Carbon emission monitoring and early warning analysis system based on edge calculation

Abstract

The invention relates to the technical field of carbon emission monitoring, and discloses a carbon emission monitoring and early warning analysis system based on edge calculation. The system comprises a dynamic topology mapping module which is used for establishing dynamic topology mapping of an emission source in a monitoring area and recording a state transition path; the self-adaptive probe deployment module is used for deploying a probe associated with the monitoring dimension and the path according to the mapping at the edge node; the early warning identification module is used for comparing the time slice sequence with a path identification early warning point; the evolution trajectory generation module is used for generating a parameter evolution trajectory; the stage analysis and traceability module is used for distinguishing a steady state stage from a transient stage and executing inertia and sudden change analysis; and the autonomous optimization module is used for constructing a double-layer parameter correction model and driving the monitoring process to be optimized.

Description

Technical Field

[0001] This invention relates to the field of carbon emission monitoring technology, specifically to a carbon emission monitoring, early warning and analysis system based on edge computing. Background Technology

[0002] Existing carbon emission monitoring systems largely rely on fixedly deployed sensors and centralized data processing architectures within a region to collect the identity information and spatial location of each emission source, and then complete status assessment and early warning on a central platform. Conventional practices often treat emission sources as static nodes, recording only their location and basic attributes, failing to form a dynamic structure that reflects changes in operating modes, and making it difficult to capture the sequence and conditions of transitions between different operating conditions. When emission sources switch between multiple operating states, existing systems lack continuous tracking and representation of state change paths, resulting in limited ability to predict potential abnormal shifts.

[0003] In the deployment and data acquisition phases of monitoring probes, conventional systems generally adopt a uniform dimension and fixed frequency acquisition strategy. All nodes operate with the same parameter set and sampling period, ignoring the differentiated requirements of different emission sources for monitoring dimensions during state transitions. Furthermore, acquisition tasks are often scheduled by cloud or remote servers, requiring long-distance data transmission for processing, increasing transmission load and response latency. For application scenarios requiring the capture of subtle parameter changes before and after state transitions, existing solutions lack a dimension adjustment mechanism associated with the state transition path, potentially missing crucial transient information and generating redundant data in unnecessary stages. This application aims to address the issue of dynamically characterizing the emission source state transition logic, enabling the monitoring dimensions of probes deployed at edge nodes to match changes in the state transition path, thus making the monitoring process more closely reflect the actual operating characteristics of the emission source. Summary of the Invention

[0004] The purpose of this invention is to provide a carbon emission monitoring, early warning and analysis system based on edge computing to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a carbon emission monitoring, early warning, and analysis system based on edge computing, the system comprising: The dynamic topology mapping module is used to establish dynamic topology mappings of different carbon emission sources within the monitoring area, and to record the identity attributes, spatial location and state transition path of each emission source. An adaptive probe deployment module is used to deploy adaptive monitoring probes on each edge computing node according to the dynamic topology mapping, and adjust the monitoring dimension sampling strategy of the adaptive monitoring probes according to the state transition path of the carbon emission source. The early warning identification module is used to capture the time slice sequence of the emission flow through the adaptive monitoring probe, and compare the time slice sequence with the state transition path to identify potential state switching early warning points; The evolution trajectory generation module is used to generate parameter evolution trajectories within the edge computing node based on the state switching early warning point, so as to describe the continuous change process of the monitoring parameters before and after the state switching. The phase analysis and tracing module is used to extract the steady-state phase and transient phase in the parameter evolution trajectory, perform parameter inertial learning in the steady-state phase, and perform parameter mutation tracing in the transient phase. The autonomous optimization module is used to integrate the results of parameter inertia learning and the results of parameter mutation tracing to construct a two-layer parameter correction model on the edge side, and use the two-layer parameter correction model to drive the autonomous optimization of the subsequent carbon emission monitoring process.

[0006] Preferably, establishing a dynamic topological mapping of different carbon emission sources within the monitoring area specifically includes: The static identification features and dynamic behavioral features of each carbon emission source are collected. The static identification features include the equipment code and its category, and the dynamic behavioral features include the operating cycle and emission intensity pattern. The static identification features and dynamic behavioral features of each emission source are fused and encoded to generate topological nodes with spatiotemporal attributes. The correlation between different topological nodes on the time axis is analyzed. Based on the causality and temporal proximity of emission events, state transition paths between nodes are drawn. The state transition paths are marked with source state, target state and triggering conditions. All topological nodes and their corresponding state transition paths are integrated to form the dynamic topological mapping.

[0007] Preferably, the step of deploying adaptive monitoring probes on each edge computing node according to the dynamic topology mapping specifically includes: The monitoring dimensions associated with each topology node in the dynamic topology mapping are analyzed, including concentration gradient, flow fluctuation and temperature field distribution; Based on the monitoring dimensions, a corresponding sensing unit group and signal conditioning unit are configured for each edge computing node to form the basis of the hardware probe. A logic control unit is embedded on the hardware probe. The logic control unit loads the state transition path and dynamically adjusts the sampling strategy of the sensing unit group according to the preset trigger conditions in the path. The combination of the hardware probe base and the logic control unit is defined as the adaptive monitoring probe, and it is bound to the corresponding edge computing node.

[0008] Preferably, the step of capturing time-slice sequences of the emission stream using the adaptive monitoring probe specifically includes: The logic control unit in the adaptive monitoring probe obtains the corresponding sampling frequency and sampling duration from the loaded state transition path according to the current state of the carbon emission source; The control sensor unit group synchronously collects concentration gradient, flow pulsation and temperature field distribution according to the sampling frequency and sampling duration to obtain raw data blocks within a complete monitoring cycle. The original data block is divided into segments according to a fixed time window to generate a series of data segments with time-series labels, and each data segment is a time slice; The consecutive time slices are arranged in chronological order to form the time slice sequence.

[0009] Preferably, the step of comparing the time slice sequence with the state transition path to identify potential state transition warning points specifically includes: Extract the characteristic patterns that should be exhibited in the monitoring dimension when all source states transition to the target state from the state transition path, and construct a characteristic pattern library; The data of each time slice in the time slice sequence is matched point by point with the feature patterns in the feature pattern library, and the matching confidence is calculated. If the matching confidence of several consecutive time slices shows a trend of convergence toward a certain target state feature pattern, and the trend strength exceeds a preset convergence threshold, then it is determined that there is a potential state switch within the time window corresponding to the several consecutive time slices. The starting time point at which a state transition is determined is marked as the state transition warning point.

