Situation awareness and decision optimization method and system for rail transit emergency command

Abstract

The present application relates to the field of rail transit technology, and provides a situation awareness and decision optimization method and system for rail transit emergency command, wherein the method comprises: acquiring a multi-source heterogeneous data stream of a rail transit emergency event; performing information fusion and feature extraction on the multi-source heterogeneous data stream by using a data processing strategy, and outputting key situation features; performing situation evolution prediction based on the key situation features, and generating a preliminary situation prediction result; quantitatively evaluating the preliminary situation prediction result according to a preset decision utility index, and generating a comprehensive decision cost; taking the comprehensive decision cost as a feedback signal, dynamically optimizing process parameters of the data processing strategy and the situation evolution prediction, and outputting a final situation prediction result based on the optimized process parameters, so that the ability of dynamically balancing compression efficiency and decision fidelity in a complex and changeable event is achieved, the problem of disconnection between technical optimization and actual demand is alleviated, and the risk of missing a situation inflection point and a best disposal opportunity is reduced.

Description

Technical Field

[0001] This invention relates to the field of rail transit technology, and in particular to a situational awareness and decision optimization method and system for rail transit emergency command. Background Technology

[0002] With the rapid expansion of rail transit networks and the frequent occurrence of extreme disasters, rail transit emergencies place extremely high demands on operation and dispatch command centers. During the handling of major incidents, command centers must rapidly extract key situational information from massive, multi-source, heterogeneous data streams within minutes or even seconds, achieving a closed loop of perception, decision-making, and action to minimize casualties and operational disruptions. This extremely high timeliness requirement presents significant challenges to the processing speed of multi-source, heterogeneous information streams, the accuracy of situational evolution prediction, and the reliability of decision-making responses within the command center.

[0003] To cope with massive concurrent data and emergencies, existing emergency command and information fusion technologies typically incorporate data time window compression and situation prediction models to assist in dispatching decisions. In practice, the system primarily aggregates multi-source data from various business nodes and performs information filtering and time window compression based on preset static thresholds or offline trained models. Meanwhile, in terms of situation evolution prediction, existing technologies mainly rely on and heavily depend on historical big data statistical models. By learning from and fitting a large number of historical routine samples, they infer the evolution trend of emergencies, thereby providing dispatching reference information for cross-departmental collaboration and the command center.

[0004] While the aforementioned traditional technologies have alleviated the pressure of data aggregation and processing to some extent, the evaluation and optimization frameworks of existing systems are limited to single indicators such as information fidelity or traditional statistical accuracy, failing to incorporate core practical demands such as dispatcher task completion time, cost of misjudgment, and cost of missed opportunities into the system closed loop. This severe disconnect between technological optimization and practical needs prevents the system from dynamically balancing compression efficiency and decision fidelity in complex and ever-changing events, ultimately causing the command center to miss turning points in the situation and the best opportunity to respond. Summary of the Invention

[0005] This invention provides a situational awareness and decision optimization method and system for emergency command in rail transit, which solves the shortcomings of existing systems that cannot dynamically balance compression efficiency and decision fidelity in complex and ever-changing events, ultimately leading to the command center missing the turning point of the situation and the best time to deal with it. It realizes the dynamic balance of parameters and closed-loop optimization of situational awareness based on the reverse drive of actual combat needs.

[0006] This invention provides a situational awareness and decision optimization method for emergency command in rail transit, executed by a computing device, comprising: Acquire multi-source heterogeneous data streams of rail transit emergencies; A data processing strategy is employed to perform information fusion and feature extraction on the multi-source heterogeneous data stream, and key situational features are output. Based on the aforementioned key situational characteristics, situational evolution prediction is performed to generate preliminary situational prediction results; The preliminary situation prediction results are quantitatively evaluated based on preset decision utility indicators to generate a comprehensive decision cost. The comprehensive decision-making cost is used as a feedback signal to dynamically optimize the data processing strategy and the process parameters of the situation evolution prediction, and the final situation prediction result is output based on the optimized process parameters.

[0007] According to the situational awareness and decision optimization method for emergency command in rail transit provided by the present invention, the step of acquiring multi-source heterogeneous data streams of rail transit emergencies includes: It can acquire in real time on-board video surveillance data and platform voice communication data, and simultaneously acquire train sensor operating status data, passenger social media text data, and alarm correlation data from cross-department business systems; The on-board video surveillance data, the platform voice communication data, the train sensor operating status data, the passenger social media text data, and the alarm-related data are collectively constructed into the multi-source heterogeneous data stream.

[0008] According to the situational awareness and decision optimization method for emergency command in rail transit provided by the present invention, the data processing strategy is a time window compression strategy; the step of using the data processing strategy to perform information fusion and feature extraction on the multi-source heterogeneous data stream and output key situational features includes: Based on a pre-established unified spatiotemporal and semantic benchmark, the multi-source heterogeneous data stream is spatiotemporally synchronized and semantically mapped to obtain an aligned multi-source heterogeneous data stream. Real-time monitoring of the evolution speed and information density of the aforementioned rail transit emergencies; Based on the evolution speed and information density, the length of the time window and the feature extraction threshold are dynamically adjusted to truncate and reduce the feature dimensionality of the aligned multi-source heterogeneous data stream, and output the key situation features.

[0009] According to the situational awareness and decision optimization method for emergency command of rail transit provided by the present invention, the step of performing spatiotemporal synchronization and semantic feature mapping on the multi-source heterogeneous data stream based on a pre-established unified spatiotemporal and semantic benchmark to obtain an aligned multi-source heterogeneous data stream includes: Extract the timestamps, spatial coordinates, and business semantic tags of various types of data from the multi-source heterogeneous data stream; By comparing the timestamps and spatial coordinates of each type of data, the time sampling deviations and spatial coordinate system differences between the various types of data are identified. The time sampling deviation is aligned by calling a time-series interpolation algorithm, and the spatial coordinate system difference is transformed by calling a spatial mapping matrix to generate spatiotemporal synchronization data. The spatiotemporal synchronization data is combined with the business semantic tags for vectorization and mapped to a pre-constructed unified representation matrix to output the aligned multi-source heterogeneous data stream.

[0010] According to the situational awareness and decision optimization method for emergency command of rail transit provided by the present invention, the step of performing situational evolution prediction based on the key situational features and generating preliminary situational prediction results includes: extracting nonlinear transition mode features from the key situational features; inputting the nonlinear transition mode features into a pre-trained dynamic situational inference network to identify inflection point nodes of situational evolution; inferring the state evolution path of future multiple time steps based on the inflection point nodes, and generating the preliminary situational prediction results according to the state evolution path.

[0011] According to the situational awareness and decision optimization method for emergency command of rail transit provided by the present invention, the preset decision utility indicators include task completion time, cost of misjudgment, and cost of missed opportunity. The step of quantifying and evaluating the preliminary situation prediction results based on preset decision utility indicators to generate a comprehensive decision cost includes: Using a preset time penalty function, calculate the time delay penalty value from acquiring the multi-source heterogeneous data stream to generating the preliminary situation prediction result; The misjudgment cost is calculated based on the deviation level of the preliminary situation prediction results and the preset first weighting coefficient. Based on the timeliness characteristics of the preliminary situation prediction results and the preset second weighting coefficient, the value of the missed opportunity is calculated; The comprehensive decision cost is generated by weighting and summing the time delay penalty value, the misjudgment cost value, and the missed opportunity cost value.

[0012] According to the situational awareness and decision optimization method for emergency command of rail transit provided by the present invention, after the step of generating the comprehensive decision cost, the method further includes: extracting the rate of change of situational features from the preliminary situational prediction result; if the rate of change of situational features is greater than a preset rate of change threshold, increasing the second weight coefficient and decreasing the first weight coefficient; if the rate of change of situational features is less than or equal to the preset rate of change threshold, increasing the first weight coefficient and decreasing the second weight coefficient.

[0013] According to the situational awareness and decision optimization method for emergency command of rail transit provided by the present invention, the process parameters include the length of the time window, the feature extraction threshold, and the weight of the situational evolution prediction model. The process of using the comprehensive decision-making cost as a feedback signal to dynamically optimize the data processing strategy and the situation evolution prediction parameters includes: If it is determined that the comprehensive decision cost does not meet the preset minimum cost convergence condition, the gradient information of the comprehensive decision cost with respect to the length of the time window, the feature extraction threshold, and the weights of the situation evolution prediction model is calculated. The length of the time window, the feature extraction threshold, and the weights of the situation evolution prediction model are updated in reverse based on the gradient information. Based on the updated time window length, feature extraction threshold, and situation evolution prediction model weights, the process of performing information fusion and feature extraction on the multi-source heterogeneous data stream using data processing strategies is repeated to output key situation features. Situation evolution prediction is then performed based on these key situation features to generate preliminary situation prediction results. The preliminary situation prediction results are then quantitatively evaluated according to preset decision utility indicators to generate a comprehensive decision cost. This process continues until the comprehensive decision cost meets the minimum cost convergence condition, thus determining the current optimal parameters.

