A metro power corridor multi-parameter redundant safety early warning control method and system

Abstract

The application discloses a kind of subway power corridor multi-parameter redundancy safety early warning control method and system.The method is received by double main concentrators parallelly multi-parameter sensor signal and each independent maintenance sliding time window calculation parameter change rate, from concentrator to double main concentrator output is carried out consistency check and majority vote error correction, and effective trend criterion is output, just display controller will parameter instantaneous value and change rate jointly into multi-level safety threshold model and fuse determination to distinguish multi-level risk state and drive graded early warning output, graded early warning result is reversely transmitted to concentrator layer dynamic adjustment sliding time window length, data reporting frequency and trend operation model selection, form the two-way cross coupling closed loop between redundancy check layer and graded early warning layer, without relying on special alarm control host computer, can realize trend identification and graded early warning.

Description

Technical Field

[0001] This invention belongs to the field of environmental safety monitoring technology for subway power corridors, specifically relating to a multi-parameter redundancy safety early warning control method and system for subway power corridors. Background Technology

[0002] Subway power corridors are narrow, enclosed underground spaces that carry high-voltage cables, communication cables, and auxiliary pipelines. They suffer from poor ventilation, restricted personnel access, and complex environmental risks. Dangerous conditions such as the accumulation of carbon monoxide, hydrogen sulfide, and flammable gases, or a decrease in oxygen concentration, may arise within these corridors due to factors like aging cable insulation, overheated joints, groundwater seepage, or leaks from external pipelines. A common characteristic of these risks is that they typically evolve gradually from non-existent to increasingly severe conditions, rather than occurring abruptly.

[0003] Current safety management technologies for subway power corridors primarily rely on manual inspections and single-point gas detection. In a typical scenario, personnel carrying handheld gas detectors enter the corridor periodically for spot checks, obtaining instantaneous readings of various parameters and manually determining whether ventilation equipment needs to be activated or deactivated. This method only obtains discrete numerical snapshots within a limited time window, failing to cover the gradual environmental deterioration process outside of inspection periods, resulting in significant monitoring blind spots. While some systems incorporate fixed gas or temperature sensors for continuous acquisition of basic parameters, the data processing layer still uses a linear architecture with a single hub for aggregation and a single communication link for transmission, centralizing all threshold calculations and alarm judgments in a dedicated alarm control host. In this centralized architecture, the firmware of the dedicated alarm control host is designed with fixed-value comparison logic as its core objective. With limited hardware resources, it can only perform one-by-one comparisons of instantaneous parameter values ​​with fixed thresholds, outputting a binary state of "normal" or "over-limit alarm." It cannot perform continuous calculations on parameter change rates, nor can it achieve multi-parameter correlation analysis or risk level segmentation. For slowly accumulating risks and phased deterioration processes, the architecture is completely insensitive during the long deterioration period before the parameters exceed the absolute threshold. It only triggers a single-level alarm without differentiation after the parameters jump over the threshold, and cannot provide on-site personnel with differentiated response guidance for minor anomalies, general hidden dangers and emergency situations.

[0004] In attempting to address these issues, the industry has long held a technological bias: the belief that reliable trend assessment and tiered early warning in the field necessitates the introduction of higher-performance dedicated alarm control hosts or independent early warning control hardware to perform local threshold calculations and alarm actuation. This bias locks trend analysis, tiered assessment, and alarm actuation into a single dedicated hardware entity, increasing system cost and complexity while failing to address the single point of failure issue where a malfunction of this dedicated hardware can lead to the loss of all early warning capabilities. Furthermore, the existing architecture lacks a two-way feedback mechanism between the acquisition layer and the early warning layer. The acquisition layer does not participate in any trend pre-calculation, and the early warning layer cannot detect distortions or frame drops during data transmission. This information gap further weakens the reliability of the early warning results. Moreover, the trend calculation algorithm model in the existing centralized architecture remains unchanged after factory installation, failing to dynamically select a calculation method more suited to the current evolutionary characteristics based on the real-time risk situation. This results in the same algorithm logic being used to process parameter changes with drastically different characteristics at different risk stages, further restricting the accuracy and timeliness of trend identification and prediction. Summary of the Invention

[0005] To address the technical problem of existing subway power corridor environmental monitoring systems lacking trend identification and hierarchical early warning capabilities for abnormal environmental changes, this invention provides a trend identification and hierarchical early warning method and system for a multi-parameter redundant safety early warning system for subway power corridors. By establishing a bidirectional cross-coupled closed-loop linkage between the data verification process of the redundant hub and the early warning judgment process of the local display controller, trend identification and hierarchical early warning functions are achieved without relying on a dedicated alarm control host.

[0006] The trend identification and hierarchical early warning method provided by this invention includes: dual main hubs receiving environmental signals continuously output by multi-parameter sensors in a corridor in parallel, each independently maintaining a parameter sequence within a sliding time window, and calculating the parameter change rate based on the currently effective model among multiple preset trend calculation models; the hubs simultaneously perform consistency verification on the current frame values ​​and change rates output by the dual main hubs, and if the verification passes, outputs a valid trend criterion confirmed by redundancy, and if the verification fails, initiates data backtracking and majority voting error correction; after receiving the valid trend criterion, the local display controller substitutes the instantaneous parameter values ​​and change rates into a multi-level safety threshold model for fusion judgment, distinguishes multi-level risk states based on the combination relationship between the degree of instantaneous value exceeding the limit and the degree of rate exceeding the limit, and drives the corresponding level of hierarchical early warning output; the output result of the hierarchical early warning is fed back to the hub layer, and the hub dynamically adjusts the length of the sliding time window and the data reporting frequency according to the early warning level, and synchronously switches the currently effective trend calculation model, forming a bidirectional cross-coupled closed loop between the redundancy verification layer and the hierarchical early warning layer.

[0007] Furthermore, the multi-parameter sensors include carbon monoxide sensors, hydrogen sulfide sensors, oxygen sensors, combustible gas sensors, and temperature sensors. Each sensor is deployed at key points along the power corridor, converting environmental parameters into standard electrical signals and continuously outputting them to the dual master hubs and slave hubs, achieving uninterrupted multi-parameter environmental sensing coverage across the entire area.

[0008] Furthermore, the dual hubs synchronously receive all sensor signals in a parallel hot-standby manner and independently analyze them. Each hub maintains a configurable sliding time window locally, sampling each parameter periodically within the window to form a continuous time series. Based on this series, the rate of change of each parameter is calculated according to the currently active trend calculation model. Multiple trend calculation models are available, including linear fitting models, exponentially weighted models, and nonlinear prediction models containing acceleration terms. The dual hubs always synchronously use the same trend calculation model. This design embeds trend calculation capabilities into the data acquisition layer, eliminating the need for additional computing hardware.

[0009] Furthermore, during consistency verification at the hub, the rate of change of the same parameter output by the two main hubs is compared for deviation. If the deviation is within a preset tolerance range, the average value is taken as the valid trend criterion and marked with a high confidence level. If the deviation exceeds the tolerance range, the original sampled data for the corresponding time period is backtracked, and the valid value is determined by majority voting and the rate of change is recalculated. This mechanism gives the trend calculation redundant fault-tolerant properties.

[0010] Furthermore, the effective trend criterion includes the current instantaneous value of each parameter, the rate of change, the confidence level, and the current computational model identifier. When the nonlinear prediction model is enabled, the effective trend criterion also includes change acceleration information and short-term prediction values. The above information is output as a complete data packet from the hub to the local display controller, providing sufficient decision information for subsequent hierarchical fusion judgment.

[0011] Furthermore, the multi-level safety threshold model sets two sets of criteria: a threshold sequence for instantaneous parameter values ​​and a threshold sequence for the rate of change. Based on the combination of exceeding limits for both instantaneous parameter values ​​and rates of change, it triggers three levels of early warning: minor anomaly, general hidden danger, and emergency. Specifically, cases where the instantaneous parameter value does not exceed the limit but the rate of change does, and cases where the instantaneous value exceeds the first-level threshold but the rate of change does not, are both classified as minor anomalies. When the effective trend criterion includes a short-term predicted value and this predicted value exceeds the second-level threshold within the prediction time domain, an emergency warning is also triggered. This model overcomes the limitations of traditional single-threshold binary judgment and achieves forward-looking, tiered risk identification.

[0012] Furthermore, the local display controller drives differentiated audible and visual alarm responses based on multi-level risk states, and synchronizes the warning level information to the remote cloud platform and the subway BAS system via a serial gateway and a multi-protocol parallel transmission channel, ensuring that all levels of management receive consistent risk status information.

[0013] Furthermore, the reverse feedback adjustment strategy includes two dimensions: quantitative adjustment of collected parameters and qualitative switching of the trend calculation model. It is implemented in four coordinated levels according to the warning level: In normal conditions, the regular sampling period and standard reporting frequency are maintained, and a linear fitting model is used to perform equal-weighted linear regression on the sampled sequences within the window to output the rate of change; for minor anomalies, the regular sampling period, standard reporting frequency, and sliding window length remain unchanged, but an exponentially weighted model is switched to calculate the rate of change by assigning time-decreasing weight coefficients to each sample point within the window; for general hidden dangers, the sliding window length is shortened and the reporting frequency is increased, the exponentially weighted model continues to be used to calculate the rate of change, and a nonlinear prediction model including an acceleration term is simultaneously activated to extract the acceleration of change and extrapolate and predict the short-term trend of parameters; for emergency situations, the highest frequency collection mode is switched, and a nonlinear prediction model is used as the currently effective model to output the rate of change, acceleration of change, and short-term predicted values. This mechanism matches the algorithmic logic of trend calculation with the parameter evolution characteristics of the current risk stage, deepening the technical essence of the cross-coupling between the redundant verification layer and the hierarchical warning layer.

