A lightweight station-city operation and maintenance intelligent agent system based on bidirectional semantic filtering

Abstract

This invention relates to the field of intelligent risk operation and maintenance technology, specifically a lightweight station operation and maintenance intelligent agent system based on bidirectional semantic filtering. The system includes: collecting multi-source operation and maintenance data during the transition between underground and above-ground operation phases of a subway system; analyzing the multi-source operation and maintenance data to generate sensor equipment performance reduction risk data; identifying risk response data based on the equipment performance reduction risk data; performing preventative processing on sensors based on the risk response data; obtaining door detection data based on the sensor data; and identifying the reliability of the decision-making basis based on the door detection data, equipment performance reduction risk data, and risk response data. This invention obtains an equipment maintenance priority index by performing time alignment and weighted fusion identification on the door detection data and the reliability of the decision-making basis, based on the data changes of sensors during the transition phases and the reliability of the decision-making basis.

Description

Technical Field

[0001] This invention relates to the field of intelligent risk operation and maintenance technology, specifically a lightweight station and city operation and maintenance intelligent agent system based on bidirectional semantic filtering. Background Technology

[0002] As a critical infrastructure for modern urban operations, the operational efficiency, safety, and service continuity of urban rail transit systems are essential for public travel and urban management. In daily operation and management, ensuring the stable operation of key equipment such as subway car doors is a core element in guaranteeing smooth and safe passenger flow. These devices heavily rely on various sensors for status monitoring and control; the accuracy and reliability of sensor data directly form the cornerstone of the entire operational decision-making and safety assurance system.

[0003] Current operational decision-making models primarily balance two strategies: one is preventative maintenance based on fixed time cycles. While this fixed resource allocation model can reduce the rate of sudden failures, it often leads to over-maintenance because it doesn't consider the actual health status of equipment and external risk factors, resulting in significant resource allocation inefficiency and cost waste. The other is reactive maintenance, which, while saving on upfront preventative investment, exposes the organization to unpredictable service interruption risks. Failures at critical nodes can trigger substantial economic losses and negative social impacts.

[0004] To address this, a lightweight station operation and maintenance intelligent agent system based on bidirectional semantic filtering is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a lightweight station maintenance intelligent agent system based on bidirectional semantic filtering. By performing time alignment and weighted fusion recognition on the reliability of door detection data and decision-making basis, the system obtains the equipment maintenance priority index based on the data changes of door sensors in the alternating phases and the reliability of decision-making basis.

[0006] To achieve the above objectives, a lightweight station and city operation and maintenance intelligent agent system based on bidirectional semantic filtering is provided, comprising:

[0007] The operation risk assessment module collects multi-source operation and maintenance data during the transition between underground and above-ground operation of the subway; it calls a pre-trained risk semantic filtering model to analyze the multi-source operation and maintenance data, generating equipment performance reduction risk data related to subway door sensors; and it identifies risk response data based on the equipment performance reduction risk data.

[0008] The risk response execution module performs preventative processing on the door sensors based on the risk response data, and obtains door detection data based on the data collected by the subway door sensors.

[0009] The decision information reliability assessment module identifies the reliability of the decision basis of the data collected by the door sensor based on the door detection data, equipment efficiency reduction risk data and risk response data.

[0010] The maintenance priority assessment module performs time alignment and weighted fusion identification on the door detection data and the reliability of the decision basis. Based on the data changes of the door sensors in the alternating phases and the reliability of the decision basis, it obtains the equipment maintenance priority index.

[0011] The multi-source operation and maintenance data includes service environment and energy consumption data, vehicle external environment data, and dynamic geographic data.

[0012] The service environment and energy consumption data include the average temperature inside the carriage, the average relative humidity, the set power of the air conditioning system, and the real-time passenger load factor.

[0013] The external environmental data includes the ambient temperature, relative humidity, and atmospheric pressure outside the vehicle compartment during underground and above-ground operation.

[0014] The dynamic geographic data includes the train's precise geographic coordinates, speed, direction of travel, and route.

[0015] The risk semantic filtering model is a hybrid prediction model built on physical mechanisms and data-driven approaches.

[0016] The risk semantic filtering model includes a physical mechanism loss function, which includes a physical mechanism penalty term based on heat transfer and phase transition theory to describe the interaction between the sensor surface temperature and the ambient dew point temperature.

[0017] The process of identifying and generating equipment performance reduction risk data based on the risk semantic filtering model includes:

[0018] The service environment and energy consumption data, vehicle external environment data and dynamic geographic data from the multi-source operation and maintenance data are input into the risk semantic filtering model; the risk semantic filtering model performs forward reasoning and outputs equipment performance reduction risk data including risk warning time window, risk level coefficient and sensor surface dew point prediction temperature.

[0019] The process of identifying risk response data based on the aforementioned equipment performance reduction risk data includes:

[0020] The operational risk assessment module has a built-in multi-objective operational decision optimization model. The input of the multi-objective operational decision optimization model is structured equipment performance reduction risk data, including risk warning time window, risk level coefficient and sensor surface dew point prediction temperature.

[0021] The multi-objective operation decision optimization model takes minimizing operation and maintenance energy consumption and maximizing service continuity as its dual optimization objectives. Based on a preset knowledge base of operation strategies containing energy consumption parameters with different avoidance measures and corresponding risk mitigation rates, it performs constraint solving and outputs time series risk response data.

