A digital twin system and method for communication tower equipment operation monitoring

Abstract

This invention relates to the field of communication equipment operation and maintenance management and digital twin technology, and discloses a digital twin system and method for monitoring the operation of communication tower equipment. The method includes multi-source heterogeneous sensor data acquisition, spatiotemporal alignment and four-dimensional reliability scoring, state representation vector generation, fault probability prediction based on a bidirectional gated cyclic network, root cause diagnosis driven by an extreme gradient boosting model, and dual-deep Q-network optimization of programmable optical transceiver configuration. The effectiveness of the strategy is verified through bidirectional LSTM simulation, and finally, a full lifecycle archive is constructed, enabling collaborative traffic offloading between neighboring towers. This application can achieve high-precision early warning, automatic root cause localization, and cross-tower resource coordination, improving the autonomy and resilience of optical communication networks.

Description

Technical Field

[0001] This invention relates to the field of communication equipment operation and maintenance management and digital twin technology, and more specifically, it relates to a digital twin system and method for monitoring the operation of communication tower equipment. Background Technology

[0002] With the deep integration of optical communication networks and deep learning technology, higher demands are being placed on the intelligence, foresight, and autonomy of communication tower equipment operation monitoring. Currently, optical communication network management is gradually evolving from passive response-based alarms to proactive predictive optimization. While traditional network management methods have made some progress in multi-source heterogeneous data acquisition, basic parameter monitoring, and equipment configuration management, they still have significant shortcomings in areas such as dynamic assessment of data reliability, forward-looking fault prediction, closed-loop optimization of configuration strategies, and cross-node collaboration. Therefore, they cannot effectively support intelligent operation and maintenance throughout the entire lifecycle of communication tower equipment and enhance network resilience.

[0003] Specifically, existing solutions largely rely on static threshold comparisons and preset rule engines, lacking effective real-time evaluation and adaptive fusion mechanisms for dynamic reliability changes of multi-source heterogeneous sensor data in complex environments. Furthermore, after hardware configuration strategies are executed, simulation verification results and actual operational feedback are rarely used for reverse correction or dynamic adjustment of resource configuration strategies. This results in lags in fault warning timeliness and configuration adaptation accuracy when dealing with environmental disturbances, equipment performance degradation, or sudden link anomalies, limiting the depth of network autonomous optimization and the breadth of cross-node collaboration.

[0004] Therefore, how to construct a digital twin method with closed-loop evolution capabilities from multi-source sensor data to reliability assessment, fault prediction, root cause diagnosis, configuration optimization and simulation verification, and realize continuous learning of equipment degradation patterns from the operating state, and dynamically optimize resource allocation, fault early warning and cross-tower collaboration strategies accordingly, has become an urgent technical problem to be solved in the field of optical communication network operation and maintenance. Summary of the Invention

[0005] This invention provides a digital twin method for monitoring the operation of communication tower equipment, which solves the technical problems in the prior art such as lack of dynamic reliability assessment and adaptive fusion of multi-source heterogeneous sensor data, lag in fault early warning timeliness and configuration adaptation accuracy, lack of closed-loop evolution capability, and insufficient network autonomous optimization and cross-node collaboration capability.

[0006] This invention provides a digital twin system and method for monitoring the operation of communication tower equipment, comprising: Firstly, a digital twin method for monitoring the operation of communication tower equipment includes: Collect multi-source heterogeneous sensor data deployed on communication towers to obtain monitoring datasets of equipment operation; The monitoring dataset is processed by timestamp alignment, unit normalization and missing value imputation to generate a standardized dataset with a unified spatiotemporal reference. Real-time reliability scoring is performed on each sensor data item in the standardized dataset; The credibility scores are dynamically weighted and fused to output the state representation vector of the unified communication tower digital twin; The state representation vector is input into a bidirectional gated loop model, which learns the device degradation trend based on historical time series data and outputs the failure probability prediction result within a preset time window in the future; when the failure probability prediction result exceeds a preset safety probability threshold, an early warning information is triggered. The state representation vector is then input into the extreme gradient boosting model to output the root cause diagnosis result of the fault. The extreme gradient boosting model, combined with a historical fault sample library, sorts the features by importance and outputs a set of key factors leading to the current abnormal state and the root cause localization results. Based on the fault warning signal and root cause localization results, the dual-deep Q network model is activated. This model takes the current network load, link quality and device health as state inputs, and the modulation format, symbol rate and forward error correction coding strength as action space. Through interactive learning with the digital twin virtual environment, it generates the optimal configuration strategy. The optimal configuration strategy is sent to the programmable optical transceiver for execution and simultaneously input into the bidirectional long short-term memory simulation network. This network uses real-time optical signal waveforms and link state data to perform amplitude and phase fitting on the adjusted transmission process and outputs simulation verification values ​​of bit error rate and optical signal-to-noise ratio. The validity of the configuration is determined by the consistency between the simulation verification value and the actual measurement value. If the consistency error is less than the preset tolerance, the configuration is confirmed to be effective; otherwise, the configuration is returned to be reconfigured to generate a new configuration strategy. By integrating effective configuration strategies, failure probability prediction results, and simulation verification values ​​into the digital twin platform, a full lifecycle operation archive for communication towers can be constructed. Based on the full lifecycle operation records, the similarity of operation modes among different communication towers is identified, and a cross-tower collaborative optimization mechanism is established. When a potential risk occurs in a certain tower, the resource reservation and traffic diversion plan of the nearby healthy towers is automatically triggered.

[0007] Furthermore, a real-time reliability score is performed on each sensor data item in the standardized dataset, including: Sliding window detection is performed on each time series data item in the standardized dataset; detection results are obtained; wherein, when the absolute value of the standardized dataset detection of consecutive sampling points exceeds a preset threshold, it is identified as a mutation point or an anomaly point that exceeds the physical reasonable range, and an anomaly detection result is obtained; Based on the anomaly detection results, the data quality status corresponding to each sensor is marked to obtain the marking results; Calculate the sharpness, noise level, and occlusion rate of the image data in the standardized dataset to obtain quantitative indicators; Based on the quantitative indicators, the image data in the standardized dataset are subjected to sharpness, noise level and occlusion rate calculations, and an internal consistency score is generated by combining the data quality status labeling results. Meteorological station and vibration monitoring data are used as environmental context information; a context correction coefficient is generated based on the environmental context information; wherein, when the visibility is lower than a preset threshold (500 meters), the credibility of the sensor data in the internal consistency score is set to a preset standard threshold (0.3). When the wind speed exceeds the preset wind speed threshold (15m / s), the drift compensation algorithm is enabled to calculate the compensation residual for the vibration sensor data in the marking results, and the compensation residual is used as the context correction coefficient of the vibration sensor. Perform reading difference calculation on adjacent sensors of the same type in the same physical area to obtain the reading difference between adjacent sensors; A cross-validation score is generated based on the reading difference, the anomaly detection result, and the data quality status. Based on the cross-validation score and the difference in readings, a preliminary confidence score is obtained between each adjacent sensor pair; A sensor steady-state distribution model is constructed by calling the historical normal operation database. The KL divergence between the current observation and the steady-state distribution model is calculated. The historical pattern matching score is determined by combining the data quality status identification results to verify the preliminary confidence score in the historical data dimension. The internal consistency score, the cross-validation score, the historical pattern matching score, and the context correction coefficient are weighted and summed to form four dimensions for weighted correction of the initial credibility score. The weights are dynamically adjusted according to the data quality status and the environmental context information to generate a real-time credibility score in the range of 0 to 1. The real-time credibility score is used as a dynamic weight in the data fusion stage. When the quality of sensor data decreases due to abnormal data quality status or extreme changes in environmental context information, its influence on the fusion result is automatically reduced according to the real-time credibility score.

[0008] Furthermore, the credibility scores are dynamically weighted and fused to output a state representation vector for the unified communication tower digital twin, including: Spatial feature vectors are extracted from the image data in the standardized dataset using a convolutional neural network; Long Short-Term Memory (LSTM) networks are used to extract temporal feature vectors from time series data. For structured log data, a fully connected deep neural network is used to extract semantic feature vectors; The spatial feature vector, the temporal feature vector, and the semantic feature vector are grouped according to the sensor source. Each group of feature vectors is multiplied by the real-time reliability score of the corresponding sensor as a dynamic weight and then normalized to generate a weighted feature vector. The weighted feature vectors are concatenated and then input into a multi-head attention mechanism for fusion, outputting a state representation vector of the digital twin. The state representation vector is used as the unified input for both the bidirectional gated loop model and the extreme gradient boosting model.

[0009] Furthermore, the state representation vector is input into a preset bidirectional gated loop model, and the fault probability prediction result within a preset future time window is output, including: The state representation vector is input into a bidirectional gated loop model to learn the trend prediction of equipment state parameters (laser bias current, housing temperature, optical output power parameters) from the historical operating data of the equipment, and the result of the trend prediction is obtained. The bidirectional gated recurrent model is trained using the cross-entropy loss function and the Adam optimizer, and whether performance degradation occurs within a preset future time period (24 hours) is set as the prediction label. The trained bidirectional gated recurrent model is deployed on an edge inference node. It receives the state representation vector at a preset time (10 minutes) and performs rolling prediction based on the change trend prediction results, outputting the fault probability prediction results. When the failure probability prediction result exceeds the preset safety probability threshold, an early warning message containing the failure probability prediction result and the change trend prediction result is generated and the early warning is triggered. The failure probability prediction results are used as a condition for activating the dual deep Q network model.

[0010] Furthermore, the state representation vector is input into a preset extreme gradient boosting model, and the fault root cause diagnosis result is output, including: The state representation vector is input into the extreme gradient boosting model to analyze historical fault data and real-time network status, and to identify the key features that lead to the fault. The extreme gradient boosting model is trained using an objective function that includes a regularization term, and classification and regression analysis are performed based on the key features. The model complexity is limited by the regularization term. The trained extreme gradient boosting model is combined with a historical fault sample library to rank the key features by importance, generating a root cause localization result that includes the key features and importance assessment. The root cause localization results and the fault probability prediction results are used as conditions for activating the dual deep Q network model.

[0011] Furthermore, based on the failure probability prediction results and root cause localization results, a dual-depth Q-network model is activated to generate an optimal configuration strategy, including: A virtual model of a programmable optical transceiver is constructed, and a dual-depth Q-network model is activated based on the failure probability prediction results and the root cause localization results. The state representation vector is input into the dual-deep Q network model as the state space, and the modulation format, symbol rate, and forward error correction coding strength are used as the action space. Reinforcement learning is performed using a reward function to optimize the dual-deep Q network model. The optimized dual-deep Q-network model is interactively learned with the virtual model. Based on the root cause localization results and the current network status, hardware configuration actions are selected to generate the optimal configuration strategy. The current network status includes device operating parameters, link health status, historical performance indicators, and real-time alarm information.