[0010] Preferably, the step of generating a parameter evolution trajectory within the edge computing node based on the state switching early warning point specifically includes: Using the aforementioned state switching warning point as a time reference, a fixed-length time period is traced back as the starting point of the evolution, and a fixed-length time period is extended forward as the ending point of the evolution. Extract parameter values ​​for concentration gradient, flow fluctuation, and temperature field distribution from all time slice sequences from the start to the end of the evolution; Plot the continuous change curve of each monitoring dimension from the starting point to the end point of evolution with time as the horizontal axis and the parameter values ​​of each monitoring dimension as the vertical axis. The change curves of all monitoring dimensions are superimposed and aligned to form a comprehensive map of the evolution of multidimensional parameters over time, which is the trajectory of the parameter evolution.

[0011] Preferably, the extraction of the steady-state and transient phases from the parameter evolution trajectory specifically includes: Smoothness analysis is performed on the change curve of each monitoring dimension in the parameter evolution trajectory, and the first derivative variance of the curve within the sliding time window is calculated. If the variance of the first derivative remains below the steady-state threshold, the curve segment corresponding to the sliding time window is determined to be in the steady-state stage. If the variance of the first derivative increases sharply in a short period of time and exceeds the transient threshold, then the curve segment near the time point corresponding to the sharp increase in the variance of the first derivative is determined to be the transient stage. Record the start and end timestamps of all steady-state phases and the center timestamps of all transient phases.

[0012] Preferably, performing parametric inertial learning during the steady-state phase specifically includes: For each steady-state stage, extract the set of all monitored dimension parameter values ​​within that steady-state stage; Calculate the statistical characteristics of the parameter values ​​for each monitoring dimension in the set, wherein the statistical characteristics include the mean, median, and standard deviation; Using the aforementioned statistical characteristics, an inertial model is established to describe the normal fluctuation range of parameters within the steady-state phase. The inertial model consists of parameter baseline values ​​and allowable deviation bands. Multiple inertial models learned at different steady-state stages are archived to form a parametric inertial knowledge base describing the normal behavior of carbon emission sources in the monitoring area under various steady-state conditions.

[0013] Preferably, the step of performing parameter mutation tracing during the transient phase specifically includes: For each transient phase, its central time marker is located, and detailed data of the parameter evolution trajectory within a period before and after the central time marker are extracted; The analysis of the detailed data reveals the order and magnitude of abrupt changes in parameters across different monitoring dimensions. The order and magnitude ratio are reverse-matched with the state transition paths recorded in the dynamic topology mapping to find the source state and target state combination most likely to trigger this parameter mutation. Record the information on the successfully matched state transition path, as well as the specific manifestation pattern of this parameter mutation in each monitoring dimension, to form a complete mutation source tracing record.

[0014] Preferably, the fusion of the results of parameter inertial learning and the results of parameter mutation tracing constructs a two-layer parameter correction model on the edge side, specifically including: The first layer of the model is the steady-state correction layer. The steady-state correction layer calls the inertial model in the parameter inertial knowledge base to perform consistency verification on the data in the steady-state stage during real-time monitoring. If the data deviates from the allowable deviation band defined by the inertial model, correction suggestions based on historical steady-state behavior are generated. The second layer model is the transient correction layer. The transient correction layer calls the state transition information and parameter behavior patterns in the mutation source record. When a new transient event is detected, it is matched with the historical record. If the match is successful, the parameter change path verified in the historical record is used for predictive correction. The outputs of the steady-state correction layer and the transient correction layer are prioritized, arbitrated, and their results are fused to generate the final control commands applied to the adjustment of monitoring parameters, thereby completing the autonomous optimization of the carbon emission monitoring process.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By recording the identity attributes, spatial location, and state transition paths of emission sources and constructing a dynamic topology mapping, the structure of the monitoring area can be updated as the operating mode of the emission sources changes. The system directly holds information on the order and conditions of state transitions, enabling it to identify impending state changes locally at edge nodes. This allows for early focus on potentially changing segments without relying on centralized analysis at remote locations, thus obtaining continuous changes in relevant parameters before the state transition fully manifests. This makes the monitoring behavior closer to the actual dynamic behavior logic of the emission sources.

[0016] Adaptive monitoring probes are deployed on edge computing nodes based on dynamic topology mapping, and the monitoring dimensions of the probes are correlated with the state transition paths of the emission source. This allows the probes to automatically adjust the types of physical quantities collected and the sampling rhythm at different state stages. This enables the probes to have higher dimensionality and density of data collection when the emission source enters a state requiring detailed observation, and to reduce unnecessary data volume when it enters a stable state. Data acquisition is completed directly at the edge side, matching the path, reducing the scale and latency of data transmission across the network. This ensures that the monitoring process can acquire sufficient information to characterize continuous parameter changes both before and after state transitions. Attached Figure Description

[0017] Figure 1 This is a timing diagram of the carbon emission monitoring, early warning and analysis system based on edge computing described in this invention; Figure 2 A flowchart for establishing dynamic topology mapping; Figure 3 Example diagram of encoding for dynamic topology nodes; Figure 4 A flowchart for capturing time-slice sequences; Figure 5 A phased analysis chart of the comprehensive trend of carbon emission monitoring parameters; Figure 6 This is a graph showing the trend of boiler flue gas oxygen content and load changes throughout the entire process. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 This invention provides a carbon emission monitoring and early warning analysis system based on edge computing. The system includes: a dynamic topology mapping module that establishes a dynamic topology mapping of different carbon emission sources within a monitoring area, recording the identity attributes, spatial location, and state transition path of each emission source; an adaptive probe deployment module that deploys adaptive monitoring probes on each edge computing node according to the dynamic topology mapping, and adjusts the monitoring dimension sampling strategy of the adaptive monitoring probes according to the state transition path of the carbon emission source; an early warning identification module that captures time slice sequences of emission flows through adaptive monitoring probes and compares the time slice sequences with the state transition paths to identify potential state switching early warning points; an evolution trajectory generation module that generates parameter evolution trajectories within the edge computing nodes based on the state switching early warning points to describe the continuous change process of monitoring parameters before and after the state switch; a stage analysis and source tracing module that extracts the steady-state stage and transient stage from the parameter evolution trajectory, performs parameter inertia learning in the steady-state stage, and performs parameter mutation source tracing in the transient stage; and an autonomous optimization module that integrates the results of parameter inertia learning and parameter mutation source tracing to construct a two-layer parameter correction model on the edge side and uses this model to drive the autonomous optimization of the subsequent carbon emission monitoring process.

