Carbon emission calculation method based on distributed monitoring

Abstract

The present application relates to the technical field of carbon emission monitoring and calculation, and discloses a carbon emission calculation method based on distributed monitoring. The method deploys hierarchical edge monitoring units in the monitoring area, dynamically slices and separates the initial monitoring flow, obtains the original energy flow, activity event flow and source feature flow. The edge side generates a steady-state energy consumption feature sequence through a feature evolution engine and reconstructs the causal chain of the activity event flow to form a feature map. The two are fused into an edge fusion feature body. The cloud performs spatial topology fusion on each edge feature body to form a global collaborative feature field, and traces and quantitatively maps the source feature flow according to the feature field to generate a quantitative emission spectrum, and finally completes the calculation of the total regional carbon emission and the generation of the list. The method improves the efficiency and quantitative accuracy of carbon emission calculation through the collaborative processing of the edge and the cloud.

Description

Carbon emission calculation method based on distributed monitoring Technical Field

[0001] This invention relates to the field of carbon emission monitoring and calculation technology, specifically a method for calculating carbon emissions based on distributed monitoring. Background Technology

[0002] Traditional regional carbon emission monitoring generally employs a centralized processing model. Raw multimodal data collected by various sensors are directly aggregated and uploaded to a remote cloud server for unified processing. Existing technical solutions mainly perform basic aggregation and statistics on sensor data. This method not only puts enormous pressure on network bandwidth and strains cloud computing resources, but more importantly, it has limitations in data feature extraction. Simple numerical processing cannot effectively capture the complex time-varying patterns and inherent correlations contained in the data.

[0003] The limitation of existing technologies lies in their insufficient depth of data processing. For continuously monitored energy consumption data, conventional methods lack the ability to model its dynamic characteristics over time, making it difficult to extract robust features from fluctuating data that reflect the essential operating patterns of equipment or processes. For discrete event data, existing methods can only present the sequence of occurrences, completely ignoring the possible causal relationships between events, resulting in a lack of logical support for subsequent emission source tracing. These superficial features directly restrict the accuracy and interpretability of carbon emission quantification calculations.

[0004] The core problem that this invention aims to solve is: how to achieve deep evolution extraction of energy flow temporal characteristics and automated reconstruction of event flow causal relationships at the edge of a distributed monitoring network, thereby generating high-quality, highly interpretable fused features to provide a reliable basis for the final high-precision carbon emission calculation in the cloud. Summary of the Invention

[0005] The purpose of this invention is to provide a method for calculating carbon emissions based on distributed monitoring, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for calculating carbon emissions based on distributed monitoring, the method comprising:

[0007] Hierarchical edge monitoring units are deployed within the monitoring area. Each edge monitoring unit receives an initial monitoring stream from the sensor cluster under its jurisdiction, which contains multimodal sensing data related to emissions.

[0008] Within the edge monitoring unit, dynamic slicing and intra-flow feature separation operations are performed on the initial monitoring stream to extract the original energy flow features, activity event flow features, and source feature stream from the multimodal sensing data;

[0009] The original energy flow features are iteratively evolved through multiple rounds using the feature evolution engine built into the edge monitoring unit to generate a steady-state energy consumption feature sequence with time context dependence.

[0010] The event causal chain of the activity event flow features is reconstructed synchronously to form an activity evolution feature map with causal labels;

[0011] The steady-state energy consumption feature sequence and the activity evolution feature map are nested within the edge monitoring unit to generate an edge fusion feature body.

[0012] The edge fusion feature volumes from all edge monitoring units are uploaded to the cloud center, where spatial topology fusion of cross-unit feature volumes is performed to generate a global collaborative feature field.

[0013] At the cloud center, source feature streams are traced and quantified based on a global collaborative feature field to generate a quantified emission spectrum.

[0014] The total regional carbon emissions are calculated based on the emission calculation kernel driven by the quantitative emission spectrum, and a structured emission inventory is generated.

[0015] Preferably, the step of performing dynamic slicing and intra-stream feature separation operations on the initial monitoring stream within the edge monitoring unit includes:

[0016] An adaptive sliding time window is configured for the initial monitoring stream, the width of which is dynamically adjusted based on the instantaneous data entropy of the initial monitoring stream;

[0017] Within the sliding time window, the multimodal sensing data is decoupled to separate the time-series mode corresponding to energy consumption, the state mode corresponding to equipment or personnel activities, and the component mode corresponding to direct emissions.

[0018] The time-series modes are segmented into subsequences to form the original energy flow characteristics;

[0019] The state modes are subjected to event boundary detection and event fragment extraction to form active event flow features;

[0020] Concentration and component characteristics are extracted from the component modes to form a source feature stream;

[0021] The dynamic slicing and intra-stream feature separation operation further includes internal timestamp alignment and stream identifier marking for each separated feature stream.

[0022] Preferably, the step of using the feature evolution engine built into the edge monitoring unit to perform multiple rounds of iterative evolution on the original energy flow features to generate a steady-state energy consumption feature sequence with time context dependence includes:

[0023] The original energy flow features are input into the primary evolution layer of the feature evolution engine to perform local pattern learning and extract preliminary energy consumption pattern fragments.

[0024] The initial energy consumption pattern fragments are fed into the secondary evolution layer to mine the pattern correlation across time slices, forming an intermediate energy consumption pattern chain with time dependence.

[0025] Perform mode stability testing on the intermediate energy consumption mode chain and filter out transient mode nodes whose fluctuations exceed a preset threshold.

[0026] The intermediate energy consumption pattern chain that has passed the stability test is input into the final evolution layer for pattern convergence and feature serialization reconstruction, and outputs a smooth and context-coherent steady-state energy consumption feature sequence.

[0027] Each iteration of the feature evolution engine incorporates feedback from the previous iteration's output to ensure the continuity and convergence of the evolution direction.

[0028] Preferably, the synchronization involves reconstructing the event causal chain of the activity event stream features to form an activity evolution feature map with causal labels, including:

[0029] Analyze each individual event node in the activity event flow characteristics to identify the type attribute, timestamp, and associated resource consumption marker of each event node;

[0030] Guided by a pre-defined causal rule knowledge base, the temporal adjacency and logical triggering relationships between event nodes are analyzed to establish preliminary event causal relationship edges.

[0031] The confidence level of the initial causal relationship edges is evaluated, strong causal edges with confidence levels higher than the set value are selected, and the event nodes connected by the strong causal edges are logically integrated to form a composite event cluster.

[0032] Using clusters of complex events as nodes and strong causal edges as connections, a directed acyclic graph structure is constructed, and each edge in the graph is labeled with its causal type and strength weight, thereby forming the activity evolution feature graph.

[0033] After the construction is completed, loop detection and pruning are performed on the activity evolution feature map to ensure the consistency of causal logic.