[0012] Furthermore, the optimal configuration strategy is sent to a programmable optical transceiver for execution, and simultaneously input into a bidirectional long short-term memory simulation network, outputting simulation verification values ​​of bit error rate and optical signal-to-noise ratio, including: The modulation parameters in the optimal configuration strategy are preprocessed with the real-time acquired optical signal waveform and link status data to generate input data for a bidirectional long short-term memory simulation network. The input data is fed into a bidirectional long short-term memory simulation network for amplitude and phase fitting. The bidirectional long short-term memory simulation network includes a forward long short-term memory layer, a backward long short-term memory layer, and a fully connected output layer. The amplitude and phase of the optical signal are fitted together by the fully connected output layer to output the simulated verification values ​​of bit error rate and optical signal-to-noise ratio. The simulation verification value is compared with the actual measurement value. If the consistency error is less than the preset tolerance, the optimal configuration strategy is confirmed as a valid configuration strategy. If the consistency error is equal to or greater than the preset tolerance, the process returns to the reconfiguration strategy step to regenerate the configuration strategy.

[0013] Further, the simulation verification value is compared with the actual measured value for consistency. If the consistency error is less than a preset tolerance, the optimal configuration strategy is confirmed as a valid configuration strategy; if the consistency error is equal to or greater than the preset tolerance, the process returns to the reconfiguration strategy step to regenerate the configuration strategy, including: The bit error rate and optical signal-to-noise ratio in the simulation verification values ​​are compared with the actual measured values ​​obtained from the multi-source heterogeneous sensing data to obtain the comparison results. The comparison results include consistency errors; When the consistency error is less than the preset tolerance, the optimal configuration strategy is confirmed as a valid configuration strategy, and a configuration validity evaluation report containing the fault probability prediction result, the root cause location result, the valid configuration strategy, and the simulation verification value is generated. When the consistency error is equal to or greater than the preset tolerance, the reconfiguration process is triggered, the simulation verification value and the consistency error are fed back to the dual deep Q network model, and the configuration strategy generation step is returned to generate a new configuration strategy until an effective configuration strategy is obtained. The effective configuration strategy, the failure probability prediction results, the root cause location results, and the simulation verification values ​​are integrated into the digital twin platform to construct a full lifecycle operation archive for communication towers.

[0014] Furthermore, based on the full lifecycle operation records, the similarity of operation modes among different communication towers is identified, and a cross-tower collaborative optimization mechanism is established. When a potential risk occurs in a certain tower, resource reservation and traffic diversion plans for nearby healthy towers can be triggered, including: The entire lifecycle operation file is stored based on a time-series database, associated with a unique device identifier, and a snapshot of the device's status from installation to maintenance is recorded. The trajectory distance between the state snapshots of different towers is calculated using a dynamic time warping algorithm. When the trajectory distance is less than a preset threshold, the towers are determined to have similar operating modes. A topology map of neighboring towers is constructed based on geographical distance and similar operating modes, and a cross-tower collaborative optimization mechanism is established. Based on the full lifecycle operation archive, deep reinforcement learning algorithms are used to dynamically adjust network parameters and optimize resource allocation according to the neighboring tower topology map. When the predicted failure probability of a certain tower exceeds a preset safety probability threshold, based on the root cause location result and the topology map of neighboring towers, a portion of the traffic of the current tower is routed to the healthy tower link. Trigger the resource reservation and traffic diversion plan for healthy towers in the neighboring tower topology map to achieve load balancing; The effective configuration strategy is shared among neighboring towers through the cross-tower collaborative optimization mechanism; The traffic diversion results and operation and maintenance reports are fed back to the digital twin platform to form a closed-loop optimization.

[0015] Secondly, a digital twin system for monitoring the operation of communication tower equipment includes: Data acquisition module: used to collect multi-source heterogeneous sensor data deployed on communication towers to obtain monitoring datasets of equipment operation; Data preprocessing module: used to preprocess the monitoring dataset to generate a standardized dataset with a unified spatiotemporal reference; The weighted fusion module is used to perform real-time credibility scoring on each sensor data item in the standardized dataset; and to dynamically weight and fuse the credibility scores to output the state representation vector of the unified communication tower digital twin. Early warning module: used to input the state representation vector into a preset bidirectional gated loop model and output the fault probability prediction result within a preset time window; when the fault probability prediction result exceeds a preset safety probability threshold, an early warning message is triggered. Fault diagnosis module: Based on the acquired fault probability prediction results, the state representation vector is input into a preset extreme gradient boosting model, and the fault root cause diagnosis results are output.

[0016] Strategy generation module: used to activate the dual deep Q network model and generate the optimal configuration strategy based on the failure probability prediction results and root cause localization results; Strategy verification module: This module is used to send the optimal configuration strategy to the programmable optical transceiver for execution, and simultaneously input the bidirectional long short-term memory simulation network, outputting the simulation verification values ​​of bit error rate and optical signal-to-noise ratio; it compares the simulation verification values ​​with the actual measured values, judges the validity of the configuration strategy based on the comparison results, and obtains the effective configuration strategy; Tower traffic diversion module: It is used to integrate effective configuration strategies, failure probability prediction results and simulation verification values ​​into the digital twin platform to build a full life cycle operation file of communication towers; based on the full life cycle operation file, it identifies the similarity of operation modes between different communication towers, establishes a cross-tower collaborative optimization mechanism, and triggers resource reservation and traffic diversion plans for nearby healthy towers when a potential risk occurs in a certain tower.

[0017] The beneficial effects of this invention are as follows: It achieves deep fusion representation of the physical and network layer states through unified spatiotemporal benchmark alignment and a four-dimensional confidence scoring mechanism for multi-source heterogeneous sensor data; it utilizes a bidirectional gated recurrent network to perform long-term time-series modeling of device degradation trends, supporting high-confidence fault warnings more than 24 hours in advance; it combines an extreme gradient boosting model to achieve automated root cause localization, eliminating reliance on manual experience; it constructs a perception-decision-execution-verification closed loop through a dual-deep Q-network and a bidirectional LSTM simulation network, enabling programmable optical transceiver configuration strategies to dynamically adapt to real-time network states; and it establishes a cross-tower collaborative optimization mechanism based on a full lifecycle archive and a dynamic time warping algorithm, completing resource reservation and traffic diversion before single-point risks occur, significantly improving the autonomy and resilience of optical communication networks. Therefore, this invention systematically breaks through the functional bottlenecks of existing optical communication network management systems through the organic integration of digital twin modeling, deep learning inference, and closed-loop control technologies. Attached Figure Description

[0018] Figure 1 This is a schematic flowchart of a digital twin method for monitoring the operation of communication tower equipment provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a digital twin system module for monitoring the operation of communication tower equipment, provided in an embodiment of the present invention. Detailed Implementation

[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0020] At least one embodiment of the present invention discloses a digital twin system and method for monitoring the operation of communication tower equipment, comprising: like Figure 1 As shown, a digital twin method for monitoring the operation of communication tower equipment includes the following steps: Step 1: Collect multi-source heterogeneous sensor data deployed on the communication tower to obtain a monitoring dataset of equipment operation; Step 2: Preprocess the monitoring dataset to generate a standardized dataset with a unified spatiotemporal reference; Step 3: Perform real-time credibility scoring on each sensor data item in the standardized dataset; dynamically weight and fuse the credibility scores to output the state representation vector of the unified communication tower digital twin; Step 4: Input the state representation vector into a preset bidirectional gated loop model and output the fault probability prediction result within a preset time window; when the fault probability prediction result exceeds a preset safety probability threshold, trigger an early warning message; Step 5: Based on the obtained fault probability prediction results, input the state representation vector into the preset extreme gradient boosting model and output the fault root cause diagnosis results.

[0021] Step 6: Based on the failure probability prediction results and root cause localization results, activate the dual-depth Q-network model to generate the optimal configuration strategy; Step 7: Send the optimal configuration strategy to the programmable optical transceiver for execution, and simultaneously input it into the bidirectional long short-term memory simulation network to output the simulation verification values ​​of bit error rate and optical signal-to-noise ratio; compare the simulation verification values ​​with the actual measured values, and determine the effectiveness of the configuration strategy based on the comparison results to obtain an effective configuration strategy; Step 8: Integrate the effective configuration strategy, failure probability prediction results and simulation verification values ​​into the digital twin platform to construct a full life cycle operation file for communication towers; based on the full life cycle operation file, identify the similarity of operation modes between different communication towers, establish a cross-tower collaborative optimization mechanism, and when a tower has a potential risk, trigger the resource reservation and traffic diversion plan of the neighboring healthy towers.

[0022] The present invention provides a digital twin method for monitoring the operation of communication tower equipment, and the specific implementation of this method is as follows: In practical deployment scenarios, communication towers are equipped with various types of sensors and monitoring devices, including but not limited to optical performance monitoring (OPM) modules, digital coherent receivers with built-in diagnostic units, CMOS (Complementary Metal-Oxide-Semiconductor) image sensors, MEMS (Micro-Electro-Mechanical Systems) vibration sensors, temperature and humidity sensors, and wind speed sensors. These sensing devices are connected to an edge gateway deployed at the base of the tower or in a data center via industrial Ethernet or RS-485 bus. This edge gateway integrates a GPS timing module and a coordinate calibration unit to assign a unified timestamp to raw data from different sensors and convert the spatial location information of each sensor to the same geographic coordinate system, such as the WGS-84 (1984 World Geodetic System) coordinate system, thereby achieving timestamp alignment and spatial coordinate unification of multi-source heterogeneous sensor data. The data stream output from the edge gateway is transmitted to an edge computing node deployed at the same site. This node serves as the front-end processing unit of the entire digital twin system, undertaking core tasks such as data preprocessing, feature extraction, model inference, and policy distribution.

[0023] Upon entering the edge computing node, the raw sensor data is first standardized by the data preprocessing module. This module performs missing value imputation, filling in data gaps caused by communication interruptions or sensor malfunctions using linear interpolation or methods based on the historical sliding window mean. Subsequently, the modules normalize the units of data with different physical dimensions; for example, optical power is converted to dBm, vibration signals to m / s², and temperature to degrees Celsius. Simultaneously, to address inconsistent sampling frequencies, resampling technology is used to unify all data to a 5-minute sampling period, thereby generating a standardized dataset with a unified spatiotemporal reference. This standardized dataset serves as the foundational input for subsequent processing and is fed into the weighted fusion module.

[0024] The weighted fusion process comprises two steps: a credibility scoring step and a feature fusion step. The credibility scoring step independently calculates a real-time credibility score for each class of sensor data items in the standardized dataset, with a score range of 0 to 1. This step employs a four-dimensional dynamic credibility evaluation mechanism. The first dimension is single-sensor internal consistency detection: For time-series data such as optical power and OSNR (optical signal-to-noise ratio), a sliding window Z-score detection algorithm is used to identify abrupt change points. The calculation formula is Z=(x-μ) / σ to calculate how many standard deviations the current value deviates from the historical mean. Here, Z represents the standard deviation multiple of the current data point from the mean, used to measure the degree of anomaly of the data point. The larger the absolute value of Z, the more abnormal the current value is: |Z|<2 is normal fluctuation, |Z|>3 is significant anomaly. x represents the raw data value collected by the sensor at the current time (time t), such as optical power -25.3dBm or OSNR value 28.5dB; σ is the arithmetic mean of the sensor's output data over the past 24 hours (288 5-minute sampling points); and σ is the standard deviation, representing the dispersion of the data over the past 24 hours. When the absolute value of the Z-score of 3 consecutive sampling points is greater than 3.0, it is marked as an internal anomaly, and the internal consistency score drops to the range of 0.4-0.6. For images of the tower structure captured by the image sensor, image quality quantification indicators are generated by calculating image sharpness (gradient magnitude calculated using the Laplacian operator), contrast (standard deviation of grayscale histogram), and signal-to-noise ratio (ratio of signal power to noise power). The weighted combination of these three indicators is mapped to the 0-1 range as an internal consistency score.