[0014] This invention also provides a situational awareness and decision optimization system for emergency command in rail transit, executed by a computing device, comprising: The data acquisition module is used to acquire multi-source heterogeneous data streams of rail transit emergencies; The feature output module is used to perform information fusion and feature extraction on the multi-source heterogeneous data stream using data processing strategies, and output key situation features. The result generation module is used to perform situation evolution prediction based on the key situation features and generate preliminary situation prediction results. The comprehensive decision-making module is used to quantitatively evaluate the preliminary situation prediction results based on preset decision utility indicators and generate a comprehensive decision cost. The optimization module is used to dynamically optimize the data processing strategy and the process parameters of the situation evolution prediction by taking the comprehensive decision cost as a feedback signal, and output the final situation prediction result based on the optimized process parameters.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the situational awareness and decision optimization method for rail transit emergency command as described above.

[0016] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the situational awareness and decision optimization method for rail transit emergency command as described above.

[0017] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the situational awareness and decision optimization method for rail transit emergency command as described above.

[0018] The present invention provides a situational awareness and decision optimization method and system for rail transit emergency command. First, it acquires multi-source heterogeneous data streams of rail transit emergencies, extracts key situational features using data processing strategies, and performs predictions to generate preliminary situational prediction results. Then, it introduces preset decision utility indicators to quantitatively evaluate the preliminary situational prediction results, generating a comprehensive decision cost. Finally, it uses this comprehensive decision cost as a feedback signal to dynamically optimize the underlying data processing strategies and situational evolution prediction process parameters, and outputs the final prediction result. Through this reverse-driven feature based on comprehensive decision cost, the system possesses the ability to dynamically balance compression efficiency and decision fidelity in complex and ever-changing events, alleviating the problem of the disconnect between technical optimization and practical needs. This helps to shorten data processing delays and reduce the risk of the command center missing situational turning points and optimal response opportunities. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the situational awareness and decision optimization method for emergency command in rail transit provided by the present invention.

[0021] Figure 2 This is a schematic diagram of the situational awareness and decision optimization system for emergency command of rail transit provided by the present invention.

[0022] Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] Before introducing the present invention in detail, the prior art related to this application will be described in more detail in order to better understand the technical solution of the present invention and its beneficial effects.

[0025] In existing rail transit emergency command and situational awareness systems, to cope with the pressure of massive, multi-source, heterogeneous data streams, a data preprocessing module is typically set up at the front end or data aggregation layer. The core task of this module is to unify the format, align the temporal and spatial dimensions, and perform preliminary screening of data from onboard video, trackside sensors, station voice, passenger social media, and external business systems such as public security and fire departments. Common techniques include: coordinate projection transformation based on Geographic Information System (GIS) and timestamps to map data from different coordinate systems to a unified latitude and longitude grid; resampling time-series data with inconsistent sampling frequencies using interpolation algorithms; and extracting situational elements from unstructured text or speech using pre-set keyword dictionaries or rule templates. However, these preprocessing operations are often performed independently, lacking the ability to differentiate the importance of the command tasks behind the data. For example, video frames from high-traffic transfer platforms and sensor values ​​from equipment in tunnel sections are treated equally in subsequent compression and prediction stages, leading to the potential misfiltering of crucial information that truly affects decision-making due to the sheer volume of data.

[0026] In terms of data compression and feature extraction, existing systems widely adopt time-window-based compression strategies. A typical approach is to set a fixed time window length (e.g., 5 seconds, 30 seconds, or 1 minute) and perform statistical aggregation (e.g., mean, maximum, minimum) or sample retention on the data within the window. Some improved solutions introduce adaptive window adjustment mechanisms, such as dynamically changing the window size based on data fluctuations (e.g., variance changes), or using time-frequency analysis methods like discrete wavelet transform to retain abrupt changes while compressing. However, the optimization target of these adaptive mechanisms is usually information fidelity metrics, such as root mean square error (RMSE) or peak signal-to-noise ratio (PSNR), and does not establish a direct correlation with the actual costs of command and decision-making (e.g., delays, consequences of misjudgments). Therefore, when the speed of an emergency changes drastically (e.g., from a slow passenger flow congestion to a sudden stampede risk), existing compression strategies often react with lag, either smoothing out critical abrupt signals due to an excessively large window or retaining a large amount of redundant noise due to an excessively small window.

[0027] In the situation evolution prediction stage, existing technologies heavily rely on offline-trained big data statistical models. Commonly used models include time series prediction models based on Long Short-Term Memory (LSTM) networks, state estimation models based on Hidden Markov Models (HMMs), and spatiotemporal correlation prediction models based on Graph Neural Networks (GNNs). These models are typically trained using massive amounts of historical event data to learn the data distribution under normal operating conditions and the evolution patterns of common faults. However, major rail transit events are characterized by significant "small sample size, suddenness, and nonlinearity," such as panicked passenger evacuation, foreign object intrusion onto tracks, and chain-like train delays. These patterns occur with extremely low frequency in historical data, or may never have occurred at all. Statistical models relying on offline training lack the ability to generalize to such unknown patterns, and their long update cycles (usually requiring weeks or even months of data accumulation and retraining) prevent them from adapting to the dynamic evolution of events online. Furthermore, the outputs of existing prediction models are mostly probabilistic state estimates or numerical predictions, and their evaluation metrics are limited to traditional machine learning metrics such as accuracy, recall, and AUC. They lack a quantitative bridge to map these prediction results to the actual command and decision-making effectiveness (such as whether to start bus connections or whether to stop trains on adjacent lines).

[0028] Regarding the transmission and sharing of prediction results, the rail transit network involves multiple heterogeneous nodes such as dispatch centers, stations, trains, public security, fire protection, and medical facilities. Existing systems typically use standard communication protocol stacks based on IP networks (such as TCP / IP and HTTP / HTTPS) for message exchange. To ensure the priority delivery of critical information, some systems implement Quality of Service (QoS) mechanisms based on Differentiated Service Code Points (DSCP) or 802.1p priorities. This involves marking different types of messages with different priorities at the network or data link layer, with switches or routers prioritizing queuing and forwarding. However, this traditional QoS mechanism only works at the queue scheduling level. Once the link itself experiences severe congestion or quality deterioration (such as 4G / 5G signal attenuation in tunnels or a sudden drop in bandwidth due to equipment failure), high-priority messages will still be lost or experience unacceptable delays due to queue overflow. More seriously, existing systems do not couple the message importance level with the dynamic changes in link quality. That is, regardless of the link condition, the message encoding method, redundancy strategy, and physical channel selection are statically configured. When network quality deteriorates rapidly from excellent to poor, the system cannot automatically switch to a more robust transmission mode (such as forward error correction, fountain codes, dual-link concurrency) within milliseconds, resulting in the loss or severe delay of critical situation prediction results in the "last mile," forming information silos.

[0029] Finally, in terms of system evaluation and optimization, existing technologies lack a closed-loop mechanism to feed back the effectiveness of real-world decision-making to the front-end processing stage. Whether it's adjusting compression strategy parameters, updating prediction models, or setting transmission priorities, the optimization goals are disconnected from the actual effectiveness of dispatching and command. For example, a system that performs well in compression ratio and prediction accuracy in offline testing might fail to predict the risk of train derailment in a major real-world event because it filters out a slight sensor fluctuation, leading to huge costs due to misjudgment. Existing offline evaluation systems cannot capture this causal chain of "information loss—decision failure—cost incurred," causing the system to freeze after deployment and unable to evolve with the accumulation of real-world cases.

[0030] To address the numerous shortcomings of the existing technologies, this invention provides a situational awareness and decision optimization method and system for emergency command in rail transit. The scheme first establishes a unified spatiotemporal reference and introduces a command task grid (CM-Grid) to assign different command priority weights to different data. Then, using an online decision utility optimizer, it employs reinforcement learning to solve for the optimal compression ratio and jump threshold online, using task completion time, misjudgment cost, and missed opportunity cost as reward signals. A two-layer jump detection mechanism combining statistical cumulative sum detection and large model semantic mutation degree detection is used for collaborative decision-making. A priority channel with network quality adaptation is used to achieve low-latency, high-reliability transmission of critical messages under extreme network conditions. Finally, a utility evaluation and feedback module continuously adjusts system parameters, forming a self-evolving loop. The specific implementation of this application is described in detail below with reference to the accompanying drawings.

[0031] Before describing the technical solutions of the embodiments of the present invention, the terms and concepts involved in the embodiments of the present invention will be explained illustratively.

