A method for comprehensive management of carbon emissions of internal combustion engine based on multi-source data fusion

Abstract

This invention discloses a comprehensive carbon emission management method for internal combustion engines based on multi-source data fusion, belonging to the field of carbon emission management technology. Specifically, it includes: acquiring real-time operating data of the internal combustion engine and environmental parameters; performing parallel carbon balance calculations through dual channels to generate a theoretical fuel consumption estimation sequence and an environmentally corrected measured fuel consumption back-calculation sequence; comparing the two to generate a real-time deviation sequence, and using a dynamic baseline learner to generate a dynamic deviation reference interval; when the deviation continuously exceeds the limit, determining that the monitor is drifting and triggering online calibration, while simultaneously freezing the measured channel; inputting the theoretical sequence and environmental parameters into a compensation function network calibrated with historical data to output a shadow carbon emission intensity sequence; during calibration, using this sequence as continuous accounting data and generating a calibration report. This invention can ensure the continuity and reliability of carbon emission accounting results even when emission monitoring data is abnormal.

Description

Technical Field

[0001] This invention relates to the field of carbon emission management technology, specifically to a comprehensive carbon emission management method for internal combustion engines based on multi-source data fusion. Background Technology

[0002] Precise, continuous, and legally binding monitoring and management of carbon emissions from internal combustion engines in road vehicles and industrial facilities has become a crucial step in achieving energy conservation and emission reduction goals. Current technological trends are driving a shift in regulatory models from fragmented, reactive approaches relying on annual inspections or periodic spot checks to a comprehensive, digital management model based on real-time online data.

[0003] In current technological practices, the monitoring of carbon emissions from internal combustion engines primarily relies on two types of data sources: one is to obtain engine operating parameters through on-board diagnostic systems and indirectly estimate carbon emissions based on theoretical models or empirical formulas; the other is to directly measure the concentration of key components such as carbon dioxide in exhaust gas by installing sensors such as exhaust gas analyzers to calculate carbon emissions. These methods provide the foundation for obtaining emission data. When the monitoring data is complete and valid, the relevant management platform can record, statistically analyze, and display the emission data.

[0004] However, the aforementioned methods face a core challenge in long-term practical operation: vehicle-mounted monitoring equipment may experience data gaps or severe distortions during specific time periods due to hardware failures, communication interruptions, or calibration drift caused by prolonged use. Current technologies typically lack effective mechanisms to address such incomplete data. Common practices include interrupting the calculation process or using fixed historical averages or empirical coefficients to fill in the missing periods. This simple filling method fails to reconstruct the actual vehicle operating conditions, the real traffic environment (e.g., congestion levels, average vehicle speed), and meteorological conditions (e.g., temperature, pressure) during the specific missing period. This results in a disconnect between the supplemented data and the actual situation, significantly reducing the accuracy of carbon emission accounting results across the entire time window. This makes it difficult to support refined carbon asset management, precise government regulation, and carbon trading and auditing scenarios requiring high-reliability data. Therefore, there is an urgent need for a technical solution that can automatically activate and generate high-confidence alternative data when critical monitoring data is unreliable, thereby ensuring the continuity and reliability of carbon emission accounting results. Summary of the Invention

[0005] The purpose of this invention is to provide a comprehensive management method for carbon emissions from internal combustion engines based on multi-source data fusion, addressing the following technical problems:

[0006] Current technologies lack effective solutions when internal combustion engine carbon emission monitoring data is missing or distorted. The commonly used methods of filling in historical averages or fixed coefficients cannot recreate the actual operating conditions, traffic environment, and meteorological conditions during the missing period, resulting in insufficient accuracy of carbon emission accounting results and making it difficult to meet the needs of refined carbon management, precise supervision, and highly reliable carbon audits.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A comprehensive management method for carbon emissions from internal combustion engines based on multi-source data fusion includes the following steps:

[0009] S1. Obtain real-time operating data and environmental dynamic parameters of the target internal combustion engine. The operating data includes the engine operating condition sequence and the exhaust gas concentration sequence uploaded by the on-board emission monitor. The environmental dynamic parameters include the intake air temperature sequence and the atmospheric pressure sequence.

[0010] S2. Perform carbon balance dual-channel parallel calculation. The working condition fuel consumption estimation channel queries the dynamically calibrated body fuel consumption mapping model based on the working condition sequence and outputs the theoretical fuel consumption estimation sequence. The exhaust gas back calculation channel outputs the measured fuel consumption back calculation sequence based on the exhaust gas concentration sequence and the exhaust flow rate corrected by environmental parameters.

[0011] S3. Compare the theoretical fuel consumption estimation sequence with the measured fuel consumption back-calculation sequence to generate a real-time deviation sequence, and input the sequence into a dynamic baseline learner. The learner outputs a dynamic deviation reference interval based on recent historical deviation characteristics.

[0012] S4. When the real-time deviation sequence continues to exceed the dynamic deviation reference range, it is determined that the on-board emission monitor has drifted, triggering an online calibration event and freezing the exhaust gas back-calculation channel output;

[0013] S5. Input the theoretical fuel consumption estimation sequence and environmental dynamic parameters into the compensation function network. The parameters of the compensation function network are calibrated based on the historical operating data of the target internal combustion engine, and the shadow carbon emission intensity sequence is calculated and output.

[0014] S6. During the calibration event, output the shadow carbon emission intensity sequence as continuous carbon accounting data and generate a calibration report that records drift characteristics and the identifier of the compensation network used.

