediction framework for time-series residuals. By transforming the one-dimensional residual sequence into a two-dimensional state transition matrix, MTF not only preserves the temporal dependencies of the time series but also visualizes the transition patterns between port operating states, enhancing the spatial expressive power of features; the LSTM model performs deep learning on the historical state transition map, explores its evolution patterns on the time axis, and outputs the probability distribution of future abnormal states, thereby realizing a forward-looking inference from "current deviation" to "future risk". Compared with the traditional method that only judges based on instantaneous threshold exceedance, this design significantly improves the early identification capability of slowly deteriorating faults, and can issue warnings before performance indicators exceed limits, gaining a valuable window of opportunity for operation and maintenance intervention, and significantly improving the timeliness and accuracy of fault prediction; (3) This invention generates a comprehensive fault occurrence probability score by fusing the comparison results of LSTM prediction probability and adaptive threshold, and outputs a stable prediction curve by combining exponential smoothing technology, forming a closed-loop, interpretable, and manual parameter tuning-free intelligent diagnostic process. The entire method is based entirely on conventional monitoring data such as optical power, bit error rate, and temperature collected by the existing operation and maintenance system, without the need to deploy additional sensors or upgrade hardware, and has good engineering feasibility and system compatibility; at the same time, the dynamic baseline construction, state graph transformation, and probability fusion mechanism jointly ensure the algorithm's generalization ability under different equipment models, load conditions, and climatic environments, significantly enhancing the system's scalability and deployment flexibility. Attached Figure Description

[0014] Figure 1 This is a flowchart of a method for predicting and supporting PON port failures in an optical access network according to the present invention. Figure 2This is a sub-flowchart of a method for predicting and supporting PON port failures in an optical access network according to the present invention. Figure 3 This is another sub-flowchart of the method for predicting and supporting PON port failures in optical access networks according to the present invention. Detailed Implementation

[0015] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0016] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0017] like Figure 1 As shown, this invention provides a method for PON port fault prediction and operation and maintenance support in optical access networks, specifically including: S1: Collect PON port runtime sequence data, including optical power, bit error rate, temperature, frame loss rate and service load indicators, to form a set of raw observation sequences; S2: Perform variational mode decomposition (VMD) on the original observation sequence. Based on the preset number of decomposition layers K=5 and penalty factor α=2000, separate the low-frequency dominant component as the dynamic baseline. S3: Calculate the residual sequence between the current observation and the dynamic baseline, and convert the residual sequence into a two-dimensional state transition map based on the Markov Transition Field (MTF) algorithm, wherein the state division adopts the equal-width binning method to set N=8 discrete states; S4: Input the state transition map into the Long Short-Term Memory (LSTM) network model, and train the neural network with 2 hidden layers (64 neurons per layer) to predict the state transition probability distribution for the next T=3 time steps; S5: Construct a quantile regression model based on historical residual sequences, use linear programming to estimate the 95th quantile threshold, and dynamically adjust the quantile level according to the variance change rate Δσ² within the sliding window (window length W=30): switch to the 90th quantile when Δσ²≥0.15, and switch to the 98th quantile when Δσ²≤0.05. S6: Combine the abnormal state probability P output by LSTM with the adaptive threshold Th to calculate the comprehensive fault probability score FP=α·P + (1-α)·(1-|e| / Th), where α=0.7 is the weighting coefficient and e is the current residual value; S7: Perform exponential smoothing on the fault probability score sequence, use a smoothing factor β=0.3 to generate a stationary prediction curve, and trigger a progressive fault warning when the predicted value exceeds the threshold Th for M=5 consecutive time steps.

[0018] Step S1: Collect PON port runtime sequence data, including optical power, bit error rate, temperature, frame loss rate, and service load indicators, to form a raw observation sequence set. Specifically, this includes: S1.1: Based on the PON port operation and maintenance management system interface, acquire optical power monitoring signals to collect real-time data of downlink optical signal strength and uplink received optical signal strength, and obtain optical power time series to reflect the working status of port optical modules; Based on the interface call configuration of the PON port operation and maintenance management system, a bidirectional data channel is established with the OLT device and ONU port to realize the synchronous acquisition of downlink optical signal strength and uplink received optical signal strength. A high-precision optical power monitoring method (sampling period Δt=1s, quantization accuracy 0.01dBm) is adopted to obtain the instantaneous values ​​of downlink and uplink optical power, and record them in the form of a unified timestamp at the interface layer to achieve time consistency of cross-directional signals; Furthermore, a sliding window filtering algorithm (parameter: window length W=10) is used to perform low-pass smoothing on the acquired instantaneous optical power signal and remove isolated noise points exceeding three times the median absolute deviation to generate a robust optical power base sequence. Furthermore, by using the standard deviation normalization method, the smoothed downlink and uplink optical power sequences are converted into zero-mean and unit-variance forms, respectively, to obtain bidirectional optical power normalized sequences that are comparable and fusionable. Furthermore, the difference sequence between downlink and uplink optical power is calculated to capture potential optical loss asymmetry characteristics at the port on the bidirectional link. This calculation is performed using the following formula:

[0019] in, This is the normalized value of the downlink optical power. This is the normalized value of the uplink optical power. It is a two-way power difference sequence; By calculating the statistical characteristics of the difference sequence, the operational stability of the port optical module is initially determined, and the above results are formed into an optical power time series, which is used as part of the subsequent multidimensional original observation sequence to achieve an effective characterization of the operating status of the port optical module. For example, in a gigabit PON link, the port device's interface call configuration sets the sampling period Δt=1s and sampling accuracy to 0.01dBm, continuously collecting downlink and uplink optical power signals for 300s. During the sampling phase, the instantaneous downlink optical power fluctuates between -2.34dBm and -2.29dBm, and the instantaneous uplink optical power fluctuates between -2.36dBm and -2.31dBm. After a moving average filter with a window length W=10, the short-time standard deviation of the downlink optical power sequence is reduced to 0.002dBm, and the short-time standard deviation of the uplink optical power sequence is reduced to 0.003dBm. After standard deviation normalization, the mean of the bidirectional sequence is 0, and the variance is 1. (The formula is used to...) The calculated bidirectional power difference sequence remained within a narrow fluctuation range of ±0.005, indicating that the port optical module was in a highly stable state. This optical power time series, as the raw observation data for a unified time reference, was then used in subsequent signal decomposition and feature extraction stages, effectively improving the sensitivity and accuracy of progressive fault detection. S1.2: Using the bit error rate statistics module built into the OLT device, obtain the bit error count before and after forward error correction (FEC), and perform sliding window averaging on the bit error rate to generate a bit error rate time series for evaluating the stability of data transmission. S1.3: The temperature sensor integrated into the PON port collects internal temperature data of the optical module and generates a temperature time series based on the temperature sampling period to monitor the thermal stability and aging trend of the optical module. S1.4: Based on the network layer performance monitoring protocol, obtain the frame loss rate index, normalize the number of frames lost in a continuous time period, and generate a frame loss rate time series to identify abnormal fluctuations in the data transmission link. S1.5: Collect service load data of PON ports through the traffic statistics module, including uplink and downlink bandwidth utilization, and generate service load time series based on the timestamp alignment mechanism to analyze the operating pressure of the port under different service cycles; S1.6: Perform timestamp alignment and missing value interpolation on the collected optical power, bit error rate, temperature, frame loss rate and service load time series to generate a set of original observation sequences under a unified time reference, which will serve as input data for subsequent signal decomposition and feature extraction.