[0032] Multi-source heterogeneous data streams: These refer to the collection of data acquired in real time from multiple data acquisition devices or business systems of different sources and structures during the emergency command process of rail transit. These data sources include, but are not limited to, onboard cameras, platform microphones, train operation sensors, social media platforms used by passengers, and business databases from departments such as transportation, fire protection, and medical services; data formats include video streams, audio streams, structured numerical sequences, and unstructured text.

[0033] Key situational characteristics refer to a set of quantitative or symbolic indicators obtained from the original multi-source heterogeneous data stream through information fusion and feature extraction. These indicators characterize the core state of the current emergency (such as event type, scope of impact, and evolution trend). Key situational characteristics typically have low dimensionality but retain the most valuable information for command and decision-making, such as abnormal train deceleration, sudden changes in platform passenger density, and abnormal changes in track temperature.

[0034] Preliminary situation forecast results: These refer to the situation estimates for one or more future moments derived from key situation characteristics at the current and historical moments through a predictive model. These results can be quantitative numerical predictions (such as the estimated cumulative passenger flow over the next 5 minutes) or qualitative situation predictions (such as the probability of passenger panic and evacuation within the next minute).

[0035] Data processing strategy: refers to a set of rules used to perform a series of operations such as preprocessing, compression, and feature extraction on raw, multi-source, heterogeneous data streams. In this application, the data processing strategy is specifically embodied in a time window compression strategy, including parameter settings such as the length of the time window, the threshold for feature extraction, and whether to prioritize the retention of specific types of data.

[0036] Decision utility indicators are quantitative evaluation scales used to measure the effectiveness of situational awareness and forecasting results on actual command and decision-making. Unlike traditional forecasting accuracy indicators, decision utility indicators are directly linked to the commander's actual combat mission execution results, including but not limited to mission completion time (the time from the occurrence of an event to its resolution), misjudgment cost (losses caused by unnecessary resource input or erroneous operations due to incorrect forecasting), and missed opportunity cost (losses caused by failure to take the best response measures in a timely manner due to forecast lag or underreporting).

[0037] Overall decision cost: This refers to a single value obtained by weighting and summing the above-mentioned decision utility indicators according to certain weights. This value comprehensively reflects the total experience cost generated by the current data processing strategy and prediction model in the handling of this event; the higher the value, the worse the system performance. The overall decision cost serves as a feedback signal to drive the dynamic optimization of system parameters.

[0038] Situation evolution prediction: This refers to the process of extrapolating the development and changes of a situation over a future period based on currently observed situation characteristics and the dynamic or statistical laws governing the events themselves. This process aims to identify inflection points in the situation (such as a shift from stable development to rapid deterioration) and predict its evolution path.

[0039] Onboard video surveillance data refers to video stream data collected by cameras installed inside train carriages or in the driver's cab. This data can be used to monitor passenger behavior (such as crowding levels, abnormal running), equipment status (such as abnormal door opening and closing), and driver operation behavior.

[0040] Platform voice communication data: This refers to audio data collected in the station platform area using audio pickup devices, as well as voice communication records between the station control room and the driver / dispatch center. This data can be used to identify passenger cries for help, unusual noises, and staff commands.

[0041] Train sensor operating status data refers to the real-time numerical data collected by various sensors installed on key train components (such as traction motors, braking systems, bogies, and doors), including data such as speed, acceleration, temperature, pressure, current, and voltage. This data reflects the train's real-time health status and operating parameters.

[0042] Passenger social media text data: This refers to text information related to rail transit incidents posted by passengers through social or service platforms. This data can serve as an auxiliary source for incident awareness; for example, a passenger's description of a burning smell in the carriage may indicate a potential malfunction or emergency.

[0043] Alarm-related data refers to alarm information or event notifications from cross-departmental business systems such as public security, fire protection, medical, telecommunications, and power supply. Examples include smoke alarms from fire alarm systems, security incident reports from public security systems, and emergency calls from medical systems. This data typically has high reliability and clear business implications.

[0044] Time sampling bias: This refers to the time offset in timestamps of data from different data sources at the same physical point in time due to asynchronous sampling clocks or inconsistent sampling frequencies. For example, video data is spaced 40 milliseconds apart (25 frames / second), while sensor data is collected every 10 milliseconds; the alignment error between the two is the time sampling bias.

[0045] Temporal interpolation algorithms are mathematical methods used to interpolate non-uniformly sampled time series data to generate data with consistent time intervals at a target time point. Commonly used algorithms include linear interpolation, polynomial interpolation, and spline interpolation. In this application, a temporal interpolation algorithm is used to correct time sampling deviations between different data sources.

[0046] Spatial mapping matrix: refers to a mathematical transformation matrix that converts coordinate points in one spatial coordinate system to another. For example, converting a track kilometer marker (such as K100+500) into geographic latitude and longitude coordinates, or converting the internal coordinate system of a station (such as column No. 3 in area A of platform level) into unified GIS coordinates, can all be achieved through a pre-calibrated spatial mapping matrix.

[0047] A unified representation matrix is ​​a data structure with fixed dimensions and a fixed field order. It is used to transform heterogeneous data from different sources and formats into data records with the same organizational form after spatiotemporal alignment and semantic mapping. Each row of the matrix corresponds to a spatiotemporal sampling point, and each column corresponds to a standardized feature (such as timestamp, longitude, latitude, data type encoding, feature value, etc.).

[0048] Nonlinear abrupt change pattern characteristics: These refer to the dramatic fluctuations in certain key indicators during the evolution of a situation, which do not conform to linear or smooth change patterns within a short period of time. Examples include a train speed suddenly dropping to zero from normal operation, a surge in the number of people on a platform several times over within seconds, or a sudden change in track temperature exceeding a threshold. These characteristics are often important precursors to major events that are about to occur or have already occurred.

[0049] Dynamic situational awareness network: This refers to a computational model used to simulate and predict the evolution of a situation over time. It typically employs deep learning architectures such as recurrent neural networks, graph neural networks, or temporal convolutional networks. This network can receive key current and historical situational features as input, output predicted states for multiple future steps, and capture the nonlinear coupling relationships between different features.

[0050] Time delay penalty value: This refers to the cost calculated using a preset time penalty function based on the time delay from acquiring raw data to generating preliminary situational prediction results. The longer the time delay, the larger the penalty value, reflecting the timeliness requirements of decision-making response.

[0051] Misjudgment cost: This refers to the quantifiable loss incurred when the system's situation prediction results do not match the actual situation, leading commanders to take unnecessary or erroneous actions due to the incorrect prediction. For example, if the system falsely reports a train fire, causing the dispatcher to stop all trains and resulting in delays across the entire line, the operational loss caused by this delay is the misjudgment cost.

[0052] Missed opportunity cost: This refers to the quantifiable loss incurred when the system fails to predict a real-time situational shift (missed report), causing commanders to miss the optimal window for action. For example, if the system fails to predict that track water accumulation reaches a dangerous threshold, forcing trains to stop in one section and causing passenger panic, the resulting rescue costs and public opinion losses constitute the missed opportunity cost.

[0053] Time window length: This refers to the size of the time range covered by each time window when using a time window compression strategy. For example, a time window length of 1 second means that data within each second is aggregated into a representative value. This parameter directly affects the compression ratio and the degree of information retention: the longer the time window, the higher the compression ratio, but rapidly changing information may be lost; the shorter the time window, the more complete the information retention, but the larger the data volume.

[0054] Feature extraction threshold: This refers to the threshold value used during time window compression to determine whether a certain value represents a critical change and should be retained. For example, mutation information is only retained when the difference between the maximum and minimum values ​​within the window exceeds this threshold; otherwise, only the statistical average is retained. This threshold can be dynamically adjusted according to the speed of event evolution.

[0055] Situation evolution prediction model weights: These refer to the parameter values ​​of each network layer or neuron connection in prediction models such as dynamic situation simulation networks. These weights determine how the model maps key input situation features to predicted output results. By adjusting the weights, the model's predictive behavior can be changed to better adapt to the evolutionary pattern of the current event.

[0056] The execution entity of the situational awareness and decision optimization method for rail transit emergency command provided in this invention is typically a computing device with powerful data processing and logical operation capabilities. In real rail transit operation scheduling and emergency command scenarios, this execution entity can specifically manifest as a central server, cloud computing cluster, or edge computing node deployed within the operation scheduling and command center. These physical hardware devices not only possess extremely high data throughput and concurrent processing performance, enabling real-time access and aggregation of massive multi-source heterogeneous data streams from onboard monitoring, platform sensors, and social media, but also embed advanced deep learning frameworks and parallel computing engines to support efficient extraction of underlying features, alignment and dimensionality reduction, and complex situational evolution inference calculations.