[0022] The process by which the decision information reliability assessment module identifies the reliability of decision-making basis includes:

[0023] Feedforward semantic prediction: Based on the risk level coefficient contained in the equipment performance reduction risk data and the response instruction strength contained in the risk response data, forward reasoning is performed to generate an initial decision basis reliability benchmark value;

[0024] Posterior semantic correction: Continuously analyze the real-time door detection data obtained after the risk response execution module performs processing, and use a pre-trained noise pattern recognizer to identify signal smoothing drift and high-frequency noise increase caused by residual water vapor, thereby obtaining data quality degradation characteristics; based on the identified data quality degradation characteristics, dynamically posteriorly correct the initial decision basis reliability benchmark value to generate the final decision basis reliability time series.

[0025] The specific identification process for the equipment maintenance priority index obtained by the maintenance priority assessment module is as follows:

[0026] The baseline model of equipment operating status, built on a deep autoencoder, is trained using historical normal operation data.

[0027] The equipment operating status baseline model can reconstruct the input door detection data to obtain baseline reconstruction data; and obtain operating status deviation data by comparing the difference between the baseline reconstruction data and the door detection data.

[0028] The deviation data of operating status is weighted and fused with the time series of the reliability of the decision basis to output the equipment maintenance priority index.

[0029] The intelligent agent system features a lightweight deployment architecture.

[0030] Edge computing nodes are deployed in the edge computing gateways and on-board controllers of subway stations. They contain the parts responsible for real-time data processing and inference from the operation risk assessment module, risk response execution module, and decision information reliability assessment module. The edge computing nodes are responsible for low-latency multi-source operation and maintenance data collection, rapid prediction of operation risks, immediate issuance of response instructions, and real-time calculation of the reliability of decision-making basis at the local time of data generation.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0032] This invention integrates physical mechanisms as soft constraints into the training of data-driven models. On the one hand, it significantly improves the generalization ability and prediction accuracy of the models, especially when training data is limited. Physical laws provide prior knowledge for the models, preventing them from overfitting to data noise. On the other hand, it ensures the physical interpretability of the model output, making the prediction results richer and more comprehensive.

[0033] This invention introduces a multi-objective optimization decision-making model to achieve automated, optimized, and refined risk response; it avoids the fixed, static threshold-based extensive management in traditional operations and maintenance, and can dynamically adjust the intensity of response strategies according to the real-time changes in risk levels; it can not only effectively mitigate operational risks with minimal energy consumption and cost, ensuring service continuity and security, but also greatly reduce the decision-making burden of back-end monitoring personnel, and improve the intelligence level and operational efficiency of the entire operations and maintenance system.

[0034] This invention intelligently integrates the physical state anomalies of equipment with the reliability of information at the data level, which can effectively filter out false alarms caused by temporary sensor problems, while accurately identifying real faults with reliable data and truly abnormal states. This frees up maintenance teams from massive amounts of data and complex alarms, allowing them to focus on addressing the issues that truly need attention. It achieves a fundamental shift from passive, extensive maintenance to predictive, precise maintenance, significantly improving maintenance efficiency and reducing operating costs. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the structure of a lightweight station and city operation and maintenance intelligent agent system based on bidirectional semantic filtering according to the present invention;

[0036] Figure 2 This is a logical schematic diagram of a lightweight station and city operation and maintenance intelligent agent system based on bidirectional semantic filtering according to the present invention.

[0037] Figure 3 This is a schematic diagram of the decision information reliability assessment module of the present invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] During the transition between underground and above-ground operation of the subway, the high speed of the subway and the influence of temperature, air pressure and light inside and outside the tunnel can easily cause vibration and condensation inside the subway train, affecting the operational safety of the subway. Therefore, it is necessary to focus on the identification and detection of the transition phase.

[0040] Example 1:

[0041] This application proposes a lightweight station and city operation and maintenance intelligent agent system based on bidirectional semantic filtering. The structure of the system is as follows: Figure 1 As shown, it includes: an operational risk assessment module, a risk response execution module, a decision information reliability assessment module, and a maintenance priority assessment module; the system's data logic is as follows: Figure 2 As shown.

[0042] The operational risk assessment module collects multi-source operation and maintenance data during the transition between underground and above-ground operation of the subway; it then uses a pre-trained risk semantic filtering model to analyze the multi-source operation and maintenance data, generating risk data related to the reduced equipment performance of subway door sensors; and finally, it identifies risk response data based on the reduced equipment performance risk data.

[0043] The multi-source operation and maintenance data includes service environment and energy consumption data, vehicle external environment data, and dynamic geographic data.

[0044] The service environment and energy consumption data include the average temperature inside the carriage, the average relative humidity, the set power of the air conditioning system, and the real-time passenger load factor.

[0045] The external environmental data includes the ambient temperature, relative humidity, and atmospheric pressure outside the vehicle compartment during underground and above-ground operation.

[0046] The dynamic geographic data includes the train's precise geographic coordinates, speed, direction of travel, and route.