[0034] Preferably, the step of performing alignment and nesting processing of the steady-state energy consumption feature sequence and the activity evolution feature map within the edge monitoring unit to generate an edge fusion feature body includes:

[0035] In the time dimension, the steady-state energy consumption feature sequence is precisely aligned with the event nodes in the activity evolution feature map to establish a time alignment mapping table;

[0036] According to the time alignment mapping table, the steady-state energy consumption feature vector at each time point is injected into the event node in the corresponding time interval of the activity evolution feature map, as the energy consumption attribute embedding of the event node.

[0037] Based on energy consumption attribute embedding, the correlation strength between event nodes in the activity evolution feature map is recalculated, the weights of causal edges are updated, and an energy consumption-enhanced activity evolution feature map is formed.

[0038] The evolution feature map of energy-intensive activities is encoded using graph structure, and then transformed into a graph feature vector of fixed dimensions.

[0039] The graph feature vector is fused with the global statistical feature vector of the steady-state energy consumption feature sequence through cross-attention fusion to generate a multi-dimensional edge fusion feature body containing spatiotemporal and causal information.

[0040] Preferably, the step of uploading edge fusion feature volumes from all edge monitoring units to the cloud center, and performing spatial topology fusion of cross-unit feature volumes at the cloud center to generate a global collaborative feature field includes:

[0041] The cloud center receives edge fusion feature data uploaded by each edge monitoring unit and adds the spatial location code of its source unit to each edge fusion feature data.

[0042] Based on spatial location coding, a topological connection map representing the spatial adjacency relationship of all edge monitoring units within the monitoring area is constructed;

[0043] Under the constraints of the topological connection graph, a graph neural network is used to perform message passing and feature aggregation on each edge fusion feature body, so that each edge fusion feature body can absorb the feature information of its spatial neighbors.

[0044] After multiple rounds of message passing, spatial feature diffusion smoothing is performed on each updated edge fusion feature volume to eliminate feature mutation boundaries between units.

[0045] All smoothed edge-fused features are superimposed and interpolated in the feature space to form a continuous and consistent global collaborative feature field covering the entire monitoring area.

[0046] Preferably, the step of tracing and quantifying the source feature stream based on a global collaborative feature field at the cloud center to generate a quantified emission spectrum includes:

[0047] Source feature streams are received synchronously from each edge monitoring unit, and the source feature streams contain direct emission concentration and composition data from different spatial points;

[0048] Project the data points in the source feature stream onto the spatial locations corresponding to the global collaborative feature field;

[0049] Guided by the global collaborative feature field, emission source analysis is performed on each projection point. The emission source analysis infers the source category and activity process most likely to cause the emission at the projection point by matching the energy consumption feature pattern and activity feature pattern of the location point of the projection point in the global collaborative feature field.

[0050] Based on the emission source tracing analysis results, each source characteristic flow data point is assigned a source category label and an activity process label;

[0051] Based on the preset emission factor mapping table, source feature stream data points with source category labels and activity process labels are transformed into standardized carbon emission intensity units.

[0052] All carbon emission intensity units are organized and arranged according to source category and spatial location to generate a quantitative emission spectrum, which records the emission intensity of different spatial locations and different source categories in matrix form.

[0053] Preferably, the calculation of total regional carbon emissions based on the emission calculation kernel driven by the quantified emission spectrum includes:

[0054] The quantified emission spectrum is input into the emission calculation core, which has built-in calculation logic paths corresponding to different source categories;

[0055] The emission calculation kernel first analyzes and quantifies the matrix structure of the emission spectrum, identifying all non-zero emission intensity units and their corresponding spatial coordinates and source categories;

[0056] For each identified emission intensity unit, the corresponding calculation logic path is activated according to its source category. The logic path combines the value of the emission intensity unit and the activity intensity characteristics of its corresponding spatial location in the global collaborative feature field to calculate the cumulative emission of the emission intensity unit within the calculation period.

[0057] The cumulative emissions of all emission intensity units within the monitoring area are spatially integrated and summed to obtain the classified emissions by source category.

[0058] The total carbon emissions of the monitoring area during the calculation period are obtained by summing the emissions of all source categories.

[0059] Preferably, generating the structured emissions inventory includes:

[0060] Using the total carbon emissions as the overarching framework, the emissions of each source category are listed in a hierarchical structure.

[0061] For the categorized emissions of each source category, we further correlate and trace back to the corresponding emission intensity units in the quantified emission spectrum, and extract the spatial location information and activity process tags corresponding to the emission intensity units;

[0062] Spatial location information, activity process tags, and corresponding emission intensity and cumulative emission are structurally bound to form emission detail entries under the source category;

[0063] Sort and organize the emission details of all source categories according to emission volume or spatial distribution;

[0064] The total regional carbon emissions, emissions by source category, and detailed emission entries are integrated into a hierarchical tree-structured document, which is the structured emissions inventory.

[0065] Preferably, the method further includes a step of dynamically updating and verifying the structured emission inventory:

[0066] A dynamic emission baseline model corresponding to the monitoring area is established. The dynamic emission baseline model is trained based on the historical structured emission inventory.

[0067] Once the new structured emissions inventory is generated, it is compared with the predicted inventory generated by the dynamic emissions baseline model to identify significant deviations.

[0068] The significant deviation terms are traced in reverse to locate the quantized emission spectrum unit that caused the deviation and the related edge fusion feature.

[0069] Trigger a re-verification process for the relevant edge monitoring unit data. If the deviation disappears or decreases after re-verification, update the dynamic emission baseline model with a new inventory. If the deviation persists, generate and record a data anomaly alarm.

[0070] Compared with the prior art, the beneficial effects of the present invention are:

[0071] Within the edge monitoring unit, multiple rounds of iterative evolution are performed on the original energy flow characteristics. Through temporal modeling and context learning, the discrete and fluctuating original characteristics are gradually converged into a steady-state sequence with long-term dependencies. Conventional techniques such as simple aggregation or sliding window averaging cannot effectively separate noise from essential patterns, while this method can extract the inherent, recurring regularities in the energy consumption process. The generated steady-state energy consumption characteristic sequence reduces data volatility and enhances consistency and continuity over time. This makes subsequent energy consumption-based emission factor mapping more stable, reduces the deviation in quantification results caused by instantaneous data fluctuations or random interference, and improves the long-term reliability of regional carbon emission calculation.