[0020] In one embodiment of the present invention, see [reference] Figure 2The dynamic topology mapping module collects static identification features and dynamic behavioral features of each carbon emission source. Static identification features include equipment codes and categories, while dynamic behavioral features include operating cycles and emission intensity patterns. The module fuses and encodes the static identification features and dynamic behavioral features of each emission source to generate topology nodes with spatiotemporal attributes. It analyzes the correlation between different topology nodes on the time axis and draws state transition paths between nodes based on the causality and temporal proximity of emission events. These state transition paths are labeled with source states, target states, and triggering conditions. All topology nodes and their corresponding state transition paths are integrated to form a dynamic topology map. The adaptive probe deployment module parses the monitoring dimensions associated with each topology node in the dynamic topology map. These monitoring dimensions include at least concentration gradients, flow pulsations, and temperature field distribution. Based on the monitoring dimensions, it configures corresponding sensor unit groups and signal conditioning units for each edge computing node to form a hardware probe base. A logic control unit is embedded on top of the hardware probe. The logic control unit loads the state transition paths and dynamically adjusts the sampling strategy of the sensor unit groups according to the preset triggering conditions in the paths. The combination of the hardware probe base and the logic control unit is defined as an adaptive monitoring probe and bound to the corresponding edge computing node.

[0021] In a specific implementation, the dynamic topology mapping module and the adaptive probe deployment module work together on the edge computing nodes. The input of the module is the raw characteristic data of carbon emission sources in the monitoring area, and the output is the adaptive monitoring probes deployed on each edge computing node and bound with specific state transition paths.

[0022] See Figure 3In practical implementation, the dynamic topology mapping module collects the static identification features and dynamic behavioral features of each carbon emission source. The static identification features include the equipment code "F-001" of the combustion furnace and its category "stationary combustion source". The dynamic behavioral features include the operating cycle of the steam turbine "once a day" and the emission intensity mode "sharp concentration increase during startup and gradual concentration change during stable operation". In practical implementation, the static identification features and dynamic behavioral features of each emission source are fused and encoded to generate a topology node with spatiotemporal attributes. For example, for a boiler with equipment code "B-005", after fusing its category "fuel boiler", operating cycle "continuous operation", and emission intensity mode "three-stage fluctuation based on load", a topology node identifier "N_B-005_CF_TripleMode" is generated. In practice, the correlation between different topological nodes on the time axis is analyzed. Based on the causality and temporal proximity of emission events, state transition paths between nodes are drawn. For example, the analysis found that the combustion furnace ignition state represented by the topological node "N_F-001_FC_Startup" always occurs before the steam turbine load increase state represented by the topological node "N_ST-002_Turbine_LoadRamp", and the time interval is within the range of 5 to 10 minutes. Based on this, a state transition path from the source state "ignition" to the target state "load increase" is drawn. The trigger condition marked on this path is "flue gas temperature exceeds 400 degrees Celsius and lasts for 120 seconds".

[0023] In some embodiments, the adaptive probe deployment module parses the monitoring dimensions associated with each topology node in the dynamic topology map. These monitoring dimensions include concentration gradient, flow pulsation, and temperature field distribution. In some embodiments, a corresponding sensing unit group and signal conditioning unit are configured for each edge computing node according to the monitoring dimensions, forming the hardware probe foundation. For example, for a topology node associated with "concentration gradient" and "temperature field distribution," an infrared gas sensor array and a thermocouple array are configured as sensing unit groups on the corresponding edge computing node, and a multi-channel analog-to-digital converter is configured as a signal conditioning unit. It can be understood that a logic control unit is embedded on the hardware probe foundation. The logic control unit loads the state transition path associated with that node in the dynamic topology map and dynamically adjusts the sampling strategy of the sensing unit group according to the preset trigger conditions in the path.

[0024] To illustrate how the logic control unit adjusts the sampling strategy based on the state transition path, an operational relationship is introduced. When the state transition path indicates a target state of "high load operation," the logic control unit calculates and sets a new sampling frequency based on the parameter change characteristics recorded in the path. This process can be represented by the following formula:

[0025] in: The sampling frequency set by the logic control unit for the sensor unit group. It depends on the adjustment coefficient of the monitoring dimension type. It is the magnitude of the reference parameter change recorded in the state transition path. It is the standard state transition time length recorded in the state transition path. It is a compensation factor that characterizes the stability of the current monitoring environment.

[0026] In one embodiment of the present invention, see [reference] Figure 4 The logic control unit in the adaptive monitoring probe obtains the corresponding sampling frequency and sampling duration from the loaded state transition path according to the current state of the carbon emission source. The control sensor unit group synchronously collects the concentration gradient, flow pulsation and temperature field distribution according to the sampling frequency and sampling duration to obtain the raw data block within a complete monitoring cycle. The raw data block is divided into a series of data segments with time sequence markings according to a fixed time window. Each data segment is a time slice. The continuous time slices are arranged in time order to form a time slice sequence.

[0027] In practical implementation, the logic control unit in the adaptive monitoring probe obtains the corresponding sampling frequency and sampling duration from the loaded state transition path based on the current state of the carbon emission source. For example, if the logic control unit detects that the bound industrial reactor is in the "catalytic pretreatment" state, it queries the state transition path to find the preparatory stage for switching to the "main reaction" state, requiring a sampling frequency of 10 Hz and a sampling duration of 180 seconds. In practice, the logic control unit controls the sensor unit group to synchronously collect concentration gradients, flow pulsations, and temperature field distributions according to the acquired sampling frequency and sampling duration, obtaining a raw data block for a complete monitoring cycle. Taking the industrial reactor numbered "R-007" as an example, its configured sensor unit group synchronously collects data at a frequency of 10 Hz within 180 seconds, ultimately generating a raw data block containing 5400 concentration gradient readings, 5400 flow pulsation readings, and 5400 temperature field distribution readings.