[0047] During the transition phase when a subway train is about to enter or leave an underground tunnel, the system collects multi-source operation and maintenance data in real time through a sensor network deployed in the train and stations, including three dimensions:

[0048] Service environment and energy consumption data: including average temperature and average relative humidity inside the vehicle, reflecting the environmental conditions inside the vehicle; the set power of the air conditioning system, reflecting the cooling parameters to maintain this environment; and real-time passenger load factor, as passenger density significantly affects the temperature, humidity and carbon dioxide concentration inside the vehicle.

[0049] External environmental data: including atmospheric temperature, relative humidity and atmospheric pressure outside the train; the stark contrast between these data and the data inside the train is the fundamental physical reason for the risk of condensation.

[0050] Dynamic geographic data: The precise geographic coordinates, speed, direction, and route of the train, obtained through GPS and inertial navigation systems, are used to accurately pinpoint when the train will enter a transitional phase of dramatic environmental change.

[0051] Traditional operation and maintenance systems may only focus on sensor data from the equipment itself, while this solution recognizes that reduced equipment performance, especially sensor malfunction caused by condensation, is a physical process strongly correlated with the environment. Therefore, it comprehensively collects key variables that lead to this physical process. The temperature and humidity difference between the inside and outside of the vehicle is the direct driving force of condensation; air conditioning power and passenger load are key variables affecting the in-vehicle environment; and geographic location information provides precise temporal and spatial trigger points for risk prediction.

[0052] This invention collects and integrates multi-dimensional operation and maintenance data to construct a comprehensive and dynamic operational environment profile that far exceeds that of a single data source. This is the foundation for achieving high-precision risk prediction. It enables the model to fundamentally understand the physical causes of condensation risks, rather than merely performing superficial data correlations. This significantly improves the accuracy and robustness of risk prediction, avoids misjudgments and omissions caused by missing key information, and provides high-quality data input for subsequent intelligent decision-making.

[0053] The risk semantic filtering model is a hybrid prediction model built on physical mechanisms and data-driven approaches.

[0054] The risk semantic filtering model includes a physical mechanism loss function, which includes a physical mechanism penalty term based on heat transfer and phase transition theory to describe the interaction between the sensor surface temperature and the ambient dew point temperature.

[0055] The process of identifying and generating equipment performance reduction risk data based on the risk semantic filtering model includes:

[0056] The service environment and energy consumption data, vehicle external environment data and dynamic geographic data from the multi-source operation and maintenance data are input into the risk semantic filtering model; the risk semantic filtering model performs forward reasoning and outputs equipment performance reduction risk data including risk warning time window, risk level coefficient and sensor surface dew point prediction temperature.

[0057] The risk semantic filtering model in this embodiment adopts a hybrid modeling method of "physical mechanism-data driven".

[0058] Data-driven component: Long Short-Term Memory (LSTM) network is used as the basic architecture. LSTM is a special type of recurrent neural network, suitable for processing and predicting time series data. It can effectively learn the complex nonlinear relationships of multi-source operation and maintenance data over time. The input of the network is the time series vector of the multi-source operation and maintenance data as defined above.

[0059] The physical mechanism component is integrated into the model training process as a "physical mechanism loss function." This loss function includes a physical mechanism penalty term based on heat transfer and phase transition theory. Specifically, this penalty term calculates the theoretical sensor surface temperature and ambient dew point temperature using classical dew point temperature calculation formulas (such as the Magnus equation) and a heat conduction model of the sensor surface. During model training, if the LSTM network's predictions contradict physical laws—for example, it predicts no condensation but its internal prediction of a sensor surface temperature lower than the theoretical dew point temperature—this penalty term will generate a large value, thus penalizing and correcting the model's weights during backpropagation.

[0060] The training set of the model contains historical multi-source operation and maintenance data and corresponding real condensation event labels obtained through manual annotation or post-analysis; the training objective is to minimize the composite loss function; the composite loss function includes a traditional supervised learning loss term (such as cross-entropy) to measure the difference between the model prediction and the real label; and a physical mechanism penalty term.

[0061] In actual operation, the collected real-time multi-source operation and maintenance data is input into the trained risk semantic filtering model; the model outputs structured equipment performance reduction risk data through forward reasoning, including:

[0062] Risk warning time window: Predict the time period in the future (e.g., the next 3-5 minutes) when the risk of condensation is highest.

[0063] Risk level coefficient: A quantitative value (such as between 0 and 1) that represents the probability and severity of condensation.

[0064] Sensor surface dew point predicted temperature: a specific physical quantity that provides a more refined basis for subsequent decision-making.

[0065] This invention integrates physical mechanisms as soft constraints into the training of data-driven models. On the one hand, it significantly improves the generalization ability and prediction accuracy of the models, especially when training data is limited. Physical laws provide prior knowledge for the models, preventing them from overfitting to data noise. On the other hand, it ensures the physical interpretability of the model output, making the prediction results richer and more comprehensive.

[0066] The process of identifying risk response data based on the aforementioned equipment performance reduction risk data includes:

[0067] The operational risk assessment module has a built-in multi-objective operational decision optimization model. The input of the multi-objective operational decision optimization model is structured equipment performance reduction risk data, including risk warning time window, risk level coefficient and sensor surface dew point prediction temperature.

[0068] The multi-objective operation decision optimization model takes minimizing operation and maintenance energy consumption and maximizing service continuity as its dual optimization objectives. Based on a preset knowledge base of operation strategies containing energy consumption parameters with different avoidance measures and corresponding risk mitigation rates, it performs constraint solving and outputs time series risk response data.