[0020] Step S2: Perform variational mode decomposition (VMD) on the original observation sequence. Based on a preset decomposition level K=5 and a penalty factor α=2000, separate the low-frequency dominant component as the dynamic baseline. Specifically, this includes: S2.1: Based on the raw runtime sequence data collected from the PON port, construct the raw observation signal sequence X(t) as the input signal of the variational mode decomposition algorithm. X(t) includes multi-dimensional indicators such as optical power, bit error rate, temperature, frame loss rate and service load. After normalization, it forms a single variable time series input. Based on multi-dimensional runtime sequence data such as optical power, bit error rate, temperature, frame loss rate and service load collected by PON port, a unified time base sequence after timestamp alignment and missing value interpolation is used as the input set. The feature fusion method (parameter: the weight of multi-dimensional index synthesis can be set to equal weight) is adopted to realize the synthesis of multi-dimensional time-series signals. The optical power, bit error rate, temperature, frame loss rate and service load sequence are spliced ​​and weighted point by point according to the time index to obtain a preliminary fusion signal reflecting the overall operating status of the port. Furthermore, by using a normalization method (parameter: Z-score normalization), scale unification is achieved for numerical signals of different dimensions and ranges, based on the formula...

[0021] in, These are the original values. This is the mean of the indicator. The standard deviation of the index is used to obtain a standardized signal sequence with a mean of zero and a variance of one, thus eliminating amplitude deviation between different indices; Furthermore, by using principal component analysis (PCA) (parameter: retaining the top principal components with a cumulative contribution rate of 0.9), we can achieve dimensionality reduction of the multidimensional standardized signal, construct a mapping matrix, compress the original multidimensional data into a univariate time series, retain the main trend information and remove redundant noise features; Furthermore, a bandpass filtering algorithm (parameters: low cutoff frequency 0Hz, high cutoff frequency half the sampling frequency) is used to preprocess the univariate time series, suppressing high-frequency noise components and retaining the effective information frequency bands for subsequent VMD decomposition, resulting in a smooth and continuous original observation signal sequence. ; Through the above algorithms and processing methods, multidimensional indicator data are transformed into well-structured univariate time series, realizing a unified data entry point before dynamic baseline extraction; For example, in a gigabit PON access scenario, the sampling period for port operation indicators is 1 minute, and the total sampling time is 1440 minutes, resulting in an optical power sequence of length 1440. The bit error rate, temperature, frame loss rate, and service load sequences have consistent lengths. The five sequences are weighted and fused (each weight is set to 0.2) to obtain a fused signal sequence with a mean of 17.5 and a standard deviation of 4.1. Z-score normalization is applied to the fused signal to obtain a standardized sequence with a mean of 0 and a standard deviation of 1. PCA dimensionality reduction is used, retaining the first principal component, which has a cumulative contribution rate of 0.92, mapping the five-dimensional data into a univariate sequence of length 1440 with an amplitude range between -3 and 3. Bandpass filtering is applied with a low cutoff frequency of 0Hz and a high cutoff frequency of 0.5Hz (corresponding to half of the sampling frequency of 1Hz) to remove high-frequency noise, resulting in a smoothed signal. The sequence is used as input for the VMD decomposition algorithm in S2.2, which can stably extract the low-frequency dominant component Baseline(t) in the subsequent decomposition steps, significantly improving the separation effect of trend components. S2.2: Set the number of decomposition layers K=5 and the penalty factor α=2000 for the variational mode decomposition (VMD) algorithm to construct a multimodal function decomposition model. The value of K is set according to the operating cycle characteristics of the PON port and the empirical signal complexity. The value of α is used to control the constraint strength of the mode function bandwidth to ensure the stable extraction of low frequency components. S2.3: Based on the constructed VMD model, non-recursive mode decomposition is performed on the original observed signal sequence X(t) to generate K intrinsic mode functions, where each intrinsic mode function corresponds to a signal sub-band with different frequency components. The bandwidth constraints and reconstruction errors of each mode function are minimized through iterative optimization. S2.4: Identify the dominant low-frequency component from the generated K intrinsic mode functions, which corresponds to the long-term trend component of the PON port operating status. The identification method is based on spectral energy distribution analysis and autocorrelation coefficient criterion to screen out the mode functions with the minimum volatility and maximum periodicity characteristics. K eigenmode functions obtained from variational mode decomposition The spectral energy distribution analysis method (parameters: frequency resolution of 0.01Hz, energy calculation range of the full frequency domain) is used to achieve a quantitative assessment of the energy concentration of each modal frequency component; Furthermore, by calculating each The power spectral density curve was analyzed, and the energy ratio in the low-frequency band (frequency < 0.1 Hz) was extracted using the integration method to obtain the modal low-frequency energy ratio index. ; Furthermore, the autocorrelation coefficient criterion (parameter: delay range τ = 1~50 sampling periods) is used to measure the periodic characteristics of each modality time series and generate the autocorrelation peak coefficient. As a result of periodic measurement; Furthermore, by constructing a joint decision function, combined with the low-frequency energy ratio... With autocorrelation peak coefficient The overall trend score is calculated using a weighted scoring method. The formula is as follows:

[0022] in, and These are the weighting coefficients for energy and periodicity criteria, respectively, with the weights set based on the sensitivity priority of actual operational fluctuations. Furthermore, by iterating through all modes from k=1 to K, the mode functions with the largest trend score Score(k) and the proportion of low-frequency energy higher than 0.8 are selected. As the dominant low-frequency component, it was confirmed that this component has a single main peak and a small spectral spread in the power spectrum; By combining spectral energy distribution analysis and autocorrelation coefficient determination, the mode decomposition result of the previous step is transformed into a unique dynamic baseline candidate component, thereby achieving accurate identification of the long-term trend signal of PON port operation. For example, in the operation and maintenance scenario of a certain PON port, the VMD decomposition layer number K is set to 5, and spectral energy analysis is performed on the optical power time series with a collection period of 1 minute to obtain... Low-frequency energy ratio =0.92, peak autocorrelation coefficient =0.87. Set the weighting coefficient. =0.6, =0.4, substitute the value into the trend scoring formula:

[0023] The calculated trend score is 0.902, which is the highest among all modes, and it meets the screening criterion that the proportion of low-frequency energy is greater than 0.8. Therefore... Selected as After subsequent moving average smoothing, this It successfully reflected the slow degradation trend of the port over the past 180 days, and the output dynamic baseline signal significantly improved the stability of the residual signal and the early identification capability of progressive faults in field testing. S2.5: The identified dominant low-frequency modes are smoothed by moving average to remove residual transient disturbances and obtain the final dynamic baseline signal. This signal is used to construct the subsequent residual sequence as a reference benchmark for judging whether the operating status of the PON port deviates from the normal trend.

[0024] like Figure 2 As shown, step S3 involves calculating the residual sequence between the current observation and the dynamic baseline, and converting the residual sequence into a two-dimensional state transition map based on the Markov Transition Field (MTF) algorithm. The state partitioning uses an equal-width binning method to set N=8 discrete states. Specifically, this includes: S3.1: Perform point-by-point difference calculation on the optical power, bit error rate, temperature, frame loss rate and service load indicators in the dynamic baseline and the original observation sequence to obtain a residual sequence that reflects the degree of deviation of the PON port's operating status, which serves as the initial feature input characterizing the port degradation trend. Based on the dynamic baseline signal Baseline(t) obtained in the previous step and the original observation sequence X(t), a multi-index residual calculation framework is constructed to accurately characterize the instantaneous deviation of the PON port operating status from the long-term trend. A point-by-point interpolation algorithm (parameter: sampling time alignment accuracy δt = 1 min) is used to perform Baseline(t) on the optical power time series and compare it with the original downlink optical signal intensity series. The difference is calculated to generate the optical power residual sequence. This is used to reflect the fluctuation anomalies of the optical module in the dimension of optical signal intensity; Furthermore, by employing a pointwise interpolation algorithm (parameter: window smoothing length L=5), the difference between the Baseline(t) and the moving average-calculated bit error rate (BER) sequence is calculated on the bit error rate time series to generate the bit error rate residual sequence. It is used to quantify the instantaneous degradation of the data transmission link in terms of inter-symbol error rate; Furthermore, the same point-by-point interpolation algorithm is used (parameter: temperature sampling frequency). =1Hz), perform the difference calculation between Baseline(t) and the internal temperature sequence T(t) of the temperature time series to generate the temperature residual sequence. It is used to capture the instantaneous offset of changes in ambient temperature and thermal stability. Using a similar point-by-point interpolation algorithm (parameter: frame loss statistics period ΔF = 1 min), the difference between Baseline(t) and the normalized frame loss rate sequence FLR(t) is calculated on the frame loss rate time series to generate a frame loss rate residual sequence. This characterizes the instability of the data link at the frame transmission level. By using a point-by-point interpolation algorithm (parameter: service load sampling precision δB = 1 min), the difference between the service load time series Baseline(t) and the uplink / downlink bandwidth utilization sequence Load(t) is calculated to generate the service load residual sequence. It is used to characterize the real-time deviation of port operating pressure under different business cycles; Using the residual vector synthesis method, , , , and The residuals are concatenated into a multidimensional residual vector R(t) according to the time step, and the residuals of each dimension are normalized (method: Z-score) to eliminate the differences in the dimensions of different indicators and enhance the uniformity of residual characteristics. By using the above algorithm, the Baseline(t) is compared with various original observation sequences one by one to generate a standardized multi-index residual sequence R(t), which realizes a refined quantitative characterization of the deviation of the PON port's operating state, and provides a unified initial feature input for subsequent state discretization and Markov transition map construction. For example, in the operating environment of a city access network, Baseline(t) is a low-frequency trend signal extracted by VMD, and the optical power sequence is... At a certain moment, the value is -22.5 dBm, the baseline is -22.0 dBm, and the residual calculation formula is: The result was -0.5 dB; the moving average of the bit error rate (BER(t)) at the same time point was 1.0 × 10⁻⁶. -6 The baseline is 0.8 × 10 -6 The residual is 0.2 × 10⁻⁶. -6 Temperature T(t) is 45.3°C, baseline is 44.0°C, residual is 1.3°C; frame loss rate (FLR) (t) is 0.0045, baseline is 0.0030, residual is 0.0015; load (Load) (t) is 80%, baseline is 75%, residual is 5%. The residual sequences, after Z-score normalization, are combined into a five-dimensional residual vector R(t) = [-0.5, 0.2 × 10⁻⁶]. -6 [1.3, 0.0015, 5], used for subsequent discretization. This processing significantly improves the sensitivity to gradual degradation trends in long-term operation monitoring, and can output stable and consistent multidimensional residual inputs without hardware modifications, laying a solid foundation for Markov state modeling; S3.2: Based on the equal-width binning method, the residual sequence is discretized, and the continuous residual values ​​are divided into N=8 mutually exclusive state intervals to generate a discrete state sequence, so as to reduce the influence of noise and facilitate the construction of state transition maps. S3.3: Using the discrete state sequence as input, a state transition probability matrix is ​​constructed based on the Markov Transition Field (MTF) algorithm, where the matrix elements represent the normalized frequency of the transition from state i to state j in adjacent time steps, so as to quantify the dynamic transition relationship of the PON port state evolution. For the state sequence generated by the equal-width binning method, a Markov transition field (MTF) is used to construct a state transition probability matrix to quantify the time evolution relationship between discrete states. A state counting method (parameter: window span is the length of the entire sequence) is used to count the states of adjacent time steps. To state Complete statistics on the number of transfers; Furthermore, the normalized frequency calculation method (parameter: total number of transitions per state) is used. (total outward transfers), achieving the transfer probability The calculation, where For state arrive Number of transfers, For state Total number of transfers; Furthermore, a matrix filling algorithm (parameter: matrix dimension N×N, where N is the number of state categories = 8) is used to achieve the following: The mapping of values ​​to their positions in the matrix yields the unnormalized Markov state transition matrix; Furthermore, by performing boundary condition correction processing (parameter: zero probability mapping rule is 0), explicit marking of state pairs that have not undergone transition is achieved, eliminating the interference of pseudo probability values; By using the above discretization counting and normalization processing methods, the discrete state sequence of the previous step is transformed into a two-dimensional matrix containing all state transition probabilities, thereby realizing a quantitative characterization of the PON port operating state evolution. For example, during a monitoring cycle of a PON port, the equal-width binning method divides the residual value into 8 state intervals, resulting in a state sequence length of 1000. The MTF counting method is used to count the number of adjacent state transitions; for example, the transition from state 2 to state 5 is 42 times, and the total outward transitions for state 2 are 120 times. The transition probability is then calculated. The value is 0.35. The probabilities of all state pairs are filled into an 8×8 matrix, with 0s placed at the positions of missing transition pairs. This matrix is ​​output as the unnormalized Markov state transition matrix and applied to subsequent row normalization and state transition map generation. Results show that, compared to directly counting state frequencies, the matrix generated by this method significantly improves the ability of the subsequent LSTM model to capture state evolution features and exhibits more stable trend prediction performance on the validation set. S3.4: Perform row normalization on the state transition probability matrix to ensure that the sum of each row vector is 1, and obtain a standardized Markov state transition matrix as a stable representation of the port running state evolution path. S3.5: The standardized Markov state transition matrix is ​​mapped into a two-dimensional image-based state transition map to visualize the temporal evolution trajectory of the PON port's operating state, serving as a graphical temporal feature representation for subsequent LSTM model inputs.