[0028] In some embodiments, the original data block is divided into fixed time windows to generate a series of data segments with time-series labels, each data segment being a time slice. In some embodiments, the length of the fixed time window is dynamically set by the logic control unit based on the parameter change characteristics marked in the state transition path, rather than a fixed value. It can be understood that consecutive time slices are arranged in chronological order to form a time slice sequence, which is stored in the buffer of the edge computing node as a linked list or array structure. Each time slice element contains its start and end timestamps and the data array for each monitoring dimension. Optionally, a sequence identifier is appended to the time slice sequence after generation for association with the state transition path early warning analysis task.

[0029] To illustrate the dynamic setting logic of the fixed time window length, an operational relationship in practice is introduced. The logic control unit determines the fixed time window length for segmenting the original data blocks based on the pre-stored characteristic pattern time scales in the state transition path. This process can be represented by the following formula:

[0030] in: This represents the fixed time window length set by the logic control unit. It is the baseline window length coefficient related to the number of monitoring dimensions. It is the typical time span required to complete one feature pattern evolution as described in the state transition path. This is a dispersion index of parameter values ​​within the past monitoring period, calculated by the logic control unit based on the current real-time data stream. The reference window length coefficient is preset by the logic control unit during initialization based on the time granularity of the parameter characteristic patterns marked in the loaded state transition path, and in combination with the sampling frequency range supported by the sensor unit group. According to this relationship, when the state transition path indicates that the characteristic patterns evolve rapidly or the real-time data dispersion is high, the logic control unit automatically shortens the fixed time window length to generate more refined time slices; conversely, it extends the fixed time window length to generate more statistically representative time slices. In specific implementation, after generating the time slice sequence, the logic control unit calculates a set of basic statistics for each time slice in the sequence. The basic statistics include the average value of the concentration gradient readings, the range of the flow pulsation readings, and the spatial standard deviation of the temperature field distribution readings within that time window.

[0031] In one embodiment of the present invention, the early warning identification module extracts the feature patterns that should be exhibited in the monitoring dimension when the source state transitions to the target state from the state transition path to construct a feature pattern library. The data of each time slice in the time slice sequence is matched point by point with the feature patterns in the feature pattern library and the matching confidence is calculated. If the matching confidence of several consecutive time slices shows a trend of convergence towards a certain target state feature pattern and the trend strength exceeds the preset convergence threshold, it is determined that there is a potential state switch within the time window corresponding to several consecutive time slices. The starting time point at which the state switch is determined to exist is marked as the state switch early warning point. The evolution trajectory generation module uses the state switching warning point as the time base to trace back a fixed-length time period as the evolution starting point and extend it forward a fixed-length time period as the evolution ending point. It extracts the parameter values ​​of concentration gradient, flow pulsation and temperature field distribution from all time slice sequences between the evolution starting point and the evolution ending point. It plots the continuous change curve of each monitoring dimension from the evolution starting point to the evolution ending point with time as the horizontal axis and the parameter values ​​of each monitoring dimension as the vertical axis. The change curves of all monitoring dimensions are superimposed and aligned to form a comprehensive map of the evolution of multi-dimensional parameters over time. This comprehensive map is the parameter evolution trajectory.

[0032] In a specific implementation, an industrial reactor under operational monitoring is used as a concrete example. The preset states of the industrial reactor include "preheating", "feeding" and "main reaction". The input of the early warning identification module is the generated time slice sequence, and the output is the state switching early warning point; the input of the evolution trajectory generation module is the state switching early warning point and the associated time slice sequence data, and the output is the parameter evolution trajectory.

[0033] In practical implementation, the early warning identification module extracts the characteristic patterns that should be exhibited in the monitoring dimensions when transitioning from the source state to the target state from the state transition path, and constructs a characteristic pattern library. For example, for the transition from the "preheating" state to the "feeding" state, the predefined characteristic patterns in the state transition path are: within 60 seconds, the concentration gradient monitoring value needs to rise from 20-30 ppm to 80-100 ppm, the flow pulsation frequency needs to increase from below 1 Hz to 2-3 Hz, and the standard deviation of the temperature field distribution in the middle of the reactor needs to expand from less than 2 degrees Celsius to 5-8 degrees Celsius. This set of constraints with threshold ranges and time-series relationships is integrated to form a characteristic pattern named "pattern_preheating_feeding" and stored in the characteristic pattern library. In practical implementation, the data of each time slice in the time slice sequence is matched point by point with the characteristic patterns in the characteristic pattern library, and the matching confidence is calculated. The matching process compares the average value and trend of each monitoring dimension parameter within the time slice with the threshold range and direction of change defined by the characteristic pattern.

[0034] In some embodiments, calculating the match confidence involves a quantitative evaluation operation. The match confidence is calculated using the following formula:

[0035] in: This represents the confidence level of the match between the current time slice data and a specific feature pattern, with a value range of [0,1]. and These are weighting coefficients, satisfying... . It is the Pearson correlation coefficient between the time slice data sequence and the preset trend of the feature pattern. It is the Mahalanobis distance between the feature vectors and feature pattern center vectors of the time slice data in the multidimensional parameter space. This is a scaling parameter used to normalize Mahalanobis distance. The scaling parameter is the trace of the overall covariance matrix of the dataset formed by the center vectors of all feature patterns in the feature pattern library, calculated when constructing the feature pattern library. For a given time slice, the matching confidence score needs to be calculated for each feature pattern in the feature pattern library. In some embodiments, if the matching confidence scores of several consecutive time slices show a trend of convergence towards a certain target state feature pattern, and the trend strength exceeds a preset convergence threshold, then it is determined that there is a potential state transition within the time window corresponding to the several consecutive time slices. For example, when monitoring an industrial reactor, the matching confidence scores for "pattern_preheat_feed" for five consecutive time slices are 0.35, 0.52, 0.68, 0.79, and 0.85, respectively, showing a monotonically increasing trend. Furthermore, the slope value of 0.125 obtained using linear fitting exceeds the preset convergence threshold of 0.08. Therefore, it is determined that there is a potential process of switching from the "preheat" state to the "feed" state within this 50-second time window.