[0015] As a further aspect of the present invention: in S2, the process of establishing the dynamically calibrated volumetric fuel consumption mapping model is as follows:

[0016] During an initialization period after the target internal combustion engine starts, the instantaneous fuel consumption sequence from the high-precision fuel flow meter, the speed sequence from the crankshaft position sensor, and the load rate sequence from the intake manifold pressure sensor are collected simultaneously, and the timestamps of the three sequences are aligned.

[0017] A two-dimensional operating condition plane is constructed with engine speed as the horizontal axis and load rate as the vertical axis. This plane is then uniformly divided into multiple rectangular grid cells. Each data point within the initialization period is mapped to the corresponding operating condition grid cell based on its engine speed and load rate values. For each grid cell containing data points, the median of all instantaneous fuel consumption values ​​within it is calculated and used as the reference fuel consumption feature value for that cell. Using all grid cells containing the reference fuel consumption feature value as control points, a continuous fuel consumption mapping surface is generated using a surface fitting algorithm, serving as the dynamically calibrated ontological fuel consumption mapping model. In real-time calculations, each pair of engine speed and load rate data from the engine operating condition sequence is used as coordinate points, and interpolation calculations are performed on the fuel consumption mapping surface to obtain theoretical fuel consumption estimates and output the sequence.

[0018] As a further aspect of the present invention: in step S2, the process of correcting the exhaust flow rate with environmental parameters and using it for exhaust gas back-calculation is as follows:

[0019] Read the fixed displacement parameters of the engine from the electronic control unit of the target internal combustion engine, and calculate the uncorrected theoretical exhaust volume flow rate sequence based on the real-time engine speed sequence and the displacement parameters;

[0020] The intake air temperature sensor reading sequence and atmospheric pressure sensor reading sequence, which are aligned with the engine speed sequence timestamp, are acquired synchronously. A gas density correction model is established, with real-time intake air temperature value, atmospheric pressure value and standard state reference value as inputs and gas density correction coefficient as output. Each data point of the theoretical exhaust volume flow rate sequence is multiplied by the real-time gas density correction coefficient calculated by the model at the corresponding time to obtain the corrected exhaust mass flow rate sequence. The corrected exhaust mass flow rate sequence is combined with the exhaust carbon dioxide concentration sequence and the carbon mass balance equation to calculate and back-calculate the measured fuel consumption back-calculation sequence.

[0021] As a further aspect of the present invention: in step S3, the process by which the dynamic baseline learner outputs a dynamic deviation reference interval based on recent historical deviation characteristics is as follows:

[0022] A fixed-capacity circular buffer is initialized as a historical deviation queue, storing the latest real-time deviation data points in chronological order. A statistical trigger mechanism is set up so that when a preset statistical interval is reached, the distribution characteristics of all data points in the queue are calculated. The arithmetic mean and standard deviation of all data points in the queue are calculated. The interval width coefficient bound to the target internal combustion engine model and emission level is retrieved from the vehicle attribute association table. The product of the interval width coefficient and the standard deviation is subtracted from the average value, and the result is assigned as the current lower boundary of the dynamic deviation reference interval. The product of the interval width coefficient and the standard deviation is added to the average value, and the result is assigned as the current upper boundary. When a new real-time deviation data point is generated, it is appended to the head of the queue, the old data points at the tail are removed, and a new round of feature calculation and boundary update is triggered.

[0023] As a further aspect of the present invention: in step S4, the condition for determining that the real-time deviation sequence continuously exceeds the dynamic deviation reference interval is:

[0024] A minimum sustained over-limit time threshold is set, which is determined based on the shortest time required for the engine to recover from transient fluctuations to steady state. Based on a fixed sampling frequency, the time threshold is converted into a corresponding threshold for the number of consecutive data points. The real-time deviation sequence is scanned in reverse from the current moment, and the latest consecutive data points are checked. Each data point value is determined to be greater than the current upper boundary or less than the current lower boundary, and the number of points continuously outside the interval is recorded. When the number of points continuously outside the interval reaches the threshold, and the values ​​of these points are in the same direction, the sustained over-limit condition is determined to be met.

[0025] As a further aspect of the present invention: in step S5, the calibration process of the compensation function network parameters is as follows:

[0026] Select a segment of historical data from the target internal combustion engine historical database as the calibration set. The historical real-time deviation sequence of this segment of data must be within the historical dynamic deviation reference range of the corresponding period throughout the entire process.

[0027] Historical theoretical fuel consumption estimates are extracted from the calibration set as the target value sequence for network training. Simultaneously, the intake air temperature, atmospheric pressure, and coolant temperature sequences matched with timestamps from the calibration set are extracted as feature input sequences. A compensation function network framework is constructed, which contains three independent sub-functions, each corresponding to one of the three feature inputs. Each sub-function is a polynomial with unknown coefficients. The target value sequence is used as the regression target, and the feature input sequence is used as the independent variable. The least squares method is used to solve for the unknown coefficients of all polynomials. The set of coefficients obtained by solving is associated with a version identifier and stored.

[0028] As a further aspect of the present invention: in S5, the process of calculating and outputting the shadow carbon emission intensity sequence by the compensation function network is as follows:

[0029] After the online calibration event is triggered, the system receives data points from the real-time theoretical fuel consumption estimation sequence. For each data point, the system simultaneously acquires the intake air temperature, atmospheric pressure, and engine coolant temperature at the same time. The intake air temperature is input into the calibrated intake air temperature compensation sub-function to obtain the temperature compensation factor. The atmospheric pressure is input into the calibrated atmospheric pressure compensation sub-function to obtain the pressure compensation factor. The coolant temperature is input into the calibrated coolant temperature compensation sub-function to obtain the thermal state compensation factor. The three compensation factors are multiplied together to obtain the comprehensive compensation coefficient.