[0069] The operational risk assessment module passes the generated structured risk data to the built-in multi-objective operational decision optimization model. The core of this model is an optimization algorithm, such as the non-dominated sorting genetic algorithm (NSGA-II).

[0070] Input: Structured data on the risk of reduced equipment performance.

[0071] Optimization goals: 1. Minimize operation and maintenance energy consumption and intervention costs: Any countermeasures (such as heating the sensors or enhancing ventilation) have energy costs and potential equipment lifespan losses; 2. Maximize service continuity: that is, eliminate the risk of condensation to the greatest extent, ensure the normal operation of sensors, and avoid train delays or safety hazards caused by equipment failure.

[0072] Constraints and Knowledge Base: The algorithm's solution process is constrained by a pre-set operational strategy knowledge base, which stores various countermeasures (such as "heater power 50%" and "ventilation system wind speed increased by 2m / s") and their corresponding energy consumption parameters and expected risk mitigation rates under different risk levels.

[0073] Output: After the algorithm solves the problem, it outputs a time series of risk response data, which is the specific operation instructions that should be executed at each step within the future risk window.

[0074] This design aims to shift risk response from experience-driven to model-optimization-driven. In real-world operations, cost and safety are often mutually constraining objectives, and finding the optimal balance between them is a complex multi-objective optimization problem. By establishing a mathematical model and utilizing mature optimization algorithms, the most cost-effective combination of response strategies can be found within milliseconds under the current risk level, achieving refined resource management and maximizing utility.

[0075] This invention introduces a multi-objective optimization decision-making model to achieve automated, optimized, and refined risk response; it avoids the fixed, static threshold-based extensive management in traditional operations and maintenance, and can dynamically adjust the intensity of response strategies according to the real-time changes in risk levels; it can not only effectively mitigate operational risks with minimal energy consumption and cost, ensuring service continuity and security, but also greatly reduce the decision-making burden of back-end monitoring personnel, and improve the intelligence level and operational efficiency of the entire operations and maintenance system.

[0076] Furthermore, the multi-objective operation decision optimization model also includes minimizing the health loss of critical equipment as a third optimization objective; the operation strategy knowledge base further stores the execution intensity of each avoidance measure and the corresponding equipment health loss model; when solving constraints, the multi-objective operation decision optimization model simultaneously optimizes operation and maintenance energy consumption, service continuity and equipment health loss, and outputs risk response data that can balance short-term risks and long-term asset value.

[0077] Health modeling: It is necessary to establish a health loss model for key actuators (such as sensor heating film, ventilation system motor). The health loss model can be built based on the life curves provided by the manufacturer, material fatigue theory or historical maintenance data, which quantifies the impact of different intensities of operation (such as heating power, switching frequency) on the remaining service life of the equipment.

[0078] Knowledge base expansion: Integrate the health loss model into the operational strategy knowledge base. Each strategy in the knowledge base includes not only energy consumption and risk mitigation rate, but also a health cost or life loss coefficient.

[0079] Three-objective optimization: The original two-objective optimization algorithm is upgraded to a three-objective optimization algorithm. In the solution process, the algorithm will search for a set of optimal solutions (Pareto front) in a three-dimensional space. This set of solutions represents the best balance strategy among the three conflicting objectives of energy consumption, risk and service continuity, and long-term equipment health.

[0080] This design enables the system to consider multiple dimensions, achieving more refined decision-making and allowing it to make wiser choices when facing similar risks. For example, for a non-extremely critical risk, the system might choose a milder strategy with slightly higher energy consumption but minimal wear and tear on equipment, rather than an aggressive strategy with the lowest energy consumption but potentially impacting equipment lifespan. Simultaneously, it supports predictive maintenance; accumulated health degradation data can serve as input for more accurate predictive maintenance, providing more scientific data support for maintenance planning and spare parts inventory management.

[0081] The risk response execution module performs preventative processing on the door sensors based on the risk response data, and obtains door detection data based on the data collected by the subway door sensors.

[0082] The decision information reliability assessment module identifies the reliability of the decision basis based on the door detection data, equipment performance reduction risk data, and risk response data.

[0083] The structure of the decision information reliability assessment module is as follows: Figure 3 As shown, the process of identifying the reliability of decision-making criteria includes:

[0084] Feedforward semantic prediction: Based on the risk level coefficient contained in the equipment performance reduction risk data and the risk response data, forward reasoning is performed to generate an initial decision basis reliability benchmark value;

[0085] First, a quick preliminary assessment is made, using forward reasoning based on the known risk level coefficient and the risk response data to be issued. This reasoning process can be a simple rule engine or a small feedforward neural network. The logic is: the higher the risk level, or the stronger the intervention measures (e.g., strong heating may introduce thermal noise), the lower the initial reliability baseline value of the data. The strength of the response instructions is obtained based on the difference between the risk response data and the historical average level, which is used to measure the intensity of the intervention measures.

[0086] Posterior semantic correction: Continuously analyze the real-time door detection data obtained after the risk response execution module performs processing, and use a pre-trained noise pattern recognizer to identify signal smoothing drift and high-frequency noise increase caused by residual water vapor, thereby obtaining data quality degradation characteristics; based on the identified data quality degradation characteristics, dynamically posteriorly correct the initial decision basis reliability benchmark value to generate the final decision basis reliability time series.