[0072] The event causal chain is reconstructed based on the characteristics of the event flow. Based on causal reasoning and graph structure learning, directional causal relationships are identified and established from discrete event flows, forming an event evolution feature map with clear logical labels. Conventional methods treat events as independent records or records with only temporal sequence relationships; this method constructs a network of driving and driven relationships between events. The resulting feature map not only describes the order of events but also reveals the intrinsic influence mechanisms and logical transmission paths between key links in the production process. This provides a clear causal logical basis for accurately linking specific carbon emissions to specific production activities, operational behaviors, or abnormal states, giving emission inventories stronger interpretability and traceability. Attached Figure Description

[0073] Figure 1 is a schematic diagram illustrating the working principle of the carbon emission calculation method based on distributed monitoring described in this invention.

[0074] Figure 2 is a flowchart of the dynamic slicing and intra-flow feature separation operation;

[0075] Figure 3 is a flowchart of the event causal chain reconstruction generating activity evolution feature map;

[0076] Figure 4 is a pie chart showing the percentage of carbon emission source categories;

[0077] Figure 5 is a bar chart comparing the causal strength of events. Detailed Implementation

[0078] 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.

[0079] Referring to Figure 1, this invention provides a carbon emission calculation method based on distributed monitoring. The method includes: deploying hierarchical edge monitoring units within a monitoring area. Each edge monitoring unit is responsible for receiving and processing the initial monitoring stream uploaded by a sensor cluster within its jurisdiction. This initial monitoring stream contains multimodal sensor data related to emissions. Within each edge monitoring unit, dynamic slicing and intra-stream feature separation operations are performed on the initial monitoring stream to separate the original energy flow features, activity event flow features, and source feature stream from the multimodal sensor data. The feature evolution engine built into the edge monitoring unit is used to perform multiple rounds of iterative evolution on the original energy flow features to generate a steady-state energy consumption feature sequence with time context dependence. Simultaneously, the event causal chain of the activity event flow features is reconstructed to form an activity evolution feature map with causal labels. The steady-state energy consumption feature sequence and the activity evolution feature map are nested within the edge monitoring unit to generate an edge fusion feature body. Subsequently, the edge fusion feature bodies generated by all edge monitoring units are uploaded to a cloud center, where spatial topology fusion of cross-unit feature bodies is performed to generate a global collaborative feature field. The cloud-based center traces and quantifies the source feature streams from each edge monitoring unit based on a global collaborative feature field, generating a quantified emission spectrum. This quantified emission spectrum drives the emission calculation kernel to calculate the total regional carbon emissions, ultimately generating a structured emission inventory.

[0080] In one embodiment of the present invention, referring to Figure 2, dynamic slicing and intra-flow feature separation operations are performed on the initial monitoring stream within the edge monitoring unit. An adaptive sliding time window is configured for the initial monitoring stream, and the width of the sliding time window is dynamically adjusted according to the instantaneous data entropy of the initial monitoring stream. Within the sliding time window, modal decoupling is performed on the multimodal sensing data to separate the time-series mode corresponding to energy consumption, the state mode corresponding to equipment or personnel activities, and the component mode corresponding to direct emissions. The time-series mode is segmented into subsequences to form the original energy flow features. Event boundary detection and event fragment extraction are performed on the state mode to form the activity event flow features. Concentration and component features are extracted from the component mode to form the source feature stream. The dynamic slicing and intra-flow feature separation operation also includes internal timestamp alignment and flow identifier marking for each separated feature stream. The original energy flow features are iteratively evolved multiple times using the feature evolution engine built into the edge monitoring unit. The original energy flow features are input into the primary evolution layer of the feature evolution engine for local pattern learning to extract preliminary energy consumption pattern fragments. The initial energy consumption pattern fragments are fed into the secondary evolution layer for cross-time-slice pattern correlation mining, forming an intermediate energy consumption pattern chain with time dependencies. The intermediate energy consumption pattern chain undergoes a pattern stability check, filtering out transient pattern nodes whose fluctuations exceed a preset threshold. The stability-checked intermediate energy consumption pattern chain is then input into the final evolution layer for pattern convergence and feature sequence reconstruction, outputting a smooth and context-coherent steady-state energy consumption feature sequence. Each iteration of the feature evolution engine incorporates feedback from the previous iteration's output to ensure the continuity and convergence of the evolution direction.

[0081] In practical implementation, an industrial park monitoring scenario covering multiple buildings is used as an example. The edge monitoring unit deployed in the monitoring area receives the initial monitoring stream from the sensor cluster under its jurisdiction. The initial monitoring stream includes power time-series data from the power metering module, equipment vibration and infrared sensor status change data, and component concentration data from the flue gas analyzer installed near the emission outlet.

[0082] In practical implementation, the edge monitoring unit performs dynamic slicing and intra-stream feature separation operations on the initial monitoring stream, configuring an adaptive sliding time window for the initial monitoring stream. The width of the sliding time window is dynamically adjusted based on the instantaneous data entropy of the initial monitoring stream. When calculating the instantaneous data entropy, the numerical distribution of multimodal sensor data is statistically analyzed within a basic time slice, and the dispersion of the numerical distribution is quantified using a dedicated formula. In a specific data comparison, when the production line startup is detected, causing drastic fluctuations in power data, the calculated instantaneous data entropy increases, and the width of the sliding time window decreases accordingly to capture rapidly changing details; when the equipment is detected to be in standby mode and the data is stable, the instantaneous data entropy decreases, and the width of the sliding time window automatically expands to integrate stable patterns over a longer period. The dedicated formula for adjusting the window width is:

[0083]

[0084] in: Represents the final width of the sliding time window. Represents the preset base window width. This represents the window width adjustment factor. This represents the calculated instantaneous data entropy. Within a sliding time window, modal decoupling is performed on the multimodal sensor data to separate the time-series mode corresponding to energy consumption, the state mode corresponding to equipment or personnel activities, and the component mode corresponding to direct emissions. The time-series mode is segmented into subsequences to form the original energy flow features organized at fixed time intervals. Event boundary detection and event fragment extraction are performed on the state mode, such as identifying discrete events like "motor start," "conveyor belt operation," and "valve closure" from continuous vibration signals, forming the activity event flow features. Concentration and component features are extracted from the component mode, such as extracting instantaneous concentration readings of carbon dioxide and carbon monoxide from flue gas analysis data, forming the source feature flow. The dynamic slicing and intra-flow feature separation operations also include internal timestamp alignment and flow identifier marking for each separated feature flow, enabling energy consumption, equipment activity, and emission concentration data from the same moment to be correlated through timestamps.