[0025] The second dimension involves multi-sensor cross-validation: For sensors of the same type deployed in the same tower area (e.g., one temperature sensor on each of the north and south sides), the system calculates the difference in readings. If the difference between any two sensor readings exceeds twice the historical standard deviation of the physical quantity (e.g., a temperature difference greater than 4°C, a wind speed difference greater than 2 m / s), the cross-validation score for both sensors is reduced to the range of 0.5-0.7. If only a single sensor reading is abnormal, its score is reduced to 0.3.

[0026] The third dimension is historical data pattern matching: The system maintains a sensor steady-state distribution model based on 30 days of historical normal operation data. This model records the data distribution characteristics (mean, variance, kurtosis, skewness) of each sensor in different time periods (0-6, 6-12, 12-18, 18-24). KL divergence is used to measure the deviation of the current observation distribution P(X) from the historical steady-state distribution Q(X). The calculation formula is KL(P||Q)=Σ[P(X)log(P(X) / Q(X))]; where KL(P||Q) is the deviation of the current observation data from the historical steady-state distribution, P represents the probability distribution of the data collected by the sensor in the current time period (e.g., today 14:00-18:00), X represents the possible output values ​​of the sensor (e.g., OSNR: 24dB, 25dB, 26dB...), Σ represents the summation over all possible X values, and log is the logarithmic operator. When the KL divergence is less than 0.5, the historical pattern matching score is 1.0; when the KL divergence is in the range of 0.5-1.5, the score linearly decays to 0.5; when the KL divergence is greater than 1.5, the score drops to 0.3.

[0027] The fourth dimension is environmental context awareness and correction: The system accesses local weather station data (including rainfall intensity, visibility, wind speed, temperature, and humidity) via an API (Application Programming Interface) and reads the vibration RMS value (root mean square value) output from the triaxial accelerometer installed at the tower base. Based on the extreme levels of environmental parameters, the system dynamically adjusts the context correction coefficient according to preset rules. The specific rules are as follows: (1) When the visibility is less than 500 meters, the confidence level of all CMOS image sensors is multiplied by a correction factor of 0.3; (2) When the wind speed exceeds 15 m / s, the drift compensation algorithm is enabled for the vibration sensor data (using Kalman filter to estimate the system state from the measurement data containing noise and eliminate the baseline drift of wind-induced vibration), and the residual after compensation (actual vibration minus the predicted value of wind-induced vibration model) is used as the confidence correction factor. (3) When the rainfall intensity exceeds 50 mm / h, the reliability of the outdoor temperature and humidity sensor is multiplied by a correction factor of 0.6; (4) When the ambient temperature exceeds 45℃ or is below -25℃, the reliability of the temperature sensor inside the optical module is multiplied by a correction factor of 0.7.

[0028] The scores of the four dimensions mentioned above are weighted and summed with weight coefficients of 0.3, 0.25, 0.25, and 0.2 to generate the final real-time credibility score ∈ [0, 1]. This score is updated every 5 minutes and serves as the dynamic weight coefficient for subsequent feature fusion stages.

[0029] The feature fusion step receives a standardized dataset and its corresponding credibility score, and calls convolutional neural networks (CNN), long short-term memory networks (LSTM), and deep neural networks (DNN) to extract features from the three types of data.

[0030] CNN image feature extraction is used to process tower structure state images (including weld areas, bolt connections, and the appearance of fiber optic cable junction boxes) output by image sensors. The input size is uniformly 224×224 pixels (RGB three-channel: red, green, and blue color image). High-level semantic features are extracted through a ResNet-18 (18-layer residual network) backbone network (pre-trained on a large-scale image recognition dataset containing 14 million images on ImageNet, and further trained on the tower image dataset after learning basic visual capabilities on ImageNet). The ResNet-18 contains four residual blocks, with skip connections within each block to mitigate the vanishing gradient problem. The final global average pooling layer outputs a 512-dimensional feature vector, Fimg, which contains high-level semantic information such as structural damage, foreign object occlusion, corrosion, and rust in the image. To adapt to the resource constraints of edge computing, the model adopts INT8 quantization (compressing 32-bit floating-point parameters into 8-bit integers) and weight pruning (removing parameters with absolute values ​​less than 0.01), compressing the model size from 44MB to 11MB and reducing the inference latency from 28ms to 7ms.

[0031] LSTM temporal feature extraction is used to process temporal sensor data, including nine types of signals: optical power, OSNR (optical signal-to-noise ratio), bit error rate, dispersion parameters, polarization mode dispersion, temperature, humidity, wind speed, and vibration RMS value. Each type of signal constitutes a univariate time series, with input data in 5-minute granularity over the past 24 hours, totaling 288 time steps. The nine time series are first input into independent LSTM encoders (64-dimensional hidden layers), with each LSTM capturing the temporal dependence features of its respective signal (e.g., the diurnal periodicity of temperature and the slow drift trend of OSNR). The final hidden states (64 dimensions each) of the nine LSTM encoders are concatenated and compressed to 128 dimensions through a fully connected layer, yielding the temporal feature vector Fts. This vector contains the joint evolution pattern of the multivariate temporal data (e.g., the correlation between rising temperature and decreasing OSNR).

[0032] DNN-based structured log feature extraction is used to process device logs and network alarm information. A log sample might be: Event Occurred (2024-01-15 14:23:07) | Alarm Level {INFO, WARNING, ERROR} | Which Device Had the Problem (Optical Transceiver No. 3) | Specific Problem (Laser bias current exceeds threshold 85 mA). The system uses a pre-trained BERT-tiny model (bidirectional Transformer encoder representation model) (12-layer Transformer transformer, 768-dimensional hidden layer) to encode the log text, extracting the 768-dimensional vector at the [CLS] (classification label) position as a sentence-level semantic representation. Since multiple logs (0-50) may be generated within a single time window, the system uses an attention pooling mechanism to perform a weighted average of all log vectors. The weights are determined by the log severity (mapped to 0.3 / 0.6 / 1.0) and the time decay factor (earlier logs have lower weights, with a decay half-life of 6 hours). The pooled 768-dimensional vector is compressed into a 128-dimensional semantic feature vector Flog through a three-layer fully connected network (768→512→256→128, each layer uses ReLU activation (modified linear unit activation function) and Dropout (random deactivation regularization technique) with a regularization rate of 0.3).

[0033] Dynamic weighted fusion: Three types of feature vectors, Fimg (512-dimensional), Fts (128-dimensional), and Flog (128-dimensional), are grouped according to their sensor sources. For example, camera images, sensor time-series data, and equipment logs are dynamically weighted according to their respective credibility. An 8-head multi-attention mechanism automatically discovers the correlation between different data (such as the correspondence between image shaking and vibration readings, and the confirmation between log alarms and actual temperatures). Finally, they are intelligently fused into a 512-dimensional digital twin state vector. This vector fully represents the current health status of the communication tower in a compact digital form (including equipment status, environmental impact, network load, and potential risks). It can be directly used as standard input for subsequent fault prediction models (BiGRU), root cause diagnosis models (XGBoost), and configuration optimization models (DDQN). The normalized 768-dimensional feature vector (512+128+128) is input into the 8-head multi-attention mechanism, with each attention head having a dimension of 96 (768 / 8). The attention mechanism uses Query-Key-Value (QKV) computation to capture the correlation between different modal features (e.g., spatial-temporal alignment of fiber optic cable swaying and vibration sensor readings in images, semantic-physical correspondence between temperature alarms and actual temperature time-series curves in logs). The outputs of the eight attention heads are concatenated and projected to 512 dimensions through a fully connected layer. This is then transformed nonlinearly through two layers of residual fully connected networks (each layer using normalization and Gaussian error activation), ultimately outputting a unified 512-dimensional digital twin state representation vector, Sdt. This vector comprehensively represents the current operating status of the communication tower in a compact numerical form, including multi-dimensional information such as equipment health, environmental disturbances, network load, and potential risks. It serves as a universal input interface for all subsequent AI models (BiGRU, XGBoost, DDQN).

[0034] The state representation vector is periodically (every 5 minutes) input to the early warning module. The early warning module is deployed on the edge proxy server, and its core is a bidirectional gated recurrent neural network (BiGRU) trained offline. During the training phase, the model uses historical running data to construct a sample set. Each sample contains a sequence of state representation vectors for 7 consecutive days × 288 time steps (5-minute granularity per day). The label y indicates whether a performance degradation event (defined as a bit error rate exceeding 1e-6 or an optical power drop exceeding 3dB for more than 10 minutes) will occur within the next 24 hours. y∈{0,1}, where 0 represents normal and 1 represents failure.

[0035] The BiGRU (Bidirectional Gated Recurrent Unit) model structure comprises a forward GRU layer (gated recurrent unit) and a backward GRU layer, each with 128 hidden units, enabling it to simultaneously capture the historical trends and future evolution patterns of device state parameters. The model uses the cross-entropy loss function L=-[y·log(p)+(1-y)·log(1-p)] to measure the difference between the predicted fault probability p and the true label y. The optimizer is Adam (Adaptive Moment Estimator), with an initial learning rate of 0.001, β1=0.9, β2=0.999, and ε=1e-8. During training, mini-batch stochastic gradient descent with a batch size of 32 is used, decaying the learning rate to 10% of its original value every 50 epochs. Simultaneously, the early-stopping metric is monitored on the validation set; training is terminated early if the validation loss does not decrease for 10 consecutive epochs to prevent overfitting. The final trained model achieved a prediction accuracy of 99.2%, a recall of 98.7%, and an F1 score of 98.9% on the test set.

[0036] The specific explanations of the cross-entropy loss function and the parameters of the Adam (adaptive estimation) optimizer are as follows: In the cross-entropy loss function, L represents the loss value, ranging from [0, +∞). The smaller the loss value, the closer the model prediction is to the true label; y is the true label, taking values ​​of 0 (normal state) or 1 (fault state); p is the model's output fault probability prediction value, ranging from [0, 1], representing the probability of a fault occurring within the next 24 hours; log is the natural logarithm function. When y = 1 (actual fault occurs), the loss function simplifies to L = -log(p). In this case, if the model predicts p close to 1 (correctly predicting a fault), the loss is close to 0; if p is close to 0 (incorrectly predicting normal), the loss tends towards infinity, and a stronger penalty is applied. When y = 0 (actually normal), the loss function simplifies to L = -log(1-p). In this case, if the model predicts p close to 0 (correctly predicting normal), the loss is close to 0; if p is close to 1 (incorrectly predicting a fault), the loss tends towards infinity. This loss function applies a non-linear penalty to incorrect predictions in logarithmic form, forcing the model output to be closer to the true probability distribution.