[0087] After the risk response command is executed, the module continuously analyzes real-time door detection data. It identifies specific data quality degradation features using a pre-trained noise pattern recognizer, which is built upon a convolutional neural network and excels at extracting pattern features from signal sequences. Data quality degradation features are subtle signal changes caused by residual moisture or side effects of intervention measures, such as a gradual shift in the overall signal or a significant increase in high-frequency noise. Once these data contaminations are identified, the system dynamically and in real-time adjusts the initial reliability benchmark value generated in the previous step according to its intensity and type, ultimately generating a time-series curve of the reliability of the decision-making basis.

[0088] The specific process of posterior semantic correction involves employing a noise pattern recognizer built on a deep convolutional neural network and pre-trained on a training set containing physical simulation enhancement data. This physical simulation enhancement data is generated by injecting various simulated disturbances conforming to the physical characteristics of water vapor effects into historical normal door detection data. The noise pattern recognizer can identify and decouple a multi-dimensional data quality degradation feature vector from the real-time door detection data. This vector includes four dimensions: signal smoothness drift, high-frequency noise energy, signal amplitude attenuation factor, and nonlinear distortion coefficient. Based on the values ​​of each component of this feature vector, the initial decision reliability benchmark value is dynamically and nonlinearly corrected to achieve more refined posterior calibration.

[0089] This invention upgrades the original, general identification of fuzzy features to a refined decoupling and quantization of signal degradation patterns.

[0090] Constructing a physical simulation enhancement training set: Collecting standard door sensor signals operating in a dry environment; based on physical principles, using programming to enhance the standard door sensor signals to simulate various degradation effects caused by residual moisture.

[0091] Injection drift: Superimposing a very low-frequency sine wave or random walk sequence onto a clean signal simulates a slow drift of the signal baseline;

[0092] Injected noise: noise superimposed with specific frequency bands and power spectral densities (such as Gaussian white noise, pink noise), analog circuit or environmental interference;

[0093] Simulated attenuation: The entire signal is multiplied by a coefficient less than 1 to simulate the decrease in signal strength caused by blurring of optical lenses or weakening of electromagnetic induction;

[0094] Simulated distortion: passing a signal through a nonlinear transfer function (such as a sigmoid or polynomial function) to simulate waveform distortion caused by component saturation or nonlinear parameter changes;

[0095] The above method generates training samples with precise degradation labels, namely specific values ​​such as drift degree and noise energy.

[0096] Design and training of a noise pattern recognizer:

[0097] Model architecture: It adopts a convolutional neural network, which is good at automatically extracting local pattern features from time series data; its input is a short segment (e.g., 1-2 seconds) of real-time door detection data;

[0098] Model output: The network's output layer is not a simple classification result, but a regression layer with 4 nodes, corresponding to the four quantitative indicators of "drift", "noise energy", "attenuation factor" and "distortion coefficient".

[0099] Training process: The model is trained end-to-end on the simulation enhancement training set constructed above, so that it learns to accurately infer the intensity of various degradation components contained in a seemingly complex signal.

[0100] Online correction: During actual operation, the pre-trained recognizer continuously analyzes real-time signals, and the output four-dimensional degradation feature vector is fed into a correction function (such as a small neural network or lookup table model). This function dynamically and non-linearly adjusts the initial reliability benchmark value generated by the feedforward prediction according to the different influence weights of different degradation components on data reliability, and finally generates a more accurate reliability time series.

[0101] This design greatly deepens the technical implications of post-hoc semantic correction and brings significant performance improvements:

[0102] Significantly Improved Assessment Accuracy: Compared to identifying general drift and noise, this solution decouples complex signal degradation phenomena into multiple orthogonal-dimensional quantitative indicators. This allows the system not only to recognize data anomalies but also to accurately understand the meaning and degree of these anomalies, thus achieving a leap from qualitative to quantitative assessment of data reliability, significantly improving assessment accuracy and granularity. Provides In-Depth Fault Diagnosis Basis: The output four-dimensional degradation feature vector itself is a highly valuable diagnostic report. For example, if the system continuously detects an abnormally high signal amplitude attenuation factor in a sensor, it can directly infer that the physical cause is most likely that the optical probe is obstructed by dirt, providing maintenance personnel with extremely specific and accurate diagnostic clues. This upgrades traditional "parts replacement repair" to "precision inspection," significantly improving maintenance efficiency and accuracy.

[0103] Feedforward prediction provides a priori judgment based on causal reasoning, while posterior correction provides posterior calibration based on real-time evidence. This combination makes reliability assessment both predictive and real-time, quantifying data uncertainty.

[0104] This invention provides a basis for subsequent maintenance decisions by dynamically quantifying the reliability of the data itself, effectively avoiding incorrect judgments made by the system based on contaminated or distorted data, significantly improving the signal-to-noise ratio of anomaly identification, greatly reducing the false alarm rate, and ensuring the effective utilization of maintenance resources.

[0105] Furthermore, the feedforward semantic prediction process also includes: maintaining a metadata archive that records all multi-source operation and maintenance data from sensors, the metadata archive containing the sensor's service life, calibration history, and historical failure rate; when performing feedforward semantic prediction, firstly, based on the metadata of each data source sensor participating in this risk data calculation, a priori credibility weight of the data source is generated; the priori credibility weight is used to weight the risk level coefficient and the response instruction strength, and then forward reasoning is performed to generate the initial decision basis reliability benchmark value.