[0036] In practical implementation, the evolution trajectory generation module uses the state transition warning point as the time base, tracing back a fixed-length time period as the evolution starting point and extending forward a fixed-length time period as the evolution ending point. The fixed length is usually set according to the duration of typical transition processes in the state transition path. For example, for the transition from "feed" to "main reaction," the tracing and extension durations are each set to 300 seconds. In practical implementation, parameter values ​​related to concentration gradient, flow pulsation, and temperature field distribution are extracted from all time slice sequences from the evolution starting point to the evolution ending point. Optionally, the extracted parameter values ​​are statistically representative values ​​of each monitoring dimension within each time slice, such as the median. This can be understood as plotting a continuous change curve for each monitoring dimension from the evolution starting point to the evolution ending point, with time as the horizontal axis and the parameter values ​​of each monitoring dimension as the vertical axis. For example, a time axis in seconds can be generated, plotting the curves of concentration gradient change over time, flow pulsation frequency change over time, and reactor axial temperature difference change over time from 300 seconds before the warning point to 300 seconds after the warning point. Optionally, the change curves of all monitoring dimensions can be superimposed and aligned to form a comprehensive map of the evolution of multi-dimensional parameters over time, which is the parameter evolution trajectory.

[0037] In one embodiment of the present invention, the stage analysis and tracing module performs smoothness analysis on the change curve of each monitoring dimension in the parameter evolution trajectory, calculates the first derivative variance of the curve within the sliding time window, and determines that the curve segment corresponding to the sliding time window is a steady-state stage if the first derivative variance is continuously lower than the steady-state threshold. If the first derivative variance rises sharply in a short period of time and exceeds the transient threshold, the curve segment near the time point corresponding to the sharp rise in the first derivative variance is determined to be a transient stage. The start and end time markers of all steady-state stages and the center time marker of all transient stages are recorded. For each steady-state stage, a set of parameter values ​​of all monitoring dimensions within the steady-state stage is extracted, and the statistical characteristics of the parameter values ​​of each monitoring dimension in the set are calculated. The statistical characteristics include the mean, median, and standard deviation. An inertial model describing the normal fluctuation range of parameters within the steady-state stage is established using the statistical characteristics. The inertial model consists of parameter baseline values ​​and allowable deviation bands. Multiple inertial models learned from different steady-state stages are archived to form a parameter inertial knowledge base describing the normal behavior of carbon emission sources in the monitoring area under various steady states. For each transient phase, the central time marker is located and detailed data of parameter evolution trajectory within a period before and after the central time marker are extracted. The order and magnitude of parameter mutations in different monitoring dimensions in the detailed data are analyzed. The order and magnitude are then reverse-matched with the state transition paths recorded in the dynamic topology mapping to find the source and target state combinations that triggered the parameter mutation. The information of the successfully matched state transition paths and the specific manifestation patterns of the parameter mutation in each monitoring dimension are recorded to form a complete mutation source tracing record.

[0038] In practical implementation, the stage analysis and tracing module performs smoothness analysis on the change curves of each monitoring dimension in the parameter evolution trajectory and calculates the variance of the first derivative of the curve within the sliding time window. For example, for the concentration gradient change curve, the sliding time window length is set to 30 seconds, and the variance of the first derivative of the concentration gradient with respect to time within each window is calculated. In some embodiments, if the variance of the first derivative is consistently lower than the steady-state threshold, the curve segment corresponding to the sliding time window is determined to be a steady-state stage. The steady-state threshold is set according to the historical noise level of the monitoring dimension. For concentration gradient monitoring of industrial reactors, the steady-state threshold is set to 0.5. In some embodiments, if the variance of the first derivative rises sharply in a short period of time and exceeds the transient threshold, the curve segment near the time point corresponding to the sharp rise in the variance of the first derivative is determined to be a transient stage. The transient threshold is set to 10 times the steady-state threshold, i.e., 5. It can be understood that the start and end timestamps of all steady-state stages and the center timestamps of all transient stages are recorded. The start and end timestamps are recorded in timestamp format, and the center timestamps are recorded at the time corresponding to the peak variance. See Table 1.

[0039] Table 1: Examples of Determining Steady-State and Transient Phases

[0040] In practical implementation, for each steady-state stage, a set of all monitored dimension parameter values ​​is extracted within that stage. Taking the steady-state stage from the beginning of 1520 seconds to the end of 1550 seconds in Table 1 as an example, the specific values ​​of concentration gradient, flow pulsation frequency, and axial temperature difference for all time slices within this 30-second period are extracted to form a parameter value set. In practical implementation, the statistical characteristics of each monitored dimension parameter value in the set are calculated. These statistical characteristics include the mean, median, and standard deviation. For example, the calculated mean of the concentration gradient for this steady-state stage is 125.6 ppm, the median is 124.8 ppm, and the standard deviation is 3.2 ppm. It can be understood that, using these statistical characteristics, an inertial model describing the normal fluctuation range of parameters within the steady-state stage is established. This inertial model consists of the parameter baseline value and the allowable deviation band. The parameter baseline value is usually the statistical mean or median, and the allowable deviation band is usually defined as the range of k times the standard deviation above and below the baseline value. The coefficient k is set according to the monitoring accuracy requirements, for example, k=2. For the aforementioned steady-state phase, the established concentration gradient inertial model is as follows: baseline value 125.6 ppm, allowable deviation band [119.2 ppm, 132.0 ppm]. Optionally, multiple inertial models learned from different steady-state phases can be archived to form a parametric inertial knowledge base describing the normal behavior of carbon emission sources within the monitoring area under various steady-state conditions. The parametric inertial knowledge base is stored in the form of database tables, with each record associated with a steady-state phase identifier, monitoring dimension name, parameter baseline value, and upper and lower bounds of the allowable deviation band.