[0085] In some embodiments, the feature evolution engine built into the edge monitoring unit performs multiple rounds of iterative evolution on the original energy flow features. The original energy flow features are input into the primary evolution layer of the feature evolution engine. The primary evolution layer performs local pattern learning, extracting preliminary energy consumption pattern fragments such as "constant base load," "linear rise," and "periodic pulse" from the power time series data. The preliminary energy consumption pattern fragments are fed into the secondary evolution layer, which performs pattern correlation mining across time slices, analyzing the transition probabilities and dependencies between pattern fragments in consecutive time slices, forming an intermediate energy consumption pattern chain with time dependencies. The intermediate energy consumption pattern chain is subjected to pattern stability testing, filtering out transient pattern nodes whose fluctuations exceed a preset threshold, such as filtering out power spike patterns with extremely short durations caused by measurement noise. The intermediate energy consumption pattern chain that has passed stability testing is input into the final evolution layer, which performs pattern convergence and feature sequence reconstruction. Through smoothing filtering and context encoding, it outputs a context-coherent steady-state energy consumption feature sequence, where each feature vector represents a stable and interpretable energy consumption state.

[0086] It is understandable that each iteration of the feature evolution engine incorporates feedback from the previous iteration to ensure the continuity and convergence of the evolution direction. The preliminary results of the original energy flow features processed by the primary evolution layer serve as constraint information, feeding back into the next round of reading and segmenting the original energy flow features. The intermediate pattern chain structure output by the secondary evolution layer also serves as prior knowledge, guiding the primary evolution layer in extracting pattern fragments within subsequent time windows, forming a progressively refined feature learning loop. Optionally, the calculation of instantaneous data entropy in dynamic slicing operations considers not only numerical changes in the data but also the mutual information between different sensor data streams. When multiple sensor data streams exhibit highly synchronized changes, the calculation result of instantaneous data entropy will be adjusted accordingly, leading the sliding time window to adopt a width strategy that is more inclined to capture collaborative events. Optionally, the pattern stability check in the feature evolution engine, in addition to threshold judgment based on fluctuation amplitude, also introduces a duration judgment condition. A pattern node must be repeatedly identified in multiple consecutive time slices to be confirmed as a stable pattern rather than random fluctuation, thus being retained in the intermediate energy consumption pattern chain.

[0087] In one embodiment of the present invention, referring to Figure 3, the event causal chain is reconstructed synchronously on the activity event flow characteristics. Each independent event node in the activity event flow characteristics is analyzed, and the type attribute, timestamp, and associated resource consumption marker of each event node are identified. Guided by a preset causal rule knowledge base, the temporal adjacency relationship and logical triggering relationship between event nodes are analyzed to establish preliminary event causal relationship edges. The confidence of the preliminary event causal relationship edges is evaluated, and strong causal edges with confidence higher than a set value are selected. The event nodes connected by the strong causal edges are logically integrated to form composite event clusters. Using composite event clusters as nodes and strong causal edges as connections, a directed acyclic graph structure is constructed, and each edge in the graph is labeled with causal type and strength weight, thereby forming an activity evolution feature graph. After construction, loop detection and pruning are performed on the activity evolution feature graph to ensure the consistency of causal logic. Within the edge monitoring unit, the steady-state energy consumption feature sequence and the activity evolution feature map are nested and aligned. In the temporal dimension, the steady-state energy consumption feature sequence and the event nodes in the activity evolution feature map are precisely aligned, establishing a time alignment mapping table. Based on this table, the steady-state energy consumption feature vector at each time point is injected into the event node within the corresponding time interval in the activity evolution feature map, serving as the energy consumption attribute embedding for the event node. Based on this energy consumption attribute embedding, the correlation strength between event nodes in the activity evolution feature map is recalculated, and the weights of causal edges are updated, forming an energy-enhanced activity evolution feature map. This enhanced activity evolution feature map is then graph-encoded into a fixed-dimensional graph feature vector. The graph feature vector is then fused with the global statistical feature vector of the steady-state energy consumption feature sequence using cross-attention, generating a multi-dimensional edge fusion feature body containing spatiotemporal and causal information.

[0088] In specific implementation, taking the industrial park monitoring scenario as an example, after the edge monitoring unit completes dynamic slicing and feature separation, it obtains the activity event flow features and steady-state energy consumption feature sequences. The activity event flow features include a series of discrete event nodes with timestamps, such as "air compressor starts", "painting line conveyor belt moves forward", "oven heating starts", and "cooling fan stops".

[0089] In practical implementation, the event causal chain of the activity event flow characteristics is reconstructed, each independent event node in the activity event flow characteristics is analyzed, and the type attribute, timestamp, and associated resource consumption marker of each event node are identified. Guided by a pre-set causal rule knowledge base, the temporal adjacency and logical triggering relationships between event nodes are analyzed to establish preliminary event causal relationship edges. The causal rule knowledge base includes domain logic such as "the start-up of an air compressor usually leads to an increase in compressed air pipeline pressure, which in turn triggers the start-up of pneumatic tools or spraying equipment." The confidence of the preliminary event causal relationship edges is evaluated, and strong causal edges with confidence scores higher than the set value are selected. The confidence evaluation is based on the event co-occurrence frequency, temporal tightness, and matching degree of the rule knowledge base. The event nodes connected by the strong causal edges are logically integrated to form composite event clusters. For example, the frequently occurring consecutive event nodes "spraying robot arm in place," "paint valve open," and "spray head move" are integrated into the composite event cluster "spraying operation execution." Using clusters of complex events as nodes and strong causal edges as connections, a directed acyclic graph (DAG) structure is constructed. Each edge in the graph is labeled with its causal type and strength weight, thus forming an activity evolution feature graph. This graph depicts causal chains such as "boiler ignition" leading to "steam generation," which in turn drives "turbine operation." After construction, loop detection and pruning are performed on the activity evolution feature graph to ensure the consistency of causal logic. For example, this eliminates logical loops caused by data noise or mislabeling, such as "event A leads to event B, and event B leads to event A."

[0090] In some embodiments, the steady-state energy consumption feature sequence and the activity evolution feature map are nested within the edge monitoring unit to achieve precise alignment between the steady-state energy consumption feature sequence and the event nodes in the activity evolution feature map in the time dimension, establishing a time alignment mapping table. The energy consumption feature vector corresponding to each time slice in the steady-state energy consumption feature sequence is associated with event nodes or composite event clusters occurring in the same time period in the activity evolution feature map through a timestamp. According to the time alignment mapping table, the steady-state energy consumption feature vector at each time point is injected into the event node within the corresponding time interval in the activity evolution feature map as an energy consumption attribute embedding for the event node. This ensures that the "oven heating" event node not only contains an event type label but also embeds vectorized features such as the average power and energy consumption curve shape during the event's duration. Based on the energy consumption attribute embedding, the correlation strength between event nodes in the activity evolution feature map is recalculated, and the weights of causal edges are updated to form an energy-enhanced activity evolution feature map. The weight update of the causal edges follows a dedicated formula that considers energy consumption transfer or sharing. The dedicated formula for updating the weights of causal edges is:

[0091]

[0092] in: This represents the updated causal edge weights. This represents the causal edge weights before the update. Represents the energy consumption impact coefficient. Represents event nodes Energy consumption attribute embedding vector With event nodes Energy consumption attribute embedding vector The cosine similarity between them is used. Graph structure encoding is performed on the energy-enhancing activity evolution feature map, transforming it into a fixed-dimensional graph feature vector. The graph feature vector is then fused with the global statistical feature vector of the steady-state energy consumption feature sequence through cross-attention, generating a multi-dimensional edge-fused feature body containing spatiotemporal and causal information.