[0106] This design adds a traceability dimension to data reliability assessment.

[0107] Metadata archive establishment: Establish a sensor device ledger database to record the static information of each sensor, such as model and installation date; dynamic information, such as last calibration date and cumulative running time; and historical performance, including the number of fault alarms and maintenance records.

[0108] Prior confidence calculation: Design a scoring model to calculate a health score or confidence level for each sensor in real time based on the aforementioned metadata. For example, a sensor that has just been calibrated, has a short service life, and no history of failure will receive a high score, and vice versa.

[0109] Weighted Inference: When performing feedforward semantic prediction, the system first identifies the set of upstream sensors upon which the risk of reduced device performance depends. Then, it takes the weighted average of the confidence scores of these sensors to form the prior confidence of the data source for this prediction. This confidence level serves as an important adjustment factor, inputting into the inference model along with the risk level coefficient to jointly determine the initial reliability baseline value.

[0110] When assessing data reliability, this design not only considers potential contamination during data transmission and use, but also proactively considers the inherent quality of the data at its source. By establishing a full-link data quality assessment system, the assessment results are made more accurate and robust, thus avoiding overreactions by the system. At the same time, sensor reliability scores can be output back to the maintenance management system to automatically identify and warn of sub-optimal sensors, providing precise data guidance for equipment calibration and replacement plans, and ensuring the stability of the entire system.

[0111] The maintenance priority assessment module performs time alignment and weighted fusion identification on the door detection data and the reliability of the decision basis. Based on the data changes of the door sensors in the alternating phases and the reliability of the decision basis, it obtains the equipment maintenance priority index.

[0112] The specific identification process for the equipment maintenance priority index obtained by the maintenance priority assessment module is as follows:

[0113] The baseline model of equipment operating status, built on a deep autoencoder, is trained using historical normal operation data.

[0114] The equipment operating status baseline model can reconstruct the input door detection data to obtain baseline reconstruction data; and obtain operating status deviation data by comparing the difference between the baseline reconstruction data and the door detection data.

[0115] The deviation data of operating status is weighted and fused with the time series of the reliability of the decision basis to output the equipment maintenance priority index.

[0116] Historical door detection data under the health status of the equipment is used as historical normal operation data to train a baseline model of the equipment operation status based on a deep autoencoder. The autoencoder consists of an encoder and a decoder, and its training goal is to make the output (reconstructed data) as equal to the input (original data) as possible.

[0117] State deviation calculation: In actual operation, real-time door detection data is fed into this trained baseline model. Since the model has only seen and learned how to reconstruct normal data, when the input data is abnormal, that is, deviates from the normal operating state, the model's reconstruction effect will be very poor. By calculating the difference between the original input data and the model's reconstructed data, such as the mean squared error, a quantitative operating state deviation can be obtained; the larger the deviation, the more abnormal the equipment state.

[0118] Weighted fusion: The deviation data of the operating status is weighted and fused with the reliability time series of the decision-making basis generated in the previous module. This design is based on the idea of ​​unsupervised anomaly detection. The deep autoencoder can automatically learn the essential characteristics of normality from complex high-dimensional data, thereby constructing a health baseline that does not require manual rule definition. More importantly, the weighted fusion step reflects a prudent decision-making logic: the value of an anomaly signal depends on the reliability of the signal itself. A highly reliable signal that deviates significantly from the normal baseline is the highest priority maintenance alarm.

[0119] This invention intelligently integrates the physical state anomalies of equipment with the reliability of information at the data level, which can effectively filter out false alarms caused by temporary sensor problems, while accurately identifying real faults with reliable data and truly abnormal states. This frees up maintenance teams from massive amounts of data and complex alarms, allowing them to focus on addressing the issues that truly need attention. It achieves a fundamental shift from passive, extensive maintenance to predictive, precise maintenance, significantly improving maintenance efficiency and reducing operating costs.

[0120] The intelligent agent system features a lightweight deployment architecture.

[0121] Edge computing nodes are deployed in the edge computing gateways and on-board controllers of subway stations. They contain the parts responsible for real-time data processing and inference from the operation risk assessment module, risk response execution module, and decision information reliability assessment module. The edge computing nodes are responsible for low-latency multi-source operation and maintenance data collection, rapid prediction of operation risks, immediate issuance of response instructions, and real-time calculation of the reliability of decision-making basis at the local time of data generation.

[0122] The subway operating environment changes rapidly; the process of moving from above ground to underground can take only tens of seconds. The generation and response to risks must be completed within seconds or even sub-seconds. The traditional model of uploading all data to the cloud for processing suffers from unacceptable network latency. By deploying lightweight inference models at the edge, the closed-loop time of "perception-decision-action" can be shortened to the minimum.

[0123] This invention, through a lightweight edge computing deployment architecture, ensures the system possesses extremely high responsiveness and real-time decision-making capabilities to dynamic changes in the operating environment, meeting the stringent real-time requirements of subway safety operations. Simultaneously, it significantly conserves network bandwidth resources, reduces reliance on continuous connections to the central server, and enhances the robustness and scalability of the entire system. Even in the event of a temporary network outage, edge nodes can still independently complete core risk assessment and response tasks, ensuring basic operational safety.