[0037] The parameters in the Adam optimizer have the following meanings: β1=0.9 is the exponential decay rate of the first moment (mean gradient), controlling the degree of retention of historical gradient information. 0.9 indicates that the current gradient accounts for 10% of the weight, and the accumulated gradient in the past accounts for 90% of the weight, which is used to smooth the gradient update direction and avoid the influence of single gradient noise; β2=0.999 is the exponential decay rate of the second moment (mean squared gradient), controlling the degree of retention of historical gradient fluctuation information. 0.999 indicates a smoother estimate of the long-term trend of gradient changes, which is used to adaptively adjust the learning rate of each parameter; ε=1e-8 (i.e. 0.00000001) is a numerical stability term, added to the denominator to prevent division by zero error. When the estimated value of the second moment of a parameter's gradient is close to 0, ε ensures that the update step size does not tend to infinity; the initial learning rate of 0.001 is the baseline update step size. The Adam optimizer adaptively adjusts the actual learning rate for each parameter based on this. For parameters with drastic gradient changes, the learning rate is reduced to maintain stability, and for parameters with gentle gradient changes, the learning rate is increased to accelerate convergence. After deployment, the BiGRU model receives the latest state representation vector in a streaming inference manner. It updates the internal hidden state every time it receives a new 5-minute time step of data and outputs the failure probability prediction result p∈[0,1] for the next 24 hours. When the probability exceeds the safety threshold of 0.85, the system generates a structured early warning message. The message is encapsulated in JSON format and includes the following fields: (1) unique identifier of the tower; (2) early warning trigger timestamp (UTC format); (3) failure probability value; (4) time series array of the past 7 days of change trends of key state parameters (optical signal-to-noise ratio, optical power, effective vibration value, temperature); (5) possible failure time points based on trend extrapolation; (6) prediction confidence interval. The message is pushed to the monitoring screen of the operation and maintenance center and the mobile terminal APP of the on-duty personnel via the MQTT (Internet of Things Message) protocol.

[0038] If an alert is triggered, the state representation vector is simultaneously sent to the fault diagnosis module. This module loads an extreme gradient boosting model trained on a historical fault sample database. The XGBoost model uses the Gradient Boosting Decision Tree (GBDT) ensemble framework, whose objective function not only includes a loss term to measure prediction error but also introduces a regularization term to prevent model overfitting. The objective function is defined as: Obj(θ) = ΣiL(yi, Yi) + ΣkΩ(fk); where Obj(θ) represents the overall optimization objective of the objective function, θ represents the entire set of model parameters (including the structural parameters of all decision trees and the weights of leaf nodes); Σi is the summation symbol, indicating the accumulation of all training samples, i is the sample index, ranging from 1 to N (N is the total number of training samples, in this implementation N = 840, i.e., the number of training set samples); L(yi, Yi) is the loss function, measuring the loss of a single sample. The objective function is defined as follows: yi is the true label of the i-th sample (e.g., fault type: 0 indicates laser aging, 1 indicates junction box micro-bending, etc., 12 categories); Yi is the model's predicted output value for the i-th sample (a 12-dimensional probability vector in multi-class tasks); Σk is the summation over all decision trees, where k is the tree index, ranging from 1 to K (K is the total number of decision trees, K=200 in this implementation); Ω(fk) is the regularization term, penalizing the complexity of the k-th decision tree fk, where fk represents the structure and parameters of the k-th decision tree. The objective function ensures model prediction accuracy (making the predicted value close to the true label) through the first term ΣiL(yi, Yi), and controls model complexity (preventing overfitting) through the second term ΣkΩ(fk). The model is optimal when the sum of the two terms is minimized. The first term, ΣiL(yi, Yi), measures the difference between the model's predicted value Yᵢ and the true label yᵢ. The second term, ΣkΩ(fk), penalizes the complexity of the k-th decision tree, defined as: Ω(f) = γ·T + 1 / 2λΣ|wj|². Here, Ω(f) represents the complexity penalty for a single decision tree f; a larger value indicates a more complex tree. The optimization process tends to select simpler tree structures with smaller Ω(f), T is the number of leaf nodes, wj is the output weight of the j-th leaf node, γ is the penalty coefficient for the number of leaf nodes (set to 0.2 in this implementation), and λ is the L2 regularization coefficient (set to 1.0 in this implementation). This regularization mechanism forces the model to learn simpler, more general decision rules by limiting the complexity (number of nodes) and the magnitude of the output weights, avoiding over-memorizing noisy features or random associations in the training data.

[0039] The specific training configuration is as follows: maxdepth=8 (maximum tree depth), minchildweight=5 (minimum sum of child node sample weights), minsplitloss=0.1 (split gain threshold, also known as gamma parameter), subsample=0.8 (randomly sample 80% of the samples during training for each tree), colsamplebytree=0.8 (randomly sample 80% of the features during training for each tree), learningrate=0.1 (learning rate), and nestimators=200 (number of decision trees). The training samples are derived from 1200 real-world fault event records from the past two years. Each record contains 512-dimensional state features and fault type labels (categorized into 12 types, including laser aging, junction box micro-bending, fiber breakage, power module failure, and temperature control system anomaly) for the seven days prior to the fault. The training, validation, and test sets are divided in a 7:2:1 ratio.

[0040] The model input consists of the original features unpacked from the current state representation vector (not the fused 512-dimensional vector, but a concatenation of features extracted by CNN, LSTM, and DNN, totaling approximately 300 dimensions). The output consists of the importance ranking of each feature (i.e., the average information gain brought by each feature when splitting across all tree nodes) and a 12-dimensional probability vector (corresponding to 12 fault types). The system extracts the top 5 features with the highest Gain values, combines them with the degree to which their current values ​​deviate from the normal range (quantized using Z-score), and matches them against a predefined failure mode rule base (this rule base was written by domain experts based on physical mechanisms and contains 180 condition-action rules) to generate a root cause diagnosis result containing the following fields: (1) Root cause device / component name (e.g., trunk fiber optic splice box). (2) Failure mode description (e.g., increased microbending loss of optical fiber in the junction box may be caused by loose fastening bolts or aging of the colloid). (3) Diagnostic confidence ∈ [0, 1] (determined by the difference between the maximum and second largest probabilities of XGBoost output; the larger the difference, the higher the confidence). (4) The names of the first 5 key features and their current anomaly measures.

[0041] Based on the failure probability prediction and root cause diagnosis results, the policy generation module is activated. This module includes a Double Deep Q-Network (DDQN) reinforcement learning engine and a programmable optical transceiver (POT) virtual simulation model. DDQN is an improved version of Deep Q-Learning (DQN). By decoupling action selection and value evaluation, it effectively alleviates the problem of overestimation of Q-values ​​in DQN, improving learning stability and policy quality.

[0042] Action Space Definition: The system combines the three types of configurable parameters of the POT (Programmable Optical Transceiver) into a discrete action space A. There are three options for the modulation format M ∈ {QPSK (Quadrature Phase Shift Keying), 16-QAM (16th-order Quadrature Amplitude Modulation), 64-QAM (64th-order Quadrature Amplitude Modulation)}, three options for the symbol rate R ∈ {32 GBaud (32 gigabits / second), 40 GBaud (40 gigabits / second), 64 GBaud (64 gigabits / second)}, and three options for the FEC coding strength F ∈ {Soft Decision 7% Overhead SD-FEC (Soft Decision Forward Error Correction), Hard Decision 15% Overhead HD-FEC (Hard Decision Forward Error Correction), Concatenated Code 25% Overhead Concatenated-FEC (Concatenated Forward Error Correction)}. The action space A = M (modulation format set) × R (symbol rate set) × F (FEC coding strength set), resulting in 27 combinations. Each action a=m (the specific modulation format selected from the modulation format set M), r represents the specific symbol rate value selected from the symbol rate set R, and f represents the specific coding scheme selected from the FEC coding strength set F, representing a complete configuration switching operation.

[0043] State space definition: State s is composed of a current 512-dimensional digital twin state representation vector, link health indicators (current BER: bit error rate, the proportion of erroneous bits in received bits, measured OSNR, measured optical signal-to-noise ratio, OSNR margin, i.e., measured value minus the theoretical requirement of modulation format), predicted remaining equipment lifetime, link traffic utilization, and historical 24-hour configuration switching counts, totaling 520 features. The state space has high dimensionality but contains rich decision-making basis.

[0044] Reward function design: The reward function R characterizes the degree of improvement in system performance after performing action a, and is defined as: R = α·ΔOSNR + β·(-log) 10 BERnew / BERold)+γ·ηspectral-δ·Cswitch-ε·Overhead; Where R represents the reward value, measuring the degree of improvement in system performance after executing a certain configuration action; a larger R value indicates a better action. α is the weighting coefficient for OSNR improvement, set to 0.4 in this implementation, controlling the contribution ratio of optical signal-to-noise ratio improvement to the total reward. ΔOSNR represents the change in OSNR after configuration adjustment (in dB), with positive values ​​indicating improvement and negative values ​​indicating deterioration. β is the weighting coefficient for bit error rate improvement, set to 0.3 in this implementation, controlling the contribution ratio of BER reduction to the total reward. -log10 is the logarithmic reduction in BER, measuring the degree of bit error rate improvement; BERnew is the new bit error rate after the action, BERold is the original bit error rate before the action, and log10 is the logarithmic function with base 10: 10(1e-9 / 1e-6) = log10(10-3) = -3 (with a negative sign, it becomes 3, indicating a BER reduction of 3 orders of magnitude). γ is the weighting coefficient for spectral efficiency, set to... The value is 0.2, which controls the contribution ratio of spectrum utilization to the total reward; ηspectral is the spectral efficiency, which represents the transmission rate per unit spectrum bandwidth (unit: bit / s / Hz, bits per second per hertz), reflecting the efficiency of spectrum resource utilization. The higher the modulation order, the higher the spectral efficiency; δ is the weighting coefficient for configuring handover penalties, which is set to 0.05 in this implementation to suppress frequent handovers; Cswitch is the configuration handover penalty item. If the interval between two adjacent configurations is less than 30 minutes, Cswitch=1 (penalty is applied), otherwise Cswitch=0 (no penalty is applied) to avoid frequent handovers causing service interruptions and affecting user experience; ε is the weighting coefficient for FEC overhead, which is set to 0.05 in this implementation to penalize excessive error correction overhead; Ooverhead is the FEC overhead rate, which represents the redundancy ratio introduced by forward error correction coding (7%, 15%, or 25% in this implementation). The higher the overhead, the stronger the error correction capability, but the lower the spectrum efficiency. The reward function comprehensively considers the balance of multiple objectives, including link quality improvement (OSNR enhancement, BER reduction), spectrum efficiency improvement, and configuration handover costs (handover penalty, FEC overhead). The weighting coefficients α=0.4, β=0.3, γ=0.2, δ=0.05, and ε=0.05 are optimized and determined in the validation environment through hyperparameter grid search to ensure that the DDQN strategy improves link quality while taking into account spectrum efficiency and configuration stability.