[0093] It is understandable that the establishment of the time alignment mapping table relies on the internal timestamp alignment and flow identifier marking already completed by all feature streams during the separation phase, which provides a temporal basis for subsequent precise alignment nesting processing. It is also understandable that the weight update of causal edges reflects the modulation of the causal relationship strength between events by the similarity of actual energy consumption patterns. If two events have highly similar energy consumption patterns, there is a stronger potential causal relationship between them; for example, different workstations on the same production line exhibit synchronized energy consumption patterns. Optionally, when constructing the activity evolution feature map, the causal rule knowledge base can be dynamically updated. New causal association patterns can be learned from historical data through association rule mining algorithms and added to the knowledge base.

[0094] In one embodiment of the present invention, edge fusion feature bodies from all edge monitoring units are uploaded to a cloud center. The cloud center receives the edge fusion feature bodies uploaded by each edge monitoring unit and adds a spatial location code of its source unit to each edge fusion feature body. Based on the spatial location code, a topological connection graph representing the spatial adjacency relationship of all edge monitoring units within the monitoring area is constructed. Under the constraints of the topological connection graph, a graph neural network is used to perform message passing and feature aggregation on each edge fusion feature body, enabling each edge fusion feature body to absorb feature information from its spatial neighbors. After multiple rounds of message passing, spatial feature diffusion smoothing is performed on each updated edge fusion feature body to eliminate feature abrupt boundary between units. All smoothed edge fusion feature bodies are superimposed and interpolated in the feature space to form a continuous and consistent global collaborative feature field covering the entire monitoring area. At the cloud center, the source feature stream is traced and quantized based on the global collaborative feature field. The source feature stream is synchronously received from each edge monitoring unit, and the source feature stream contains direct emission concentration and composition data from different spatial points. The data points in the source feature stream are projected onto the spatial locations corresponding to the global collaborative feature field. Guided by a global collaborative feature field, emission source tracing analysis is performed on each projection point. This analysis infers the source category and activity process causing emissions at that projection point by matching the energy consumption and activity patterns of the projection point's location within the global collaborative feature field. Based on the emission source tracing analysis results, each source feature flow data point is assigned a source category label and an activity process label. According to a pre-defined emission factor mapping table, the source feature flow data points with source category and activity process labels are transformed into standardized carbon emission intensity units. All carbon emission intensity units are organized and arranged according to source category and spatial location to generate a quantified emission spectrum. This quantified emission spectrum records the emission intensity at different spatial locations and from different source categories in matrix form.

[0095] In specific implementation, taking an industrial park covering multiple functional areas as an example, edge monitoring units deployed in the painting workshop, assembly workshop, boiler room and office building each generate edge fusion feature bodies. The edge fusion feature bodies contain the spatiotemporal causal information of each unit after the fusion of steady-state energy consumption feature sequence and activity evolution feature map.

[0096] In practice, edge fusion feature volumes from all edge monitoring units are uploaded to the cloud center. The cloud center receives these feature volumes from each monitoring unit and adds a spatial location code to each feature volume, based on pre-mapped grid coordinates or GPS coordinates of the monitoring area. Based on these codes, a topology graph representing the spatial adjacency relationships of all edge monitoring units within the monitoring area is constructed. Nodes in the graph represent edge monitoring units, and edges represent spatial adjacency or functional connectivity between units. For example, although the edge monitoring units in the painting workshop and the boiler room are some distance apart, they are considered to have a functional connectivity edge due to their connection via steam pipes. Under the constraints of the topology graph, a graph neural network is used for message passing and feature aggregation of each edge fusion feature volume. This allows each feature volume to absorb feature information from its spatial neighbors. After multiple rounds of message passing, the edge fusion feature volume of the office building located at the edge of the park can integrate energy consumption activity pattern features from the internal production area. Each updated edge-fused feature volume undergoes spatial feature diffusion smoothing to eliminate abrupt feature boundaries between units, such as smoothing out sharp changes in nighttime energy consumption characteristics between the painting workshop and adjacent warehouse areas due to different production processes. All smoothed edge-fused feature volumes are then superimposed and interpolated in the feature space to form a continuous and consistent global collaborative feature field covering the entire monitoring area. This global collaborative feature field is spatially continuous, and its feature vector can be obtained from any location through interpolation.

[0097] In some embodiments, the source feature stream is traced and quantified at the cloud center based on a global collaborative feature field. The source feature stream is received synchronously from each edge monitoring unit. This source feature stream contains direct emission concentration and composition data from different spatial points, such as carbon dioxide and non-methane total hydrocarbon concentration readings uploaded by sensors deployed at boiler chimneys, painting workshop exhaust outlets, and factory boundary monitoring points. Data points in the source feature stream are projected onto their corresponding spatial locations in the global collaborative feature field, with each concentration reading associated with a specific geographic coordinate. Guided by the global collaborative feature field, emission source tracing analysis is performed on each projection point. This analysis infers the source category and activity process causing the emission at the projection point by matching the energy consumption and activity patterns of the location point in the global collaborative feature field. For example, if a peak volatile organic compound concentration is detected at an exhaust outlet, and the global collaborative feature field shows that the location exhibits high energy consumption and "spraying operation" activity characteristics at that time, then the emission source category is inferred to be "surface coating process." Based on the emission source tracing analysis results, each source feature stream data point is assigned a source category label and an activity process label.

[0098] Based on a pre-defined emission factor mapping table, source characteristic flow data points labeled with source category and activity process are transformed into standardized carbon emission intensity units. The transformation process uses emission factors specific to different source categories and activity processes for calculation. All carbon emission intensity units are organized and arranged according to source category and spatial location to generate a quantified emission spectrum. This quantified emission spectrum records the emission intensity at different spatial locations and from different source categories in matrix form. The specific formula used to transform source characteristic flow data points into standardized carbon emission intensity units is:

[0099]

[0100] in: This represents the calculated carbon emission intensity unit value. The concentration of greenhouse gases detected in the representative source characteristic stream. This represents the basic emission factor corresponding to the source category label. This represents the process correction factor corresponding to the activity process label. This represents the standard emission flow rate calculated based on the type of monitoring point and operating conditions.