[0014] Furthermore, the rate outputs of various trend calculation models are all normalized to the rate of change under a unified physical dimension and are calibrated for factory consistency. The first three sampling periods after model switching are the transition period. During the transition period, the rate outputs of the old and new models are weighted and fused according to linearly changing weights. The rate exceeding the limit for the first time during the transition period needs to be confirmed for two consecutive sampling periods before the warning level is upgraded, so as to ensure the consistency of the judgment results of the fixed rate threshold sequence under different model states.

[0015] Furthermore, the local display controller generates dynamic change curves of each parameter in real time based on effective trend criteria and retains historical trend data. The dynamic change curves, early warning status and fault information are synchronized to the remote cloud platform and the subway BAS system through a multi-protocol parallel method to achieve full-domain information sharing and provide a data foundation for post-event traceability and operation and maintenance decisions.

[0016] The present invention also provides a multi-parameter redundant safety early warning system for subway power corridors implementing the above method, including a multi-parameter sensor group, two master and one slave redundant hub units, a local display controller and a multi-terminal data synchronization channel, wherein the redundant hub units and the local display controller form a bidirectional cross-coupled closed loop.

[0017] The beneficial effects of this invention include: 1. By embedding trend calculation into the data verification process of redundant hubs as a natural extension of verification, continuous tracking of environmental parameter change trends is achieved, and the rate of change of continuous time series can be calculated without configuring additional dedicated computing hardware.

[0018] 2. By using a fusion judgment mechanism of instantaneous parameter values ​​and rate of change, a low-level warning can be issued in advance when the parameter has not yet exceeded the absolute threshold but the trend of change has shown an abnormal acceleration. This achieves forward-looking hierarchical identification of risks, and the granularity and targeting of the response strategy are significantly enhanced compared to traditional binary alarms.

[0019] 3. The trend calculation function is distributed in the redundant hub layer, and the hierarchical early warning judgment is undertaken by the local display controller. The two layers are cross-coupled and mutually verified. The failure of any single node will not cause the trend recognition or hierarchical early warning capability to be completely interrupted. The overall fault tolerance of the system is significantly better than that of the centralized single host solution.

[0020] 4. Through a closed-loop coupling mechanism that drives the collection density in reverse by the early warning level, the system automatically improves the data capture accuracy when the risk increases, and maintains an economical sampling cycle under normal conditions, thus realizing the intelligent dynamic allocation of monitoring resources and the adaptive improvement of trend capture accuracy.

[0021] 5. The reverse feedback of the early warning level not only drives the quantitative adjustment of the sampling density, but also drives the qualitative switching of the trend calculation algorithm model, enabling the hub to adopt the calculation method that best matches the parameter evolution characteristics at different risk stages, improving the information quality and prediction timeliness of the trend criterion in the closed-loop coupling, and realizing the deep linkage of cross-coupling from the parameter level to the algorithm level. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall process of the multi-parameter redundancy safety early warning control method for subway power corridors according to the present invention.

[0023] Figure 2 This is a schematic diagram of the multi-parameter redundancy safety early warning system architecture for subway power corridors according to the present invention.

[0024] Figure 3 This is a schematic diagram showing the comparison results of the verification experiment for this invention. Detailed Implementation

[0025] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0026] Example 1 This embodiment corresponds to the following content: "A multi-parameter redundant safety early warning control method for subway power corridors, including: dual main hubs receiving environmental signals continuously output by multi-parameter sensors in the corridor in parallel, each independently maintaining the parameter sequence within a sliding time window, and calculating the parameter change rate based on the currently effective model among a variety of preset trend calculation models; The hub performs a consistency check on the current frame value and change rate output by the dual main hubs. If the check passes, it outputs a valid trend criterion that has been redundantly confirmed. If the check fails, it initiates data backtracking and majority voting error correction. After receiving the effective trend criteria, the local display controller substitutes the instantaneous value of the parameter and the rate of change into the multi-level safety threshold model for fusion judgment. Based on the combination relationship between the degree of instantaneous value exceeding the limit and the degree of rate exceeding the limit, it distinguishes the multi-level risk status and drives the corresponding level of graded early warning output. The dual master hubs and the slave hubs together constitute a hub layer. The output of the hierarchical early warning is fed back to the hub layer. The hub layer dynamically adjusts the length of the sliding time window and the data reporting frequency according to the early warning level, and synchronously switches the currently effective trend calculation model, forming a bidirectional cross-coupled closed loop between the redundancy verification layer and the hierarchical early warning layer. like Figure 1 This document details the complete process of trend identification and hierarchical early warning methods for a multi-parameter redundancy safety early warning system used in subway power corridors.

[0027] The overall architecture of this method is built around a bidirectional cross-coupling closed loop between the redundancy verification layer and the hierarchical early warning layer, and specifically includes the following steps.

[0028] Step S1: The dual hubs receive the environmental signals continuously output by the multi-parameter sensors in the corridor in parallel, each independently maintains the parameter sequence within the sliding time window, and calculates the parameter change rate based on the currently effective model among the preset multiple trend calculation models.

[0029] It should be noted that in this step, dual hubs refer to two hubs with equal functionality operating in parallel. They work simultaneously in a hot standby mode, rather than a primary / standby switching mode. Each hub synchronously receives signals from all environmental sensors within the corridor, independently performs signal analysis locally, and writes the analyzed parameter values ​​into its own maintained sliding time window in chronological order. The sliding time window is a sequence of accumulated parameter samples over a fixed time interval. This window continuously slides forward over time, with new data written to the end of the window while expired data at the beginning is removed. Each hub independently calculates the rate of change of each parameter based on the parameter sequence within its window, i.e., the amount of change in the parameter value per unit time. The two hubs independently maintain their respective windows and rate calculation results, embedding trend calculation capabilities into the data acquisition layer as a natural extension of the acquisition process, rather than relying on downstream hardware.

[0030] It should be noted that the calculation method for the rate of change is not fixed, but can dynamically select different calculation models based on the current operating state of the system. The hub has multiple pre-set trend calculation models, including linear fitting models, exponentially weighted models, and nonlinear prediction models containing acceleration terms. Each model has different adaptability and predictive timeliness under different parameter evolution characteristics. The specific model used is determined by the downstream early warning layer through the back-transmitted early warning level signal, a mechanism detailed in step S4.

[0031] Step S2: Simultaneously perform consistency verification on the current frame value and change rate output by the dual main hubs from the hub. If the verification passes, output the valid trend criterion confirmed by redundancy. If the verification fails, start data backtracking and majority voting error correction.

[0032] It should be noted that the slave hub is the third hub in a two-master-one-slave redundant hub architecture. Its core function is to cross-check and verify the data output by the two master hubs. The consistency check includes not only the instantaneous values ​​of each parameter in the current frame, but also the change rate values ​​calculated by each of the two master hubs. The slave hub calculates the deviation of the change rates output by the two master hubs for the same parameter. When the deviation is within a preset allowable range, the check is considered successful, and the average of the two rate values ​​is taken as the final valid trend criterion for output. When the deviation exceeds the allowable range, the check is considered abnormal, and the slave hub immediately retrieves the original sampled data from the two master hubs within the corresponding time period for backtracking and verification. The valid data is determined by a majority vote based on the consistency of two out of the three data sources, and the change rate is recalculated based on the corrected valid data. This design gives the trend calculation process redundant fault-tolerant properties. That is, the correctness of the trend calculation is guaranteed through redundant verification, and the verification process is further enhanced by incorporating the comparison of the trend calculation results. The two verify and reinforce each other.

[0033] Step S3: After receiving the valid trend criteria, the local display controller substitutes the instantaneous value of the parameter and the rate of change into the multi-level safety threshold model for fusion judgment. Based on the combination relationship between the degree of instantaneous value exceeding the limit and the degree of rate exceeding the limit, it distinguishes the multi-level risk status and drives the corresponding level of graded early warning output.

[0034] It should be noted that the multi-level safety threshold model built into the local display controller differs from traditional single threshold comparison in that it simultaneously evaluates both the absolute position and the trend of change of the parameters. The instantaneous value of the parameter reflects the absolute safety status of the current environment, while the rate of change reflects how quickly the environment evolves towards a dangerous direction.

[0035] Different combinations of the two dimensions correspond to different risk levels: when the instantaneous value is still within the safe range but the rate has accelerated abnormally, it means the environment is in the early stages of deterioration, triggering a lower-level warning for early prevention; when the instantaneous value has exceeded the safe range and the rate is still accelerating, it means the risk is rapidly expanding, triggering a higher-level warning to drive emergency response. The local display controller directly drives the corresponding graded warning output based on the determined risk level, including differentiated audible and visual alarm signals and warning status information transmitted to external systems.

[0036] Step S4: The output results of the hierarchical early warning are sent back to the hub layer. The hub dynamically adjusts the length of the sliding time window and the data reporting frequency according to the early warning level, and simultaneously switches the currently effective trend calculation model, forming a two-way cross-coupling closed loop between the redundancy verification layer and the hierarchical early warning layer.

[0037] It should be noted that this step is crucial for achieving a bidirectional cross-coupling closed loop. After the local display controller completes the graded early warning determination, it transmits the current early warning level information back to the hub layer via the communication link. Upon receiving the early warning level information, the hub performs dynamic adjustments at both levels accordingly.

[0038] The first level involves the quantitative adjustment of the collected parameters: when the warning level is low, a longer sliding time window and a regular data reporting frequency are maintained to balance the economic efficiency of monitoring; when the warning level is high, the length of the sliding time window is shortened to enable it to capture rapid changes more sensitively, while the data reporting frequency is increased to provide more intensive criterion input to the downstream warning layer.

[0039] The second level is the qualitative switching of the trend calculation model: the hub selects the rate of change calculation model that best matches the evolution characteristics of the parameters in the current risk stage based on the warning level signal.