[0124] Example 2:

[0125] This invention proposes a lightweight station and city operation and maintenance intelligent agent system based on bidirectional semantic filtering. The solution will be introduced below in conjunction with a specific operation scenario.

[0126] Specifically, a subway train is running in an underground tunnel and is about to exit the tunnel and enter the high-temperature and high-humidity elevated track; the onboard system collects the following real-time data.

[0127] Current environmental data (inside the tunnel): The ambient temperature inside the tunnel is 26℃, and the relative humidity is 65%.

[0128] Target environmental data: Based on ground meteorological sensors and weather forecast data along the route, the predicted ground environment temperature is 34℃, relative humidity is 82%, and atmospheric pressure is 100.2kPa.

[0129] Service environment and energy consumption data: The average temperature inside the carriage is 24℃, the average relative humidity is 55%, the air conditioning system is set to 80% power, and the real-time passenger load factor assessed by the onboard vision sensor is 85%.

[0130] Dynamic geographic data: The train is currently traveling at 55 km / h and is approximately 1.2 km from the underground tunnel exit. It is expected to exit in 80 seconds.

[0131] The operational risk assessment module collects multi-source operation and maintenance data during the transition between underground and above-ground operation of the subway; it then uses a pre-trained risk semantic filtering model to analyze the multi-source operation and maintenance data, generating risk data related to the reduced equipment performance of subway door sensors; and finally, it identifies risk response data based on the reduced equipment performance risk data.

[0132] The multi-source operation and maintenance data collected above, especially the huge temperature and humidity differences inside and outside the tunnel, are input into the operation risk assessment module in real time.

[0133] Model Invocation and Risk Prediction: The module invokes a pre-trained risk semantic filtering model.

[0134] In this embodiment, the model is specifically configured as follows:

[0135] Model architecture: Long Short-Term Memory (LSTM) network with 3 hidden layers, each containing 128 neurons, using ReLU as the activation function.

[0136] Physical mechanism penalty term: The physical mechanism penalty term included in its loss function.

[0137] Risk data generation: The model performs forward inference based on the input data and outputs structured equipment performance reduction risk data within 300 milliseconds. The content is as follows:

[0138] Risk warning time window: within the next 10-90 seconds.

[0139] Risk level coefficient: 0.90 (judged as extremely high risk).

[0140] Predicted dew point temperature on sensor surface: 27.0℃ (far lower than the ambient dew point temperature).

[0141] Risk response decision-making: The module's built-in multi-objective operation decision optimization model receives the above risk data. Based on a preset operation strategy knowledge base, the model solves the constraints with the objectives of minimizing operation and maintenance energy consumption and maximizing service continuity.

[0142] Since a risk level coefficient of 0.90 indicates extremely high risk, the model selects a strategy with maximizing service continuity as the primary objective; it generates time-series risk response data: the instruction "45 seconds before the train exits the tunnel, activate the sensor heating module of the target station door, set the power to 80%, and continue for 120 seconds" to preheat the sensor surface to a temperature higher than the outside dew point temperature.

[0143] The risk response execution module performs preventative processing on the door sensors based on the risk response data, and obtains door detection data based on the data collected by the subway door sensors.

[0144] This module receives risk response data from the operational risk assessment module and executes instructions precisely at a specified time via the vehicle and station control bus to preheat the heating film of the target door sensor.

[0145] Meanwhile, the module continuously collects infrared beam signals, ultrasonic echo signals, electrical signals, and vibration recognition signals output by the door sensors after preventative processing, forming real-time door detection data.

[0146] The decision information reliability assessment module identifies the reliability of the decision basis of the data collected by the door sensor based on the door detection data, equipment efficiency reduction risk data and risk response data.

[0147] Feedforward semantic prediction: Based on extremely high risk level coefficients and high-intensity response instructions, the system predicts that the operation may cause interference to the signal due to factors such as thermal noise, and generates an initial reliability benchmark value for decision-making basis through forward inference.

[0148] Posterior semantic correction: During the execution of heating commands and when the train exits the tunnel, the module continuously analyzes real-time door detection data. A pre-trained noise pattern recognizer identifies and decouples data quality degradation feature vectors from the signal.

[0149] The maintenance priority assessment module performs time alignment and weighted fusion identification on the door detection data and the reliability of the decision basis. Based on the data changes of the door sensors in the alternating phases and the reliability of the decision basis, it obtains the equipment maintenance priority index.

[0150] The processed door detection data is input into a baseline model of the equipment's operating status built based on a depth autoencoder. Due to the combined effects of strong internal and external temperature differences, thermal noise, and possibly unavoidable transient micro-condensation, the sensor signals exhibit significant distortion, resulting in poor model reconstruction. The deviation data of the operating status is obtained by comparing the root mean square error between the original input and the reconstructed output.

[0151] Weighted fusion and priority rating: The module weights and fuses the deviation of the operating status with the reliability of the decision basis to calculate the final equipment maintenance priority index.

[0152] Final output: The maintenance priority assessment module outputs a device maintenance priority index, which is interpreted by the system as "high priority". An electronic work order is automatically generated and marked "may be subject to instantaneous overload due to drastic environmental changes". It is recommended that maintenance personnel conduct a key inspection of the sensitivity and circuit stability of the door sensor within 24 hours.