[0057] Furthermore, the executing entity can also be a distributed collaborative computing system built upon a high-speed rail transit network. In this architecture, front-end edge computing devices can be responsible for performing spatiotemporal synchronization and time window compression / truncation of multi-source data, which have extremely high timeliness requirements. Meanwhile, the back-end cloud-based core computing cluster concentrates superior computing power to specifically perform mathematical cost quantification evaluation of real decision-making utility and gradient backpropagation updates of global parameters. Through this coordinated physical arrangement, the computing system can support the reverse control closed-loop logic proposed in this embodiment of the invention, thereby ensuring that the emergency dispatch system stably and efficiently completes the adaptive optimization of underlying algorithm parameters and the output of timely decisions.

[0058] Figure 1 This is a flowchart illustrating the situational awareness and decision optimization method for emergency command in rail transit provided in an embodiment of the present invention. The method includes the following: Step 101: Obtain multi-source heterogeneous data streams of rail transit emergencies.

[0059] The computing device accesses multiple data sources deployed in the rail transit system in real time via wired or wireless communication interfaces. These data sources include, but are not limited to: train-mounted sensors (such as speed, acceleration, temperature, and pressure sensors), video surveillance equipment in carriages and platforms, voice acquisition equipment in platforms and the driver's cab, social media texts posted by passengers via mobile terminals, and alarm information pushed by cross-departmental business systems such as public security, fire protection, and medical services. The computing device receives these data streams from different sources and in various formats, temporarily stores them in a memory buffer, and uses them as raw input for subsequent processing.

[0060] This step enables computing devices to comprehensively capture various types of information related to the event, providing a sufficient data foundation for subsequent fusion analysis.

[0061] Step 102: Use data processing strategies to perform information fusion and feature extraction on the multi-source heterogeneous data streams, and output key situational features.

[0062] The computing device first preprocesses the multi-source heterogeneous data within the buffer, including spatiotemporal alignment and semantic normalization. Specifically, the device unifies the timestamps from different data sources to the same time base, maps spatial locations in different coordinate systems to unified geographic information system coordinates, and performs entity extraction and vectorization representation on unstructured text and speech data. After alignment, the computing device fuses and reduces the dimensionality of the normalized data according to the currently configured data processing strategy. This data processing strategy can be a dynamic time window compression strategy, which adaptively adjusts the length of the time window and the feature extraction threshold based on the current information density and the speed of event evolution, aggregating, filtering, or extracting features from the data within the window.

[0063] After the above processing, the computing device outputs a set of key situational features with low dimensionality and high information density, such as quantitative indicators reflecting train operation status, passenger flow density, environmental parameters, and confidence levels of abnormal events.

[0064] This step transforms the original high-dimensional redundant data into compact key situational features, reducing the computational load on subsequent prediction modules while preserving core event evolution information.

[0065] Step 103: Perform situation evolution prediction based on the key situation features to generate preliminary situation prediction results.

[0066] The computing device inputs key situational feature sequences from the current and historical moments into a pre-built situational evolution prediction model. This model can be a dynamic inference network based on architectures such as recurrent neural networks, temporal convolutional networks, or graph neural networks, capable of capturing temporal dependencies and nonlinear coupling relationships between features. Based on the input feature sequences, the model infers situational changes at one or more future moments and outputs preliminary situational prediction results. These results can be expressed as numerical predictions (such as estimates of future passenger flow), state classifications (such as the probability of a certain event occurring within a future period), or trajectory predictions (such as the evolution curves of situational indicators).

[0067] Through this step, computing devices can extrapolate future evolutionary trends based on current observations, providing forward-looking reference information for command and decision-making.

[0068] Step 104: Quantitatively evaluate the preliminary situation prediction results based on preset decision utility indicators to generate a comprehensive decision cost.

[0069] The computing device acquires a preset decision utility index, which includes one or more of the following: task completion time, cost of misjudgment, and cost of missed opportunity.

[0070] Task completion time measures the total duration from the occurrence of an event to its resolution; misjudgment cost quantifies the losses caused by unnecessary resource investment or erroneous operations due to incorrect predictions; missed opportunity cost quantifies the losses caused by missed optimal response windows due to prediction lags or underreporting. Based on the current preliminary situation forecast, the computing device estimates the expected costs of making command decisions based on this forecast, and then weights and sums these costs according to preset weights to generate a scalar-form comprehensive decision cost. This comprehensive decision cost reflects the overall performance of the current data processing strategy and prediction model in this prediction.

[0071] Through this step, the computing device establishes a quantitative mapping relationship between the technical predictions and the actual command effects.

[0072] Step 105: Use the comprehensive decision cost as a feedback signal to dynamically optimize the data processing strategy and the process parameters of the situation evolution prediction, and output the final situation prediction result based on the optimized process parameters.

[0073] The computing device inputs the comprehensive decision cost as a feedback signal into the policy optimizer. This optimizer aims to minimize the comprehensive decision cost by dynamically adjusting the parameters of the data processing strategy (such as time window length and feature extraction threshold) and the internal parameters of the situation evolution prediction model (such as network weights). The policy optimizer can employ reinforcement learning algorithms (such as policy gradient methods or PPO) or gradient-based optimization methods to calculate the update direction of the current parameters relative to the comprehensive decision cost and iteratively update the parameters. After the update is complete, the computing device re-executes steps 102-104 using the updated parameters to generate a new round of preliminary situation prediction results, which are then output as the final situation prediction results to the command and display terminal or downstream decision-making system.

[0074] Through this step, the computing device forms a closed-loop mechanism of evaluation-feedback-optimization, enabling the system to adapt to the evolutionary characteristics of different events, continuously improve the prediction quality, and finally output the situation prediction results by incorporating the optimization strategies learned from the effects of past decisions.

[0075] The situational awareness and decision optimization method for rail transit emergency command provided in this invention first acquires multi-source heterogeneous data streams of rail transit emergencies, extracts key situational features using data processing strategies, and performs predictions to generate preliminary situational prediction results. Then, a preset decision utility index is introduced to quantitatively evaluate the preliminary situational prediction results, generating a comprehensive decision cost. Finally, this comprehensive decision cost is used as a feedback signal to dynamically optimize the underlying data processing strategies and process parameters for situational evolution prediction, and outputs the final prediction result. Through this reverse-driven feature based on comprehensive decision cost, the system possesses the ability to dynamically balance compression efficiency and decision fidelity in complex and ever-changing events, alleviating the problem of the disconnect between technical optimization and practical needs. This helps to shorten data processing delays and reduce the risk of the command center missing situational turning points and optimal response opportunities.

[0076] In a specific implementation process, the computing device acquires various types of data related to rail transit emergencies in parallel through multiple data acquisition channels. Specifically, the computing device receives real-time onboard video monitoring data from cameras inside the train carriages and in the driver's cab. These video streams are typically encoded in H.264 or H.265 format and transmitted at a rate of 25 or 30 frames per second, used to monitor passenger behavior, equipment status, and driver operations within the carriages. Simultaneously, the computing device acquires platform voice communication data, including ambient audio collected by platform microphones and recorded intercom conversations between the station control room and the driver / dispatch center. This audio data is transmitted in compressed formats (such as AAC or Opus) after necessary noise reduction and voice activity detection. In addition, the computing device continuously receives train sensor operating status data, such as traction motor current and temperature, brake cylinder pressure, bogie vibration acceleration, door opening and closing status, and real-time train speed. This sensor data is typically reported at millisecond intervals in a structured numerical sequence format.

[0077] Furthermore, the computing device accesses passenger social media text data via web crawlers or API interfaces. This includes text descriptions related to the incident posted by passengers on social media platforms or public information channels (such as "there's a burning smell in the carriage," "the train suddenly stopped," etc.). While this text data has a low sampling frequency and is subjective, it often provides abnormal clues that are difficult for sensors and videos to directly capture. Simultaneously, the computing device also interfaces with cross-departmental business systems through dedicated data interfaces to obtain alarm-related data in real time, including but not limited to: smoke alarms and sprinkler status from fire protection systems, security incident reports and personnel deployment information from public security systems, emergency calls and casualty information from medical systems, communication link status from signaling systems, and voltage and current anomaly alarms from power supply systems. These various types of data differ significantly in timestamps, spatial coordinates, data formats, and update frequencies. For example, sensor data is high-frequency values ​​at the millisecond level, video data is frame sequences, text data is unstructured strings at the minute level, and alarm data is structured records triggered by the event. After receiving these data streams from different sources and in different formats, the computing device creates buffers in memory according to the data source type and adds a unified time reception tag to them, thus constructing a multi-source heterogeneous data stream for subsequent processing.