[0040] Under normal conditions, parameter fluctuations are gentle and the trend is linear. The hub uses a linear fitting model to calculate the rate of change. This model outputs the rate using the slope of an equally weighted linear regression of the sampled sequences within the window. It has low computational cost and high fitting accuracy for stable changes. When the warning level is slightly abnormal, although the instantaneous parameter values ​​are still within the safe range, the rate of change has exceeded the normal fluctuation range. Recent data is more valuable for trend judgment than historical data. While maintaining the normal collection parameters, the hub first switches to an exponentially weighted model. This model assigns a weight coefficient that decays over time to each sampling point within the window, making the calculation result of the rate of change more focused on the contribution of recent changes. Its response sensitivity to accelerated deterioration is significantly higher than that of the equally weighted linear model. When the warning level rises to a general hidden danger, the hub shortens the sliding time window and increases the collection and reporting frequency. It continues to calculate the rate of change using the exponentially weighted model, while simultaneously activating a nonlinear prediction model that includes an acceleration term. It extracts the acceleration of change in parallel and extrapolates and predicts the short-term trend of the parameters, so that the short-term predicted value becomes a trend criterion from the general hidden danger stage. When the warning level rises to an emergency, the parameters may be in a nonlinear rapid change phase. The hub switches the parallel nonlinear prediction model containing acceleration terms to the currently active main operation model. This model further extracts the acceleration information based on the rate of change and extrapolates and predicts the short-term trend of the parameters. This makes the trend criteria not only reflect the current speed of change, but also carry the prediction information of the upcoming change situation, providing forward-looking support for emergency decision-making at the downstream warning level.

[0041] The synergistic effect of these two levels creates a deeply cross-coupled closed loop within the entire system: higher risks require more precise trend calculation models and denser data collection; more precise calculations result in richer information dimensions for trend criteria; more comprehensive criteria lead to more accurate tiered early warnings; and more accurate early warnings result in more targeted reverse-driving mechanisms. The redundancy verification layer and the tiered early warning layer, through positive output of trend criteria and negative feedback of early warning levels, form a bidirectional cross-coupling from parameter adjustment to algorithm switching. This makes trend identification and tiered early warning a system-level collaborative capability, completely eliminating dependence on a dedicated alarm control host.

[0042] Example 2 This embodiment corresponds to the following: "The multi-parameter sensor includes a carbon monoxide sensor, a hydrogen sulfide sensor, an oxygen sensor, a combustible gas sensor, and a temperature sensor. Each sensor is deployed at key points along the power corridor to convert environmental parameters into standard electrical signals and continuously output them to the dual master hub and the slave hub." The multi-parameter sensors include five types: carbon monoxide sensors, hydrogen sulfide sensors, oxygen sensors, combustible gas sensors, and temperature sensors. Carbon monoxide sensors detect the concentration of carbon monoxide gas produced by the thermal decomposition or incomplete combustion of cable insulation materials, covering a complete concentration range from trace leaks to dangerous accumulations that may occur within the corridor. Hydrogen sulfide sensors detect the concentration of hydrogen sulfide gas that may be carried by groundwater seepage or pipeline leaks; this gas is extremely toxic, and even trace amounts require attention. Oxygen sensors monitor the oxygen concentration within the corridor; when other gases accumulate in a confined space, the oxygen concentration will decrease accordingly, and personnel entering below the safe limit face a risk of asphyxiation. Combustible gas sensors detect the concentration of combustible gases such as methane; these gases can ignite and explode when exposed to open flames or high temperatures when they reach a certain concentration range. Temperature sensors monitor the ambient temperature within the corridor and the local temperature near cable joints; abnormally high temperatures are precursor signals of cable overload or insulation deterioration.

[0043] Each sensor is deployed at key locations along the power corridor. The selection criteria for key locations include: areas with dense cable joints, corners or slope changes in the corridor structure, areas with poor ventilation, areas where pipelines cross, and near entrance / exit access control points. Each sensor converts environmental parameters into standard electrical signals. In this embodiment, a 4-20mA analog current signal or an RS485 digital signal is used, and the signal is continuously output to a dual master hub and a slave hub. The slave hub synchronously records a copy of the sensor data for majority voting error correction in case of consistency verification anomalies. Continuous output means that the sensors operate continuously with a fixed sampling period, without interruption windows such as intermittent sleep or manually triggered sampling, ensuring the temporal continuity of environmental parameters.

[0044] Example 3 This embodiment corresponds to the following: "The dual hubs synchronously receive all sensor signals in a parallel hot-standby mode and analyze them independently. Each hub maintains a configurable sliding time window locally, sampling each parameter periodically within the window to form a continuous time series, and calculating the rate of change of each parameter based on the series according to the currently effective trend calculation model; the various trend calculation models include a linear fitting model, an exponential weighted model, and a nonlinear prediction model containing an acceleration term, and the dual hubs always synchronously use the same trend calculation model."

[0045] In this system, the dual master hubs simultaneously receive all sensor signals in a parallel hot-standby mode and independently process them. Parallel hot standby means that both master hubs are in normal working condition at all times, simultaneously receiving signals from the same set of sensors, and independently completing signal analysis and subsequent processing, rather than a master-standby switching mode where one works while the other is on standby. The two master hubs establish connections with the sensors through their respective signal input ports. Sensor signals are simultaneously sent to the inputs of both master hubs and the slave hub via a splitter or bus broadcast, allowing the slave hub to simultaneously record a copy of the sensor data.

[0046] Each hub maintains a configurable sliding time window locally. The configurable length means the window's time span can be set according to actual engineering needs, and can also be dynamically adjusted by the downstream early warning layer via reverse feedback signals during system operation. In this embodiment, the default window length is 5 minutes. Within the window, the values ​​of each parameter are recorded point-by-point according to the sampling period, forming a continuous time series. Taking carbon monoxide concentration as an example, if the sampling period is 10 seconds and the window length is 5 minutes, the window stores the concentration values ​​of the most recent 30 consecutive sampling points, constituting a concentration change sequence over a period of 5 minutes.

[0047] The main hub calculates the rate of change of each parameter based on a continuous time series within a window. The calculation of the rate of change selects the appropriate calculation method from several preset trend calculation models based on the current operating state of the system. The hub has three preset trend calculation models, each suitable for the parameter evolution characteristics at different risk stages.

[0048] The first type is the linear fitting model, suitable for scenarios where parameters fluctuate smoothly under normal conditions. This model performs an equally weighted linear regression on all sampling points within the window, outputting the slope of the regression line as the rate of change. Specifically, suppose there are N sampling points within the window, and the time corresponding to each point is from t1 to t2. N The parameter values ​​are x1 to x N The linear fitting model uses the least squares method to find the slope of the best-fitting line as the rate of change v. l Its calculation is expressed as: Where N is the total number of sampling points within the window, t i Let x be the time of the i-th sampling point. i Let be the parameter measurement value at the i-th sampling point. This model has low computational cost and high fitting accuracy for stationary sequences, but its response to recent abrupt changes is delayed.

[0049] The second type is the exponentially weighted model, suitable for scenarios where parameters show abnormal trends under minor anomalies and general hidden dangers. This model assigns weights to each sampling point within the window that decay exponentially over time, making the contribution of recent data to the rate calculation significantly greater than that of earlier data, thus more sensitively reflecting the latest trends. Its calculation is expressed as follows: Where α is the decay factor, with a value greater than 0 and less than 1, and is set to 0.85 in this embodiment; the terms in the numerator are the instantaneous rate of change between adjacent sampling points multiplied by the corresponding exponential weight, and the denominator is the weight normalization factor. The smaller α is, the more concentrated the weights of recent data are, and the more sensitive the response to accelerated changes. This model can capture signals of accelerated change earlier than the equal-weighted linear model when the parameters have already shown a trend of shift.

[0050] The third type is a nonlinear prediction model that includes an acceleration term. It is used in parallel with the exponentially weighted model from the general hazard state and is suitable for scenarios where parameters may undergo nonlinear rapid changes. This model calculates the rate of change, extracts the acceleration information, and extrapolates and predicts the short-term future values ​​of the parameters based on the current rate and acceleration.

[0051] The calculation method for changing acceleration involves performing a difference operation on the velocity values ​​over a continuous time period to obtain the rate of change of the velocity itself. Specifically, the changing acceleration a(t) is expressed as: a(t) = (v(t) − v(t − Δt)) / Δt, where v(t) is the rate of change calculated based on the exponential weighted model in the current sampling period, v(t − Δt) is the rate of change calculated in the previous sampling period, and Δt is the current sampling period interval. Under normal hazardous conditions, Δt is 5 seconds, and under emergency conditions, Δt is 2 seconds.

[0052] The predicted output is the estimated value of the parameters that will be reached in the specified prediction time domain τ in the future, assuming that the current rate and acceleration remain constant. Its calculation is expressed as: Where x(t) is the parameter value at the current moment, v(t) is the current rate of change, a(t) is the current acceleration, and τ is the prediction time domain length. The trend criterion output by this model includes three pieces of information: rate v(t), acceleration a(t), and predicted value, providing forward-looking support for decision-making by the downstream early warning layer in emergency situations.

[0053] It should be noted that the system initially uses a linear fitting model by default upon power-up. Subsequent model switching is performed based on the warning level signals relayed back from the downstream warning layer. Both main hubs always use the same computational model synchronously to ensure that consistency checks on the slave hubs are performed using the same algorithm, avoiding false verification anomalies caused by algorithm differences.

[0054] Example 4 This embodiment corresponds to the following: "When the consistency check is performed by the slave hub, the rate of change of the same parameter output by the dual master hubs is compared for deviation; if the deviation is within the preset allowable range, the average of the two rate values ​​is taken as the valid trend criterion and marked with a high confidence level; if the deviation exceeds the allowable range, the original sampling data of the corresponding time period is traced back, the valid value is determined by majority voting, and the rate of change is recalculated and marked as the trend criterion confirmed by error correction."