[0041] In practical implementation, for each transient phase, its central time marker is located, and detailed data of the parameter evolution trajectory within a certain period before and after the central time marker are extracted. For example, for a transient phase with a central time of 1595 seconds, the original parameter evolution trajectory data stored at a higher time resolution within the interval from 1580 seconds to 1610 seconds is extracted as detailed data. In practical implementation, the order and magnitude ratio of parameter mutations in different monitoring dimensions in the detailed data are analyzed. For example, by comparing the times when parameters in the detailed data exceed twice the standard deviation of their respective steady-state baseline values, it is found that the concentration gradient mutation begins at 1590.2 seconds, the flow pulsation frequency mutation begins at 1590.8 seconds, and the axial temperature difference mutation begins at 1591.5 seconds, with magnitude ratios of 1.8:1.5:0.9, respectively. In practical implementation, the order and magnitude ratio are reverse-matched with the state transition paths recorded in the dynamic topology mapping to find the source state and target state combination that triggered this parameter mutation. It is understandable that the matching process calculates the similarity between the mutation patterns observed in the detailed data and the expected mutation patterns described by each state transition path in the dynamic topology map. To quantify the severity of mutations during the transient phase, an operational relationship is introduced:

[0042] in: The sharpness index represents the transient phase. It represents the number of monitoring dimensions involved in the analysis. It is the first The weighting coefficients of each monitoring dimension. It is the first in the detailed data of the transient phase. The peak value reached by each monitoring dimension parameter. It is the first steady-state phase before the transient phase occurs. The baseline values ​​of each monitoring dimension parameter. It is the first in the same steady-state stage Standard deviation of each monitoring dimension parameter. Sharpness index. Used to assist in determining the confidence level of mutation tracing results, high sharpness index values ​​typically correspond to more explicit state transition events. In practice, the information of successfully matched state transition paths and the specific manifestation patterns of this parameter mutation in each monitoring dimension are recorded to form a complete mutation tracing record. For example, if the matching result is that the state transition path corresponding to this transient is "from 'heat preservation standby' to 'programmed heating'", then the path ID, the observed sequence of mutations (concentration gradient -> flow pulsation -> axial temperature difference), the amplitude ratio of each dimension, and the calculated sharpness index are recorded. The above information together constitutes a mutation tracing record, which is stored in the local storage of the edge computing node.

[0043] See Figure 5 The comprehensive trend analysis chart for carbon emission monitoring parameters presents the comprehensive evolution trend of the normalized values ​​(after normalization) of three types of parameters—concentration gradient, flow fluctuation (scaled × 1 / 0.7), and axial temperature difference (scaled × 3)—over time in carbon emission monitoring and early warning analysis. This trend is visualized by combining steady-state and transient phase divisions. Specifically, the chart uses time (seconds) as the horizontal axis and the normalized parameter values ​​(normalized to the [0,200] range) as the vertical axis. Different colored curves distinguish the change trajectories of the three monitoring dimensions: the blue curve corresponds to the concentration gradient, the green curve to the flow fluctuation, and the red curve to the axial temperature difference. Light green and pink backgrounds respectively indicate the steady-state and transient phases. In actual analysis, the stage division is based on the variance of the first derivative of the parameter curve within the sliding time window: when the variance is consistently below the steady-state threshold, the corresponding time interval (e.g., 1520-1580 seconds, after 1640 seconds) is determined to be the steady-state stage. At this time, the normalized values ​​of the three types of parameters fluctuate less and are in a relatively stable range. When the variance rapidly exceeds the transient threshold in a short period of time, the corresponding time interval (e.g., 1580-1640 seconds) is determined to be the transient stage. At this time, the normalized values ​​of the three types of parameters all show significant fluctuations, and the axial temperature difference (red curve) has more prominent visual characteristics of its change amplitude due to the scaling ×3 processing.

[0044] In one embodiment of the present invention, the autonomous optimization module integrates the results of parameter inertia learning and parameter mutation tracing to construct a two-layer parameter correction model on the edge side. The first layer is a steady-state correction layer, which calls the inertial model in the parameter inertia knowledge base to perform consistency verification on the data in the steady-state stage during real-time monitoring. If the data deviates from the allowable deviation band defined by the inertial model, a correction suggestion based on historical steady-state behavior is generated. The second layer is a transient correction layer, which calls the state transition information and parameter performance patterns in the mutation tracing record. When a new transient event is detected, it is matched with the historical record for pattern matching. If the match is successful, the parameter change path verified in the historical record is used for predictive correction. The outputs of the steady-state correction layer and the transient correction layer are prioritized and fused to generate the final control command applied to the adjustment of monitoring parameters, thereby completing the autonomous optimization of the carbon emission monitoring process.

[0045] In a specific implementation, a cluster of gas-fired boilers in an industrial park is used as an example scenario to illustrate the process by which the autonomous optimization module constructs and applies a two-layer parameter correction model based on the output of the stage analysis and traceability modules. The input to the autonomous optimization module is the inertial model in the parameter inertia knowledge base and the newly added mutation traceability records, and the output is the control commands used to adjust the monitoring parameters.

[0046] In practical implementation, the autonomous optimization module integrates the results of parameter inertia learning and parameter mutation tracing to construct a two-layer parameter correction model on the edge side. The first layer is the steady-state correction layer. This layer calls upon the inertial model in the parameter inertia knowledge base to perform consistency checks on the data in the steady-state phase during real-time monitoring. If the data deviates from the allowable deviation band defined by the inertial model, a correction suggestion based on historical steady-state behavior is generated. For example, for a gas-fired boiler numbered "GB-03", when its real-time data is determined to be in the "low-load steady-state" phase, the steady-state correction layer calls the "GB-03_low-load steady-state" inertial model in the parameter inertia knowledge base. This model defines a baseline value of 4.2% for flue gas oxygen content and an allowable deviation band of [3.8%, 4.6%]. If the real-time monitored oxygen content value is 4.9% for five consecutive sampling periods, exceeding the allowable deviation band, the steady-state correction layer generates a correction suggestion: "Adjust the calibration offset of the oxygen content sensor 'OX_GB-03' to -0.6%."