[0078] Through the above methods, computing devices can comprehensively capture information related to rail transit emergencies from multiple dimensions, forming a three-dimensional perception of the event's state. Onboard video and sensors provide objective physical measurements, platform voice and social media text supplement subjective feelings and descriptions of anomalies, while cross-departmental alarm data incorporates the judgment results of professional systems. This multi-source, complementary data acquisition strategy helps overcome the limitations of single data sources in terms of coverage, real-time performance, and reliability, laying a more complete data foundation for subsequent information fusion and situation prediction.

[0079] Furthermore, in this embodiment, the data processing strategy employed by the computing device is specifically embodied as a dynamic time window compression strategy. The execution process of this strategy includes the following steps: First, the computing device performs spatiotemporal synchronization and semantic feature mapping on the acquired multi-source heterogeneous data streams based on a pre-established unified spatiotemporal and semantic benchmark. Specifically, the computing device extracts the timestamp information carried by various types of data, unifying the time benchmark of all data streams to the same clock source (e.g., GPS clock or Network Time Protocol server clock). For data streams with inconsistent sampling frequencies, interpolation or resampling methods are used to align the time dimension. In the spatial dimension, the computing device converts spatial location information under different coordinate systems (e.g., track kilometer markers, station internal numbers, equipment installation point identifiers) into unified geographic coordinate system coordinates through a preset spatial mapping relationship. Simultaneously, the computing device performs semantic feature mapping on unstructured text, speech, and video data. For example, it converts text into semantic vectors using a pre-trained language model, converts audio into text using a speech recognition model and further extracts keywords, and extracts entity categories and action labels from images using a video analysis model. After the above processing, the original multi-source heterogeneous data is transformed into an aligned data stream with a unified time axis, unified spatial reference, and unified semantic representation framework, providing standardized input for subsequent compression operations.

[0080] Secondly, the computing equipment monitors the evolution speed and information density of rail transit emergencies in real time. Evolution speed can be quantified by analyzing the time derivatives of key indicators (such as train speed change rate, passenger flow density change rate, and sensor value fluctuation amplitude). Information density reflects the abundance of effective information per unit time; for example, during morning and evening rush hours, the values ​​of various data streams change frequently, resulting in higher information density, while during late-night off-peak hours, data changes are gradual, leading to lower information density. The computing equipment continuously calculates these indicators using a sliding window approach and uses the results as the basis for dynamically adjusting compression parameters.

[0081] Finally, the computing device dynamically adjusts the length of the time window and the feature extraction threshold based on the monitored evolution speed and information density, and performs truncation and feature dimensionality reduction on the aligned multi-source heterogeneous data stream. Specifically, when the evolution speed is fast or the information density is high, the computing device automatically shortens the time window length and lowers the feature extraction threshold to capture rapidly changing key signals more precisely and avoid smoothing or filtering out important information. Conversely, when the evolution speed is slow and the information density is low, the computing device appropriately extends the time window length and increases the feature extraction threshold to reduce data redundancy and improve compression efficiency. Within each dynamically adjusted time window, the computing device performs aggregation operations (such as calculating the mean, maximum, minimum, or rate of change) and threshold filtering (retaining only mutation information exceeding a set threshold) on the data within the window, thereby compressing the high-dimensional original data stream into low-dimensional key situational features. These key situational features retain the most representative indicators during the event evolution process, such as the peak value of a certain sensor within the window, the rate of change of passenger flow density in a certain area, and the triggering flags of a certain type of abnormal event.

[0082] Through the aforementioned dynamic time window compression strategy, the computing device can adaptively adjust the compression intensity according to the real-time evolution of events, achieving a dynamic balance between information fidelity and data compression rate. Compared to static compression methods using fixed time windows, the strategy in this embodiment can effectively avoid the problems of critical alarms being falsely filtered during peak periods and redundant logs being retained during off-peak periods, thereby significantly reducing the data processing load of subsequent prediction modules while ensuring the integrity of key situational characteristics.

[0083] Furthermore, in this embodiment, the specific process by which the computing device performs spatiotemporal synchronization and semantic feature mapping on multi-source heterogeneous data streams is as follows: First, computing devices extract metadata carried by various types of data from multi-source heterogeneous data streams, including timestamps, spatial coordinates, and business semantic tags. Timestamps record the moment the data was generated; different data sources may use different time bases (such as the device's local clock, network time, or GPS clock). Spatial coordinates vary in form; for example, train sensors use track kilometer markers (e.g., K100+500), station equipment uses station grid coordinates, and public security and fire protection data uses latitude and longitude coordinates. Business semantic tags identify the type of data content, such as train speed, door status, smoke alarm, and passenger text. These tags can be obtained from structured fields in the data stream or through rule parsing.

[0084] Secondly, the computing device compares the timestamps and spatial coordinates of various data types to identify time sampling biases and spatial coordinate system differences between different data sources. Time sampling biases manifest as follows: some sensors report data every 10 milliseconds, while video data is reported as a frame every 40 milliseconds, and social media text may only have second-level time precision. Spatial coordinate system differences manifest as: there is no direct correspondence between track kilometer markers and geographical latitude and longitude, and there are translational and rotational relationships between the station's internal coordinate system and the city's coordinate system. The computing device generates a bias description table by comparing each pair of data sources, recording the time offset and spatial transformation relationship between them.

[0085] Next, the computing device uses a temporal interpolation algorithm to align the time sampling deviations. Specifically, the computing device selects a unified time reference (such as a GPS clock) and a unified time step (such as 10 milliseconds). For data sources with sampling frequencies lower than this step step, linear interpolation or spline interpolation methods are used to generate estimated values ​​at missing time points; for data sources with sampling frequencies higher than this step step, downsampling or extraction methods are used to retain data at representative moments. Through interpolation and alignment operations, all data streams are synchronized in the time dimension. Simultaneously, the computing device uses a pre-calibrated spatial mapping matrix to transform spatial coordinate system differences. For example, for the transformation from track kilometer markers to latitude and longitude coordinates, the computing device uses a track mileage-latitude and longitude mapping table to construct a transformation matrix; for coordinates within stations, an affine transformation matrix of the station plan is used. Multiplying the spatial coordinates in various types of data by the corresponding mapping matrix yields the spatial location in a unified geographic coordinate system. After spatiotemporal alignment processing, spatiotemporally synchronized data is generated.

[0086] Finally, the computing device vectorizes the spatiotemporal synchronized data and its business semantic tags, mapping it to a pre-constructed unified representation matrix. The vectorization method depends on the data type: for numerical data, the numerical value is directly used as a one-dimensional feature vector; for categorical data (such as the open / closed status of a car door), one-hot encoding or embedded vector representation is used; for textual semantic tags, a pre-trained word embedding model converts them into fixed-dimensional vectors. The unified representation matrix is ​​a predefined data structure where rows correspond to spatiotemporal sampling points (each sampling point is determined by a unified time step and spatial grid cells), and columns correspond to various standardized features (such as timestamps, longitude, latitude, data type encoding, feature values, or vectors). The computing device fills the corresponding positions in the matrix with the vectorized results of each spatiotemporal synchronized data point. For missing positions in the matrix (i.e., a spatiotemporal point lacking information from a certain data source), null values ​​or forward padding can be used. After completing the mapping, the computing device outputs the fully filled unified representation matrix, i.e., the aligned multi-source heterogeneous data stream.

[0087] Through the aforementioned spatiotemporal synchronization and semantic feature mapping process, data from different sources, in different formats, and with different spatiotemporal precisions are transformed into a unified reference framework, eliminating fusion barriers caused by sampling bias and coordinate system differences. The output format of the unified representation matrix provides standardized data input for subsequent dynamic time window compression, enabling compression operations to be performed based on unified time and spatial dimensions, avoiding the loss or misjudgment of key information due to inaccurate data alignment.

[0088] Furthermore, in this embodiment, the computing device performs situation evolution prediction based on the key situation features output in step 102, and the specific process is as follows: First, the computing device extracts nonlinear abrupt change pattern features from key situational characteristics. These key characteristics are typically presented as time-varying sequence data, such as train speed sequences, carriage density sequences, and track temperature sequences. The computing device performs nonlinear analysis on these sequences to identify abrupt change patterns that deviate from normal evolutionary patterns. Specifically, the computing device can calculate the higher-order differences, local rates of change, or cumulative sum statistics of the sequences, marking fluctuations exceeding preset dynamic thresholds as nonlinear abrupt change candidates. For example, when a train speed suddenly drops to zero from a normal operating state in a very short time, and its deceleration rate exceeds the maximum rate of change under normal braking conditions, the speed sequence segment within that period is extracted as a nonlinear abrupt change pattern feature. Similarly, for passenger flow density sequences, if a surge from a sparse state to a congested state within seconds, with the growth curve exhibiting an exponential rather than nonlinear trend, this segment is also marked as an abrupt change pattern feature. The computing device packages these extracted abrupt change segments with corresponding spatiotemporal labels to form an abrupt change pattern feature set.