[0030] Multiply the current theoretical fuel consumption estimate by the comprehensive compensation coefficient to obtain the shadow fuel consumption value, and then multiply it by the fuel carbon content conversion coefficient to obtain the shadow carbon emission intensity value. Repeat the above process for each data point and output the shadow carbon emission intensity sequence in chronological order.

[0031] As a further aspect of the present invention: in step S6, the content and format of the generated calibration report record are as follows:

[0032] When a calibration event is triggered, the Coordinated Universal Time (UTC) timestamp of the event is recorded, and the original data segment of the real-time deviation sequence before the trigger is extracted. This data segment contains the values ​​of each point, millisecond-level timestamps, and position status codes relative to the boundary at that time. The extracted data segment is subjected to pattern analysis, and it is marked as positive monotonic drift, negative monotonic drift, or non-monotonic oscillation based on monotonicity and fluctuation characteristics. The pattern markings are written into the report.

[0033] Record the network parameter version number of the compensation function called in this event, record the start and end times of the event, calculate the total duration, perform online statistics on the output shadow carbon emission intensity sequence during the event duration, calculate its average, maximum, minimum and standard deviation, and enter the statistical summary value into the report. Organize the timestamp, trigger data segment, mode tag, version number, time span and statistical summary according to the predefined key-value pair format, encapsulate and generate a calibration report file, calculate the hash check value of the file content and append it to the end of the report.

[0034] The beneficial effects of this invention are:

[0035] This invention, through parallel dual-channel carbon balance calculation and dynamic monitoring of real-time deviation sequences, can automatically trigger online calibration events when vehicle emission monitoring data drifts or fails. It immediately activates a compensation function network calibrated based on historical operating data, generating a high-confidence shadow carbon emission intensity sequence using theoretical fuel consumption estimation sequences and real-time environmental parameters, thus ensuring uninterrupted carbon emission accounting. This method abandons simple fixed-value filling, systematically compensating by integrating real-time operating conditions, traffic environment, and precise meteorological parameters. This allows the supplemented data to accurately reflect the actual operating status within a specific time window, significantly improving the accuracy of reconstructing carbon emission data for missing or distorted periods.

[0036] The dynamic baseline learner constructed in this invention can adaptively learn the individualized deviation characteristics of the learning device, achieving precise customization of anomaly criteria and further enhancing the reliability of state judgment. Throughout the process, key data, calibration events, and the model version used are fully recorded and a structured report with hash verification values ​​is generated. Combined with blockchain evidence storage capabilities, a complete, tamper-proof, and auditable data traceability chain is formed, providing solid legal validity and audit support for carbon emission accounting results. Ultimately, it achieves continuous, accurate, and reliable management of internal combustion engine carbon emissions under complex real-world operating environments. Attached Figure Description

[0037] The invention will now be further described with reference to the accompanying drawings.

[0038] Figure 1 This is a flowchart illustrating the present invention. Detailed Implementation

[0039] 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.

[0040] Please see Figure 1 As shown, this invention is a comprehensive management method for carbon emissions from internal combustion engines based on multi-source data fusion, comprising the following steps:

[0041] S1. Obtain real-time operating data and environmental dynamic parameters of the target internal combustion engine. The operating data includes the engine operating condition sequence and the exhaust gas concentration sequence uploaded by the on-board emission monitor. The environmental dynamic parameters include the intake air temperature sequence and the atmospheric pressure sequence.

[0042] S2. Perform carbon balance dual-channel parallel calculation. The working condition fuel consumption estimation channel queries the dynamically calibrated body fuel consumption mapping model based on the working condition sequence and outputs the theoretical fuel consumption estimation sequence. The exhaust gas back calculation channel outputs the measured fuel consumption back calculation sequence based on the exhaust gas concentration sequence and the exhaust flow rate corrected by environmental parameters.

[0043] S3. Compare the theoretical fuel consumption estimation sequence with the measured fuel consumption back-calculation sequence to generate a real-time deviation sequence, and input the sequence into a dynamic baseline learner. The learner outputs a dynamic deviation reference interval based on recent historical deviation characteristics.

[0044] S4. When the real-time deviation sequence continues to exceed the dynamic deviation reference range, it is determined that the on-board emission monitor has drifted, triggering an online calibration event and freezing the exhaust gas back-calculation channel output;

[0045] S5. Input the theoretical fuel consumption estimation sequence and environmental dynamic parameters into the compensation function network. The parameters of the compensation function network are calibrated based on the historical operating data of the target internal combustion engine, and the shadow carbon emission intensity sequence is calculated and output.

[0046] S6. During the calibration event, output the shadow carbon emission intensity sequence as continuous carbon accounting data and generate a calibration report that records drift characteristics and the identifier of the compensation network used.

[0047] In a preferred embodiment of the present invention, the process of establishing the dynamically calibrated bulk fuel consumption mapping model in step S2 is as follows:

[0048] First, a specific initialization period is selected. This period begins when the target internal combustion engine completes a cold start and the engine coolant temperature sensor reading stabilizes between 85 and 95 degrees Celsius, and its duration is typically set to 30 minutes. This condition is chosen to ensure the engine enters a thermally stable operating state, eliminating the impact of abnormal conditions such as incomplete fuel evaporation and unstable combustion during the cold start phase on basic data acquisition. During this period, the system simultaneously triggers three high-frequency data acquisition threads.