[0045] DDQN Network Architecture and Training: DDQN consists of two structurally identical but parameter-independent deep neural networks: the main network Qmain(s, a; θ) and the target network Qtarget(s, a; θ). The main network is responsible for online interaction and policy iteration, while the target network is used to compute the temporal difference target value. Its parameter θ is softly updated every 200 steps from the main network parameter θ (using a parameter smoothing update method with τ=0.01). Both networks have the following structure: 520-dimensional input layer → 512-dimensional fully connected layer (ReLU activation) → 256-dimensional fully connected layer (ReLU activation) → 128-dimensional fully connected layer (ReLU activation) → 27-dimensional output layer (linear activation, each dimension corresponds to the Q-value of an action).

[0046] Training employs an ε-greedy exploration strategy (an ε-exploration strategy is a strategy that balances exploration and exploitation, randomly selecting actions for exploration with probability ε and exploiting actions with probability 1-ε, where ε decays exponentially from an initial value of 1.0 to a minimum of 0.05 (decay factor 0.995), ensuring sufficient exploration of the action space early on and convergence to the optimal strategy later. An experience replay buffer (storing historical interaction experiences for offline learning) has a capacity of 10,000 quadruplets (s, a, r, s') (state, action, reward, next state), and 32 experiences are randomly sampled from the buffer for mini-batch updates each time. The loss function uses the Huber loss piecewise function, combining the advantages of MSE (sensitive to small errors, fast convergence), mean squared error and MAE (robust to large errors, stable training), and mean absolute error, resulting in a loss function robust to outliers: Training employs an ε-greedy exploration strategy, with ε decaying exponentially from an initial value of 1.0 to a minimum of 0.05 (decay factor 0.995), ensuring sufficient exploration of the action space early on and convergence to the optimal strategy later. The experience replay buffer capacity is set to 10,000 quadruplets (s, a, r, s'), and 32 experiences are randomly sampled from the buffer for mini-batch updates each time. The loss function used is HuberLoss (combining the advantages of MSE and MAE, robust to outliers): The temporal difference objective value is calculated as yi = r-i + γ·Q-target(s'-i, argmax-a'Q-main(s'-i, a';θ);θ -), where ri is the immediate reward of the i-th sample, γ is the discount factor, set to 0.95 in this implementation to balance the importance of immediate reward and future reward (γ close to 1 emphasizes long-term reward, γ close to 0 emphasizes short-term reward); s'-i is the next state transitioned to after the i-th sample performs the action; argmax-a'Q-main(s'-i,a';θ) represents the action a' (the action that maximizes Q value) that the main network Q-main considers optimal in the next state s'-i; Q-target(s'-i,argmax-a'Q-main(s'-i,a';θ);θ - The expression () indicates that the Q-value of the optimal action is evaluated using the target network (Q-target). This embodies the core idea of ​​Double-DQN: using the main network to select actions, but using the target network to evaluate their value, decoupling the selection and evaluation processes, effectively avoiding the problem of overestimation of Q-values ​​in standard DQN, and improving policy quality and training stability. Note that this reflects the core of Double DQN: the temporal difference target value embodies the core of Double DQN: using the main network to select actions (argmax-a'Q-main), but using the target network to evaluate the value of the action (Q-target), decoupling the selection and evaluation processes, and avoiding overestimation. The discount factor γ = 0.95.

[0047] The training environment was a POT (Programmable Optical Transceiver) virtual simulation model, built using the physical layer link simulation tool VPItransmissionMaker (optical communication system simulation software developed by VPIphotonics). This model can quickly calculate the output BER (Bit Error Rate) and OSNR (Optical Signal-to-Noise Ratio) based on the input modulation format, symbol rate, FEC (Forward Error Correction) parameters, and link status (fiber length, dispersion, nonlinear coefficients). The simulation latency is approximately 50 milliseconds per iteration. The DDQN agent interacted with the virtual environment for 100,000 episodes (rounds, complete interaction sequences from the initial state to the final state in reinforcement learning), with a maximum of 50 configuration adjustments per episode. The total training time was approximately 72 hours. After training convergence, the DDQN policy, on average, improved the system's overall reward by 18% per configuration adjustment in the test environment, outperforming manual rule-based policies (5% improvement) and traditional DQN (12% improvement).

[0048] The optimal configuration strategy is executed by the programmable optical transceiver at the physical layer via the southbound interface. Meanwhile, the strategy parameters (modulation format, symbol rate, FEC type) and the real-time acquired optical signal waveforms (I / Q data) and link status data (such as received optical power and polarization mode dispersion) must undergo rigorous data preprocessing before they can be spliced ​​into an input sequence and sent to the strategy verification module.

[0049] Data preprocessing comprises four key steps: First, the real-time acquired I / Q waveform data comes from the high-speed ADC (analog-to-digital converter) built into the digital coherent receiver, with a sampling rate of 80 GSa / s. The raw data inevitably contains thermal noise, ADC quantization noise, and high-frequency electromagnetic interference. The system first applies a digital low-pass filter (Butterworth 5th order, cutoff frequency set to 1.2 times the signal bandwidth; for example, for a 40 GBaud symbol rate signal, the cutoff frequency is set to 48 GHz) to filter out high-frequency noise components exceeding the signal bandwidth. Simultaneously, a moving average filter (window length of 5 sampling points) is used to smooth ADC glitches. After pre-filtering, the signal-to-noise ratio (SNR) is improved from the original 18 dB to 25 dB.

[0050] Because different batches or models of coherent receivers may have different sampling rates (e.g., 80GSa / s vs. 100GSa / s), and the subsequent BiLSTM network requires the input time series to have a fixed time resolution, the system uses a cubic spline interpolation algorithm to resample all waveform data to a standard sampling rate of 100GSa / s. This interpolation algorithm maintains the smoothness of the signal envelope while avoiding time alignment errors caused by inconsistent sampling rates. For link state data (such as OSNR, dispersion, and polarization mode dispersion), these parameters are typically sampled once per second. The system uses linear interpolation to extend them to the same time resolution as the I / Q waveforms (one data point every 10 nanoseconds), ensuring that all input features are strictly aligned on the time axis. I / Q waveform data is essentially a complex signal z(t), and in-phase or quadrature waveform data is essentially a complex signal, which can be expressed as the complex signal value at time t equal to the in-phase component plus the imaginary unit multiplied by the quadrature component. The in-phase component is the real part of the complex signal, the quadrature component is the imaginary part, and the imaginary unit is a mathematical constant whose square equals -1. The system extracts the signal amplitude and phase angle through a mathematical transformation from rectangular coordinates to polar coordinates. The signal amplitude is calculated as follows: the amplitude value at time t equals the square of the in-phase component plus the square of the quadrature component, and the square root of the sum. The phase angle is calculated as follows: the phase value at time t equals the quadrature component divided by the in-phase component, and the arctangent function of the quotient. To prevent large amplitude features from dominating the neural network training process, the system performs zero-to-one normalization on the amplitude sequence. The specific normalization formula is: the normalized amplitude value at time t equals the original amplitude value minus the minimum value of the entire amplitude sequence, and then this difference is divided by the range obtained by subtracting the minimum value from the maximum value of the entire amplitude sequence. The minimum and maximum values ​​of the entire amplitude sequence are the minimum and maximum amplitude values ​​of all sampling points in the sequence, respectively. The phase sequence ranges from negative to positive pi, i.e., from -180 degrees to +180 degrees. This phase sequence is normalized using a linear mapping method. The specific normalization formula is: the normalized phase value at time t equals the original phase value plus pi, and then this sum is divided by twice pi, thus normalizing the phase sequence to the zero-to-one interval. Link state parameters are also normalized. For example, the range of the receiver optical power is from -30 dBmW to 0 dBmW (a dBmW is the logarithmic unit relative to one milliwatt of optical power), and the range of the dispersion coefficient is from 0 to 2,000 picoseconds per nanometer (a picosecond per nanometer is the physical unit of the dispersion coefficient). These parameters are all mapped to the range of 0 to 1 using the minimum-maximum value normalization method.

[0051] Bidirectional Long Short-Term Memory (LSTM) networks (a variant of recurrent neural networks that can simultaneously learn forward and backward dependencies in time series) have fixed input data length requirements, and the input time window must also be of fixed length. In this implementation, it is set to 2,048 time steps (corresponding to a signal period of 20.48 microseconds at a sampling rate of 100 billion times per second). The system uses a sliding window segmentation technique to divide the entire waveform sequence, with each window moving in 512 time steps (this step size ensures 75% data overlap between adjacent windows), generating multiple independent input samples through the sliding window method. Each input sample contains multiple feature data: normalized amplitude time series, normalized phase time series, one-hot encoded representation of modulation format (one-hot encoding is a categorical variable encoding method, using a three-dimensional vector to represent three modulation format options), one-hot encoded representation of symbol rate (using a three-dimensional vector to represent three symbol rate options), one-hot encoded representation of forward error correction type (using a three-dimensional vector to represent three error correction scheme options), scalar value of received optical power (the scalar is a single numerical value), scalar value of dispersion coefficient, and scalar value of polarization film dispersion. The total feature dimensions are 2048 multiplied by 2 plus 9, equaling 4105-dimensional feature vectors. The strategy verification module includes a bidirectional long short-term memory simulation network designed with a dual-model linkage fitting architecture for amplitude and phase. The core idea behind this architecture is that the amplitude and phase of an optical signal do not evolve independently during fiber optic transmission. Instead, they are simultaneously influenced by multiple physical effects, including dispersion (the difference in propagation speeds of different wavelengths of light in the fiber material, leading to pulse broadening in the time domain), self-phase modulation (the modulation of the optical signal's phase by its own intensity changes through fiber nonlinearity), and cross-phase modulation (the mutual modulation of phases between optical signals from different wavelength channels through fiber nonlinearity). This results in a complex physical coupling relationship between amplitude and phase. Traditional modeling methods that treat amplitude and phase evolution as two independent processes and establish separate mathematical models may lead to asynchronous amplitude and phase trends in simulation predictions, thus violating the true physical laws of fiber optic transmission.

[0052] The specific network structure is designed as follows: The input layer receives the preprocessed amplitude sequence and phase sequence, and the two are respectively input to the independent Embedding layer (each containing 64 neurons, which maps the original time series to the high-dimensional feature space).

[0053] Amplitude modeling branch: This branch consists of a forward LSTM layer (128 hidden units) and a backward LSTM layer (128 hidden units). The forward layer learns the cumulative amplitude impairments from the transmitter to the receiver (such as fiber attenuation and splitter insertion loss), while the backward layer learns the global constraints of amplitude across the entire link (such as total power conservation). The hidden states from both directions are concatenated and then output as the amplitude distortion prediction value ΔA(t) through a fully connected layer.

[0054] Phase modeling branch: This branch also includes a forward LSTM layer (128 hidden units) and a backward LSTM layer (128 hidden units). The forward layer learns the cumulative phase rotation from the transmitter to the receiver (e.g., linear phase shift caused by dispersion, nonlinear phase shift caused by nonlinear effects), while the backward layer learns the global coherence constraints of the phase throughout the link. The hidden states from both directions are concatenated and then output as a phase rotation prediction value Δφ(t) via a fully connected layer.