[0089] Secondly, the computing device inputs the extracted nonlinear transition pattern features into a pre-trained dynamic situational awareness network. This network can employ a temporal neural network architecture, such as a Long Short-Term Memory network or a gated recurrent unit network. Its training process uses historical catastrophic event data (including the complete evolution trajectories of various sudden events) for supervised learning or reinforcement learning. The input dimension of the network matches the dimension of the transition pattern feature vector, and the output layer is designed to output the state category or numerical prediction of the situational evolution. The computing device takes the transition pattern features at the current moment as input, and through the forward computation of the network, outputs a hidden state vector, which encodes the stage information of the current event evolution.

[0090] Building upon this foundation, the computing device utilizes a dynamic situational awareness network to identify inflection points in the situation's evolution. An inflection point is a critical moment when the situation transitions from one state to another, such as shifting from normal operation to an emergency state, from a controllable state to an out-of-control state, or from slow change to rapid change. The computing device analyzes the temporal rate of change of the network's hidden states and the probability distribution of various states at the output layer to determine whether it is currently at or about to enter an inflection point. For example, when the probability of "emergency state" in the network output rapidly increases from below 0.2 to above 0.8 within several consecutive time steps, and the rate of change exceeds a set threshold, the computing device marks the current time step as an inflection point. The identified inflection points record the time, location, and event type (e.g., braking anomaly inflection point, passenger flow surge inflection point).

[0091] Finally, the computing device extrapolates the state evolution path for future multiple time steps based on the identified inflection point nodes, and generates preliminary situation prediction results based on this path. Specifically, starting from the inflection point node, the computing device inputs the current key situation characteristics and inflection point information into the decoder part of the dynamic situation extrapolation network (if the network is an encoder-decoder structure), or adopts a rolling prediction method: using the prediction output of the current moment as the input of the next moment, iterating multiple times, thereby generating a sequence of situation states for multiple future time steps. This state evolution path can be represented as a numerical sequence (such as the predicted passenger flow density value per second in the next 5 seconds), a state sequence (such as the event state label per second in the next 10 seconds, including normal, warning, emergency, etc.), or a combination of both. The computing device summarizes this path information to form preliminary situation prediction results, for example: "In the next 3 seconds, the train speed will remain at 0 km / h; in the next 4 to 6 seconds, the carriage density will increase from 0.85 to 0.92; in the next 7 seconds, the probability of passenger panic will reach 0.75, and it is recommended to prepare for evacuation and guidance in advance." Through the above methods, the computing equipment can specifically model and predict nonlinear and sudden evolution patterns commonly seen in emergencies. Extracting features of nonlinear transition patterns allows the prediction process to focus on truly important signal changes, avoiding interference from data in stable periods. Identifying inflection point nodes helps to capture critical moments of qualitative changes in the situation in advance, providing commanders with valuable early warning time. Extrapolating multi-step state paths based on inflection points provides continuous estimates of the future evolution process, rather than single-point predictions, thus providing richer decision-making basis for tiered response and resource scheduling. Compared to prediction methods relying on conventional time-series modeling, the solution in this embodiment has higher sensitivity and prediction accuracy for small-sample, sudden, and nonlinear evolution patterns of major rail transit events.

[0092] Furthermore, in this embodiment, the preset decision utility indicators include three dimensions: task completion time, cost of misjudgment, and cost of missed opportunity. The computing device quantifies and evaluates the preliminary situation prediction results based on these indicators to generate a comprehensive decision cost. The specific process is as follows: First, the computing device uses a preset time penalty function to calculate the time delay penalty value from acquiring multi-source heterogeneous data streams to generating preliminary situation prediction results. Time delay reflects the system's response speed; the total time elapsed from the arrival of the first data packet at the computing device to the output of the preliminary situation prediction result in step 103 is the time delay. The time penalty function can be designed as a monotonically increasing function, such as a linear function, exponential function, or piecewise function. In one example, the time penalty function is a piecewise linear function: when the time delay is below a first threshold (e.g., 500 milliseconds), the penalty value is 0; when the time delay is between the first and second thresholds (e.g., 2 seconds), the penalty value increases linearly with the delay; when the time delay exceeds the second threshold, the penalty value rises sharply to reflect the penalty for severe delay. The computing device substitutes the actually measured time delay into this function to obtain the time delay penalty value.

[0093] Secondly, the computing device calculates the cost of misjudgment based on the deviation level of the preliminary situation prediction results and a preset first weighting coefficient. The cost of misjudgment measures the adverse consequences caused by the deviation between the predicted result and the actual situation. The computing device first determines the deviation level: it compares the preliminary situation prediction results with the actual situation (obtainable through subsequent sensor feedback or manual confirmation), classifying the deviation into several levels based on the degree of deviation (such as the relative error between the predicted and actual values, and the degree of matching between the predicted and actual state categories), for example, slight deviation, moderate deviation, and severe deviation. Each deviation level corresponds to a preset baseline cost. Then, the computing device multiplies the prediction result by the corresponding first weighting coefficient based on the decision-making domain involved (such as train scheduling, passenger evacuation, fire response, etc.) to obtain the final cost of misjudgment. The first weighting coefficient can be preset based on historical experience or expert knowledge to reflect the relative severity of different types of misjudgments.

[0094] Next, the computing device calculates the cost of missed opportunities based on the timeliness characteristics of the preliminary situation prediction results and a preset second weighting coefficient. The cost of missed opportunities measures the loss incurred due to the failure to take timely and effective measures caused by prediction lag or underreporting. Timeliness characteristics can include: the difference between the occurrence time of the critical event in the prediction results and the actual occurrence time (ahead or behind), and whether the prediction results were generated within the decision window. For example, if the dangerous event identified in the prediction results actually occurs a considerable time after the prediction time, allowing the system sufficient response time, the cost of missed opportunities is low; conversely, if the dangerous event has already occurred or is about to occur when the prediction results are generated (prediction lag), preventing commanders from deploying in advance, the cost of missed opportunities is high. The computing device quantifies the timeliness characteristics into a timeliness coefficient between 0 and 1 (a smaller timeliness coefficient indicates greater lag), and then multiplies it by a preset baseline cost and second weighting coefficient to obtain the cost of missed opportunities. The second weighting coefficient can be dynamically adjusted according to the event type; for example, for security events requiring extremely rapid response, the second weighting coefficient is set higher.

[0095] Finally, the computing device performs a weighted summation of the obtained time delay penalty value, misjudgment cost value, and missed opportunity cost value to generate a comprehensive decision cost. The weighted summation formula can be expressed as: C=w d ·P d +w m ·C m +w l ·C l Among them, P d C is the time delay penalty value. m C misjudged the value of the substitute. l To miss the opportunity and lose value, w d w m w l These are the corresponding weighting coefficients. These three weighting coefficients can be configured according to actual business needs; for example, increasing w in scenarios emphasizing rapid response. d In scenarios where accurate prediction is emphasized, improving w m In scenarios emphasizing preventative measures, improve w l The computing device outputs the scalar value obtained by weighted summation as the comprehensive decision cost.

[0096] Through the aforementioned quantitative evaluation process, computing devices transform the previously difficult-to-quantify effects of command and decision-making into concrete numerical costs, allowing for direct comparison of the merits of different data processing strategies and prediction models. The time delay penalty reflects the system's real-time response capability, the cost of misjudgment reflects the accuracy of prediction, and the cost of missed opportunities reflects the timeliness of prediction. The weighted sum of these three factors comprehensively characterizes the system's overall performance in a real-world environment, providing a clear and optimizable objective function for subsequent parameter optimization, thereby driving the system to evolve towards reducing the costs of real-world decision-making.

[0097] Furthermore, in this embodiment, after generating the comprehensive decision cost, the computing device also dynamically adjusts the weight coefficients corresponding to the misjudgment cost and the missed opportunity cost based on the rate of change of situation characteristics in the preliminary situation prediction results, so that the evaluation mechanism can adapt to different stages of event evolution.

[0098] Specifically, the computing device first extracts the rate of change of situational features from the preliminary situation prediction results. The rate of change of situational features measures the drastic change in the event state over time. One or more key situational features (such as the rate of change of train speed, the rate of change of passenger flow density, and the amplitude of sensor value fluctuations) can be selected to calculate their time derivatives. For example, the computing device can obtain a predicted sequence of passenger flow density over several future time steps from the preliminary situation prediction results, and then calculate the first difference of this sequence to obtain the rate of change of passenger flow density. The larger the rate of change, the more drastic the situation evolution, and the event may be in a stage of rapid deterioration or approaching an inflection point; the smaller the rate of change, the more stable the situation.