[0047] In some embodiments, the second-layer model is a transient correction layer. This layer calls upon state transition information and parameter behavior patterns from the mutation source record. When a new transient event is detected, it is matched against historical records. If a match is successful, the validated parameter change path from the historical record is used for predictive correction. For example, if the load of boiler "GB-03" begins to rise, exhibiting transient characteristics, the transient correction layer matches the initial pattern of the current transient with the mutation source record library. A historical record is matched: "transition from '40% load' to '70% load'". This record contains a predictive model of the typical decrease curve of flue gas oxygen content during this transition. The transient correction layer then uses this predictive model to predict the expected oxygen content value for subsequent moments within the current monitoring period. The typical decline curve prediction model is constructed by extracting parameter evolution trajectory details stored in historical mutation source records, and constructing the model by extracting parameter change sequences of monitoring dimensions such as flue gas oxygen content during specific state transitions. The model records the typical decline trend of oxygen content when transitioning from the source state to the target state, including the rate of change and morphological characteristics. These characteristics are derived from verified parameter performance patterns in historical records, such as the order and magnitude of mutations. When a new transient event is detected, the transient correction layer performs pattern matching between the initial parameter change data of the current event and the historical model. If the match is successful, the typical decline curve model is directly applied, and the expected oxygen content value for subsequent moments in the current monitoring period is predicted through time alignment and curve fitting methods, thereby achieving early correction of parameter changes. It can be understood that if the predicted value deviates significantly from the subsequent actual monitoring value, the transient correction layer will generate correction suggestions, such as: "Adjust the gain coefficient K of the oxygen content prediction model during load transients from the historical value of 1.05 to 1.02."

[0048] In practical implementation, the outputs of the steady-state correction layer and the transient correction layer are prioritized and fused to generate the final control commands applied to the adjustment of monitoring parameters. Priority arbitration follows the principle of "transient state takes precedence over steady state." When the system is determined to be in a transient phase, the correction suggestions from the transient correction layer are adopted first; when the system is in a steady-state phase, the correction suggestions from the steady-state correction layer are adopted. In practical implementation, the result fusion process involves a weighted synthesis of the parameter adjustment amounts from the two types of correction suggestions. To quantify the fusion logic, an operational relationship in the implementation is introduced:

[0049] in: This represents the parameter correction amount in the final output. This represents the amount of parameter adjustment suggested by the transient correction layer. This represents the parameter adjustment amount suggested by the steady-state correction layer. It is the confidence weighting factor for the transient phase, and its value is determined by the degree of matching between the current monitoring data and the transient pattern, as well as the sharpness index of the transient process. The decision is made jointly. For example, in a load increase event, the transient correction layer suggests lowering the oxygen content setpoint by 0.8%, while the steady-state correction layer, based on long-term statistics, suggests fine-tuning the oxygen content setpoint by 0.1%. The system calculates the current transient matching degree and sharpness index. Larger, therefore set The final correction amount obtained by fusion The value is -0.73%. This instruction will be sent to the boiler's control system to achieve autonomous optimization of the key parameter, oxygen content monitoring, in the carbon emission monitoring process. In some embodiments, the control instruction may include not only adjustments to the parameter setpoints but also optimization instructions for the adaptive monitoring probe sampling strategy. For example, when a specific transient phase is predicted, the instruction logic control unit may temporarily increase the sampling frequency.

[0050] See Figure 6This diagram illustrates the dynamic relationship between flue gas oxygen content and load throughout the boiler's entire operation. It also distinguishes three operating stages—"low load steady state," "load surge transient," and "high load steady state"—using background colors. Specifically, in the low load steady state stage (green background), the boiler load remains stable at approximately 40%, corresponding to flue gas oxygen content fluctuations around 4.2%, consistent with the baseline and deviation band characteristics of the "GB-03_low load steady state" inertial model in the parameter inertia knowledge base. Entering the load surge transient stage (pink background), the boiler load rapidly jumps from 40% to 70%, with flue gas oxygen content exhibiting a typical downward trend. This process matches the parameter evolution trajectory of "jumping from 40% load to 70% load" in the mutation source tracing record, demonstrating the application scenario of the oxygen content typical decline curve prediction model called by the transient correction layer. In the high load steady state stage (purple background), the load stabilizes at around 70%, while the flue gas oxygen content fluctuates around 3.5%, entering a new steady-state parameter range. It intuitively reflects the linkage change law between flue gas oxygen content and load during boiler operation status switching. It is a key data support carrier for the stage analysis and source tracing module to identify steady-state / transient stages and for the autonomous optimization module to build a two-layer parameter correction model.

[0051] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0052] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A carbon emission monitoring, early warning, and analysis system based on edge computing, characterized in that, Includes the following modules: The dynamic topology mapping module is used to establish dynamic topology mappings of different carbon emission sources within the monitoring area, and to record the identity attributes, spatial location and state transition path of each emission source. An adaptive probe deployment module is used to deploy adaptive monitoring probes on each edge computing node according to the dynamic topology mapping, and adjust the monitoring dimension sampling strategy of the adaptive monitoring probes according to the state transition path of the carbon emission source. The early warning identification module is used to capture the time slice sequence of the emission flow through the adaptive monitoring probe, and compare the time slice sequence with the state transition path to identify potential state switching early warning points; The evolution trajectory generation module is used to generate parameter evolution trajectories within the edge computing node based on the state switching early warning point, so as to describe the continuous change process of the monitoring parameters before and after the state switching. The phase analysis and tracing module is used to extract the steady-state phase and transient phase in the parameter evolution trajectory, perform parameter inertial learning in the steady-state phase, and perform parameter mutation tracing in the transient phase. The autonomous optimization module is used to integrate the results of parameter inertia learning and the results of parameter mutation tracing to construct a two-layer parameter correction model on the edge side, and use the two-layer parameter correction model to drive the autonomous optimization of the subsequent carbon emission monitoring process.

2. The carbon emission monitoring, early warning, and analysis system based on edge computing according to claim 1, characterized in that, The establishment of a dynamic topological mapping for different carbon emission sources within the monitoring area specifically includes: The static identification features and dynamic behavioral features of each carbon emission source are collected. The static identification features include the equipment code and its category, and the dynamic behavioral features include the operating cycle and emission intensity pattern. The static identification features and dynamic behavioral features of each emission source are fused and encoded to generate topological nodes with spatiotemporal attributes. The correlation between different topological nodes on the time axis is analyzed. Based on the causality and temporal proximity of emission events, state transition paths between nodes are drawn. The state transition paths are marked with source state, target state and triggering conditions. All topological nodes and their corresponding state transition paths are integrated to form the dynamic topological mapping.