[0049] The first thread collects instantaneous fuel consumption values ​​from a high-precision Coriolis mass fuel flow meter, which is directly connected in series after the engine's high-pressure fuel rail. Its sampling frequency is set to 10 Hz, recording a fuel consumption value every 0.1 seconds, in grams per second. The second thread collects speed signals from the engine crankshaft position sensor, calculating the instantaneous engine speed in revolutions per minute by measuring the number of crankshaft position pulses per unit time, also at a sampling frequency of 10 Hz. The third thread collects readings from the intake manifold absolute pressure sensor. This reading, combined with the current atmospheric pressure, is used to calculate the engine load rate, also at a sampling frequency of 10 Hz. To ensure the validity of subsequent analysis, a unified timestamp is generated for each sampling moment during data acquisition, with millisecond-level precision. This forces the three sets of data sequences to be perfectly aligned in the time dimension, forming a set of data points with one-to-one correspondence between timestamps, fuel consumption values, speed values, and load rate values.

[0050] After obtaining the initial dataset, the model building phase begins. First, a two-dimensional operating plane for the engine is defined. The horizontal axis of this plane represents engine speed, with its range determined based on the actual operating range of the internal combustion engine model, for example, from an idle speed of 800 rpm to the maximum torque point speed of 2200 rpm. The vertical axis represents the engine load rate, ranging from 0% to 100%. This two-dimensional plane is then uniformly discretized along the speed and load rate directions. For example, the speed axis is divided into intervals of 100 rpm, and the load rate axis is divided into intervals of 10%, thus generating multiple rectangular operating condition grid cells. Each grid cell represents a specific, narrow range of engine operating conditions.

[0051] Next, each data point collected during the initialization period is traversed. Based on the speed and load rate values ​​of each data point, it is precisely mapped to the corresponding rectangular grid cell in the two-dimensional operating plane. For example, a data point with a speed of 1250 rpm and a load rate of 65% will be assigned to a specific grid cell with a speed range of 1200-1300 rpm and a load rate range of 60%-70%.

[0052] For each operating condition grid cell containing at least one mapped data point, a baseline fuel consumption characteristic value is calculated. The calculation method is as follows: extract the instantaneous fuel consumption values ​​corresponding to all data points falling within the cell, and then calculate the median of these fuel consumption values. The median is chosen instead of the arithmetic mean to effectively resist interference from potential instantaneous measurement outliers or noise points on the baseline value. The calculated median is defined as the standard state baseline fuel consumption characteristic value for that specific operating condition grid cell. Its physical meaning can be understood as a typical and common fuel consumption rate level when the engine is running stably within that speed and load rate range.

[0053] After obtaining the baseline fuel consumption characteristic values ​​of a considerable number of grid cells, a continuous mathematical model can be constructed using these discrete baseline points. Specifically, a non-uniform rational B-spline surface fitting algorithm is employed. The center point of each grid cell containing the baseline value (the midpoint of the cell's speed and load rate range) is used as the control point for surface fitting, and its corresponding baseline fuel consumption characteristic value is used as the Z-coordinate value of that control point. The NURBS surface algorithm can generate a smooth, continuous mathematical surface that can extrapolate a certain range based on these discrete control points. This surface is the final dynamically calibrated "body fuel consumption mapping model." Essentially, it is a continuous function with speed and load rate as input variables and theoretical fuel consumption estimates as output variables. In subsequent real-time calculations, when a set of speed and load rate data from the engine's real-time operating condition sequence is obtained, such as a speed of 1520 rpm and a load rate of 48%, this data pair is treated as two-dimensional coordinate points. By querying or interpolating to calculate the corresponding Z value on the NURBS surface, the theoretical fuel consumption estimate corresponding to this operating condition, calibrated based on the machine's historical measured data, can be quickly obtained, thus forming a continuous theoretical fuel consumption estimate sequence.

[0054] In step S2, the process of correcting the exhaust flow rate with environmental parameters and using it for exhaust gas back-calculation is as follows:

[0055] The starting point of the exhaust gas reverse calculation channel is obtaining accurate exhaust mass flow rate. First, the engine's fixed structural parameters are read from the read-only memory of the target internal combustion engine's electronic control unit, namely, the displacement per cylinder, for example, 2.5 liters, and the number of cylinders, for example, 6 cylinders. From this, the total engine displacement is calculated to be 15 liters. According to the working principle of a four-stroke engine, every two revolutions of the engine, each cylinder completes one exhaust stroke, and the volume of gas discharged is theoretically equal to the total engine displacement. Therefore, the uncorrected theoretical exhaust volume flow rate can be calculated using the following relationship: exhaust volume flow rate equals the total engine displacement multiplied by the real-time engine speed, divided by 2, and then multiplied by 60 (used to convert the unit from liters per minute to cubic meters per hour). Based on this principle, the real-time collected engine speed sequence, for example, a set of speed values ​​arranged by time, is substituted into the calculation to obtain the corresponding uncorrected theoretical exhaust volume flow rate sequence, in cubic meters per hour.