[0055] In this process, when the slave hub performs consistency verification, it compares the rate of change of the same parameter output by the two master hubs. Specifically, the slave hub periodically receives the rate of change values ​​calculated by the two master hubs, and calculates the absolute value of the difference between the two rate values ​​for the same parameter at the same time as the deviation index. The slave hub has a preset allowable range value, which is determined based on the measurement accuracy of the sensor itself, the computational accuracy of the hub, and the current sampling period. Since the dimension of the rate of change is the parameter dimension divided by the time dimension, the allowable range setting needs to convert the sensor's full-scale accuracy to the rate dimension for effective comparison. In this embodiment, the allowable range of the rate for each parameter is set to twice the rate accuracy obtained by dividing the sensor's full-scale accuracy by the current sampling period. This conversion method ensures that the allowable range and the rate value being compared are consistent in dimension, giving the deviation judgment physical meaning.

[0056] If the deviation is within the preset allowable range, the trend calculation results of the two main hubs are considered consistent and reliable. The arithmetic mean of the two rate values ​​from the hubs is taken as the valid trend criterion and output externally, while marking this criterion as a high reliability level. A high reliability level indicates that the trend criterion has undergone dual protection through independent dual-path calculation and consistency verification, and the downstream early warning layer can directly adopt it and perform hierarchical judgment accordingly.

[0057] If the deviation exceeds the allowable range, it is determined that there is signal channel distortion or single hub resolution deviation. The data backtracking process is immediately initiated from the hub: the original sampled data stored by the two main hubs in the corresponding time period is retrieved, along with the data copy synchronously recorded by the hub itself. The three original data are then subject to majority voting point by point according to the sampling time. At each sampling time, at least two consistent values ​​from the three data are taken as valid values. If the three data are different, the median value is taken. The rate of change is recalculated based on the valid original data sequence determined by the majority vote. The final output trend criterion is marked as the error-corrected confirmation level, indicating that although the criterion has been corrected, its original data had inconsistencies, reminding downstream to re-examine when necessary. To this end, the same linear fitting model, exponential weighted model, and nonlinear prediction model containing acceleration terms are pre-set in the hub as in the main hub. The rate calculation in error correction recalculation and degradation mode is performed using the same algorithm as the currently effective model, ensuring the comparability of the recalculation results with the normal dual-channel output based on the algorithm.

[0058] It should be noted that when performing consistency checks, the slave hub also needs to verify whether the trend calculation models currently used by the two master hubs are consistent. Since the model switching command is forwarded uniformly from the slave hub to both master hubs, under normal circumstances, the two master hubs always operate under the same model. However, if communication delays cause one master hub to have not yet completed the model switch while the other has, the slave hub will suspend this round of checks upon detecting an inconsistency in model identifiers. It will wait until both master hubs have confirmed the switch is complete before performing the comparison, thus avoiding false anomaly detections due to algorithm differences.

[0059] It should also be noted that the slave hub monitors the online status of the two master hubs through periodic heartbeat frames. When no data frames are received from a master hub for three consecutive sampling cycles, the master hub is determined to be faulty and enters single-master degradation mode: the slave hub stops comparing the dual-path rate deviation and instead compares the instantaneous value output by the surviving master hub with its own recorded sensor data copy point by point. If the comparison matches, the trend criterion calculated by the surviving master hub is directly forwarded and the confidence mark is set to single-path confirmation. If the comparison does not match, the change rate is recalculated based on its own data copy according to the currently effective model and then output. At the same time, the slave hub reports the master hub fault information along with the trend criterion to the local display controller, and notifies the remote cloud platform and the subway BAS system to arrange maintenance through the multi-terminal data synchronization channel. After the faulty master hub is restored to online status, the slave hub automatically resumes the dual-path consistency verification mode after its sliding time window is filled with data again.

[0060] Example 5 This embodiment corresponds to the following: "The effective trend criterion includes the current instantaneous value of each parameter, the rate of change, the confidence level, and the current computational model identifier. When the nonlinear prediction model is enabled, the effective trend criterion also includes change acceleration information and short-term prediction values. The above information is output as a complete data packet from the hub to the local display controller."

[0061] The effective trend criterion includes four basic pieces of information for each parameter: the current instantaneous value, the rate of change, the confidence level, and the current computational model identifier. The current instantaneous value refers to the parameter measurement value in the latest sampling period after redundancy verification. This value has been verified through dual-master comparison to eliminate data distortion caused by single-channel signal distortion or single-hub parsing errors. The rate of change refers to the slope of the parameter's change over time, calculated based on a continuous sequence within a sliding time window, reflecting the direction and speed of the environmental parameter's evolution. The credibility flag is an identifier field characterizing the reliability of this set of criteria, divided into three levels: high credibility, error-corrected confirmation, and single-path confirmation. High credibility indicates that the dual-master calculation results directly agree. Error-corrected confirmation indicates that the result is obtained after majority voting for error correction. Single-path confirmation indicates that the criterion originates only from the calculation result of a single master hub and has not undergone dual-path consistency verification. This includes situations where the slave hub directly forwards data from the surviving master hub when the single master hub fails, and situations where the master hub directly sends data through a backup direct connection channel when the slave hub fails or the master channel is interrupted. The local display controller performs the hierarchical early warning judgment for single-path confirmation criteria as usual, while marking the single-path confirmation status in the early warning record and generating maintenance information prompting maintenance personnel to check faulty equipment or communication links. The current calculation model identifier is used to inform the downstream early warning layer which calculation model generated the current trend criterion, facilitating the downstream to correctly interpret the information dimensions and applicable scenarios of the criterion.

[0062] When the nonlinear prediction model is in operation (i.e., the warning level is general hazard or above), the effective trend criterion, in addition to the four basic information items mentioned above, also includes information on the rate of change and short-term predicted values. The rate of change reflects the speed of change of the rate itself, while the short-term predicted values ​​are parameter estimates extrapolated from the current rate and acceleration in the prediction time domain. These two additional pieces of information enable the local display controller to obtain forward-looking decision support in emergency situations, knowing not only how fast the parameters are changing, but also whether this change is intensifying and to what level it is about to reach.

[0063] The aforementioned information is packaged as a complete data packet by the hub and output to the local display controller. The data packet is encapsulated using a fixed-format frame structure. Each frame contains a parameter identifier, instantaneous value field, rate value field, confidence field, computational model identifier field, and, when the nonlinear prediction model is enabled, acceleration and predicted value fields. A checksum may be appended at the end. The data packet is transmitted at fixed intervals via the communication link between the hub and the local display controller. Upon receiving the data packet, the local display controller parses the contents of each field as input to the multi-level security threshold model. This design of packaging all information for unified output allows the local display controller to obtain complete decision information in a single reception, eliminating the need for multiple upstream requests or data splicing from different sources. This reduces communication overhead and ensures the time synchronization of all information.

[0064] Example 6 This embodiment corresponds to the following: "The multi-level safety threshold model sets two sets of criteria: a parameter instantaneous value threshold sequence and a change rate threshold sequence. When the parameter instantaneous value does not exceed the limit but the change rate exceeds the corresponding rate threshold, or when the parameter instantaneous value exceeds the first-level threshold but the change rate does not exceed the limit, a minor anomaly level warning is triggered; when the parameter instantaneous value exceeds the first-level threshold and the change rate also exceeds the limit, a general hazard level warning is triggered; when the parameter instantaneous value exceeds the second-level threshold or the change rate exceeds the emergency rate threshold, an emergency hazard level warning is triggered; when the effective trend criterion includes a short-term predicted value and the predicted value exceeds the second-level threshold within the prediction time domain, an emergency hazard level warning is also triggered."

[0065] The multi-level safety threshold model sets two sets of criteria: instantaneous value threshold sequences and rate of change threshold sequences. The instantaneous value threshold sequences include multiple progressive thresholds for each environmental parameter. For example, carbon monoxide concentration has a Level 1 threshold and a Level 2 threshold; the Level 1 threshold corresponds to a concentration level requiring attention, while the Level 2 threshold corresponds to a concentration level requiring immediate action. Similarly, the rate of change threshold sequences also have multiple progressive thresholds for each parameter. For example, the rate of change of carbon monoxide concentration has a normal rate threshold and an emergency rate threshold; the normal rate threshold corresponds to a rate of change exceeding the normal fluctuation range, while the emergency rate threshold corresponds to a rate of change indicating a rapidly deteriorating environment. The specific values ​​of each parameter threshold are determined according to relevant national standards and industry specifications, including occupational exposure limits for various hazardous gases in GBZ 2.1, oxygen safety limits in GB8958 (Safety Regulations for Oxygen-Deficient Hazardous Operations), and combustible gas alarm setting values ​​in GB 50493 (Design Standard for Combustible and Toxic Gas Detection Alarms in the Petrochemical Industry), ensuring that the system threshold settings have a legal basis and safety margin. The method for determining the rate of change threshold sequence is as follows: the normal rate threshold is set based on the long-term statistical characteristics of the fluctuation rate of each parameter under normal corridor operation conditions, taking the larger of three times the standard deviation of the normal fluctuation rate and the sensor rate resolution, to ensure that normal environmental disturbances will not trigger an early warning; the emergency rate threshold is set by reverse calculation based on the shortest emergency response time, that is, to ensure that when the parameter changes continuously at this rate, the remaining time from the first-level threshold to the second-level threshold is not less than the shortest response time required for on-site personnel evacuation and emergency linkage. Taking carbon monoxide as an example, under the test conditions of this embodiment, the first-level threshold is set to 6 ppm, the second-level threshold is set to 24 ppm with reference to the short-term exposure limit of GBZ 2.1 and with a safety margin, the normal rate threshold is set to 0.6 ppm per minute, and the emergency rate threshold is set to 12 ppm per minute; the two sets of thresholds for the other parameters are set separately using the same method, combined with the corresponding national standard limits and normal fluctuation statistical characteristics.