[0099] The computing device acquires a preset rate of change threshold, which can be pre-set based on historical event statistics or expert experience, or dynamically configured according to the event type. The computing device then compares the extracted rate of change of situational characteristics with this threshold.

[0100] If the rate of change of situational characteristics exceeds a preset threshold, it indicates that the current situation is evolving rapidly, and the event may be about to undergo a qualitative change or has already entered a phase of rapid deterioration. In this case, the consequences of missing the optimal window for action are often more severe than the consequences of making a wrong judgment, because missing the opportunity may lead to irreversible losses. Therefore, the computing device increases the second weight coefficient corresponding to the cost of missing the opportunity, while decreasing the first weight coefficient corresponding to the cost of misjudgment. The adjusted weights make the calculation of the comprehensive decision cost focus more on penalizing prediction lag and underreporting, thereby driving the subsequent parameter optimization process to prioritize the timeliness and coverage of predictions, even at the cost of sacrificing some prediction accuracy.

[0101] Conversely, if the rate of change of situational characteristics is less than or equal to a preset rate of change threshold, it indicates that the current situation is relatively stable, the event evolution is relatively slow, and the commander has relatively ample time to make judgments and decisions. In this case, the relative cost of unnecessary resource investment or erroneous operations caused by incorrect judgments may be higher than that of slightly delayed predictions. Therefore, the computing device increases the first weight coefficient corresponding to the cost of misjudgment while decreasing the second weight coefficient corresponding to the cost of missed opportunities. The adjusted weights make the calculation of the overall decision-making cost focus more on the penalty for prediction accuracy, thereby driving the system to prioritize ensuring the accuracy of predictions and reduce false alarms.

[0102] After the computing device completes the adjustment of the weight coefficients, it recalculates the comprehensive decision cost using the updated weights, or applies the adjusted weights to the comprehensive decision cost calculation in subsequent time steps. The adjustment range of the weights can be a fixed step size (such as increasing or decreasing by 0.1 each time), or it can be proportional to the degree to which the rate of change exceeds a threshold; the greater the rate of change, the greater the adjustment range.

[0103] Through the aforementioned dynamic adjustment mechanism, this adaptive weight adjustment strategy helps to make the overall decision-making cost more aligned with the actual decision-making needs in different scenarios, thereby guiding the system parameters to be optimized in a direction that better meets the requirements of actual combat.

[0104] Furthermore, in this embodiment, the computing device uses the comprehensive decision-making cost as a feedback signal to dynamically optimize the data processing strategy and the process parameters for situation evolution prediction. These process parameters include the length of the time window, the feature extraction threshold, and the weights of the situation evolution prediction model. The optimization process is as follows: First, the computing device determines whether the currently generated comprehensive decision cost meets the preset minimum cost convergence condition. This convergence condition can be set according to actual needs. For example, convergence is considered achieved when the absolute value of the difference between the comprehensive decision costs calculated in several consecutive iterations is less than a preset threshold; or when the comprehensive decision cost is less than or equal to an acceptable target value; or when the number of iterations reaches a preset maximum, optimization stops and the current parameters are taken as the optimal parameters. If the comprehensive decision cost meets the convergence condition, no further parameter updates are needed; the current parameters are the optimal parameters, and the computing device outputs the final situation prediction result based on these parameters. If the convergence condition is not met, the parameter update process begins.

[0105] Secondly, if the integrated decision cost does not meet the minimum cost convergence condition, the computing device calculates the gradient information of the integrated decision cost with respect to the time window length, feature extraction threshold, and situation evolution prediction model weights. The gradient information reflects the trend and magnitude of the change in the integrated decision cost when each parameter undergoes a small change. Specifically, for the time window length parameter, the computing device applies a small perturbation to this parameter near its current value (e.g., increasing or decreasing by a preset step size) and observes the change in the integrated decision cost, using the ratio of the change to the perturbation as the gradient estimate of this parameter. The gradients of the feature extraction threshold and situation evolution prediction model weights are calculated in the same way. If the number of situation evolution prediction model weights is large (e.g., connection weights in a deep neural network), the computing device can use a backpropagation algorithm to calculate the gradients of all weight parameters at once to improve computational efficiency. Regardless of the specific calculation method used, the final result is the gradient value of the integrated decision cost with respect to each adjustable parameter in the time window length, feature extraction threshold, and model weights.

[0106] Next, the computing device updates the time window length, feature extraction threshold, and situation evolution prediction model weights in reverse based on the calculated gradient information. The update rule follows the gradient descent principle: for each parameter, its current value is subtracted from the product of the learning rate and the corresponding gradient. The learning rate is a positive decimal used to control the step size of each update. Specifically, if the gradient of a parameter is positive, it means that increasing the parameter will lead to an increase in the overall decision cost; therefore, after updating according to the above rule, the value of the parameter will decrease, thereby reducing the overall decision cost. If the gradient is negative, it means that increasing the parameter will lead to a decrease in the overall decision cost; therefore, after updating, the value of the parameter will increase, which also helps to reduce the overall decision cost. During the update process, the computing device also needs to consider the physical boundary constraints of each parameter. For example, the length of the time window cannot be less than a single data sampling period and cannot exceed the preset maximum window length; the feature extraction threshold cannot be negative; and the model weights must be kept within a reasonable range to avoid numerical overflow. If the updated parameters exceed the boundaries, the computing device truncates them to the boundary values.

[0107] Finally, based on the updated time window length, feature extraction threshold, and situation evolution prediction model weights, the computing device repeatedly executes the following steps: using data processing strategies to fuse and extract features from multi-source heterogeneous data streams; performing situation evolution prediction based on key situation features to generate preliminary situation prediction results; and performing quantitative evaluation based on decision utility indicators to generate a comprehensive decision cost. Specifically, the computing device re-executes the dynamic time window compression operation using the updated time window length and feature extraction threshold to obtain new key situation features; then, it re-executes situation evolution prediction using the updated prediction model weights to obtain new preliminary situation prediction results; and then calculates a new comprehensive decision cost based on these prediction results. Subsequently, the computing device again determines whether the new comprehensive decision cost meets the convergence condition. If it still does not meet the condition, it repeats the above gradient calculation and parameter update process. This iterative process continues until the comprehensive decision cost meets the preset minimum cost convergence condition. At this point, the computing device determines the current time window length, feature extraction threshold, and situation evolution prediction model weights as the current optimal parameters and generates the final situation prediction result output based on these parameters.

[0108] Through the aforementioned iterative optimization process based on gradient information, the computing device enables the time window length, feature extraction threshold, and situational evolution prediction model weights to automatically adjust based on feedback from the comprehensive decision cost, gradually approaching the parameter configuration that minimizes the decision cost. This process does not rely on a specific reinforcement learning framework; it only requires calculating the gradient of the cost function with respect to the parameters, exhibiting good versatility and interpretability. After each event handling, the computing device can store empirical data (including parameter configurations and corresponding decision costs) for use in setting initial values ​​or adaptively adjusting the learning rate in subsequent iterations, thereby further accelerating the convergence speed.

[0109] Further, updating the length of the time window, the feature extraction threshold, and the weights of the situation evolution prediction model in reverse based on the gradient information includes: Obtain the physical boundary constraints of the time window length, feature extraction threshold, and situation evolution prediction model weights. The physical boundary constraints include the minimum time window length constraint and the upper limit of device computing power constraint. Based on gradient information, the length of the time window, the feature extraction threshold, and the weights of the situation evolution prediction model are updated tentatively to generate tentative update parameters. If the trial update parameters exceed the physical boundary constraints, the trial update parameters are truncated, and the truncated parameters are used as the length of the updated time window, the feature extraction threshold, and the weights of the situation evolution prediction model.

[0110] The following describes the situational awareness and decision optimization system for rail transit emergency command provided in the embodiments of the present invention. The situational awareness and decision optimization system for rail transit emergency command described below can be referred to in correspondence with the situational awareness and decision optimization method for rail transit emergency command described above.

[0111] This invention provides a situational awareness and decision optimization system for emergency command in rail transit, executed by a computing device. (See also...) Figure 2 ,include: Data acquisition module 210 is used to acquire multi-source heterogeneous data streams of rail transit emergencies; Feature output module 220 is used to perform information fusion and feature extraction on the multi-source heterogeneous data stream using data processing strategies, and output key situation features; The result generation module 230 is used to perform situation evolution prediction based on the key situation features and generate preliminary situation prediction results. The comprehensive decision-making module 240 is used to quantitatively evaluate the preliminary situation prediction results based on preset decision utility indicators and generate a comprehensive decision cost. The optimization module 250 is used to use the comprehensive decision cost as a feedback signal to dynamically optimize the data processing strategy and the process parameters of the situation evolution prediction, and output the final situation prediction result based on the optimized process parameters.