[0056] However, this volumetric flow rate does not account for changes in air density under varying environmental conditions. To convert it to a more accurate mass flow rate, environmental parameters must be introduced for correction. The system simultaneously acquires readings from the intake air temperature sensor and the atmospheric pressure sensor, ensuring that their timestamps are strictly aligned with the engine speed sequence. The intake air temperature sensor, typically located after the air filter, measures the temperature of the air entering the engine, in degrees Celsius or Kelvin. The atmospheric pressure sensor measures the ambient atmospheric pressure, in kilopascals.

[0057] Next, a "gas density correction model" is established. The physical basis of this model is the ideal gas law. The model sets a standard reference state, for example, a temperature of 25 degrees Celsius and a pressure of 101.325 kPa. The core of the model is calculating the ratio of the air density under actual conditions to the air density under the standard reference state, i.e., the gas density correction coefficient. According to the ideal gas law, with constant gas composition, density is directly proportional to absolute pressure and inversely proportional to absolute temperature. Therefore, the formula for calculating the correction coefficient can be expressed as: divide the actual atmospheric pressure by the standard reference pressure, then multiply by the standard reference temperature and divide by the actual intake air temperature (the temperature value needs to be converted to absolute temperature, i.e., Kelvin). This formula is the specific mathematical expression of the gas density correction model; its inputs are real-time intake air temperature and atmospheric pressure, and its output is a dimensionless gas density correction coefficient.

[0058] Subsequently, the theoretical exhaust volumetric flow rate sequence is corrected point by point. For each data point in the sequence, the corresponding inlet air temperature and atmospheric pressure values ​​are obtained and input into the aforementioned gas density correction model to calculate a gas density correction coefficient in real time. Then, the theoretical exhaust volumetric flow rate value is multiplied by this correction coefficient, which physically converts the volumetric flow rate into the exhaust mass flow rate based on the actual air density. For example, in high-altitude, low-pressure, and high-temperature environments, the calculated correction coefficient may be less than 1, resulting in a corrected exhaust mass flow rate that is less than the theoretical volumetric flow rate, which is consistent with reality. After performing this operation on all data points, a more physically accurate "corrected exhaust mass flow rate sequence" is obtained, expressed in kilograms per hour.

[0059] Finally, this corrected exhaust mass flow rate sequence is coupled with the exhaust carbon dioxide concentration sequence uploaded by the vehicle emission monitor for calculation. Exhaust carbon dioxide concentration is usually expressed as a volume fraction percentage. According to the carbon mass balance principle, carbon in fuel is ultimately converted into carbon dioxide and emitted. Therefore, the mass of fuel consumed can be deduced from the mass of emitted carbon dioxide. The specific calculation process is as follows: First, the carbon dioxide volume concentration, the corrected exhaust mass flow rate, and the carbon dioxide gas density (which needs to be corrected according to the exhaust temperature and pressure at the time, or an empirical value) are combined to calculate the mass of carbon dioxide emitted per unit time. Then, based on the fact that the mass fraction of carbon in a carbon dioxide molecule is approximately 27.29%, and the mass fraction of carbon in the fuel used (such as diesel) is approximately 86%, a series of proportional conversions can be used to deduce the instantaneous fuel consumption corresponding to the generation of this carbon dioxide. This calculation is performed at all time points, ultimately forming the "measured fuel consumption reverse calculation sequence." This sequence is completely synchronized in time with the "theoretical fuel consumption estimation sequence" generated in the first part, providing a basis for subsequent real-time comparison and deviation analysis.

[0060] In another preferred embodiment of the present invention, in step S3, the process by which the dynamic baseline learner outputs a dynamic deviation reference interval based on recent historical deviation characteristics is as follows:

[0061] The core concept is to establish and run a dynamic baseline learner. The fundamental task of this learner is to analyze the real-time deviation sequence generated by comparing the theoretical fuel consumption estimation sequence with the measured fuel consumption back-calculation sequence, and based on the recent historical deviation performance of the equipment, dynamically generate a reference range—the dynamic deviation reference interval—to determine whether the current data is normal. This process abandons the traditional method of using fixed thresholds, thus enabling it to adapt to the characteristic differences of different internal combustion engines and the slow performance changes that may occur in the same equipment over time.

[0062] The operation of the dynamic baseline learner begins with the initialization of the data structure. Specifically, a fixed-capacity, first-in-first-out (FIFO) circular buffer is allocated in memory, and this buffer is defined as a historical deviation queue. The queue's capacity is set to a fixed value, such as 1000 data points, meaning it always stores the 1000 most recently generated real-time deviation values. Each real-time deviation value is a signed numerical value obtained by subtracting the measured fuel consumption from the theoretical fuel consumption estimate at the same moment, and its unit can be grams per second or a percentage. New deviation data points are always added to the head of the queue in chronological order, while the oldest data points are automatically removed from the tail of the queue, thus ensuring that the queue always contains the most recent historical deviation records.

[0063] Next, a triggering mechanism for driving statistical calculations needs to be set. Instead of performing complex statistical calculations immediately every time a new data point is added, a preset statistical interval is set, such as triggering a calculation every 10 seconds. The purpose of this is to strike a balance between computational timeliness and processor resource consumption. When the 10-second interval is reached, the dynamic baseline learner performs distribution feature calculations on all data points currently stored in the historical deviation queue, assuming the queue is full (containing 1000 points).

[0064] The calculation of distribution characteristics mainly involves solving for two key statistics. The first is the arithmetic mean, which represents the overall central tendency of recent deviations. It is calculated by summing the 1000 deviation values ​​and then dividing by 1000. The second is the standard deviation, which reflects the dispersion or volatility of these deviation values ​​relative to the mean. The calculation process involves first calculating the difference between each deviation value and the mean, squared this difference, then averaging all 1000 squared values, and finally taking the square root of the average. The mean and standard deviation together describe the statistical distribution of recent deviation data.