[0025] like Figure 3 As shown, step S4 involves inputting the state transition map into a Long Short-Term Memory (LSTM) network model, and training a neural network with two hidden layers (64 neurons per layer) to predict the state transition probability distribution for the next T=3 time steps. Specifically, this includes: S4.1: Based on the two-dimensional state transition map generated by the Markov transition field (MTF), construct the input tensor of the LSTM model, where the tensor dimension is [batch size × time step × state dimension], to adapt to the input requirements of the neural network; Based on the two-dimensional state transition graph dataset generated by Markov transition fields (MTF), a tensor construction method (parameters: batch size B, time step T, number of state categories N) is used to transform the graph sequence into a high-dimensional tensor structure that can be input into a long short-term memory network (LSTM) model. Furthermore, by using the time series slicing method (parameter setting: time step T), the state transition map of each port is divided into continuous segments of length T along the time axis, generating an input sample matrix that meets the network's time-series dependency requirements; Furthermore, a dimension mapping rule (number of state categories N=8) is adopted to expand the two-dimensional map in the state category dimension, so that the probability vector of the state category corresponding to each time step is completely preserved in the last dimension of the tensor, ensuring that the network can perceive the spatial distribution characteristics of all states simultaneously during the training process. Furthermore, based on the batch loading mechanism (B is determined by the total number of training samples and the batch size), multiple port samples are stacked in the batch dimension to form an input tensor with a unified structure. This achieves dimensional matching with the input interface of the LSTM network; Using the tensor construction method described above, the two-dimensional state transition graph result from the previous step is transformed into a data input format that conforms to the temporal structure of a neural network, thereby enabling the subsequent state transition probability prediction to synchronously model spatiotemporal features. For example, for a training set containing B=64 PON port samples, a state transition map length of 240 time steps per port, and N=8 state categories, performing a time series slicing parameter T=30 yields a sample size per port of [number missing]. =211 input sequence segments, each segment having a shape of... Stack all port fragments with a batch size B=64, constructing an input tensor with dimension . It directly connects to the LSTM input layer. After performing this step, the network is verified to be able to load data stably and significantly improve its ability to capture port progressive failure trends during time series modeling. The output predicted multi-time step state probability distribution is smoother than the baseline method and the advance of abnormal states is increased. S4.2: Normalize the constructed LSTM input tensor using the Z-score normalization method to eliminate the numerical dimension differences in the state graphs between different PON ports and obtain the normalized state transition graph input sequence. S4.3: Based on the input sequence of the standardized state transition graph, a neural network model with two LSTM hidden layers is constructed, with 64 neurons in each layer. The ReLU activation function and Dropout regularization mechanism are used to enhance the model's ability to model long-term dependencies of port states. Based on the input sequence of the standardized state transition graph, a two-layer deep recurrent neural network structure (64 neurons per layer) is constructed using a long short-term memory network (LSTM) to model the temporal features of the PON port state evolution process. Furthermore, by configuring the weight matrices of the input gate, forget gate, and output gate of the gated recurrent unit, and using the Xavier uniform distribution initialization method (the range is automatically calculated based on the input feature dimension), the stability and convergence control of the network's initial parameters are achieved, and the initial weight mapping matrix data of each gated unit is obtained. Furthermore, by embedding the Rectified Linear Unit (ReLU) activation function in each LSTM hidden layer, the following is utilized: The nonlinear mapping enhances the network's nonlinear response to abnormal state transition features and generates a sequence of activated hidden state vectors. Furthermore, by introducing a Dropout regularization mechanism (parameter p=0.2) at the output of each hidden layer, 20% of the neuron outputs are randomly masked, thereby suppressing the risk of model overfitting and generating a regularized hidden layer output tensor. By using the above LSTM network structure and parameter configuration processing method, the standardized state transition graph input sequence is transformed into a deep feature representation with long-term temporal dependencies, thereby achieving high-precision prediction capability for future state evolution trends. For example, when processing the state transition graph of a PON port cluster, the input tensor dimension is configured as [128×50×8], where the batch size is 128, the time step is 50, and the number of state categories N=8. A two-layer LSTM structure is used, with 64 neurons in each layer. The weight initialization range is calculated by the Xavier method as [-0.176, 0.176]. ReLU activation function and Dropout with p=0.2 are configured at the output of both layers. The historical 50-step state transition graph within the port's operating cycle is used as the training input. After network feedforward calculation and nonlinear mapping, two layers of deep feature sequences are obtained, and the feature dimension is compressed to [128×50×64]. During training, the ReLU activation function significantly improves the network's response amplitude to nonlinear transitions in port states, and the Dropout mechanism effectively reduces overfitting in high-density port scenarios. The final output deep features show a significant improvement in the stability of the prediction curve waveform and convergence speed in subsequent tasks predicting future state distributions. S4.4: The LSTM model is trained using the historical state transition graph sequence. The cross-entropy loss function and Adam optimizer are used, and the network parameters are adjusted based on the backpropagation algorithm to obtain the trained LSTM state evolution prediction model. Based on the input sequence of the standardized state transition graph, a prediction error measurement model is constructed using the cross-entropy loss function to achieve a quantitative evaluation of the difference between the predicted probability distribution of the LSTM output and the true state distribution. Furthermore, through the Adam optimizer (parameter settings: learning rate...), First-order moment estimation of attenuation rate Second-order moment estimation of attenuation rate This enables adaptive adjustment of network weights and biases, and yields the optimal gradient update step size. Furthermore, the gradient of the cross-entropy loss function is calculated layer by layer using the backpropagation algorithm, and the connection weight matrix from the second hidden layer, the first hidden layer, and the input layer to the first hidden layer is updated sequentially to ensure that the model's ability to capture long-term dependent features is effectively trained. Furthermore, a batch training method is adopted (batch size) In each batch, the average loss value is calculated and used to drive the parameter update process to reduce gradient variance and accelerate convergence. Furthermore, model validation is performed after each iteration. The optimal iteration round and corresponding parameter snapshot are selected by evaluating the mean squared error and prediction accuracy of the predicted state transition probability distribution on an independent validation set. Through the Adam optimization driven by the cross-entropy loss and the backpropagation parameter adjustment, the standardized state transition graph features of the previous step are transformed into a trained LSTM state evolution prediction model that can be used for future state prediction, thereby achieving high-precision time series prediction of the progressive failure probability of PON ports. For example, regarding the collected PON port data in the past The state transition graph at each time step is used to construct the input tensor with dimension [missing information]. Corresponding batch size Time step Number of state categories Define the cross-entropy loss function. for: ,in The total number of samples, For the first The true state distribution vector of each sample This is the predicted probability distribution vector. During training, the number of iteration rounds is set. In each update round, the gradient is adaptively adjusted by the Adam optimizer, and the average training loss changes from the initial value. Gradually reduce to In the validation set evaluation, the model's performance in predicting the future... When the state transition distribution is calculated at each time step, the mean square error is significantly reduced and the recall rate of abnormal state identification is significantly improved, proving that the trained LSTM state evolution prediction model can stably output a high-precision future state probability distribution, providing a reliable input for subsequent fault probability scoring calculation. S4.5: Based on the trained LSTM state evolution prediction model, input the current port state transition map, output the state transition probability distribution of the next T=3 time steps, and generate a three-dimensional output tensor [batch size × number of prediction steps × number of state categories] to characterize the probability density of the port in each state at future time steps; Based on the already trained Long Short-Term Memory (LSTM) network state evolution prediction model, the input is a standardized two-dimensional state transition graph tensor of the current batch of PON ports. The tensor dimension is set to [batch size × number of time steps × number of state categories] to maintain consistency with the input structure during the training phase and to achieve the task of predicting future state distribution. A batch forward propagation computation method is adopted (parameter: batch size B is taken from the current task queue length), the input tensor is passed to the first hidden layer LSTM unit layer by layer, and the time-dependent feature extraction of the port state sequence is realized based on the combination of internal memory unit and gating mechanism, and the hidden state sequence of the first layer is output. Furthermore, the first layer hidden state sequence is subjected to another time-series abstraction process through the second hidden layer LSTM unit (parameters: number of hidden units 64, activation function ReLU), preserving the long-term state transition pattern across time steps and generating a deep temporal feature representation for probabilistic prediction; Furthermore, the final hidden state of the second hidden layer is input into the fully connected output layer. Matrix multiplication and bias addition are performed based on the weight vector and bias coefficients learned by the network to achieve preliminary calculation of the state class probability for the future prediction step T=3. N state probability values ​​are output for each prediction step, where N is the number of state classes. Furthermore, by stacking the state probability vectors of each prediction step, a three-dimensional output tensor is generated, with the structure set as [batch size × number of prediction steps × number of state categories], to achieve a probability density characterization of the different states of each batch of ports in future time steps; Through the above neural network inference processing method, the image-based temporal feature results of the previous step are transformed into quantifiable port future state probability distribution data, thereby achieving the temporal prediction technology effect required for early identification of progressive faults. For example, select a batch size B=20 PON ports in a certain network area, with an input tensor dimension of [20×15×8], where the number of time steps is 15 and the number of state categories N=8. Using a trained two-layer LSTM model (64 units per layer, ReLU activation), perform forward propagation on the input tensor to obtain a first-layer hidden state dimension of [20×15×64] and a second-layer hidden state dimension of [20×15×64]. Input the hidden state of the last time step of the second layer into a fully connected layer to calculate the state probabilities for the next T=3 time steps, outputting a preliminary prediction matrix dimension of [20×3×8]. Tensor assembly generates the final prediction tensor [20×3×8], where, for example, in the prediction of the 5th port, the probability of an abnormal state in the first prediction step is... The probability of an abnormal state in the second prediction step is The probability of an abnormal state in the third prediction step is The average deviation of the prediction results from the actual state transition distribution in the validation set is significantly reduced, demonstrating the high fidelity of the model in capturing the port state evolution trend, and providing accurate input for subsequent threshold comparison and fault probability scoring; S4.6: Perform Softmax normalization on the state transition probability distribution of the LSTM output to obtain the probability value P of the port being in an abnormal state at each prediction time step, which is used as the input parameter for subsequent comprehensive fault probability score calculation.