[0055] Linkage and Fusion Layer: This is the key innovation of this architecture. This layer performs cross-attention interaction (a mechanism for attention interaction between different modalities or sequences) on the intermediate hidden states (each 256-dimensional, derived from forward and backward concatenation) of the amplitude branch and the phase branch. Specifically: (1) Using the hidden state of the amplitude branch as Query and the hidden states of the phase branch as Key and Value, the amplitude dependence weight on the phase is calculated through the attention mechanism to generate phase-enhanced amplitude features; (2) Conversely, using the hidden state of the phase branch as Query and the hidden states of the amplitude branch as Key and Value, phase-enhanced phase features are generated. The two sets of enhanced features are fed back to the output layer of their respective branches after residual connection.

[0056] Output layer: The amplitude modeling branch outputs the amplitude distortion increment sequence at each time step (used to predict amplitude distortion), and the phase modeling branch outputs the phase rotation increment sequence at each time step (used to predict phase rotation). The system reconstructs the complex signal at the receiver based on these distortion increments. The reconstruction formula is as follows: the reconstructed complex signal value at time t is equal to the original amplitude plus the amplitude distortion increment, and then multiplied by an exponential function value with imaginary units as the base and the original phase plus the phase rotation increment as the exponent (in the complex signal reconstruction process, the imaginary unit satisfies that its square equals negative one, and the exponential function is the natural exponential function). Then, the estimated bit error rate is calculated using a standard bit error rate estimation algorithm (such as the error vector amplitude method, which assesses the signal quality deviation by measuring the Euclidean distance between the received signal vector and the ideal signal vector). Finally, the estimated optical signal-to-noise ratio is calculated using the signal-to-noise ratio calculation formula, which is: the estimated optical signal-to-noise ratio is equal to the quotient of ten times the signal power divided by the noise power, taken as the logarithm to the base ten (where the signal power is the power value of the useful optical signal, the noise power is the power value of the background noise, and the logarithm to the base ten is a commonly used logarithmic function).

[0057] During the training phase, the bidirectional long short-term memory simulation network was trained using 8,000 sets of historically accumulated experimental datasets. Each set of training data consisted of two parts: the first part was the input data, including in-phase quadrature complex signal waveforms and corresponding link physical parameters collected under different modulation format options, different symbol rate options, and different forward error correction coding configuration options; the second part was the label data, including the measured bit error rate (obtained by bit-by-bit comparison of the bit streams at the transmitting and receiving ends) and the measured optical signal-to-noise ratio (obtained directly by a spectral analyzer) under the corresponding configuration. The loss function adopted a weighted summation form jointly optimized by amplitude prediction loss, phase prediction loss, and final output loss, specifically expressed as: the total loss function equals the amplitude loss weight coefficient multiplied by the amplitude distortion prediction loss, plus the phase loss weight coefficient multiplied by the phase rotation prediction loss, plus the output layer loss weight coefficient multiplied by the final output prediction loss.

[0058] The parameters of the joint loss function for the bidirectional long short-term memory network are explained in detail below: The total loss function represents the comprehensive objective function that needs to be minimized during training. This function comprehensively measures three aspects: amplitude prediction error, phase prediction error, and final output prediction error. The first Greek letter, alpha, is the weighting coefficient for amplitude loss, set to 0.3 in this implementation scheme. This coefficient controls the contribution ratio of amplitude distortion prediction to the total loss function. Amplitude distortion prediction loss represents the prediction error of the amplitude branch. Its definition formula is that amplitude distortion prediction loss equals the mean square error between the amplitude distortion sequence predicted by the model and the actual measured amplitude distortion sequence. The mean square error is calculated by summing the squares of the differences between the predicted and true values ​​and dividing by the total number of samples. The amplitude distortion sequence predicted by the model is the amplitude distortion prediction value output by the neural network, and the actual measured amplitude distortion sequence is the true amplitude distortion value obtained through experimental data labeling. The second Greek letter, beta, is the weighting coefficient for phase loss, set to 0.3 in this implementation scheme. This coefficient controls the contribution ratio of phase rotation prediction to the total loss function. Phase rotation prediction loss represents the phase... The prediction error of the branch is defined as follows: the phase rotation prediction loss equals the mean square error between the phase rotation sequence predicted by the model and the actual measured phase rotation sequence, where the phase rotation sequence predicted by the model is the phase rotation prediction value output by the neural network, and the actual measured phase rotation sequence is the true phase rotation value obtained through experimental data annotation; the third Greek letter, gamma, is the weight coefficient of the output layer loss, which is set to 0.4 in this implementation scheme. This coefficient controls the contribution ratio of the final bit error rate and optical signal-to-noise ratio prediction to the total loss function, and the highest weight value ensures priority for the final output accuracy; the output layer prediction loss represents the prediction error of the final output, and is defined as follows: the output layer prediction loss equals the mean square error between the model estimated bit error rate and the actual measured bit error rate plus the mean square error between the model estimated optical signal-to-noise ratio and the actual measured optical signal-to-noise ratio, where the model estimated bit error rate is the bit error rate prediction value of the final output of the neural network, the actual measured bit error rate is the bit error rate value actually measured in the experiment, the model estimated optical signal-to-noise ratio is the optical signal-to-noise ratio prediction value of the final output of the neural network, and the actual measured optical signal-to-noise ratio is the optical signal-to-noise ratio value actually measured in the experiment. The joint loss function optimizes three objectives simultaneously during training by using a weighted summation of three terms: the amplitude distortion prediction loss term ensures the accuracy of amplitude distortion prediction in accordance with physical laws, the phase rotation prediction loss term ensures the accuracy of phase rotation prediction in accordance with physical laws, and the output layer prediction loss term ensures that the final bit error rate and optical signal-to-noise ratio output meet the requirements of engineering practicality.The design of weighting coefficients alpha equal to 0.3, beta equal to 0.3, and gamma equal to 0.4 ensures that the model balances the physical consistency of the intermediate layers (amplitude evolution and phase evolution conform to the real physical laws of optical fiber transmission) and the prediction accuracy of the final output (minimizing the prediction errors of bit error rate and optical signal-to-noise ratio) during the optimization process. These weighting coefficients are determined after performance tuning on the validation dataset.

[0059] The specific calculation formulas for each term in the above loss function are as follows: Amplitude distortion prediction loss equals the mean square error between the model's predicted amplitude distortion sequence and the actual measured amplitude distortion sequence; phase rotation prediction loss equals the mean square error between the model's predicted phase rotation sequence and the actual measured phase rotation sequence; output layer prediction loss equals the mean square error between the model's estimated bit error rate and the actual measured bit error rate plus the mean square error between the model's estimated optical signal-to-noise ratio and the actual measured optical signal-to-noise ratio. The weighting coefficients are set to alpha = 0.3, beta = 0.3, and gamma = 0.4, ensuring that the model balances the physical consistency of the intermediate layers and the prediction accuracy of the final output during optimization.

[0060] The neural network optimizer employs an adaptive moment estimation algorithm with an initial learning rate of 0.0005. Each batch of training samples contains 64 samples (meaning 64 samples are used simultaneously for gradient calculation and parameter updates during each training iteration), and the total number of training epochs is 200 (meaning the entire training dataset is traversed 200 times). After training, the model achieves the following performance metrics on the independent test dataset: a mean relative error of 4.2% for bit error rate prediction and a mean absolute error of 0.3 Bb for optical signal-to-noise ratio prediction. This prediction accuracy meets the engineering requirements for configuration strategy verification. The system compares the bit error rate prediction and optical signal-to-noise ratio prediction values ​​output by the simulated network with the actual measurements provided by the optical performance monitoring module and the coherent receiver diagnostic unit in the multi-source heterogeneous sensing data for consistency verification. If the absolute errors of both bit error rate and optical signal-to-noise ratio are less than the tolerance threshold of 5%, the configuration strategy is deemed effective; otherwise, the prediction error is used as a feedback signal input to the dual-depth quality network model to regenerate a new configuration strategy and repeat the above verification process until the prediction error meets the tolerance requirements.

[0061] Once the configuration strategy is confirmed as effective, the system writes the effective configuration strategy, the corresponding fault probability prediction results, fault root cause diagnosis results, and simulation verification values ​​into a time-series database built on a time-series database system (an open-source database specifically optimized for storing and querying time-series data). Each database record associates a unique device identifier (e.g., International Mobile Equipment Identity or Media Access Control address) with the communication tower, forming a full lifecycle operation archive containing equipment installation time, maintenance records, and snapshots of operational status at various times. The tower distribution module periodically calls the dynamic time warping algorithm (this algorithm is used to measure the similarity between two time series and can effectively handle sequence data of unequal length or misaligned time axes) to calculate the trajectory distance between the state representation vector sequences of any two communication towers within the last thirty days. If the calculated distance value is less than 0.3 (this threshold is the normalized Euclidean distance threshold), the two towers are considered to have similar operating modes. Meanwhile, the system obtains the latitude and longitude geographic coordinates of each tower from the geographic information system (which is used to collect, store, manage, analyze and display geospatial data), and constructs a topology map of the neighboring towers with a radius of five kilometers. The nodes in the map represent each communication tower, and the edges in the map represent the geographical proximity between the towers.

[0062] When the predicted failure probability of any tower exceeds the safety threshold, the system filters out healthy towers (those with a failure probability below 0.2 and a dynamic time warping distance of less than 0.3) based on the neighboring tower topology map. Subsequently, the system automatically generates a resource reservation instruction, which is sent to the software-defined network controller (which employs a network architecture that separates the network control plane from the data forwarding plane) via the northbound interface between the upper-layer application and the controller. This resource reservation instruction includes the bandwidth capacity to be reserved, wavelength channel identifiers, and quality of service (QoS) level parameters. Based on the Open Flow Protocol (the standard communication protocol between the controller and switching equipment in a software-defined network architecture), the software-defined network controller issues flow entries (which define the matching rules and forwarding actions for data packets) to the relevant optical switching nodes, rerouting some service traffic from high-risk links to nearby healthy links. Meanwhile, the system pushes the verified and effective configuration policies to neighboring towers with similar operating modes through a message queue system (such as a distributed stream processing platform, which is used to build high-throughput real-time data pipelines and streaming applications). The edge computing nodes of these towers preload the configuration policies so that when similar performance degradation occurs on the local tower, the configuration can be quickly switched without recalculation.

[0063] In the entire digital twin system architecture, functional modules interact with each other through standardized application programming interfaces (APIs). The data acquisition module and edge gateway transmit raw sensor data using fieldbus industrial communication protocols based on Transmission Control Protocol (TCP) and Internet Protocol (IPC) or message queue telemetry transmission protocols. The edge gateway and edge computing nodes transmit standardized datasets via Google Remote Procedure Call (RPC) (a high-performance open-source RPC framework). Each AI model is deployed on the edge AI server using an Open Neural Network Exchange Format (a standard exchange format for cross-framework neural network models) optimized by a deep learning inference optimizer (a software development kit for accelerating neural network inference). The time-series database interfaces with the tower offloading module via a Representational State Transition API (a lightweight network service interface based on Hypertext Transfer Protocol). The software-defined network controller communicates with the underlying optical transmission equipment via Open Streaming Protocol version 1.3. All software modules run on a containerized platform based on a container orchestration platform (an open-source system for automating the deployment, scaling, and management of containerized applications), ensuring high availability and elastic scalability.