[0153] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A lightweight station and city operation and maintenance intelligent agent system based on bidirectional semantic filtering, characterized in that, include: The operation risk assessment module collects multi-source operation and maintenance data during the transition between underground and above-ground operation of the subway. The pre-trained risk semantic filtering model is invoked to analyze the multi-source operation and maintenance data, and generate risk data related to the equipment effectiveness reduction of subway door sensors. Based on the risk data of reduced equipment performance, risk response data is obtained; the multi-source operation and maintenance data includes service environment and energy consumption data, vehicle external environment data and dynamic geographic data. The risk semantic filtering model is a hybrid prediction model built on physical mechanisms and data-driven approaches. The risk semantic filtering model includes a physical mechanism loss function, which includes a physical mechanism penalty term based on heat transfer and phase transition theory to describe the interaction between sensor surface temperature and ambient dew point temperature. The data on the risk of reduced equipment performance includes the risk warning time window, risk level coefficient, and predicted temperature of dew point on sensor surface. The operational risk assessment module has a built-in multi-objective operational decision optimization model. The multi-objective operational decision optimization model takes minimizing operation and maintenance energy consumption and maximizing service continuity as dual optimization objectives. Based on a preset knowledge base of operational strategies that includes energy consumption parameters with different avoidance measures and corresponding risk mitigation rates, it performs constraint solving and outputs time series risk response data. The risk response execution module performs preventative inspections of the door sensors based on the risk response data, and obtains door detection data based on the data collected by the subway door sensors. The decision information reliability assessment module identifies the reliability of the decision basis of the data collected by the door sensor based on the door detection data, equipment efficiency reduction risk data and risk response data. The process by which the decision information reliability assessment module identifies the reliability of decision-making basis includes: Feedforward semantic prediction: Based on the risk level coefficient contained in the equipment performance reduction risk data and the response instruction strength contained in the risk response data, forward reasoning is performed to generate an initial decision basis reliability benchmark value; Posterior semantic correction: Continuously analyze the door detection data obtained after the risk response execution module performs the processing, and use a pre-trained noise pattern recognizer to identify signal smoothing drift and high-frequency noise increase caused by residual water vapor, thereby obtaining data quality degradation characteristics; based on the identified data quality degradation characteristics, dynamically posteriorly correct the initial decision basis reliability benchmark value to generate the final decision basis reliability time series. The maintenance priority assessment module performs time alignment and weighted fusion identification on the door detection data and the reliability of the decision basis. Based on the data changes of the door sensors in the alternating phases and the reliability of the decision basis, it obtains the equipment maintenance priority index.

2. The lightweight station and city operation and maintenance intelligent agent system based on bidirectional semantic filtering according to claim 1, characterized in that: The service environment and energy consumption data include the average temperature inside the carriage, the average relative humidity, the set power of the air conditioning system, and the real-time passenger load factor. The external environmental data includes the ambient temperature, relative humidity, and atmospheric pressure outside the vehicle compartment during underground and above-ground operation. The dynamic geographic data includes the train's precise geographic coordinates, speed, direction of travel, and route.

3. The lightweight station and city operation and maintenance intelligent agent system based on bidirectional semantic filtering according to claim 1, characterized in that: The process of identifying multi-source operation and maintenance data and generating equipment performance reduction risk data based on the risk semantic filtering model includes: The service environment and energy consumption data, vehicle external environment data and dynamic geographic data from the multi-source operation and maintenance data are input into the risk semantic filtering model; the risk semantic filtering model performs forward reasoning and outputs equipment performance reduction risk data including risk warning time window, risk level coefficient and sensor surface dew point prediction temperature.

4. A lightweight station and city operation and maintenance intelligent agent system based on bidirectional semantic filtering according to claim 1, characterized in that: The process of identifying risk data related to reduced equipment performance and obtaining risk response data includes: The input to the multi-objective operation decision optimization model is structured equipment performance reduction risk data, including risk warning time window, risk level coefficient, and sensor surface dew point prediction temperature.

5. A lightweight station and city operation and maintenance intelligent agent system based on bidirectional semantic filtering according to claim 1, characterized in that: The specific identification process for the equipment maintenance priority index obtained by the maintenance priority assessment module is as follows: The baseline model of equipment operating status, built on a deep autoencoder, is trained using historical normal operation data. The equipment operating status baseline model can reconstruct the input door detection data to obtain baseline reconstruction data; By comparing the differences between the baseline reconstruction data and the door detection data, the deviation data of the operating status is obtained; The deviation data of operating status is weighted and fused with the time series of the reliability of the decision basis to output the equipment maintenance priority index.

6. A lightweight station and city operation and maintenance intelligent agent system based on bidirectional semantic filtering according to claim 1, characterized in that: The intelligent agent system features a lightweight deployment architecture. Edge computing nodes are deployed in the edge computing gateways and on-board controllers of subway stations. They contain the parts responsible for real-time data processing and inference from the operation risk assessment module, risk response execution module, and decision information reliability assessment module. Edge computing nodes are responsible for low-latency, multi-source operation and maintenance data collection, rapid prediction of operational risks, immediate issuance of response instructions, and real-time calculation of the reliability of decision-making data at the local location where the data is generated.

Images