[0026] Step S5: Construct a quantile regression model based on the historical residual sequence, estimate the 95th quantile threshold using linear programming, and dynamically adjust the quantile level according to the variance change rate within a sliding window (window length W=30): switch to the 90th quantile when the variance change rate ≥ 0.15, and switch to the 98th quantile when the variance change rate ≤ 0.05. Specifically, this includes: S5.1: Perform standardization on the historical residual sequence to eliminate the scaling effect caused by the difference in operating baselines between different PON ports, and obtain a normalized residual sequence for consistent input in subsequent quantile modeling; S5.2: Construct a quantile regression model based on the normalized residual sequence, and use the linear programming method to optimize the loss function to obtain the initial threshold estimate at a 95% confidence level, which serves as the initial warning boundary of the system. S5.3: Perform variance calculation on the residual sequence within the current sliding window (window length W=30) to obtain the residual fluctuation intensity index within the window, which is used to reflect the short-term stability change trend of the system's operating state; An unbiased sample variance calculation method (parameter: window length W=30) is used for the residual sequence within a sliding window length W=30 to quantify short-term operational fluctuations. Furthermore, the mean of the residuals within the window is calculated using a mean estimation method (parameter: normalized residual sequence). And provide a benchmark value for variance calculation; Furthermore, the variance within the sliding window is calculated using the accumulated squared deviation method, as follows:

[0027] in, Let be the normalized residual value at the i-th time step. Let W be the mean of the residuals within the window, and W be the window length. Furthermore, the volatility index Δσ² is calculated by the ratio of the difference between the current variance and the historical benchmark variance, specifically as follows:

[0028] in, The variance of the current window. This represents the baseline variance of historical operating conditions. By solving for the fluctuation intensity Δσ², the residual sequence from the previous step is transformed into a quantifiable short-term stability change index, enabling rapid response to sudden changes in the system's operating state. For example, in a PON port monitoring scenario, the sliding window length W=30, and the mean of the normalized residual sequence within a certain window is calculated as follows: =0.012, the variance of the current window is obtained using the variance calculation formula. =0.0045, historical baseline variance =0.0030. The fluctuation intensity calculation formula yields Δσ²=0.225. This value exceeds the preset sensitivity enhancement threshold (0.15), triggering a switch to the quantile level to 90% in subsequent steps, enabling early identification of potential progressive faults. Within this window period, after introducing the Δσ² index, the system fault risk assessment model's prediction curve exhibits a stable and controllable gradual upward trend during the smoothing phase, effectively improving the accuracy of abnormal trend capture. S5.4: Based on the comparison between the variance fluctuation intensity index and the preset threshold, a dynamic switching strategy for the quantile level is executed: when the variance fluctuation intensity index is ≥0.15, it switches to the 90th quantile to improve detection sensitivity; when the variance fluctuation intensity index is ≤0.05, it switches to the 98th quantile to enhance system stability. Based on the variance fluctuation intensity index Δσ² and the preset dynamic adjustment threshold rules, a comparison calculation method (parameters: current value of Δσ², sensitivity enhancement threshold 0.15, stability enhancement threshold 0.05) is used to achieve the matching and judgment of the system residual fluctuation characteristics and threshold rules within the current sliding window; Furthermore, a conditional branching algorithm (parameters: Δσ², 0.15, 0.05) is used to determine the switching of quantile levels and generate quantile level control variables. This is used to guide the interval setting of subsequent quantile regression models; Furthermore, the assignment operation method (parameter:) is adopted. Target quantile level set {90%, 95%, 98%}), to achieve when When Set to 90% to improve detection sensitivity; when When Set to 98% to enhance system stability; maintain the initial 95% level when within the above range to keep the threshold response neutral and balanced; Furthermore, through the state update method (parameter: The current value and historical quantile level records are used to record the time series state transition of quantile levels, which is used to analyze the statistical correlation between quantile level changes and system operating status. By combining conditional judgment and state update processing, the Δσ² fluctuation intensity result calculated in the previous step is transformed into quantile level control data, thereby achieving real-time adjustment of the adaptive threshold generation mechanism. For example, within a certain PON port operating cycle, with a sliding window length W=30, the variance of the normalized residual sequence is calculated to obtain Δσ²=0.162. Δσ² is then compared with a preset threshold. hour, The initial value was changed from 95% to 90%. At this point, the quantile regression model set the target quantile to 90% and recalculated the anomaly detection boundary. In other running cycles on the same port, Δσ² = 0.038 was obtained, and a comparison judgment was performed. hour, Switching to 98% and generating an adaptive threshold at this quantile level effectively suppressed false alarms caused by short-term fluctuations. In the above different scenarios, switching the quantile level enabled the system to achieve higher detection sensitivity in periods of higher volatility and stronger stability in periods of lower volatility, significantly improving the accuracy and robustness of progressive fault prediction. S5.5: Based on the dynamically adjusted quantile level, re-execute the quantile regression model to generate an updated adaptive threshold Th, which serves as the anomaly judgment boundary for the current operating cycle and is used as input for the subsequent fault probability scoring module.

[0029] Step S6: Fuse the abnormal state probability P from the LSTM output with the adaptive threshold Th to calculate the comprehensive fault probability score FP = α·P + (1-α)·(1-|e| / Th), where α = 0.7 is the weighting coefficient and e is the current residual value. Specifically, this includes: S6.1: Based on the state transition probability distribution of the future T=3 time steps output by the LSTM model, extract the abnormal state probability P of the current time step as the output result of the time series prediction layer; S6.2: Obtain the adaptive threshold Th, which is dynamically estimated by the quantile regression model and generated after adjusting the quantile level based on the variance change rate Δσ² within the sliding window. It is used to reflect the abnormal boundary in the current operating environment. S6.3: Collect the residual value e at the current time step. This residual value is the deviation between the PON port observation and the dynamic baseline, which is used to measure the degree to which the current operating state deviates from the normal trend. S6.4: Based on the abnormal state probability P, the adaptive threshold Th, and the residual value e, a comprehensive fault probability score is calculated, and a weighted fusion formula is used to balance the contribution ratio of prediction and threshold response. In this sub-step, the input condition is the probability of an abnormal state predicted by the Long Short-Term Memory (LSTM) network model. The adaptive threshold generated by adjusting the sliding window variance rate of the quantile regression model. and the residual value at the current time step. The objective is to obtain a comprehensive failure probability score; A weighted fusion method is adopted (with preset weight coefficients). =0.7), the abnormal state probability output by the time series prediction layer is weighted and synthesized with the threshold response factor based on the residual ratio to achieve the combination of prediction information and real-time deviation. Furthermore, a comprehensive fault probability score is defined by constructing a fusion calculation relationship through mathematical formulas. as follows:

[0030] in, This represents the probability of an abnormal state. For adaptive threshold, The residual value, These are the weighting coefficients; Furthermore, regarding the residual ratio in the formula... The calculation uses absolute value functions and real number division to ensure non-negativity and normalization within the numerical domain, and the ratio is used to characterize the deviation of the current observation from the threshold. Furthermore, using weighting factors and The ratio of prediction information to deviation response is controlled to determine the contribution ratio of prediction information to the overall score, thereby ensuring enhanced response speed under strong prediction signals and increased robustness under weak prediction signals. This fusion calculation method transforms the abnormal state probability and adaptive threshold response factor from the previous step into a comprehensive fault probability score that reflects the operational health of the PON port, thereby achieving progressive fault probability quantification while taking into account both sensitivity and stability. For example, in a certain PON port operation scenario, the LSTM model outputs the probability of an abnormal state at the current time step. = Adaptive threshold dynamically adjusted by quantile regression model = residual value = Weighting coefficient = First, calculate the residual ratio. = Thus, the deviation response factor is obtained. = Then, the overall score is calculated according to the formula: + = + = This score effectively reflects the risk of progressive failure at the port. After subsequent exponential smoothing and continuous anomaly detection, it can be used to trigger operation and maintenance early warnings. In actual verification, the score remains stable under similar load fluctuation scenarios and can show a significant improvement when degradation signals are enhanced, meeting the dual requirements of prediction sensitivity and false alarm suppression. S6.5: Output comprehensive fault probability score FP, as a quantitative indicator of progressive fault status of PON port, for use in subsequent index smoothing processing and early warning mechanism; Based on the abnormal state probability calculated in the previous sub-step Adaptive threshold and the current residual value As input conditions, the process of generating the comprehensive failure probability scoring index; A weighted fusion calculation method is adopted (parameters: This enables the quantitative fusion of time-series prediction results and threshold response results; Furthermore, by constructing a fusion calculation formula This enables the comprehensive calculation of the abnormal state probability component and the residual relative threshold deviation component, and yields a fault probability score. result; Furthermore, through the residual term Quantify the degree of deviation of the current observation value from the dynamic baseline to achieve a unified mapping of weak degradation signals and major anomaly signals as a scoring scale; Furthermore, by setting weighting coefficients The numerical basis of the model performance evaluation results is used to balance and adjust the relative contribution ratio of the model prediction part and the residual judgment part, and generate a set of scoring algorithm parameters that are stable and applicable in multi-scenario operation environment. Through the above-mentioned fusion calculation method, the probability of abnormal state and the residual deviation are transformed into a quantitative fault risk score that can be directly compared, thereby realizing a quantitative characterization index of the progressive fault state of PON port. For example, in a batch of running data, the probability of an anomalous state output by the LSTM model at the current time step. for Adaptive threshold adjusted by quantile regression model for Current residual value for Input the above parameters into the fusion formula: Calculate the residual term. have to Substituting into the formula, we get ,as well as The two parts are added together to get the comprehensive score. = The score, after subsequent S7 exponential smoothing, forms a stable risk curve. When the threshold condition is exceeded for multiple consecutive time steps, a progressive fault warning signal is successfully triggered, verifying the significantly improved sensitivity and stability of the scoring method in early fault identification.

[0031] Step S7: Perform exponential smoothing on the fault probability scoring sequence, using a smoothing factor β=0.3 to generate a stationary prediction curve, and trigger a progressive fault warning when the predicted value exceeds the threshold Th for M=5 consecutive time steps. Specifically, this includes: S7.1: Obtain the fault probability score sequence and perform recursive weighted averaging based on the exponential smoothing algorithm to reduce the short-term volatility of the score sequence and retain the long-term trend characteristics. S7.2: Perform trend stability enhancement processing on the smoothed prediction curve. Use a sliding window of W=10 time steps to perform local mean normalization on the smoothed prediction curve to eliminate the influence of periodic business load changes on the shape of the prediction curve and obtain the normalized sequence. S7.3: Based on the normalized prediction curve and the dynamic threshold, a point-by-point comparison is performed. If the prediction curve is greater than the dynamic threshold, the point at that time is marked as a potential anomaly point, and a binary anomaly label sequence is generated for subsequent continuity judgment. S7.4: Perform continuity detection processing on the abnormal label sequence, and use a sliding window W=5 time steps to count the number C of consecutive labels as 1 within the window. If C≥M=5, then it is determined that there is a continuous abnormal trend within the time window. Binarized anomaly marker sequence Using a sliding window detection method (parameter: window length) This enables the identification of continuous abnormal trends. Furthermore, through sequential traversal The data at each time step are analyzed using an element counting algorithm within a window (the statistical value is marked as ""). The number of consecutive abnormal states is used to calculate the cumulative frequency of abnormal states and obtain the number of consecutive abnormal points. As input for trend judgment; Furthermore, by comparing the number of consecutive outliers With threshold parameter To achieve the logical judgment of persistent abnormal trends, when the following conditions are met... A persistent abnormal state flag is generated in real time, which is used as a condition for triggering subsequent early warnings; Furthermore, to ensure the robustness of the detection, a window boundary condition correction method is adopted, which incorporates the number of remaining elements when the window slides to the end of the sequence into the decision logic to prevent misjudgment caused by sequence truncation and generate a corrected persistent anomaly judgment result. By using a sliding window continuity detection algorithm, the abnormal marker sequence generated in the previous step is transformed into a continuous abnormal trend judgment result, thereby achieving a stable identification function for progressive fault trends. For example, within a certain PON port operating cycle, the sequence Length is Sliding window length In the To the All marker values ​​within the time step are Then the cumulative number of outliers The conditions for determining persistence are met; Perform continuity detection on this window using the formula.

[0032] in, Let i be the i-th flag value in the window, and [condition] be an indicator function that converts the logical value to 1 or 0. The result of the formula is the total number of abnormal flags in the window. In this embodiment ,conform to The judgment conditions are used to output a continuous abnormal trend flag, which is used to drive the fault warning signal generation module of S7.5 to realize early intervention for progressive faults. S7.5: Generate a progressive fault warning signal Alert based on the results of the continuous abnormal trend determination, and push the warning information to the operation and maintenance execution module through the operation and maintenance interface module to trigger subsequent manual verification or automatic isolation mechanism to achieve pre-fault intervention.

[0033] The present invention also provides a PON port fault prediction and operation and maintenance support system for optical access networks, which uses the above-mentioned PON port fault prediction and operation and maintenance support method to perform PON port fault prediction and operation and maintenance support for optical access networks.