[0101] It is understandable that the construction of the topology connection graph considers not only physical adjacency but also process flow relationships, enabling effective feature transfer between functionally tightly coupled but spatially non-adjacent units. Similarly, emission source tracing analysis relies on the rich context provided by the global collaborative feature field. While a single concentration reading itself lacks source tracing capability, matching it with the energy consumption and activity patterns at the corresponding location in the global collaborative feature field can significantly improve the accuracy of emission source identification. Optionally, spatial feature diffusion smoothing can employ an algorithm based on the principle of heat conduction equations, allowing feature values ​​to diffuse naturally from high-value regions to low-value regions in space, forming a smoothly transitioning feature field surface.

[0102] In one embodiment of the present invention, a quantitative emission spectrum-driven emission calculation kernel is used to calculate the total regional carbon emissions. The quantitative emission spectrum is input into the emission calculation kernel, which has built-in calculation logic paths corresponding to different source categories. The emission calculation kernel first parses the matrix structure of the quantitative emission spectrum, identifying all non-zero emission intensity units and their corresponding spatial coordinates and source categories. For each identified emission intensity unit, the corresponding calculation logic path is activated according to its source category. The logic path combines the numerical value of the emission intensity unit and the activity intensity characteristics of its corresponding spatial location in the global collaborative feature field to calculate the cumulative emissions of the emission intensity unit within the calculation period. The cumulative emissions of all emission intensity units in the monitoring area are spatially integrated and summed to obtain the classified emissions by source category. The classified emissions of all source categories are summarized to obtain the total carbon emissions of the monitoring area within the calculation period. A structured emission inventory is generated, with the total carbon emissions as the general framework, and the classified emissions of each source category are listed in the lower-level structure. For the classified emissions of each source category, the corresponding emission intensity unit in the quantitative emission spectrum is further correlated and traced back, and the spatial location information and activity process tags corresponding to the emission intensity unit are extracted. Spatial location information, activity process tags, and corresponding emission intensity and cumulative emissions are structurally linked to form detailed emission entries under source categories. All source category emission entries are sorted and organized according to emission size or spatial distribution. The total regional carbon emissions, emission amounts for each source category, and detailed emission entries are integrated into a hierarchical tree-structured document, which constitutes the structured emissions inventory.

[0103] In practice, the quantitative emission spectrum generated by the aforementioned industrial park is used as input. The quantitative emission spectrum is a matrix with spatial grids as rows and emission source categories as columns. Each cell in the matrix stores the standardized carbon emission intensity cell value corresponding to the location and source category.

[0104] In practice, the emission calculation kernel, driven by a quantified emission spectrum, calculates the total regional carbon emissions. The quantified emission spectrum is input into the emission calculation kernel, which contains built-in calculation logic paths corresponding to different source categories. The emission calculation kernel first parses the matrix structure of the quantified emission spectrum, identifying all non-zero emission intensity units and their corresponding spatial coordinates and source categories. For each identified emission intensity unit, the corresponding calculation logic path is activated based on its source category. This logic path combines the emission intensity unit's numerical value and the activity intensity characteristics of its corresponding spatial location in the global collaborative feature field to calculate the cumulative emissions of the emission intensity unit within the calculation period. The specific formula for calculating the cumulative emissions of a single emission intensity unit within the calculation period is:

[0105]

[0106] in: This represents the cumulative emissions of that emission intensity unit within the calculation period. This represents the numerical value of the emission intensity unit located at spatial coordinates (x, y) and with source category s in the quantified emission spectrum. This represents the feature scalar value of the global cooperative feature field at spatial coordinates (x, y) at time slice t, which characterizes the activity intensity. Represents the total number of time slices within the computation period. This represents the duration of each time slice. The cumulative emissions of all emission intensity units within the monitoring area are spatially integrated and summed to obtain the categorized emissions by source category.

[0107] In some embodiments, a structured emissions inventory is generated, with total carbon emissions as the overall framework, and the emissions of each source category listed in a hierarchical structure. For the emissions of each source category, the corresponding emission intensity units in the quantified emissions spectrum are further linked and traced back, and the spatial location information and activity process tags corresponding to the emission intensity units are extracted. The spatial location information, activity process tags, and the corresponding emission intensity and cumulative emissions are structurally bound to form detailed emission entries under the source category. The detailed emission entries of all source categories are sorted and organized according to emission size or spatial distribution. The total regional carbon emissions, the emissions of each source category, and the detailed emission entries are integrated into a hierarchical tree structure document, which is the structured emissions inventory. See Table 1.

[0108] Table 1: Schematic diagram of structured emission inventory contents

[0109]

[0110] It is understandable that the calculation logic pathways for different source categories in the emissions calculation kernel will consider the specificity of the source category. For stationary combustion sources, the activity intensity characteristics... Directly related to fuel supply rate; for process sources, activity intensity characteristics This is related to the production line's operating load rate. It can be understood that the tree structure of the structured emissions inventory not only presents the summarized results but also retains the complete link from the total emissions down to the specific spatial location and activity process through hierarchical association. The data in the detailed emissions entries comes from the specific calculated values ​​of the quantified emissions spectrum units and the global collaborative characteristic field. Optionally, when performing spatial integration and summation, the emissions calculation kernel can use numerical integration methods, such as dividing the monitoring area into finer grids and summing the cumulative emissions of all emissions intensity units within each grid.

[0111] Referring to Figure 4, this is a pie chart showing the proportion of carbon emission sources by category, illustrating the distribution of different emission sources in overall carbon emissions. It is commonly found in carbon emission analysis reports of industrial parks and enterprises. It is typically used in carbon emission accounting and environmental management reports to visually present the contribution percentage of each emission source, assisting in the formulation of emission reduction strategies. Using different colors to distinguish categories, it is a typical "composition-based" data visualization tool, suitable for displaying the proportional relationship of different parts within an overall picture. It intuitively displays the weight of each emission source, helping managers quickly identify the "biggest contributors" of carbon emissions (such as the source category with the highest proportion) and prioritize the allocation of emission reduction resources and the implementation of control measures. As one of the core charts in carbon emission inventories and environmental reports, it enhances the professionalism and readability of the reports, meeting the needs of compliance disclosure or external presentation.

[0112] In one embodiment of the present invention, the method further includes a dynamic updating and verification step for the structured emission inventory. A dynamic emission baseline model corresponding to the monitoring area is established, trained based on historical structured emission inventories. When a new structured emission inventory is generated, it is compared with the predicted inventory generated by the dynamic emission baseline model to identify significant deviations. These significant deviations are then traced back to pinpoint the quantized emission spectrum units and related edge fusion features that caused the deviations. A re-verification process for the relevant edge monitoring unit data is triggered. If the deviation disappears or decreases after re-verification, the dynamic emission baseline model is updated with the new inventory; if the deviation persists, a data anomaly alarm is generated and recorded.

[0113] In practice, for the structured emission inventory generated in the aforementioned industrial park monitoring scenarios, dynamic updating and verification steps are established and implemented.