[0065] The mean and standard deviation alone are insufficient to directly define a reasonable reference range, because the allowable normal deviation range between the two calculation channels may differ reasonably for different models and emission standards of internal combustion engines. Therefore, an adjustment parameter, the interval width coefficient, needs to be introduced. This coefficient is not obtained through real-time calculation but is pre-calibrated based on a large database of experiments and vehicle models and stored in a vehicle attribute association table. This table uses the engine model and the factory emission certification level, such as China VI A or China VI B, as the joint search key. When the dynamic baseline learner operates, it retrieves the corresponding interval width coefficient from this table based on the specific model and emission level of the target internal combustion engine. For example, the coefficient retrieved for a China VI B engine model might be 2.0, while for another China VI A engine model, the coefficient might be 2.5. The larger the coefficient, the wider the defined interval range and the higher the tolerance for deviation.

[0066] After obtaining the mean, standard deviation, and interval width coefficient, the specific boundaries of the dynamic deviation reference interval can be calculated. The lower boundary is calculated by subtracting the product of the interval width coefficient and the standard deviation from the mean. The upper boundary is calculated by adding the product of the interval width coefficient and the standard deviation to the mean. For example, assuming the calculated recent deviation mean is 5 grams per second, the standard deviation is 3 grams per second, and the retrieved interval width coefficient is 2.0, then the lower boundary of the current dynamic deviation reference interval is 5 minus 2.0 multiplied by 3, which equals -1 gram per second; the upper boundary is 5 plus 2.0 multiplied by 3, which equals 11 grams per second. This means that, based on the recent performance of this internal combustion engine, the deviation of its two channels has a high probability of falling within the range of -1 to 11 grams per second. This interval is not constant; it is continuously updated during the learning process.

[0067] The update mechanism is closely linked to the data inflow. Whenever a new real-time deviation data point is generated and added to the head of the historical deviation queue, and an old point is removed from the tail of the queue, if the statistical trigger time of 10 seconds has been reached, the mean and standard deviation will be recalculated immediately using the new queue data, and the upper and lower boundary values ​​of the dynamic deviation reference interval will be updated accordingly. In this way, the reference interval achieves synchronous "sliding" and self-adaptation with the recent operating status of the equipment, and can follow the slow changes of the equipment itself. For example, as the engine wears, the reference deviation between the two channels may gradually increase, and the reference interval will shift and adjust accordingly.

[0068] In another preferred embodiment of the present invention, in step S4, the condition for determining that the real-time deviation sequence continuously exceeds the dynamic deviation reference interval is:

[0069] This step defines a clear and quantifiable criterion to identify whether the on-board emissions monitor has experienced a drift or malfunction requiring intervention, specifically the meaning of "continuously exceeding the dynamic deviation reference range." This criterion must be set rigorously enough to capture genuine abnormal trends while effectively filtering out transient, harmless random fluctuations or momentary disturbances.

[0070] The first step in the judgment process is to establish a basic time scale, called the minimum sustained over-limit time threshold. This threshold is determined based on engineering experience and the engine's dynamic characteristics. The fundamental principle is that this time length should be greater than the duration of brief, drastic fluctuations in operating conditions that may occur during typical operation of the target internal combustion engine. For example, a vehicle's rapid acceleration to overtake or a sudden loading of construction machinery can cause instantaneous changes in engine speed and load. This transient process may cause a temporary mismatch in the calculation results between the two channels, resulting in a large instantaneous deviation. However, this deviation will quickly disappear as the operating conditions stabilize. Therefore, the minimum sustained over-limit time threshold must be longer than this normal transient response time. For medium-sized diesel engines, this threshold is typically set between 30 and 120 seconds. In a specific embodiment, it can be set to 60 seconds, meaning that only when the abnormal deviation persists for more than 60 seconds is it considered that it may not be a transient interference, but rather a substantial sensor drift.

[0071] Since the real-time deviation sequence consists of discrete data points sampled at a fixed frequency, the aforementioned physical time threshold needs to be converted into a data point count threshold that is easier for the program logic to judge. This conversion relies on a fixed data sampling frequency. Assuming the unified sampling frequency for data acquisition and calculation throughout the method is 10 Hz, meaning a new deviation data point is generated every 0.1 seconds, then the threshold for the number of consecutive data points corresponding to a 60-second time threshold is 60 seconds multiplied by 10 Hz, equal to 600 points. This means the program will subsequently check if 600 consecutive deviation data points exhibit abnormalities.

[0072] The decision-making logic begins execution from the latest calculation time. The program scans the real-time deviation sequence backward along the time axis, that is, it checks sequentially from the latest point to slightly earlier points. It retrieves the latest continuous segment of data points, the number of which must reach at least the threshold for the number of points to be judged, such as the latest 600 points.

[0073] For each of these 600 consecutive data points, the program performs two checks. The first check is to see if the value of that point is greater than the current upper boundary of the dynamic deviation reference interval. The second check is to see if the value is less than the current lower boundary. If either condition is met, the point is considered "outside the interval." The program checks these points one by one and records how many consecutive points are outside the interval starting from the latest point. For example, if the check finds that the latest 1st to 605th points are all greater than the upper boundary, but the 606th point returns to the interval, then the number of consecutive points outside the interval is 605.