[0066] For parameters such as oxygen, where the risk direction is a decrease in concentration, "exceeding the limit" means that the instantaneous value is lower than the corresponding threshold, and "exceeding the limit in terms of rate of change" means that the absolute value of the rate of decrease exceeds the corresponding rate threshold.

[0067] The specific logic of the fusion judgment is as follows: When the instantaneous value of a parameter does not exceed the limit but the rate of change exceeds the corresponding rate threshold, a minor anomaly level warning is triggered. This situation means that the environmental parameter is currently still within the safe range, but is developing in a dangerous direction at a rate exceeding normal fluctuations. If no intervention is taken, it will exceed the safety limit within a foreseeable time. At this time, the minor anomaly level warning provides an early warning signal, which facilitates maintenance personnel to arrange inspections or preventive measures in advance. When the instantaneous value of a parameter exceeds the first-level threshold but the rate of change does not exceed the corresponding rate threshold, it indicates that the parameter is in a state of slight exceedance but the change tends to be stable. This also triggers a minor anomaly level warning, so that all combinations of instantaneous value judgment and rate judgment have a clear level classification.

[0068] When the instantaneous value of a parameter exceeds the Level 1 threshold and its rate of change also exceeds the limit, a general hazard level warning is triggered. This situation means that the environmental parameter has deviated from the safe range and is still deteriorating, requiring proactive intervention measures. When the instantaneous value of a parameter exceeds the Level 2 threshold or its rate of change exceeds the emergency rate threshold, an emergency hazard level warning is triggered. This situation means that the environment has reached or is rapidly approaching a dangerous state, requiring the immediate activation of the highest level of emergency response.

[0069] It should be noted that the emergency hazard level determination uses "OR" logic. This means that even if the instantaneous value has not yet reached the Level 2 threshold, as long as the rate of change exceeds the emergency rate threshold, the system will immediately trigger the highest level warning. The intention behind this design is that when the rate of change is extremely high, the parameter may cross all threshold levels in a very short time. Waiting for the instantaneous value to actually reach the Level 2 threshold before triggering an emergency warning would result in a loss of emergency response time. Furthermore, since the nonlinear prediction model is activated in parallel from the general hazard state, the short-term predicted value added to the trend criterion participates in the emergency hazard determination from this stage onwards: if the short-term predicted value will exceed the Level 2 threshold within the prediction time domain, even if the current instantaneous value and rate of change have not reached the emergency threshold, the system will upgrade the warning level to an emergency hazard in advance, gaining emergency response time before the actual breakthrough occurs, further enhancing its foresight. It should be noted that when multiple parameters simultaneously trigger warnings of different levels, the local display controller uses the highest-level priority principle to determine the overall system warning output level. That is, the highest risk level among the individual judgment results of all parameters is taken as the current system warning level. This overall level is used to drive both the audible and visual alarms and to send back data to the hub layer. The independent warning status of each parameter is still separately labeled on the display interface and in the reported data packet, allowing maintenance personnel to identify the specific source of risk.

[0070] Example 7 This embodiment corresponds to the following: "The local display controller drives differentiated audible and visual alarm responses according to the multi-level risk status, and synchronizes the warning level information to the remote cloud platform and the subway BAS system via a serial port gateway and a multi-protocol parallel transmission channel."

[0071] The local display controller drives differentiated audible and visual alarm responses based on multi-level risk conditions. This differentiation manifests in different alarm intensities and methods corresponding to different warning levels: minor anomalies trigger a low-frequency flashing yellow indicator light and a short, single buzzer; general hazards trigger a high-frequency flashing orange indicator light and an intermittent buzzer; and emergencies trigger a continuously lit red indicator light and a continuous high-decibel alarm. On-site personnel can quickly determine the current risk level by observing the alarm's color, flashing frequency, and sound pattern, without needing to consult specific values ​​on the display screen. This is particularly suitable for dimly lit and restrictive working environments in corridors.

[0072] While driving local audible and visual alarms, the local display controller simultaneously transmits warning level information to the remote cloud platform and the metro BAS system via a serial gateway and a multi-protocol parallel transmission channel. The serial gateway converts the communication protocol output by the local display controller into the network protocol required for remote transmission, bridging the protocol domains of the field devices and the remote platform. The multi-protocol parallel transmission channel means that warning level information is simultaneously sent to different target platforms via multiple communication protocols, including network transmission protocols for the remote cloud platform and industrial bus protocols for the metro BAS system. Upon receiving the warning level information, each platform can simultaneously display the current risk status of the corridor on its respective interface, allowing managers at different levels to make corresponding scheduling decisions.

[0073] Example 8 This embodiment corresponds to the following: "The reverse feedback adjustment strategy includes two dimensions: quantitative adjustment of the collected parameters and qualitative switching of the trend calculation model. It is implemented in four coordinated levels according to the warning level: When the warning level is normal, the hub layer maintains the regular sampling period and standard reporting frequency, and uses a linear fitting model to perform equal-weighted linear regression on the sampled sequence within the window to output the rate of change; when the warning level is slightly abnormal, the hub layer maintains the regular sampling period, standard reporting frequency, and sliding time window length unchanged, and switches to an exponentially weighted model to assign a weight coefficient that decays over time to each sample point within the window to calculate the rate of change; when the warning level rises to a general hidden danger, the sliding time window length is shortened and the reporting frequency is increased, the exponentially weighted model continues to be used to calculate the rate of change, and a nonlinear prediction model containing an acceleration term is simultaneously activated to extract the acceleration of change and extrapolate and predict the short-term trend of the parameters; when the warning level rises to an emergency danger, it switches to the highest frequency acquisition mode, using the nonlinear prediction model containing an acceleration term as the currently effective model to output the rate of change, acceleration of change, and short-term predicted values."

[0074] The reverse feedback adjustment strategy includes two dimensions: quantitative adjustment of collected parameters and qualitative switching of trend calculation model, and is implemented in four levels according to the early warning level.

[0075] When the warning level is normal, the hub maintains a regular sampling cycle and standard reporting frequency. In this embodiment, the regular sampling cycle is set to 10 seconds, meaning data is acquired and processed from each sensor every 10 seconds. The standard reporting frequency is one valid trend criterion data packet sent to the local display controller every 30 seconds. In terms of the trend calculation model, under normal conditions, the hub uses a linear fitting model to calculate the rate of change. This model performs an equally weighted linear regression on all sampling points within the window, outputting the slope of the fitted line as the rate of change. The linear fitting model is characterized by giving equal importance to data throughout the entire window period. The calculation result reflects the average trend within the window period and has a good smoothing and filtering effect on random noise and short-term impulses, making it suitable for trend representation when parameters fluctuate slightly and smoothly under normal conditions. Under normal conditions, parameter changes are slow and without a clear directional acceleration. Using the linear model with the lowest computational load satisfies the trend monitoring requirements without placing an excessive computational burden on the hub's embedded processor.

[0076] When the warning level is minor anomaly, the hub maintains the regular sampling cycle, standard reporting frequency, and sliding time window length unchanged, only switching to the exponentially weighted model in the trend calculation model dimension. The reason for this design is that minor anomalies indicate that the parameter change rate has exceeded the normal fluctuation range, but the instantaneous value is still within a safe range. At this time, the environment has not substantially deteriorated, and no additional acquisition and communication resources are needed. However, the accuracy of the rate information is crucial for subsequent trend judgment. The equal-weighted linear model treats early stable data and recent abnormal data equally, which leads to the calculated rate being lowered by the early stable data, causing the output rate value to lag behind the actual rate of change of the parameter. The exponentially weighted model assigns exponentially increasing weights to recent sampling points, making the calculation results more realistically reflect the actual rate of change of the parameter in the most recent period. In this embodiment, the attenuation factor α is set to 0.85, meaning that for each forward sampling cycle, the weight of the data point decreases to 85% of the previous point; under the regular 5-minute window configuration, the weight of the earliest difference term at the beginning of the window is only about 1% of the weight of the latest term at the end, and about 3% after the window for general hidden dangers is shortened to 2 minutes, which can be ignored. The rate of change output by this model is significantly more sensitive to recent accelerating trends than that of a linear model, enabling downstream early warning systems to more accurately assess the true rate of risk deterioration. The model's proactive switching allows the system to operate with a sensitive algorithm from the early stages of risk development, gaining an advance advantage in the accuracy of judgment criteria for potential risk escalation.

[0077] When the warning level rises to a general hazard level, the hub shortens the sliding time window and increases the reporting frequency. In this embodiment, the window length is shortened from the usual 5 minutes to 2 minutes, the sampling period is shortened from 10 seconds to 5 seconds, and the reporting frequency is increased from every 30 seconds to every 10 seconds. In terms of the trend calculation model, under the general hazard state, the exponentially weighted model, which was switched on and became effective during the minor anomaly stage, continues to calculate the rate of change. Simultaneously, a nonlinear prediction model containing an acceleration term is activated and operated in parallel with the exponentially weighted model to extract the acceleration of change and extrapolate and predict the short-term trend of the parameters, so that the short-term predicted value is output along with the trend criteria from the general hazard stage onwards. The prediction time domain of this parallel prediction is consistent with that of the emergency hazard state, both being 60 seconds.