[0114] In practice, a dynamic emission baseline model corresponding to the monitoring area is established. This model is trained based on historical structured emission inventories. The process involves collecting multi-period structured emission inventories generated within the historical period of the monitoring area as a training dataset. These inventories contain emission amounts categorized by source type, detailed emission entries, and their corresponding spatial location information and activity process labels. Using this historical inventory data, a predictive model capable of capturing the temporal patterns and characteristics of carbon emissions in the monitoring area is trained. During model training, in addition to historical emission values, input features can integrate emission-related contextual information, such as production planning, seasonal factors, or weather conditions, to enhance the model's understanding of carbon emission drivers. Through iterative learning, the model can ultimately output a predicted emission inventory for the monitoring area in the future calculation period based on a given time point or production scenario, thus providing a benchmark for subsequent deviation detection. The dynamic emission baseline model can learn the carbon emission patterns and intensities of the monitoring area under normal production and operation conditions at different time periods and in different production stages. Based on input time, planned production activities, and other information, the dynamic emission baseline model can generate a predicted emission inventory for the next calculation period. Once the new structured emissions inventory is generated, it is compared with the predicted inventory generated by the dynamic emissions baseline model. The comparison includes the total emissions for each source category and detailed emissions from key spatial locations. Significant deviations are identified. The determination of a significant deviation is based on whether the difference between a certain emission value in the new inventory and the predicted value of the corresponding item in the predicted inventory exceeds a threshold set based on historical fluctuation statistics. The specific formula used to determine significant deviations is:

[0115]

[0116] in: This represents the result of the deviation determination. This represents the actual value of a specific emission in the new structured emissions inventory. This represents the predicted value of the same emission from the dynamic emission baseline model. This represents the standard deviation of the historical values ​​of this emission. This represents the preset significance threshold coefficient.

[0117] In some embodiments, significant deviations are traced back to pinpoint the quantified emission spectrum unit that caused the deviation and its associated edge fusion feature. This back-tracing process is based on the hierarchical relationships of the structured emission inventory. For example, if a significant positive deviation is identified in the emissions of the "surface coating process" category, the process traces back to all emission intensity units in the quantified emission spectrum whose source category is "solvent use." Units with values ​​significantly higher than the predicted range are further selected, and their corresponding spatial coordinates are located. These spatial coordinates are then linked to the edge monitoring unit responsible for monitoring that area, thereby locating the original data stream upon which the associated edge fusion feature was generated. A re-verification process for the relevant edge monitoring unit data is triggered. This process includes checking the sensor's operational status, re-executing the calculation process from the initial monitoring stream to the edge fusion feature, and verifying the emission factor mapping table version. If the deviation disappears or decreases after re-verification, the dynamic emission baseline model is updated with a new inventory to adapt it to the new operating state. If the deviation persists, a data anomaly alarm is generated and recorded. The alarm information includes details of the deviation, the identified suspicious quantified emission spectrum unit, the identifier of the associated edge monitoring unit, and a cause analysis. Understandably, building a dynamic emissions baseline model is a continuous learning process. As more validated structured emissions inventories are added to the training, the dynamic emissions baseline model will become increasingly accurate in depicting carbon emission patterns in the monitored area. Optionally, the dynamic emissions baseline model can be built using time series forecasting models or machine learning models. Model inputs can include historical emissions data, production plans, weather data, and other external factors.

[0118] Referring to Figure 5, this is a comparative bar chart of causal strength, used to show the changes in the causal relationship strength of different event types before and after refactoring. It is commonly used in event analysis and process optimization scenarios in industrial systems. The causal strength of all events is higher after refactoring than before, especially for events like "environmental change" and "equipment start-up / shutdown," where the strength increase is significant. The strength of "equipment start-up / shutdown" after refactoring is close to 0.85, making it the event type with the strongest causal relationship. This type of chart typically serves industrial event causal chain optimization scenarios, demonstrating through the increased strength that the refactoring operation enhanced the accuracy of event correlation identification. High-intensity events such as equipment start-up / shutdown can be the focus of subsequent process monitoring and anomaly warnings. Based on the differences in the strength of different events, the system's analysis weight for event correlations can be adjusted to improve the decision-making efficiency of industrial systems.

[0119] 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.

[0120] 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 method for calculating carbon emissions based on distributed monitoring, characterized in that, The method includes: deploying hierarchical edge monitoring units within a monitoring area, each edge monitoring unit receiving an initial monitoring stream from its assigned sensor cluster, the initial monitoring stream containing multimodal sensor data related to emissions; performing dynamic slicing and intra-stream feature separation operations on the initial monitoring stream within the edge monitoring unit to extract raw energy flow features, activity event flow features, and source feature flow from the multimodal sensor data; and using a feature evolution engine built into the edge monitoring unit to perform multiple rounds of iterative evolution on the raw energy flow features to generate a steady-state energy consumption feature sequence with time context dependence, including: inputting the raw energy flow features into the primary evolution layer of the feature evolution engine for local pattern learning. The process involves extracting preliminary energy consumption pattern fragments; feeding these fragments into a secondary evolution layer for cross-time-slice pattern correlation mining to form an intermediate energy consumption pattern chain with time dependencies; performing a pattern stability check on the intermediate energy consumption pattern chain to filter out transient pattern nodes with fluctuations exceeding a preset threshold; inputting the stability-checked intermediate energy consumption pattern chain into the final evolution layer for pattern convergence and feature sequence reconstruction, outputting a smooth and context-coherent steady-state energy consumption feature sequence; each iteration of the feature evolution engine incorporates feedback from the previous iteration to ensure the continuity and convergence of the evolution direction; simultaneously, the event causal chain of the activity event stream features is reconstructed to form a causal chain with time dependencies. The activity evolution feature graph with causal tags includes: parsing each independent event node in the activity event flow features, identifying the type attribute, timestamp, and associated resource consumption marker of each event node; under the guidance of a pre-set causal rule knowledge base, analyzing the temporal adjacency and logical triggering relationships between event nodes, and establishing preliminary event causal relationship edges; evaluating the confidence of the preliminary event causal relationship edges, selecting strong causal edges with confidence scores higher than a set value, and logically integrating the event nodes connected by the strong causal edges to form composite event clusters; using composite event clusters as nodes and strong causal edges as connections, constructing a directed acyclic graph structure, and labeling each edge in the graph with causal type and strength weight, thereby forming a... The activity evolution feature map is constructed. After construction, loop detection and pruning are performed on the activity evolution feature map to ensure the consistency of causal logic. The steady-state energy consumption feature sequence and the activity evolution feature map are nested within the edge monitoring unit to generate an edge fusion feature body. The edge fusion feature bodies from all edge monitoring units are uploaded to the cloud center, where spatial topology fusion of cross-unit feature bodies is performed to generate a global collaborative feature field. The source feature flow is traced and quantized based on the global collaborative feature field in the cloud center to generate a quantified emission spectrum. The emission calculation kernel is driven by the quantified emission spectrum to calculate the total regional carbon emissions and generate a structured emission inventory.