3. The carbon emission monitoring, early warning, and analysis system based on edge computing according to claim 2, characterized in that, The step of deploying adaptive monitoring probes on each edge computing node according to the dynamic topology mapping specifically includes: The monitoring dimensions associated with each topology node in the dynamic topology mapping are analyzed, including concentration gradient, flow fluctuation and temperature field distribution; Based on the monitoring dimensions, a corresponding sensing unit group and signal conditioning unit are configured for each edge computing node to form the basis of the hardware probe. A logic control unit is embedded on the hardware probe. The logic control unit loads the state transition path and dynamically adjusts the sampling strategy of the sensing unit group according to the preset trigger conditions in the path. The combination of the hardware probe base and the logic control unit is defined as the adaptive monitoring probe, and it is bound to the corresponding edge computing node.

4. The carbon emission monitoring and early warning analysis system based on edge computing according to claim 3, characterized in that, The capture of time-slice sequences of the emission stream via the adaptive monitoring probe specifically includes: The logic control unit in the adaptive monitoring probe obtains the corresponding sampling frequency and sampling duration from the loaded state transition path according to the current state of the carbon emission source; The control sensor unit group synchronously collects concentration gradient, flow pulsation and temperature field distribution according to the sampling frequency and sampling duration to obtain raw data blocks within a complete monitoring cycle. The original data block is divided into segments according to a fixed time window to generate a series of data segments with time-series labels, and each data segment is a time slice; The consecutive time slices are arranged in chronological order to form the time slice sequence.

5. A carbon emission monitoring and early warning analysis system based on edge computing according to claim 4, characterized in that, The step of comparing the time slice sequence with the state transition path to identify potential state transition warning points specifically includes: Extract the characteristic patterns that should be exhibited in the monitoring dimension when all source states transition to the target state from the state transition path, and construct a characteristic pattern library; The data of each time slice in the time slice sequence is matched point by point with the feature patterns in the feature pattern library, and the matching confidence is calculated. If the matching confidence of several consecutive time slices shows a trend of converging towards a certain target state feature pattern, and the trend strength exceeds a preset convergence threshold, then it is determined that there is a potential state switch within the time window corresponding to the several consecutive time slices. The starting time point at which a state transition is determined is marked as the state transition warning point.

6. The carbon emission monitoring, early warning, and analysis system based on edge computing according to claim 5, characterized in that, The generation of parameter evolution trajectories within the edge computing node based on the state switching early warning points specifically includes: Using the aforementioned state switching warning point as a time reference, a fixed-length time period is traced back as the starting point of the evolution, and a fixed-length time period is extended forward as the ending point of the evolution. Extract parameter values ​​for concentration gradient, flow fluctuation, and temperature field distribution from all time slice sequences from the start to the end of the evolution; Plot the continuous change curve of each monitoring dimension from the starting point to the end point of evolution with time as the horizontal axis and the parameter values ​​of each monitoring dimension as the vertical axis. The change curves of all monitoring dimensions are superimposed and aligned to form a comprehensive map of the evolution of multidimensional parameters over time, which is the trajectory of the parameter evolution.

7. A carbon emission monitoring and early warning analysis system based on edge computing according to claim 6, characterized in that, The extraction of the steady-state and transient phases from the parameter evolution trajectory specifically includes: Smoothness analysis is performed on the change curve of each monitoring dimension in the parameter evolution trajectory, and the first derivative variance of the curve within the sliding time window is calculated. If the variance of the first derivative remains below the steady-state threshold, the curve segment corresponding to the sliding time window is determined to be in the steady-state stage. If the variance of the first derivative increases sharply in a short period of time and exceeds the transient threshold, then the curve segment near the time point corresponding to the sharp increase in the variance of the first derivative is determined to be the transient stage. Record the start and end timestamps of all steady-state phases and the center timestamps of all transient phases.

8. A carbon emission monitoring and early warning analysis system based on edge computing according to claim 7, characterized in that, The parameter inertial learning performed during the steady-state phase specifically includes: For each steady-state stage, extract the set of all monitored dimension parameter values ​​within that steady-state stage; Calculate the statistical characteristics of the parameter values ​​for each monitoring dimension in the set, wherein the statistical characteristics include the mean, median, and standard deviation; Using the aforementioned statistical characteristics, an inertial model is established to describe the normal fluctuation range of parameters within the steady-state phase. The inertial model consists of parameter baseline values ​​and allowable deviation bands. Multiple inertial models learned at different steady-state stages are archived to form a parametric inertial knowledge base describing the normal behavior of carbon emission sources in the monitoring area under various steady-state conditions.

9. A carbon emission monitoring and early warning analysis system based on edge computing according to claim 7, characterized in that, The execution of parameter mutation tracing during the transient phase specifically includes: For each transient phase, its central time marker is located, and detailed data of the parameter evolution trajectory within a period before and after the central time marker are extracted; The analysis of the detailed data reveals the order and magnitude of abrupt changes in parameters across different monitoring dimensions. The order and magnitude ratio are reverse-matched with the state transition paths recorded in the dynamic topology mapping to find the source state and target state combination most likely to trigger this parameter mutation. Record the information on the successfully matched state transition path, as well as the specific manifestation pattern of this parameter mutation in each monitoring dimension, to form a complete mutation source tracing record.

10. A carbon emission monitoring, early warning, and analysis system based on edge computing according to claim 1, characterized in that, The results of fusing parameter inertial learning and parameter mutation tracing are used to construct a two-layer parameter correction model on the edge side, specifically including: The first layer of the model is the steady-state correction layer. The steady-state correction layer calls the inertial model in the parameter inertial knowledge base to perform consistency verification on the data in the steady-state stage during real-time monitoring. If the data deviates from the allowable deviation band defined by the inertial model, correction suggestions based on historical steady-state behavior are generated. The second layer model is the transient correction layer. The transient correction layer calls the state transition information and parameter behavior patterns in the mutation source record. When a new transient event is detected, it is matched with the historical record. If the match is successful, the parameter change path verified in the historical record is used for predictive correction. The outputs of the steady-state correction layer and the transient correction layer are prioritized, arbitrated, and their results are fused to generate the final control commands applied to the adjustment of monitoring parameters, thereby completing the autonomous optimization of the carbon emission monitoring process.

Images