[0064] In terms of hardware deployment, edge computing nodes are typically installed in protective enclosures within the communication tower base, possessing a protection rating of 65 (dustproof rating 6 means complete protection against dust ingress, and waterproof rating 5 means protection against low-pressure water jets), with an operating temperature range of -40°C to +70°C. Sensor deployment follows industry technical specifications: the optical performance monitoring module is integrated into the output port of the optical distribution frame; the coherent receiver diagnostic interface is connected via a small-size 8639 connector (a connector standard for high-speed signal interfaces); complementary metal-oxide-semiconductor (CMOS) image sensors are installed at the top and middle crossarms of the tower, covering the welded areas and bolted connections of the main structure; microelectromechanical system (MEMS) vibration sensors are fixed to the main tower structure, with a sampling frequency of no less than 1,000 samples per second; temperature and humidity sensors and wind speed sensors are installed on meteorological supports 10 meters above the ground. All sensors are powered by the tower's DC power supply system, and communication lines use shielded twisted-pair cables or fiber optic cables to effectively suppress electromagnetic interference.

[0065] Through the above specific implementation methods, the present invention realizes a complete technology chain from multi-source data acquisition, trusted fusion, intelligent early warning, root cause diagnosis, strategy generation and verification to cross-tower collaborative optimization. The modules work together through clear data interfaces and physical connections, ensuring the feasibility and engineering implementation of the technical solution.

[0066] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principles of this invention are further supplemented below with a specific application scenario.

[0067] On a communication tower deployed in a coastal area with strong winds and high humidity, the system first collects optical power and OSNR data of the backbone optical link in real time through the OPM module. Simultaneously, the built-in diagnostic unit of the digital coherent receiver continuously outputs bit error rate, dispersion, and polarization mode dispersion parameters. CMOS image sensors installed on the tower top and middle crossarms capture structural images of the tower's welded and bolted connection areas at a frequency of one frame every 10 minutes. MEMS vibration sensors fixed to the main tower structure, with a sampling frequency of 1.28kHz, are downsampled by the edge gateway and uploaded in 5-minute granularities. Temperature, humidity, and ultrasonic wind speed sensors are integrated on a meteorological support 10 meters above the ground, simultaneously collecting environmental disturbance data. All raw sensor data is aggregated via an RS-485 bus to the edge gateway inside the tower base protective box. This gateway uses a built-in GPS timing module to add a UTC timestamp to each data point and uses a coordinate calibration unit to uniformly convert the physical locations of each sensor to the WGS-84 geographic coordinate system, completing spatiotemporal alignment.

[0068] Upon entering the edge computing node, the data preprocessing module detected that power supply fluctuations during a typhoon caused three consecutive data gaps in the MEMS vibration sensor. Therefore, it interpolated the data using the average of the sliding window over the same period in the past 72 hours. Simultaneously, it converted the mW unit optical power output of the OPM to dBm and normalized the vibration acceleration unit to m / s². 2 The system aligns non-time-series data such as images and logs into 5-minute windows to generate a standardized dataset. Subsequently, a four-dimensional evaluation is initiated for the credibility scoring step: For the OSNR data, sliding window abrupt change detection reveals a standard deviation of 0.8 dB, while the current jump is 3.1 dB, exceeding the three-fold threshold, thus reducing the internal consistency score to 0.6; image quality analysis shows a signal-to-noise ratio of 22 dB and a contrast ratio of 0.45, lower than the average for clear weather, resulting in an image credibility score of 0.7; the difference between readings from two adjacent wind speed sensors reaches 2.5 m / s, lowering the cross-validation score to 0.65; KL divergence calculation shows a significant difference between the current multidimensional distribution and the historical steady-state distribution (KL=1.8), resulting in a historical matching score of 0.55; combining the typhoon orange warning context obtained from the meteorological API, the system automatically multiplies the weights of optical sensors by a correction factor of 0.8. The final weighted average output yields a comprehensive credibility score of 0.63 for this period.

[0069] The feature fusion step uses ResNet-18 to process a 224×224 pixel tower image, extracting 512-dimensional CNN features. An LSTM network receives sequences of optical power, OSNR, BER, and wind speed over 288 time steps in the past 24 hours, outputting 128-dimensional temporal features. Device logs are processed using BERT word embeddings and a three-layer DNN to generate 256-dimensional semantic features. These three sets of features are multiplied by their corresponding confidence weights (0.63, 0.68, and 0.71) and concatenated. An 8-head attention mechanism is then used, with each head focusing on the correlation between different modalities—for example, the temporal alignment between a sudden increase in wind speed and an increase in vibration amplitude, and the spatial coupling between OSNR degradation and the swaying area of ​​the cable in the image. Finally, these features are fused to generate a 512-dimensional state representation vector.

[0070] This vector is fed into the BiGRU early warning model every 5 minutes. The model predicts the probability of failure in the next 24 hours based on the historical 7-day sequence. In this scenario, the model output probability is 0.89, exceeding the 0.85 threshold, triggering an early warning message. This message includes a trend curve showing the OSNR continuously decreasing from 28dB to 24.5dB and the vibration RMS value decreasing from 0.15m / s². 2 Rise to 0.42 m / s 2 The abnormal trajectory was then analyzed. Subsequently, the XGBoost fault diagnosis model analyzed the original features in the state vector and calculated the top three Gain values ​​of the features related to micro-bending loss of the junction box (such as a sudden drop in local optical power, a slight increase in PMD, and increased jitter in the junction box area in the image). Combined with its deviation from the normal range of 2.7σ, the failure mode of increased micro-bending loss in the rule library was matched, and the root cause was output as the trunk optical cable junction box, with a confidence level of 0.91.

[0071] The policy generation module activates the DoubleDQN engine accordingly. The state space contains the current 512-dimensional vector and the OSNR margin (only 4.5dB remaining). The Q network explores 27 action combinations. Through interactive simulation with the POT virtual model, it is found that reducing the modulation format from 64-QAM to 16-QAM, maintaining the symbol rate at 40GBaud, and switching FEC to concatenated codes can reduce the OSNR requirement by 5.2dB, decrease the BER from 1.2e-6 to 3.8e-7, and increase the bandwidth overhead by only 8%. The reward value of this policy, R = 0.5×5.2 + 0.4×(1.2-3.8) - 0.1×8, is the current maximum value, so it is selected as the optimal policy and sent to the physical POT.

[0072] The strategy verification module synchronously acquires I / Q waveforms and link parameters, inputting them into the BiLSTM simulation network. The forward LSTM captures the amplitude attenuation trend, and the backward LSTM models the phase noise accumulation, jointly outputting BER=4.1e-7 and OSNR=25.1dB. Compared with the measured OSNR of OPM=25.3dB and the measured BER of coherent receiver=3.9e-7, the absolute errors are 0.8% and 5.1% respectively, with the BER slightly exceeding the 5% tolerance. The system feeds back the error signal to the DQN, re-searches the action space, and in the second round selects the QPSK+32GBaud+concatenated code combination. The simulated BER is 2.9e-7 and OSNR=26.0dB, while the measured values ​​are 2.8e-7 and 26.2dB, respectively. Both errors are less than 5%, confirming it as an effective strategy.

[0073] This effective strategy, along with early warning and diagnostic results, is written to InfluxDB and linked to the tower's IMEI number to form a full lifecycle snapshot. The tower traffic offloading module calls the DTW algorithm to compare the status vector sequences of this tower with those of eight surrounding towers over the past 30 days. It finds that the DTW distance of tower TJ-207 is 0.26, and its GIS coordinates are 3.2 kilometers apart, indicating that they have similar operating modes and are geographically close. When the probability of failure of this tower reaches 0.89, the system automatically generates a resource reservation instruction and pushes it to the edge node of TJ-207 via Kafka to preload the QPSK+32GBaud+concatenation code strategy; at the same time, it sends an OpenFlow flow table entry to the SDN controller to reserve a 200Gbps wavelength channel at the optical switching node where TJ-207 is located, rerouting 30% of the service traffic of this tower to this healthy link to avoid regional outages caused by single-point failures.

[0074] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0075] like Figure 2 As shown, a digital twin system for monitoring the operation of communication tower equipment includes: Data acquisition module: used to collect multi-source heterogeneous sensor data deployed on communication towers to obtain monitoring datasets of equipment operation; Data preprocessing module: used to preprocess the monitoring dataset to generate a standardized dataset with a unified spatiotemporal reference; The weighted fusion module is used to perform real-time credibility scoring on each sensor data item in the standardized dataset; and to dynamically weight and fuse the credibility scores to output the state representation vector of the unified communication tower digital twin. Early warning module: used to input the state representation vector into a preset bidirectional gated loop model and output the fault probability prediction result within a preset time window; when the fault probability prediction result exceeds a preset safety probability threshold, an early warning message is triggered. Fault diagnosis module: Based on the acquired fault probability prediction results, the state representation vector is input into a preset extreme gradient boosting model, and the fault root cause diagnosis results are output.

[0076] Strategy generation module: used to activate the dual deep Q network model and generate the optimal configuration strategy based on the failure probability prediction results and root cause localization results; Strategy verification module: This module is used to send the optimal configuration strategy to the programmable optical transceiver for execution, and simultaneously input the bidirectional long short-term memory simulation network, outputting the simulation verification values ​​of bit error rate and optical signal-to-noise ratio; it compares the simulation verification values ​​with the actual measured values, judges the validity of the configuration strategy based on the comparison results, and obtains the effective configuration strategy; Tower traffic diversion module: It is used to integrate effective configuration strategies, failure probability prediction results and simulation verification values ​​into the digital twin platform to build a full life cycle operation file of communication towers; based on the full life cycle operation file, it identifies the similarity of operation modes between different communication towers, establishes a cross-tower collaborative optimization mechanism, and triggers resource reservation and traffic diversion plans for nearby healthy towers when a potential risk occurs in a certain tower.

[0077] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A digital twin method for monitoring the operation of communication tower equipment, characterized in that, include: Collect multi-source heterogeneous sensor data deployed on communication towers to obtain monitoring datasets of equipment operation; The monitoring dataset is preprocessed to generate a standardized dataset with a unified spatiotemporal reference. Real-time reliability scoring is performed on each sensor data item in the standardized dataset; The credibility scores are dynamically weighted and fused to output the state representation vector of the unified communication tower digital twin; The state representation vector is input into a preset bidirectional gated loop model, and the failure probability prediction result within a preset future time window is output. When the predicted failure probability exceeds a preset safety probability threshold, an early warning message is triggered. Based on the obtained fault probability prediction results, the state representation vector is input into a preset extreme gradient boosting model, and the fault root cause diagnosis results are output. Based on the failure probability prediction results and root cause localization results, the dual-depth Q-network model is activated to generate the optimal configuration strategy. The optimal configuration strategy is sent to the programmable optical transceiver for execution, and simultaneously input into the bidirectional long short-term memory simulation network, outputting the simulation verification values ​​of bit error rate and optical signal-to-noise ratio; The simulation verification values ​​are compared with the actual measured values, and the effectiveness of the configuration strategy is judged based on the comparison results to obtain an effective configuration strategy. By integrating effective configuration strategies, failure probability prediction results, and simulation verification values ​​into the digital twin platform, a full lifecycle operation archive for communication towers can be constructed. Based on the full lifecycle operation archive, the similarity of operation modes between different communication towers is identified, and a cross-tower collaborative optimization mechanism is established. When a potential risk occurs in a certain tower, the resource reservation and traffic diversion plan of the nearby healthy towers is triggered.