2. The carbon emission calculation method based on distributed monitoring according to claim 1, characterized in that, The dynamic slicing and intra-flow feature separation operation performed on the initial monitoring stream within the edge monitoring unit includes: configuring an adaptive sliding time window for the initial monitoring stream, the width of which is dynamically adjusted based on the instantaneous data entropy of the initial monitoring stream; within the sliding time window, performing modal decoupling on the multimodal sensing data to separate the time-series mode corresponding to energy consumption, the state mode corresponding to equipment or personnel activities, and the component mode corresponding to direct emissions; performing sub-sequence segmentation on the time-series mode to form the original energy flow features; performing event boundary detection and event fragment extraction on the state mode to form the activity event flow features; and extracting concentration and component features from the component mode to form the source feature stream; wherein, the dynamic slicing and intra-flow feature separation operation further includes internal timestamp alignment and flow identifier marking for each separated feature stream.

3. The carbon emission calculation method based on distributed monitoring according to claim 1, characterized in that, The step of performing alignment and nesting processing of the steady-state energy consumption feature sequence and the activity evolution feature map within the edge monitoring unit to generate an edge fusion feature body includes: precisely aligning the steady-state energy consumption feature sequence with the event nodes in the activity evolution feature map in the time dimension to establish a time alignment mapping table; injecting the steady-state energy consumption feature vector at each time point into the event node in the corresponding time interval of the activity evolution feature map according to the time alignment mapping table, as the energy consumption attribute embedding of the event node; recalculating the correlation strength between event nodes in the activity evolution feature map based on the energy consumption attribute embedding, updating the weights of causal edges, and forming an energy-enhanced activity evolution feature map; performing graph structure encoding on the energy-enhanced activity evolution feature map to convert it into a fixed-dimensional graph feature vector; and performing cross-attention fusion of the graph feature vector and the global statistical feature vector of the steady-state energy consumption feature sequence to generate a multi-dimensional edge fusion feature body containing spatiotemporal and causal information.

4. The carbon emission calculation method based on distributed monitoring according to claim 1, characterized in that, The process of uploading edge fusion feature bodies from all edge monitoring units to the cloud center, and performing spatial topology fusion of cross-unit feature bodies at the cloud center to generate a global collaborative feature field includes: the cloud center receiving edge fusion feature bodies uploaded by each edge monitoring unit and attaching a spatial location code of its source unit to each edge fusion feature body; constructing a topology connection graph representing the spatial adjacency relationship of all edge monitoring units within the monitoring area based on the spatial location code; under the constraints of the topology connection graph, using a graph neural network to perform message passing and feature aggregation on each edge fusion feature body, so that each edge fusion feature body can absorb the feature information of its spatial neighbors; after multiple rounds of message passing, performing spatial feature diffusion smoothing processing on each updated edge fusion feature body to eliminate feature abrupt boundary between units; and superimposing and interpolating all smoothed edge fusion feature bodies in the feature space to form a continuous and consistent global collaborative feature field covering the entire monitoring area.

5. The carbon emission calculation method based on distributed monitoring according to claim 1, characterized in that, The process of tracing and quantifying source feature streams based on a global collaborative feature field at the cloud center to generate a quantified emission spectrum includes: synchronously receiving source feature streams from each edge monitoring unit, the source feature streams containing direct emission concentration and composition data from different spatial points; projecting data points in the source feature streams onto the spatial locations corresponding to the global collaborative feature field; performing emission source tracing analysis for each projection point under the guidance of the global collaborative feature field, the emission source tracing analysis inferring the source category and activity process most likely to cause emissions at the projection point by matching the energy consumption characteristic pattern and activity characteristic pattern of the location of the projection point in the global collaborative feature field; assigning a source category label and an activity process label to each source feature stream data point based on the emission source tracing analysis results; converting the source feature stream data points with source category labels and activity process labels into standardized carbon emission intensity units according to a preset emission factor mapping table; organizing and arranging all carbon emission intensity units according to source category and spatial location to generate a quantified emission spectrum, the quantified emission spectrum recording the emission intensity of different spatial locations and different source categories in matrix form.

6. The carbon emission calculation method based on distributed monitoring according to claim 5, characterized in that, The calculation of total regional carbon emissions based on a quantitative emission spectrum-driven emission calculation kernel includes: inputting the quantitative emission spectrum into the emission calculation kernel, which has built-in calculation logic paths corresponding to different source categories; the emission calculation kernel first parses the matrix structure of the quantitative emission spectrum to identify all non-zero emission intensity units and their corresponding spatial coordinates and source categories; for each identified emission intensity unit, the corresponding calculation logic path is activated according to its source category, and the logic path combines the value of the emission intensity unit and the activity intensity characteristics of its corresponding spatial location in the global collaborative feature field to calculate the cumulative emissions of the emission intensity unit within the calculation period; the cumulative emissions of all emission intensity units in the monitoring area are spatially integrated and summed to obtain the classified emissions by source category; the classified emissions of all source categories are summarized to obtain the total carbon emissions of the monitoring area within the calculation period.

7. The carbon emission calculation method based on distributed monitoring according to claim 1, characterized in that, The generation of the structured emissions inventory includes: using the total carbon emissions as the overall framework, listing the emissions of each source category in a hierarchical structure; for each source category's emissions, further associating and tracing back to the corresponding emission intensity unit in the quantified emission spectrum, and extracting the spatial location information and activity process tags corresponding to the emission intensity unit; structurally binding the spatial location information, activity process tags, and corresponding emission intensity and cumulative emissions to form detailed emission entries under the source category; sorting and organizing the detailed emission entries of all source categories according to emission size or spatial distribution; and integrating the total regional carbon emissions, the emissions of each source category, and the detailed emission entries into a hierarchical tree structure document, which is the structured emissions inventory.

8. The carbon emission calculation method based on distributed monitoring according to claim 1, characterized in that, The method also includes a dynamic update and verification step for the structured emission inventory: establishing a dynamic emission baseline model corresponding to the monitoring area, which is trained based on historical structured emission inventories; when a new structured emission inventory is generated, it is compared with the predicted inventory generated by the dynamic emission baseline model to identify significant deviations; the significant deviations are traced back to locate the quantized emission spectrum unit that caused the deviation and the related edge fusion feature body; a re-verification process is triggered for the relevant edge monitoring unit data; if the deviation disappears or decreases after re-verification, the dynamic emission baseline model is updated with the new inventory; if the deviation persists, a data anomaly alarm is generated and recorded.

Images