[0078] When the warning level escalates to an emergency, the hub switches to the highest frequency acquisition mode. In this embodiment, the sampling period is shortened to 2 seconds, the window length is shortened to 30 seconds, and the reporting frequency is increased to once every 5 seconds. In terms of the trend calculation model, under emergency conditions, the hub switches the parallel-running nonlinear prediction model containing an acceleration term to the currently active main calculation model. The reason for using this model as the main model is that when the warning level reaches the highest level, the system needs to provide as much forward-looking information as possible for emergency decision-making, rather than just informing the current rate of change. While calculating the rate of change, the nonlinear prediction model further extracts the acceleration, i.e., whether the rate itself is still increasing or decreasing. Based on the current parameter value, rate, and acceleration, the model performs a secondary extrapolation prediction of the parameter's value to be reached within one prediction time domain. The prediction time domain τ is set to 60 seconds in this embodiment. If the prediction result shows that the parameter will break through a higher threshold within 60 seconds, this information is sent to the local display controller along with the trend criteria, enabling the latter to initiate corresponding emergency response measures before the actual breakthrough occurs. The computational load of the nonlinear prediction model is increased compared to the previous two models. However, since it is only activated from the general hazard state and the window length is shortened to 30 seconds in the emergency state and the number of data points in the window is limited, the embedded processor of the hub can complete the calculation under real-time constraints.

[0079] It should be noted that the model switching execution mechanism is as follows: the local display controller sends the warning level signal to the slave hub. After confirming the change in warning level, the slave hub synchronously sends a model switching command to the dual master hubs. Upon receiving the command, the dual master hubs use the new model to calculate the rate of change starting from the next sampling period. During the model switching process, the new model immediately begins calculation based on the data sequence currently existing in the window, without waiting for the new model to re-accumulate data from an empty window. Therefore, the delay of the first output after the switch does not exceed one sampling period. The outputs of the three trend calculation models are all normalized to the rate of change under a unified physical dimension: temperature corresponds to degrees Celsius per minute, and gas concentration corresponds to ppm per minute. During the equipment factory calibration stage, the three models are calibrated consistently using a standard ramp signal, ensuring that the steady-state rate estimation deviation for the same constant rate input does not exceed ±5%, thereby guaranteeing that the fixed rate threshold sequence is comparable for any model state. The first three sampling periods after model switching constitute a transition period. During this period, the hub weights the rate outputs of the old and new models using a linearly varying weighting process. The weights gradually transition from 100% for the old model to 100% for the new model throughout the cycle to eliminate rate jumps caused by differences in model sensitivity during the switch. After the transition period, the output is entirely based on the new model. For the first occurrence of a rate of change exceeding the limit during the transition period, the local display controller must confirm for two consecutive sampling periods before raising the warning level. Rate of change exceeding the limit outside the transition period is judged immediately without affecting the response speed under normal operating conditions.

[0080] The model downgrade strategy adopts a conservative approach: the hub only performs the model downgrade switch after receiving three consecutive feedback signals marked as lower warning levels, so as to avoid the computational jitter caused by frequent model switching due to the instantaneous fluctuation of the warning level.

[0081] The aforementioned four adjustment strategies, through the synergistic effect of both the data acquisition parameter dimension and the computational model dimension, achieve the core objective of comprehensively and adaptively improving the trend capture accuracy according to the risk level. When the environment is safe, the system operates with economical resource allocation and a simple algorithm; when risks first emerge, the system switches to a more sensitive algorithm without increasing resource consumption; when risks become apparent, the system encrypts data acquisition and initiates parallel prediction; and in dangerous situations, the system operates with the highest-precision nonlinear prediction model and outputs forward-looking predictions. This adaptive mechanism is the technical effect of the closed-loop coupling between the early warning layer and the acquisition layer: it not only changes the quantitative parameters of how fast to acquire data, but also changes the qualitative logic of how to calculate, deepening the cross-coupling from the parameter level to the algorithm level, forming a deep closed-loop linkage.

[0082] Example 9 This embodiment corresponds to the following: "The local display controller generates dynamic change curves of each parameter in real time based on the effective trend criteria and retains historical trend data. The dynamic change curves, early warning status, and fault information are synchronized to the remote cloud platform and the subway BAS system through a multi-protocol parallel method to achieve full-domain information sharing."

[0083] The local display controller generates dynamic change curves for each parameter in real time based on effective trend criteria. These dynamic change curves are line graphs formed by plotting the instantaneous values ​​of continuously received parameters in chronological order on a coordinate system, with time on the horizontal axis and parameter values ​​on the vertical axis. The curves are refreshed in real time with each new data reception, allowing on-site personnel to intuitively determine whether the parameters are stabilizing, rising, or rapidly deteriorating by observing the curve's trend. The local display controller's screen can simultaneously display curve comparison views for multiple parameters, facilitating the observation of the correlation and changes between different parameters. When the nonlinear prediction model is enabled, a short-term prediction trend line is also overlaid on the dynamic change curves in dashed form, allowing on-site personnel to visually see the predicted trend of the parameters.

[0084] The local display controller simultaneously retains historical trend data. This historical data is stored in the local storage medium of the local display controller as a time-series database, with retention covering at least the most recent 30 days of parameter records. The retention of historical trend data enables post-event retrospective analysis; when a security incident occurs, the complete evolution of each parameter before the event can be replayed, allowing for analysis of the event's causes and the timely response of the early warning system. The historical data also records the trend calculation model type used by the hub at each moment and the nodes of early warning level changes, providing complete process evidence for post-event review of the system's model switching timing and closed-loop response effectiveness.

[0085] Dynamic change curves, early warning statuses, and fault information are synchronized to the remote cloud platform and the metro BAS system via multiple parallel protocols, achieving full-domain information sharing. Upon receiving this information, the remote cloud platform provides a unified monitoring view for operation and maintenance personnel distributed across different geographical locations. The metro BAS system, upon receiving early warning statuses, can centrally display them at the integrated monitoring center and incorporate them into overall operation and scheduling considerations. This full-domain information sharing ensures that personnel at all levels, from frontline staff to the remote management center, receive consistent and synchronized risk situation information, avoiding response delays caused by information asymmetry.

[0086] Example 10 This embodiment corresponds to the following content: "A multi-parameter redundancy safety early warning system for subway power corridors, comprising: A multi-parameter sensor array is deployed at key locations along the power corridor to convert environmental parameters such as carbon monoxide, hydrogen sulfide, oxygen, combustible gas, and temperature into standard electrical signals and continuously output them. The dual-master-one-slave redundant hub unit includes two parallel hot-standby master hubs and one slave hub. Each of the two master hubs independently receives signals from the multi-parameter sensor group and maintains a sliding time window parameter sequence. Each master hub has a pre-set linear fitting model, an exponentially weighted model, and a nonlinear prediction model containing an acceleration term. It calculates the rate of change based on the currently active model. The slave hub synchronously receives signals from the multi-parameter sensor group and records one copy of the sensor data. It is used to perform consistency checks and majority voting error correction on the values ​​and rates of change output by the dual master hubs and output a valid trend criterion confirmed by redundancy. It also undertakes the unified distribution of model switching instructions. The local display controller has a built-in multi-level safety threshold model, which is used to receive the effective trend criteria and fuse the instantaneous value of the parameter with the rate of change to distinguish the multi-level risk status, drive the corresponding level of audible and visual alarm, and send the graded early warning results back to the redundant hub unit to dynamically adjust the sliding time window length, data reporting frequency and trend calculation model selection. A multi-terminal data synchronization channel, including a serial port gateway and a multi-protocol transmission link, is used to synchronize real-time parameters, dynamic change curves, early warning status and fault information to a remote cloud platform and the metro BAS system. The redundant hub unit and the local display controller form a bidirectional cross-coupled closed loop. The redundant hub unit outputs trend criteria to the local display controller, and the local display controller feeds back the warning level to the redundant hub unit to drive adaptive adjustment of the data acquisition density and dynamic switching of the trend calculation model. Figure 2 This document provides a detailed description of the overall architecture and functional units of the multi-parameter redundancy safety early warning system for subway power corridors.

[0087] The system comprises four functional units: a multi-parameter sensor group, a two-master-one-slave redundant hub unit, a local display controller, and a multi-terminal data synchronization channel.

[0088] Multi-parameter sensor arrays are deployed at key locations along the power corridor to convert environmental parameters such as carbon monoxide, hydrogen sulfide, oxygen, combustible gas, and temperature into standard electrical signals and continuously output them. Each sensor in the array simultaneously sends its output signal to both the dual master and slave hubs in a redundant master-slave hub unit via a signal splitter or bus broadcast. The number and spacing of the sensor arrays are determined based on the corridor length, ventilation zones, and risk assessment results to ensure that the environmental conditions of each section are within the sensing coverage area.

[0089] The dual-master-one-slave redundant hub unit comprises three hardware entities: two parallel hot-standby master hubs and one slave hub. Each master hub independently receives signals from a multi-parameter sensor array and maintains a sliding time window parameter sequence to calculate the rate of change. The two master hubs are physically independent and powered independently to ensure operational reliability. Each master hub's embedded processor has three pre-configured trend calculation algorithms: a linear fitting model, an exponentially weighted model, and a nonlinear prediction model. It can dynamically switch the currently active calculation model based on the warning level commands relayed back from downstream systems.

[0090] The slave hub connects to two master hubs simultaneously via an internal communication bus. It performs consistency checks and majority voting error correction on the values ​​and rates of change output by the two master hubs, ultimately outputting a valid trend criterion confirmed by redundancy. The slave hub also maintains a copy of the sensor data for reference during majority voting error correction. The slave hub also handles the unified distribution of model switching commands. After receiving the warning level signal from the local display controller, it synchronously sends model switching and parameter adjustment commands to the two master hubs, ensuring that they always operate under the same algorithm and parameter configuration. The redundant hub unit provides hardware-level redundancy for data acquisition and trend calculation. Even if one master hub fails, the system can still continue to output trend criteria through the collaborative work of the other master hub and the slave hub.