[0112] Figure 3 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3 As shown, the electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840. The processor 810, communications interface 820, and memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute a situational awareness and decision optimization method for rail transit emergency command. This method includes: acquiring multi-source heterogeneous data streams of rail transit emergencies; performing information fusion and feature extraction on the multi-source heterogeneous data streams using a data processing strategy to output key situational features; performing situational evolution prediction based on the key situational features to generate preliminary situational prediction results; quantitatively evaluating the preliminary situational prediction results according to preset decision utility indicators to generate a comprehensive decision cost; using the comprehensive decision cost as a feedback signal to dynamically optimize the data processing strategy and the process parameters of the situational evolution prediction, and outputting the final situational prediction result based on the optimized process parameters.

[0113] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0114] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the situational awareness and decision optimization method for rail transit emergency command provided by the above methods. The method includes: acquiring multi-source heterogeneous data streams of rail transit emergencies; using a data processing strategy to perform information fusion and feature extraction on the multi-source heterogeneous data streams and output key situational features; performing situational evolution prediction based on the key situational features and generating preliminary situational prediction results; quantitatively evaluating the preliminary situational prediction results according to a preset decision utility index and generating a comprehensive decision cost; using the comprehensive decision cost as a feedback signal to dynamically optimize the data processing strategy and the process parameters of the situational evolution prediction, and outputting the final situational prediction results based on the optimized process parameters.

[0115] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the situational awareness and decision optimization method for rail transit emergency command provided by the above methods. The method includes: acquiring multi-source heterogeneous data streams of rail transit emergencies; performing information fusion and feature extraction on the multi-source heterogeneous data streams using a data processing strategy to output key situational features; performing situational evolution prediction based on the key situational features to generate preliminary situational prediction results; quantitatively evaluating the preliminary situational prediction results according to a preset decision utility index to generate a comprehensive decision cost; using the comprehensive decision cost as a feedback signal to dynamically optimize the data processing strategy and the process parameters of the situational evolution prediction, and outputting the final situational prediction results based on the optimized process parameters.

[0116] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0117] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A situational awareness and decision optimization method for emergency command in rail transit, characterized in that, Performed by a computing device, including: Acquire multi-source heterogeneous data streams of rail transit emergencies; A data processing strategy is employed to perform information fusion and feature extraction on the multi-source heterogeneous data stream, and key situational features are output. Based on the aforementioned key situational characteristics, situational evolution prediction is performed to generate preliminary situational prediction results; The preliminary situation prediction results are quantitatively evaluated based on preset decision utility indicators to generate a comprehensive decision cost. The comprehensive decision-making cost is used as a feedback signal to dynamically optimize the data processing strategy and the process parameters of the situation evolution prediction, and the final situation prediction result is output based on the optimized process parameters.

2. The situational awareness and decision optimization method for emergency command of rail transit according to claim 1, characterized in that, The acquisition of multi-source heterogeneous data streams for rail transit emergencies includes: It can acquire in real time on-board video surveillance data and platform voice communication data, and simultaneously acquire train sensor operating status data, passenger social media text data, and alarm correlation data from cross-department business systems; The on-board video surveillance data, the platform voice communication data, the train sensor operating status data, the passenger social media text data, and the alarm-related data are collectively constructed into the multi-source heterogeneous data stream.

3. The situational awareness and decision optimization method for emergency command of rail transit according to claim 1, characterized in that, The data processing strategy is a time window compression strategy; The data processing strategy employed to perform information fusion and feature extraction on the multi-source heterogeneous data streams, outputting key situational features, including: Based on a pre-established unified spatiotemporal and semantic benchmark, the multi-source heterogeneous data stream is spatiotemporally synchronized and semantically mapped to obtain an aligned multi-source heterogeneous data stream. Real-time monitoring of the evolution speed and information density of the aforementioned rail transit emergencies; Based on the evolution speed and information density, the length of the time window and the feature extraction threshold are dynamically adjusted to truncate and reduce the feature dimensionality of the aligned multi-source heterogeneous data stream, and output the key situation features.

4. The situational awareness and decision optimization method for emergency command of rail transit according to claim 3, characterized in that, The process of performing spatiotemporal synchronization and semantic feature mapping on the multi-source heterogeneous data streams based on a pre-established unified spatiotemporal and semantic benchmark to obtain aligned multi-source heterogeneous data streams includes: Extract the timestamps, spatial coordinates, and business semantic tags of various types of data from the multi-source heterogeneous data stream; By comparing the timestamps and spatial coordinates of each type of data, the time sampling deviations and spatial coordinate system differences between the various types of data are identified. The time sampling deviation is aligned by calling a time-series interpolation algorithm, and the spatial coordinate system difference is transformed by calling a spatial mapping matrix to generate spatiotemporal synchronization data. The spatiotemporal synchronization data is combined with the business semantic tags for vectorization and mapped to a pre-constructed unified representation matrix to output the aligned multi-source heterogeneous data stream.

5. The situational awareness and decision optimization method for emergency command of rail transit according to claim 1, characterized in that, The step of performing situation evolution prediction based on the key situation features and generating preliminary situation prediction results includes: Extract the nonlinear jump mode features from the key situation features; The nonlinear jump mode features are input into a pre-trained dynamic situation inference network to identify inflection points in situation evolution. Based on the inflection point, the future state evolution path at multiple time steps is deduced, and the preliminary situation prediction result is generated according to the state evolution path.

6. The situational awareness and decision optimization method for emergency command of rail transit according to claim 1, characterized in that, The preset decision utility indicators include task completion time, cost of misjudgment, and cost of missed opportunity; The step of quantifying and evaluating the preliminary situation prediction results based on preset decision utility indicators to generate a comprehensive decision cost includes: Using a preset time penalty function, calculate the time delay penalty value from acquiring the multi-source heterogeneous data stream to generating the preliminary situation prediction result; The misjudgment cost is calculated based on the deviation level of the preliminary situation prediction results and the preset first weighting coefficient. Based on the timeliness characteristics of the preliminary situation prediction results and the preset second weighting coefficient, the value of the missed opportunity is calculated; The comprehensive decision cost is generated by weighting and summing the time delay penalty value, the misjudgment cost value, and the missed opportunity cost value.

7. The situational awareness and decision optimization method for emergency command of rail transit according to claim 6, characterized in that, Following the step of generating the comprehensive decision cost, the method further includes: Extract the rate of change of situation features from the preliminary situation prediction results; If the rate of change of the situation feature is greater than a preset rate of change threshold, the second weight coefficient is increased and the first weight coefficient is decreased. If the rate of change of the situation feature is less than or equal to the preset rate of change threshold, the first weight coefficient is increased and the second weight coefficient is decreased.

8. The situational awareness and decision optimization method for emergency command of rail transit according to claim 1, characterized in that, The process parameters include the length of the time window, the feature extraction threshold, and the weights of the situation evolution prediction model. The process of using the comprehensive decision-making cost as a feedback signal to dynamically optimize the data processing strategy and the situation evolution prediction parameters includes: If it is determined that the comprehensive decision cost does not meet the preset minimum cost convergence condition, the gradient information of the comprehensive decision cost with respect to the length of the time window, the feature extraction threshold, and the weights of the situation evolution prediction model is calculated. The length of the time window, the feature extraction threshold, and the weights of the situation evolution prediction model are updated in reverse based on the gradient information. Based on the updated time window length, feature extraction threshold, and situation evolution prediction model weights, the process of performing information fusion and feature extraction on the multi-source heterogeneous data stream using data processing strategies is repeated to output key situation features. Situation evolution prediction is then performed based on these key situation features to generate preliminary situation prediction results. The preliminary situation prediction results are then quantitatively evaluated according to preset decision utility indicators to generate a comprehensive decision cost. This process continues until the comprehensive decision cost meets the minimum cost convergence condition, thus determining the current optimal parameters.

9. A situational awareness and decision optimization system for emergency command in rail transit, characterized in that, Performed by a computing device, including: The data acquisition module is used to acquire multi-source heterogeneous data streams of rail transit emergencies; The feature output module is used to perform information fusion and feature extraction on the multi-source heterogeneous data stream using data processing strategies, and output key situation features. The result generation module is used to perform situation evolution prediction based on the key situation features and generate preliminary situation prediction results. The comprehensive decision-making module is used to quantitatively evaluate the preliminary situation prediction results based on preset decision utility indicators and generate a comprehensive decision cost. The optimization module is used to dynamically optimize the data processing strategy and the process parameters of the situation evolution prediction by taking the comprehensive decision cost as a feedback signal, and output the final situation prediction result based on the optimized process parameters.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the situational awareness and decision optimization method for rail transit emergency command as described in any one of claims 1 to 8.

Images