[0034] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0035] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for optical access network (PON) port fault prediction and operation and maintenance support, characterized in that, The method comprises the following steps: S1: collecting PON port runtime sequence data, including optical power, bit error rate, temperature, frame loss rate and service load indicators, to form an original observation sequence set; S2: performing variational mode decomposition processing on the original observation sequence set, separating low-frequency dominant components as dynamic baseline based on a preset decomposition layer number and a penalty factor; S3: calculating the residual sequence of the current observation value and the dynamic baseline, and converting the residual sequence into a two-dimensional state transition graph based on a Markov transition field algorithm, wherein the state division adopts an equal-width binning method to set discrete states; S4: inputting the two-dimensional state transition graph into a long short-term memory network model to predict the state transition probability distribution of future time steps through neural network training; S5: constructing a quantile regression model based on historical residual sequences, estimating quantile thresholds using linear programming methods, and dynamically adjusting quantile levels according to the variance rate within the sliding window; S6: fusing the abnormal state probability output by the long short-term memory network model and the adaptive threshold to calculate the comprehensive failure probability score; S7: performing exponential smoothing processing on the comprehensive failure probability score to generate a smoothed prediction curve, and triggering a gradual failure warning when the predicted value of consecutive time steps exceeds the threshold. 2.The method of claim 1, wherein, The step S1 specifically comprises: Based on the PON port operation and maintenance management system interface, the optical power monitoring signal is obtained, the real-time data of the downlink optical signal strength and the uplink received optical signal strength is collected, and the optical power time sequence is obtained; The bit error rate counting before and after forward error correction is obtained by using the bit error rate statistical module built-in the OLT device, the bit error rate is processed by sliding window averaging, and the bit error rate time sequence is generated; The temperature data inside the optical module is collected through the temperature sensor integrated in the PON port, and the temperature time sequence is generated based on the temperature sampling period; Based on the network layer performance monitoring protocol, the frame loss rate indicator is obtained, the frame loss number in the continuous time period is normalized, and the frame loss rate time sequence is generated; The service load data of the PON port is collected through the flow statistical module, including the uplink and downlink bandwidth utilization, and the service load time sequence is generated based on the timestamp alignment mechanism; The optical power time sequence, the bit error rate time sequence, the temperature time sequence, the frame loss rate time sequence and the service load time sequence are processed by timestamp alignment and missing value interpolation to generate the original observation sequence set. 3.The method of claim 1, wherein, The step S2 specifically comprises: Based on the original operation time sequence data collected by the PON port, an original observation signal sequence is constructed, and after normalization, a single variable time sequence input is formed; The decomposition layer number and the penalty factor of the variational mode decomposition algorithm are set to construct a multi-modal function decomposition model; Based on the multi-modal function decomposition model, the original observation signal sequence is subjected to non-recursive mode decomposition to generate an eigenmode function, wherein each eigenmode function corresponds to a signal subband of different frequency components, and the bandwidth constraint and reconstruction error of each mode function are minimized through iterative optimization. A dominant low-frequency component corresponding to a long-term trend component of the PON port operating state is identified from the eigenmode functions, and the identification method is based on spectral energy distribution analysis and autocorrelation coefficient criterion to screen out the mode functions with minimum volatility and maximum periodicity characteristics; The dominant low-frequency component is subjected to moving average smoothing processing to obtain a final dynamic baseline signal.

4. The method according to claim 1, wherein, The step S3 specifically includes: Point-by-point difference calculation is performed on the dynamic baseline and the optical power, bit error rate, temperature, frame loss rate and service load indicators in the original observation sequence to obtain a residual sequence reflecting the deviation degree of the PON port operating state; Discrete state sequence is generated by performing discrete processing on the residual sequence based on the equal-width binning method, and the continuous residual values are divided into multiple mutually exclusive state intervals; The discrete state sequence is used as input to construct a state transition probability matrix based on the Markov transition field algorithm; Row normalization processing is performed on the state transition probability matrix to obtain a standardized Markov state transition matrix; The Markov state transition matrix is mapped into a state transition atlas in the form of a two-dimensional image to visually represent the time evolution trajectory of the PON port operating state.

5. The method according to claim 4, wherein, The residual sequence includes an optical power residual sequence, a bit error rate residual sequence, a temperature residual sequence, a frame loss rate residual sequence and a service load residual sequence, and all components are subjected to Z-score normalization processing to generate a standardized multi-index residual sequence.

6. The method according to claim 1, wherein, The step S4 specifically includes: A long short-term memory network input tensor is constructed based on the two-dimensional state transition atlas generated by the Markov transition field; Normalization processing is performed on the long short-term memory network input tensor to obtain a standardized state transition atlas input sequence using a standardization method; A neural network model containing two long short-term memory network hidden layers is constructed based on the standardized state transition atlas input sequence, wherein each layer is configured with 64 neurons and uses a ReLU activation function and a Dropout regularization mechanism; The neural network model is trained using historical state transition atlas sequences, a cross-entropy loss function and an Adam optimizer, network parameters are adjusted based on a backpropagation algorithm, and a trained long short-term memory network state evolution prediction model is obtained; A three-dimensional output tensor is generated by inputting the current port state transition atlas into the trained long short-term memory network state evolution prediction model and outputting the state transition probability distribution of multiple future time steps; Softmax normalization processing is performed on the state transition probability distribution to obtain the probability value of the port being in an abnormal state at each predicted time step.

7. The method according to claim 1, wherein the method further comprises: determining whether the PON port is in a normal state or an abnormal state according to the PON port state information; and if the PON port is in the abnormal state, determining whether the PON port is in a failure state or a non-failure state according to the PON port state information. The step S5 specifically includes: Normalization processing is performed on the historical residual sequence to obtain a normalized residual sequence; A quantile regression model is constructed based on the normalized residual sequence, and a loss function is optimized and solved using a linear programming method to obtain an initial threshold estimate value; The residual sequence in the current sliding window is subjected to variance calculation to obtain a window variance change rate as a residual volatility intensity indicator; A quantile level dynamic switching strategy is executed according to the comparison result of the residual volatility intensity indicator and the preset threshold value. Based on the dynamically adjusted quantile rank, the quantile regression model calculation is re-executed to generate an updated adaptive threshold.

8. An optical access network (PON) port failure prediction and operation and maintenance support system, characterized in that, The method is used for optical access network PON port fault prediction and operation and maintenance support.

Images