[0074] Simply stating "continuous exceedances" is insufficient. To prevent misjudging an abnormal state—oscillating around the upper and lower boundaries—as a unidirectional drift requiring calibration, a "directional consistency" requirement is added to the conditions. This means that all data points continuously outside the interval must deviate in the same direction. Specifically, they must all be greater than the upper boundary or all be less than the lower boundary. Some cannot be greater than the upper boundary while others are less than the lower boundary. In the example above, the latest 605 points are all greater than the upper boundary, thus satisfying the directional consistency requirement.

[0075] Ultimately, the logic for determining whether the "continuous exceedance condition" is met is as follows: First, the number of data points continuously outside the range reaches or exceeds a preset threshold of 600; second, these continuously exceeding points maintain a consistent deviation direction. Only when both conditions are met simultaneously will the program ultimately determine that the on-board emissions monitor has experienced a drift requiring a system response, thereby triggering subsequent online calibration events. This dual-condition setting significantly improves the accuracy of status determination and the ability to resist transient interference, avoiding false alarms caused by brief changes in operating conditions.

[0076] In another preferred embodiment of the present invention, the calibration process of the compensation function network parameters in step S5 is as follows:

[0077] Parameter calibration is a prerequisite for the correct operation of the compensation function network. Its goal is to use historical operating data of the internal combustion engine in a "healthy" state (i.e. when the deviation between the two calculation channels is normal) to learn the quantitative influence relationship between environmental and thermal state parameters on the theoretical fuel consumption estimate.

[0078] First, a calibration set must be carefully selected from the long-term historical database of the target internal combustion engine. There is a golden rule for selecting the calibration set: the historical real-time deviation sequence calculated by the dynamic baseline learner within this historical period must remain entirely within the historical dynamic deviation reference range for the corresponding period. This condition ensures that the selected data completely excludes periods where the on-board emissions monitor itself exhibits significant drift or malfunction, because in such periods, the measured fuel consumption back-calculation sequence is unreliable and cannot be used as a learning benchmark. For example, the program can automatically scan data from the past six months, searching for a period of 30 consecutive days where at least 95% of the deviation data points fall within the dynamic range of that day, and using this 30-day data as the candidate calibration set.

[0079] After determining the calibration set, two main data sequences need to be extracted. The first is the sequence of historical theoretical fuel consumption estimates, which serves as the learning target. This sequence is calculated by the operating condition fuel consumption estimation channel within the historical period based on the engine operating condition data at that time. It is considered a relatively reliable fuel consumption reference benchmark for this internal combustion engine, excluding direct sensor failure factors. The second is the feature sequence, which serves as the model input and needs to be strictly synchronized with the target sequence in time. These feature sequences include: intake air temperature sensor readings, atmospheric pressure sensor readings, and engine coolant temperature sensor readings. Coolant temperature is an important new feature, characterizing the engine's thermal state, which significantly affects combustion efficiency and friction loss, thus having a systematic impact on actual fuel consumption. All feature sequences are aligned and matched with the target sequence using millisecond-level timestamps to ensure that each theoretical fuel consumption estimate corresponds to a specific set of intake air temperature, atmospheric pressure, and coolant temperature values ​​at a given moment.

[0080] Next, we construct the mathematical model framework for the compensation function network. This framework is designed to contain three parallel, independent sub-functions. The first sub-function specifically handles the influence of intake air temperature characteristics and is called the intake air temperature compensation sub-function. The second sub-function specifically handles the influence of atmospheric pressure characteristics and is called the atmospheric pressure compensation sub-function. The third sub-function specifically handles the influence of coolant temperature characteristics and is called the coolant temperature compensation sub-function. Each sub-function is expressed using a polynomial with unknown coefficients. The order of the polynomial can be predetermined based on prior knowledge or through cross-validation. For example, the intake air temperature compensation sub-function might be expressed as a quadratic polynomial, with the expression a0 + a1×T + a2×T. 2 Where T represents the intake air temperature, and a0, a1, and a2 are the unknown coefficients to be determined. The atmospheric pressure compensation subfunction may take a simpler first-order polynomial form, such as b0 + b1 × P, where P represents atmospheric pressure. The coolant temperature compensation subfunction may take a cubic polynomial form, such as c0 + c1 × Tc + c2 × Tc. 2 +c3×Tc 3 Where Tc represents the coolant temperature. The outputs of these polynomials are physically referred to as "compensation factors," which are dimensionless values ​​that characterize the adjustment factor required for the theoretical fuel consumption estimate under specific environmental or thermal conditions.

[0081] The core of calibration is to use mathematical methods to solve for the unknown coefficients contained in all the polynomials mentioned above, including a0, a1, a2, b0, b1, c0, c1, c2, c3, using the data. This is essentially a multiple regression problem. The data of the entire calibration set is organized as follows: the target value sequence (Y) consists of N historical theoretical fuel consumption estimates, and the feature matrix (X) has N rows, each containing three eigenvalues: the intake air temperature T, atmospheric pressure P, and coolant temperature Tc at the corresponding time. The total output of the compensation function network can be seen as the product of the theoretical fuel consumption estimate and the outputs of the three sub-functions. To facilitate linear solution, the logarithm of both sides of the model equation is usually taken, transforming the multiplicative relationship into an additive relationship. After this transformation, the problem can be approximately transformed into a linear regression problem about the logarithmic compensation factor. Specifically, a design matrix can be constructed, which includes the intake air temperature T, Tc, and Tc. 2 Atmospheric pressure P, coolant temperature Tc, Tc 2 Tc 3 Various factors (depending on the preset polynomial order) are considered. Then, the classic least squares method is used to solve for these factors. The goal of least squares is to find a set of coefficients that minimizes the sum of squared errors between the network's predicted "compensated fuel consumption" and the target fuel consumption value sequence. By solving a normal equation, the best estimates of all unknown coefficients can be calculated at once.