[0091] The local display controller incorporates a multi-level safety threshold model to receive valid trend criteria and fuse instantaneous parameter values ​​with their rate of change to differentiate between various risk levels. It features an industrial-grade display screen for on-site visualization, an audible and visual alarm output port to drive alarm devices at different levels, and a communication interface for data exchange with redundant hub units and external systems. The local display controller drives corresponding levels of audible and visual alarms and sends the tiered warning results back to the redundant hub unit to dynamically adjust the sliding time window length, data reporting frequency, and trend calculation model selection.

[0092] Reverse feedback is accomplished through a bidirectional communication link between the local display controller and the slave hub. After receiving the warning level information, the slave hub forwards it to the dual master hubs to achieve synchronous parameter adjustment and model switching of the three hubs. In addition, a backup direct communication channel is provided between the dual master hubs and the local display controller. This channel is in a heartbeat monitoring state and does not carry service data during normal operation. When the local display controller detects that the slave hub's main channel has not received data packets for more than three consecutive reporting cycles, it automatically switches to the backup direct communication channel to receive trend criteria directly sent by the dual master hubs. The backup direct communication channel uses the same RS485 physical interface and frame format as the main channel. The dual master hubs also automatically activate this channel for direct data transmission after detecting a communication interruption with the slave hub. In the backup channel operating mode, the trend criteria are not subject to dual-path consistency verification by the slave hub. The local display controller automatically downgrades its confidence level to single-path confirmation and then performs the graded warning judgment as usual.

[0093] The multi-terminal data synchronization channel consists of two components: a serial port gateway and a multi-protocol transmission link. The serial port gateway, acting as a protocol conversion bridging device, converts the field communication protocol output by the local display controller into a network protocol suitable for remote transmission. The multi-protocol transmission link synchronizes real-time parameters, dynamic change curves, early warning status, and fault information to the remote cloud platform and the subway BAS system, supporting the parallel operation of multiple transmission protocols to improve data arrival rate and transmission reliability.

[0094] In this system, redundant hub units and local display controllers form a bidirectional cross-coupled closed loop. On the forward path, the redundant hub units output redundantly confirmed trend criteria to the local display controller, providing a reliable decision-making basis for tiered early warning. On the reverse path, the local display controller feeds back the early warning level to the redundant hub units to drive adaptive adjustment of data acquisition density and dynamic switching of the trend calculation model, enabling dynamic matching of trend capture accuracy and algorithm adaptability with the level of risk. This bidirectional cross-coupled architecture allows trend identification and tiered early warning to no longer rely on a single dedicated hardware unit, but rather as a system-level capability operating collaboratively at two levels: the redundant hub units and the local display controller. Furthermore, this capability, through deep algorithmic linkage, possesses adaptive characteristics that automatically adjust according to the risk situation. The overall reliability and intelligence level of the system surpass traditional centralized solutions.

[0095] Example 11 like Figure 3 As shown, this embodiment is a verification experiment used to quantitatively verify the technical advantages of the method of the present invention over the traditional dedicated alarm control host solution in terms of trend recognition capability, hierarchical early warning accuracy, system reliability, adaptive response capability, and model switching effect.

[0096] The experimental environment was set up as a 200-meter-long subway power corridor simulation test field. A multi-parameter sensor was set up every 25 meters along the line, for a total of 8 sets of sensors. Each set contained 5 signals: carbon monoxide, hydrogen sulfide, oxygen, combustible gas and temperature.

[0097] The experiment was divided into a control group and an experimental group: the control group adopted the traditional scheme, which was configured with a single hub and a dedicated alarm control host, and the alarm threshold was a single-level fixed setting.

[0098] The experimental group adopted the scheme of this invention, configuring two master-slave redundant hub units and a local display controller, and enabling a bidirectional cross-coupling closed loop and a four-level linkage trend calculation model switching mechanism. Both groups used the exact same sensors and test conditions.

[0099] Experiment 1: Verification of trend recognition capability.

[0100] Carbon monoxide is slowly released at a constant rate in the middle section of the corridor via a gas release device, causing the concentration to increase linearly from 0 at a rate of 3 ppm per minute.

[0101] The control group only generated an alarm signal when the concentration reached the preset alarm threshold of 24 ppm, which took 8 minutes, and there were no abnormal indications during the period.

[0102] The experimental group detected a rate of change exceeding the normal fluctuation range 90 seconds after the release began, issuing a mild anomaly warning. At this point, the concentration was only 4.5 ppm, with a margin of 19.5 ppm from the absolute threshold. Compared to the control group, the experimental group issued its first warning signal 6 minutes and 30 seconds earlier, achieving an 81.25% improvement in response time for trend recognition.

[0103] Experiment 2: Verification of the accuracy of graded early warning.

[0104] Three different gas release modes were designed to simulate different risk scenarios: Mode A simulates slow accumulation with a low-rate continuous release, Mode B simulates a general leak with a medium-rate release, and Mode C simulates a severe leak with a high-rate burst release. Each mode was tested 20 times, for a total of 60 tests.

[0105] The control group found that all three modes could only output a single-level alarm without any difference when the concentration exceeded the threshold. The alarm outputs of the three modes were exactly the same, and it was impossible to distinguish the degree of risk.

[0106] In Mode A, the experimental group correctly triggered the minor anomaly warning 19 times, with one trigger delayed by two sampling cycles after error correction from the hub due to momentary interference in the signal channel; in Mode B, all 20 warnings correctly triggered the general hazard warning; and in Mode C, all 20 warnings correctly triggered the emergency hazard warning. The experimental group achieved a classification accuracy rate of 98.3%, with no cross-classification errors among the three risk levels.

[0107] Experiment 3: System Reliability Verification.

[0108] Artificially create faults during normal system operation: disconnect the power supply of one master hub, disconnect the communication link of one slave hub, and simultaneously disconnect one master hub and one slave hub.

[0109] In the control group, the system lost all early warning capabilities immediately after the power to the dedicated alarm control host was disconnected, and two simulated gas leak events that occurred within 15 minutes before the power was restored were not detected.

[0110] When one master hub was disconnected in the experimental group, the other master hub continued to independently calculate the trend criteria and forward them directly to the local display controller via the slave hub. The early warning function was not interrupted; only the confidence level was downgraded to single-channel confirmation. When the slave hub's communication link was disconnected, both master hubs sent the trend criteria to the local display controller through a backup direct connection channel. The early warning determination delay increased by about one sampling period, but the function was not interrupted. In the extreme case of simultaneously disconnecting one master hub and one slave hub, the remaining master hub maintained basic early warning capabilities through a backup direct connection channel with the local display controller. The classification accuracy was slightly reduced, but the system did not completely fail. During 72 hours of continuous testing, the experimental group experienced 12 simulated fault injections without a single instance of complete interruption of the early warning function.

[0111] Experiment 4: Verification of Adaptive Response Capability.

[0112] In the experimental group, the actual sampling cycle and reporting frequency of the system were observed under normal, slightly abnormal, general hidden danger, and emergency conditions.

[0113] Under normal conditions, the system operates stably with a 10-second sampling period and a 30-second reporting frequency, accumulating 25,920 data points per sensor over 72 hours. When the warning level rises to a minor anomaly, the system maintains the 10-second sampling period and 30-second reporting frequency, only switching from a linear fitting model to an exponentially weighted model. The actual model switch takes effect within one sampling period.

[0114] When the warning level rises to a general hidden danger, the system completes parameter adjustment within the next sampling cycle after receiving the reverse feedback signal. The sampling cycle is switched to 5 seconds and the reporting frequency is switched to 10 seconds. The actual switching response time is 8 to 12 seconds.

[0115] When the warning level is raised to an emergency, the sampling cycle is switched to 2 seconds and the reporting frequency is switched to 5 seconds. The actual switching response time is 4 to 6 seconds.

[0116] During the transition from emergency state to normal state after 5 minutes of continuous operation, the system restores normal parameters after confirming 3 consecutive normal judgment results to avoid frequent switching due to a single fluctuation.

[0117] The control group operated with a fixed 10-second sampling period and 30-second reporting frequency throughout the entire test, and had no adaptive adjustment capability regardless of changes in the risk situation.

[0118] Experiment 5: Verification of the effect of switching trend calculation models.

[0119] A deteriorating gas release pattern was designed: the carbon monoxide concentration increased at a constant rate of 2 ppm per minute for the first 3 minutes, and then at a rate of 6 ppm per minute from the 3rd to the 5th minute.2 The acceleration linearly increases the release rate, simulating a real-world scenario where the environment changes from slow to rapid deterioration.

[0120] Two experimental groups were set up for comparison: Experimental group A adopted the complete scheme of this invention, including the model switching mechanism, while experimental group B adopted the scheme of this invention but disabled the model switching function and used a linear fitting model throughout. Each configuration was tested 15 times.

[0121] Under the timeliness of detecting the inflection point of accelerated deterioration, experimental group A triggered a minor abnormality level warning and switched to the exponential weighted model when the rate of change exceeded the normal rate threshold during the slow deterioration stage. After the release rate accelerated, the rate output of the exponential weighted model quickly followed the real rate increase. Experimental group A was upgraded to the general hidden danger level after an average of 17 seconds because the instantaneous value of the parameter exceeded the first level threshold and the rate of change exceeded the limit. At the same time, the nonlinear prediction model was activated. After an average of another 12 seconds, or about 2 to 3 sampling cycles of rate difference accumulation, the acceleration feature was accurately captured. The total delay from the appearance of the inflection point to the accurate identification of the acceleration trend was an average of 29 seconds.

[0122] Because model switching was disabled and the linear fitting model was used to calculate the rate throughout the process, although the experimental group B also triggered a slight anomaly warning during the slow deterioration phase, the equal weight smoothing effect of the early stable data in the window after the inflection point caused the rate estimate to lag behind the actual rate of change. On average, it took 48 seconds for the increase in its calculation rate to significantly exceed the level of the slow deterioration phase and be judged as an acceleration trend. The delay in identifying the acceleration trend was 65.5% higher than that of experimental group A.