2. The digital twin method for monitoring the operation of communication tower equipment according to claim 1, characterized in that, Real-time reliability scoring is performed on each sensor data item in the standardized dataset, including: Sliding window detection is performed on each time series data item in the standardized dataset; detection results are obtained; wherein, when the absolute value of the standardized dataset detection of consecutive sampling points exceeds a preset threshold, it is identified as a mutation point or an anomaly point that exceeds the physical reasonable range, and an anomaly detection result is obtained; Based on the anomaly detection results, the data quality status corresponding to each sensor is marked to obtain the marking results; Calculate the sharpness, noise level, and occlusion rate of the image data in the standardized dataset to obtain quantitative indicators; Based on the quantitative indicators, the image data in the standardized dataset are subjected to sharpness, noise level and occlusion rate calculations, and an internal consistency score is generated by combining the data quality status labeling results. Meteorological station and vibration monitoring data are used as environmental context information; a context correction coefficient is generated based on the environmental context information; wherein, when the visibility is lower than a preset threshold, the credibility of the sensor data in the internal consistency score is set to a preset standard threshold; When the wind speed exceeds the preset wind speed threshold, the drift compensation algorithm is enabled to calculate the compensation residual for the vibration sensor data in the marking result, and the compensation residual is used as the context correction coefficient of the vibration sensor. Perform reading difference calculation on adjacent sensors of the same type in the same physical area to obtain the reading difference between adjacent sensors; A cross-validation score is generated based on the reading difference, the anomaly detection result, and the data quality status labeling result; Based on the cross-validation score and the difference in readings, a preliminary confidence score is obtained between each adjacent sensor pair; A sensor steady-state distribution model is constructed by calling the historical normal operation database. The KL divergence between the current observation and the steady-state distribution model is calculated. The historical pattern matching score is determined by combining the data quality status identification results to verify the preliminary confidence score in the historical data dimension. The initial credibility score is weighted and corrected by summing the internal consistency score, the cross-validation score, the historical pattern matching score, and the context correction coefficient into four dimensions, generating a real-time credibility score in the range of 0 to 1.

3. The digital twin method for monitoring the operation of communication tower equipment according to claim 1, characterized in that, The credibility scores are dynamically weighted and fused to output a state representation vector for the unified communication tower digital twin, including: Spatial feature vectors are extracted from the image data in the standardized dataset using a convolutional neural network; Temporal feature vectors are extracted from the time series data in the standardized dataset using a long short-term memory network. Semantic feature vectors are extracted from the structured log data in the standardized dataset using a fully connected deep neural network. The spatial feature vector, the temporal feature vector, and the semantic feature vector are grouped according to the sensor source. Each group of feature vectors is multiplied by the real-time reliability score of the corresponding sensor as a dynamic weight and then normalized to generate a weighted feature vector. The weighted feature vectors are concatenated and then input into a multi-head attention mechanism for fusion, outputting a state representation vector of the unified communication tower digital twin. The state representation vector is used as the unified input for both the bidirectional gated loop model and the extreme gradient boosting model.

4. The digital twin method for monitoring the operation of communication tower equipment according to claim 1, characterized in that, The state representation vector is input into a preset bidirectional gated loop model, and the predicted failure probability within a preset future time window is output, including: The state representation vector is input into a bidirectional gated loop model to learn the equipment state parameters in the historical operating data of the equipment and perform trend prediction to obtain the trend prediction result. The bidirectional gated recurrent model is trained using the cross-entropy loss function and the Adam optimizer, and whether performance degradation occurs within a preset future time period is set as the prediction label. The trained bidirectional gated recurrent model is deployed on an edge inference node, receives the state representation vector at a preset time interval, performs rolling prediction based on the change trend prediction result, and outputs the fault probability prediction result. When the failure probability prediction result exceeds the preset safety probability threshold, an early warning message containing the failure probability prediction result and the change trend prediction result is generated and the early warning is triggered. The failure probability prediction results are used as a condition for activating the dual deep Q network model.

5. A digital twin method for monitoring the operation of communication tower equipment according to claim 1, characterized in that, The state representation vector is input into a preset extreme gradient boosting model, and the fault root cause diagnosis results are output, including: The state representation vector is input into the extreme gradient boosting model to analyze historical fault data and real-time network status, and to identify the key features that lead to the fault. The extreme gradient boosting model is trained using an objective function that includes a regularization term, and classification and regression analysis are performed based on the key features. The model complexity is limited by the regularization term. The trained extreme gradient boosting model is combined with a historical fault sample library to rank the key features by importance, generating a root cause localization result that includes the key features and importance assessment. The root cause localization results and the fault probability prediction results are used as conditions for activating the dual deep Q network model.

6. The digital twin method for monitoring the operation of communication tower equipment according to claim 1, characterized in that, Based on the failure probability prediction results and root cause localization results, a dual-depth Q-network model is activated to generate an optimal configuration strategy, including: A virtual model of a programmable optical transceiver is constructed, and a dual-depth Q-network model is activated based on the failure probability prediction results and the root cause localization results. The state representation vector is input into the dual-deep Q network model as the state space, and the modulation format, symbol rate, and forward error correction coding strength are used as the action space. Reinforcement learning is performed using a reward function to optimize the dual-deep Q network model. The optimized dual-deep Q-network model is interactively learned with the virtual model. Based on the root cause localization results and the current network status, hardware configuration actions are selected to generate the optimal configuration strategy. The current network status includes device operating parameters, link health status, historical performance indicators, and real-time alarm information.

7. A digital twin method for monitoring the operation of communication tower equipment according to claim 4, characterized in that, The optimal configuration strategy is sent to the programmable optical transceiver for execution, and simultaneously input into the bidirectional long short-term memory simulation network. The simulation verification values ​​of the bit error rate and optical signal-to-noise ratio are output, including: The modulation parameters in the optimal configuration strategy are preprocessed with the real-time acquired optical signal waveform and link status data to generate input data for a bidirectional long short-term memory simulation network. The input data is fed into a bidirectional long short-term memory simulation network for amplitude and phase fitting. The bidirectional long short-term memory simulation network includes a forward long short-term memory layer, a backward long short-term memory layer, and a fully connected output layer. The amplitude and phase of the optical signal are fitted together by the fully connected output layer to output the simulated verification values ​​of bit error rate and optical signal-to-noise ratio. The simulation verification value is compared with the actual measurement value. If the consistency error is less than the preset tolerance, the optimal configuration strategy is confirmed as a valid configuration strategy. If the consistency error is equal to or greater than the preset tolerance, the process returns to the reconfiguration strategy step to regenerate the configuration strategy.

8. A digital twin method for monitoring the operation of communication tower equipment according to claim 7, characterized in that, The simulation verification value is compared with the actual measurement value. If the consistency error is less than the preset tolerance, the optimal configuration strategy is confirmed as a valid configuration strategy. If the consistency error is equal to or greater than the preset tolerance, return to the reconfiguration policy step to regenerate the configuration policy, including: The bit error rate and optical signal-to-noise ratio in the simulation verification values ​​are compared with the actual measured values ​​obtained from the monitoring dataset to obtain the comparison results. The comparison results include consistency errors; When the consistency error is less than the preset tolerance, the optimal configuration strategy is confirmed as a valid configuration strategy, and a configuration validity evaluation report containing the fault probability prediction result, the root cause location result, the valid configuration strategy, and the simulation verification value is generated. When the consistency error is equal to or greater than the preset tolerance, the reconfiguration process is triggered, and the simulation verification value and the consistency error are fed back to the dual deep Q network model. The process returns to the configuration strategy generation step to generate a new configuration strategy until an effective configuration strategy is obtained. The effective configuration strategy, the failure probability prediction results, the root cause location results, and the simulation verification values ​​are integrated into the digital twin platform to construct a full lifecycle operation archive for communication towers.

9. A digital twin method for monitoring the operation of communication tower equipment according to claim 6, characterized in that, Based on the full lifecycle operation records, the similarity of operation modes among different communication towers is identified, and a cross-tower collaborative optimization mechanism is established. When a potential risk occurs on a certain tower, resource reservation and traffic diversion plans for nearby healthy towers are triggered, including: The entire lifecycle operation file is stored based on a time-series database, associated with a unique device identifier, and a snapshot of the device's status from installation to maintenance is recorded. The trajectory distance between the state snapshots of different towers is calculated using a dynamic time warping algorithm. When the trajectory distance is less than a preset threshold, the towers are determined to have similar operating modes. A topology map of neighboring towers is constructed based on geographical distance and similar operating modes, and a cross-tower collaborative optimization mechanism is established. Based on the full lifecycle operation archive, deep reinforcement learning algorithms are used to dynamically adjust network parameters and optimize resource allocation according to the neighboring tower topology map. When the predicted failure probability of a certain tower exceeds a preset safety probability threshold, based on the root cause location result and the topology map of neighboring towers, a portion of the traffic of the current tower is routed to the healthy tower link. Trigger the resource reservation and traffic diversion plan for healthy towers in the neighboring tower topology map to achieve load balancing; The effective configuration strategy is shared among neighboring towers through the cross-tower collaborative optimization mechanism; The traffic diversion results and operation and maintenance reports are fed back to the digital twin platform to form a closed-loop optimization.

10. A digital twin system for monitoring the operation of communication tower equipment, used to execute a digital twin method for monitoring the operation of communication tower equipment as described in any one of claims 1-9, characterized in that, include: Data acquisition module: used to collect multi-source heterogeneous sensor data deployed on communication towers to obtain monitoring datasets of equipment operation; Data preprocessing module: used to preprocess the monitoring dataset to generate a standardized dataset with a unified spatiotemporal reference; Weighted fusion module: used to perform real-time credibility scoring on each sensor data item in the standardized dataset; The credibility scores are dynamically weighted and fused to output the state representation vector of the unified communication tower digital twin; Early warning module: used to input the state representation vector into a preset bidirectional gated loop model and output the fault probability prediction result within a preset time window in the future; When the predicted failure probability exceeds a preset safety probability threshold, an early warning message is triggered. Fault diagnosis module: Based on the acquired fault probability prediction results, the state representation vector is input into a preset extreme gradient boosting model, and the fault root cause diagnosis results are output. Strategy generation module: used to activate the dual deep Q network model and generate the optimal configuration strategy based on the failure probability prediction results and root cause localization results; Strategy verification module: used to send the optimal configuration strategy to the programmable optical transceiver for execution, and synchronously input the bidirectional long short-term memory simulation network, and output the simulation verification values ​​of bit error rate and optical signal-to-noise ratio; The simulation verification values ​​are compared with the actual measured values, and the effectiveness of the configuration strategy is judged based on the comparison results to obtain an effective configuration strategy. Tower offloading module: used to integrate effective configuration strategies, failure probability prediction results and simulation verification values ​​into the digital twin platform to build a full life cycle operation archive for communication towers; Based on the full lifecycle operation archive, the similarity of operation modes between different communication towers is identified, and a cross-tower collaborative optimization mechanism is established. When a potential risk occurs in a certain tower, the resource reservation and traffic diversion plan of the nearby healthy towers is triggered.

Images