[0123] Regarding the lead time for emergency warnings, after the warning level was raised to general hidden danger and the nonlinear prediction model was activated in parallel, experimental group A used the acceleration term for short-term extrapolation. When the short-term predicted value exceeded the second-level threshold in the prediction time domain, the warning level was upgraded in advance. On average, an emergency warning was issued 38 seconds before the parameter actually reached the emergency threshold.

[0124] Because the linear model in experimental group B could not characterize acceleration, its ability to predict rapid changes was insufficient. On average, the emergency warning was triggered only 11 seconds before the parameters actually reached the emergency threshold, which was only 28.9% of the lead time of experimental group A.

[0125] Prediction accuracy: The nonlinear prediction model of experimental group A had an average prediction deviation of 7.2% of the actual value for the parameter values ​​after 60 seconds. The maximum deviation was 12.8% and the minimum deviation was 3.1% in 15 tests. This prediction accuracy is sufficient to support the local display controller in making reliable forward-looking early warning judgments.

[0126] False alarm rate: In experimental group A, there were no false alarms due to model switching in the normal fluctuation tests and no missed alarms in the accelerated deterioration tests, with an overall false alarm rate of 0% and a missed alarm rate of 0%. Experimental group B also had no false alarms in the 15 normal fluctuation tests, but there were two instances in the 15 accelerated deterioration tests where the warning escalation was delayed by more than 60 seconds. Although the correct level warning was eventually triggered, the optimal response window had already been missed.

[0127] Based on the above experimental results, the multi-parameter redundant safety early warning control method and system for subway power corridors proposed in this invention significantly outperforms traditional schemes relying on dedicated alarm control hosts in terms of trend recognition lead time, hierarchical early warning accuracy, system fault resistance, adaptive adjustment speed of data acquisition density, and dynamic switching effect of trend calculation models. In particular, the introduction of a model switching mechanism deepens cross-coupling from the parameter level to the algorithm level, improving early warning lead time and prediction accuracy when facing rapidly deteriorating risks. This verifies the technical feasibility and superior performance of the cross-coupling architecture in replacing dedicated hardware.

[0128] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-parameter redundancy safety early warning control method for subway power corridors, characterized in that, include: The dual main hubs receive the environmental signals continuously output by the multi-parameter sensors in the corridor in parallel, each independently maintains the parameter sequence within the sliding time window, and calculates the parameter change rate based on the currently effective model among the preset multiple trend calculation models. The hub performs a consistency check on the current frame value and change rate output by the dual main hubs. If the check passes, it outputs a valid trend criterion that has been redundantly confirmed. If the check fails, it initiates data backtracking and majority voting error correction. After receiving the effective trend criteria, the local display controller substitutes the instantaneous value of the parameter and the rate of change into the multi-level safety threshold model for fusion judgment. Based on the combination relationship between the degree of instantaneous value exceeding the limit and the degree of rate exceeding the limit, it distinguishes the multi-level risk status and drives the corresponding level of graded early warning output. The dual master hubs and the slave hubs together constitute a hub layer. The output results of the hierarchical early warning are fed back to the hub layer. The hub layer dynamically adjusts the length of the sliding time window and the data reporting frequency according to the early warning level, and synchronously switches the currently effective trend calculation model, forming a bidirectional cross-coupled closed loop between the redundancy verification layer and the hierarchical early warning layer.

2. The method according to claim 1, characterized in that, The multi-parameter sensor includes a carbon monoxide sensor, a hydrogen sulfide sensor, an oxygen sensor, a combustible gas sensor, and a temperature sensor. Each sensor is deployed at key points along the power corridor, converting environmental parameters into standard electrical signals and continuously outputting them to the dual master hub and the slave hub.

3. The method according to claim 1, characterized in that, The dual hubs synchronously receive all sensor signals in a parallel hot standby mode and analyze them independently. Each hub maintains a configurable sliding time window locally, sampling each parameter periodically within the window to form a continuous time series, and calculating the rate of change of each parameter based on the series according to the currently effective trend calculation model. The various trend calculation models include a linear fitting model, an exponential weighted model, and a nonlinear prediction model containing an acceleration term. The dual hubs always synchronously use the same trend calculation model.

4. The method according to claim 3, characterized in that, When the slave hub performs consistency verification, the rate of change of the same parameter output by the dual master hubs is compared for deviation; if the deviation is within the preset allowable range, the average of the two rate values ​​is taken as the effective trend criterion and marked as a high confidence level. If the deviation exceeds the allowable range, the original sampling data for the corresponding time period is traced back, the valid value is determined by majority voting, and the rate of change is recalculated and marked as the trend criterion confirmed by error correction.

5. The method according to claim 4, characterized in that, The effective trend criterion includes the current instantaneous value of each parameter, the rate of change, the confidence level, and the current computational model identifier. When the nonlinear prediction model is enabled, the effective trend criterion also includes change acceleration information and short-term prediction values. The above information is output as a complete data packet from the hub to the local display controller.

6. The method according to claim 1, characterized in that, The multi-level safety threshold model sets two sets of criteria: a parameter instantaneous value threshold sequence and a change rate threshold sequence. When the parameter instantaneous value does not exceed the limit but the change rate exceeds the corresponding rate threshold, or when the parameter instantaneous value exceeds the first-level threshold but the change rate does not exceed the limit, a minor anomaly level warning is triggered. When the instantaneous value of a parameter exceeds the first-level threshold and the rate of change also exceeds the limit, a general hazard level warning is triggered. When the instantaneous value of a parameter exceeds the second-level threshold or the rate of change exceeds the emergency rate threshold, an emergency hazard level warning is triggered. When the effective trend criterion includes a short-term predicted value and the predicted value exceeds the second-level threshold within the prediction time domain, an emergency risk level warning is also triggered.

7. The method according to claim 6, characterized in that, The local display controller drives differentiated audible and visual alarm responses based on the multi-level risk status, and synchronizes the warning level information to the remote cloud platform and the subway BAS system via a serial gateway and a multi-protocol parallel transmission channel.

8. The method according to claim 1, characterized in that, The reverse feedback adjustment strategy includes two dimensions: quantitative adjustment of the collected parameters and qualitative switching of the trend calculation model. It is implemented in four coordinated levels according to the warning level: When the warning level is normal, the hub layer maintains the regular sampling period and standard reporting frequency, and uses a linear fitting model to perform equal-weighted linear regression on the sampled sequence within the window to output the rate of change; when the warning level is slightly abnormal, the hub layer maintains the regular sampling period, standard reporting frequency, and sliding time window length unchanged, and switches to an exponentially weighted model to assign a time-decreasing weight coefficient to each sample point within the window to calculate the rate of change; when the warning level rises to a general hidden danger, the sliding time window length is shortened and the reporting frequency is increased, the exponentially weighted model continues to be used to calculate the rate of change, and a nonlinear prediction model containing an acceleration term is simultaneously activated to extract the acceleration of change and extrapolate and predict the short-term trend of the parameters; when the warning level rises to an emergency danger, it switches to the highest frequency acquisition mode, using the nonlinear prediction model containing an acceleration term as the currently effective model to output the rate of change, acceleration of change, and short-term predicted values.

9. The method according to claim 1, characterized in that, The outputs of the various trend calculation models are all normalized to the rate of change under a unified physical dimension and are calibrated for factory consistency. The first three sampling periods after the trend calculation model is switched is a transition period. During the transition period, the dual hubs weight and fuse the rate outputs of the old and new models according to a linearly changing weight. The weight changes from 100% of the old model to 100% of the new model cycle by cycle. For the first occurrence of rate of change exceeding the limit during the transition period, the local display controller needs to confirm for two consecutive sampling periods before upgrading the warning level. Rate of change exceeding the limit outside the transition period is judged immediately. The local display controller generates dynamic change curves of each parameter in real time based on the effective trend criteria and retains historical trend data. The dynamic change curves, early warning status and fault information are synchronized to the remote cloud platform and the subway BAS system through a multi-protocol parallel method to achieve full-domain information sharing.

10. A multi-parameter redundancy safety early warning system for subway power corridors implementing the method of any one of claims 1 to 9, characterized in that, include: A multi-parameter sensor array is deployed at key locations along the power corridor to convert environmental parameters such as carbon monoxide, hydrogen sulfide, oxygen, combustible gas, and temperature into standard electrical signals and continuously output them. The dual-master-one-slave redundant hub unit includes two parallel hot-standby master hubs and one slave hub. Each of the two master hubs independently receives signals from the multi-parameter sensor group and maintains a sliding time window parameter sequence. Each master hub has a pre-set linear fitting model, an exponentially weighted model, and a nonlinear prediction model containing an acceleration term. It calculates the rate of change based on the currently active model. The slave hub synchronously receives signals from the multi-parameter sensor group and records one copy of the sensor data. It is used to perform consistency checks and majority voting error correction on the values ​​and rates of change output by the dual master hubs and output a valid trend criterion confirmed by redundancy. It also undertakes the unified distribution of model switching instructions. The local display controller has a built-in multi-level safety threshold model, which is used to receive the effective trend criteria and fuse the instantaneous value of the parameter with the rate of change to distinguish the multi-level risk status, drive the corresponding level of audible and visual alarm, and send the graded early warning results back to the redundant hub unit to dynamically adjust the sliding time window length, data reporting frequency and trend calculation model selection. A multi-terminal data synchronization channel, including a serial port gateway and a multi-protocol transmission link, is used to synchronize real-time parameters, dynamic change curves, early warning status and fault information to a remote cloud platform and the metro BAS system. The redundant hub unit and the local display controller form a bidirectional cross-coupled closed loop. The redundant hub unit outputs trend criteria to the local display controller, and the local display controller feeds back the warning level to the redundant hub unit to drive adaptive adjustment of the acquisition density and dynamic switching of the trend calculation model.

Images