[0082] Once the solution is obtained, a complete set of compensation function network parameters specific to this target internal combustion engine is obtained, namely the specific coefficients of all sub-function polynomials. This set of parameters is permanently stored in non-volatile memory. Simultaneously, for ease of management and traceability, a unique version identifier is generated and associated with it, such as "CalibVer_20231027_Device1234". This identifier and the parameter set together form the basis for subsequent online calculations.

[0083] In S5, the process of calculating and outputting the shadow carbon emission intensity sequence using the compensation function network is as follows:

[0084] For each newly arrived theoretical fuel consumption estimate data point from the operating condition fuel consumption estimation channel, such as an instantaneous fuel consumption estimate of 205 grams per second, the program immediately and synchronously captures three environmental parameters at the same sampling time. This is typically achieved by querying a shared timestamp data buffer to ensure that the acquired intake air temperature, atmospheric pressure, and engine coolant temperature values ​​strictly correspond to that fuel consumption point. Assume the captured values ​​are: intake air temperature 15 degrees Celsius, atmospheric pressure 98.0 kPa, and coolant temperature 88 degrees Celsius.

[0085] Next, these real-time parameter values ​​are input into the corresponding calibrated sub-functions for calculation. The intake air temperature value of 15 degrees Celsius is substituted into the calibrated intake air temperature compensation sub-function polynomial a0 + a1 × 15 + a2 × 225. Assuming the calibrated coefficient results in a calculation of 1.02, this 1.02 is the temperature compensation factor, meaning that under the current intake air temperature of 15 degrees Celsius, the theoretical fuel consumption estimate may need to be increased by 2% to compensate for the impact of increased cold air density.

[0086] Substitute the atmospheric pressure value of 98.0 kPa into the calibrated atmospheric pressure compensation subfunction polynomial b0 + b1 × 98.0. Assuming the calculation result is 0.99, this 0.99 is the pressure compensation factor, which means that under the current atmospheric pressure of 98.0 kPa, the theoretical fuel consumption estimate may need to be adjusted downward by 1%.

[0087] Substituting the coolant temperature value of 88 degrees Celsius into the calibrated coolant temperature compensation subfunction polynomial c0 + c1 × 88 + c2 × 7744 + c3 × 681472, and assuming the calculation result is 1.05, this 1.05 is the thermal state compensation factor. Its meaning may be that the engine has not yet reached its optimal thermal equilibrium temperature (e.g., the optimal temperature is 90-95 degrees Celsius), and the combustion efficiency is slightly lower at the current temperature of 88 degrees Celsius. Therefore, the theoretical fuel consumption estimate needs to be increased by 5% to compensate for this efficiency loss.

[0088] Then, the three compensation factors from different dimensions are multiplied together: 1.02 multiplied by 0.99 multiplied by 1.05, to obtain the comprehensive compensation coefficient, for example, approximately 1.06. This comprehensive coefficient of 1.06 represents the total adjustment required to the theoretical fuel consumption estimate under the current specific environmental and thermal conditions.

[0089] Next, the shadow fuel consumption value is calculated. The current theoretical fuel consumption estimate of 205 grams per second is multiplied by the comprehensive compensation coefficient of 1.06, resulting in approximately 217.3 grams per second. This 217.3 grams per second is the shadow fuel consumption value, which is a more realistic fuel consumption estimate obtained after systematic error correction for environmental and thermal conditions, based on the theoretical estimate.

[0090] Finally, the shadow fuel consumption value is converted into carbon emissions. There is a fixed fuel carbon content conversion factor depending on the type of fuel used. For example, for diesel, this factor is approximately 3.15 (meaning that for every 1 gram of diesel consumed, approximately 3.15 grams of carbon dioxide are produced). Multiplying the shadow fuel consumption value of 217.3 grams per second by the conversion factor 3.15 yields approximately 684.5 grams per second. This 684.5 grams per second is the final output shadow carbon intensity value, which characterizes the instantaneous carbon emission rate of the internal combustion engine at the current moment.

[0091] The above process is executed iteratively for each theoretical fuel consumption estimation data point reached during the online calibration event. Each point, based on its corresponding real-time environmental parameters, has a unique shadow carbon emission intensity value calculated through a compensation function network. All these intensity values ​​calculated in chronological order are concatenated to form a continuous sequence of shadow carbon emission intensity values. This sequence serves as the core data source, continuously outputting to the carbon accounting module until the on-board emission monitor is repaired and verified. This ensures the continuity and data reliability of carbon emission monitoring and accounting, overcoming the data stream interruption defects of traditional methods in the event of sensor failure.

[0092] In another preferred embodiment of the present invention, in step S6, the content and format of the generated calibration report record are as follows:

[0093] The generation of the calibration report begins at the instant the event is triggered. The first step is to accurately record the moment the event occurred, using a Coordinated Universal Time (UTC) timestamp with microsecond-level precision, for example, recording it as 14:35:18.239156 microseconds on October 27, 2023. This timestamp serves as the benchmark for reporting all subsequent time-related information.

[0094] Next, the program automatically extracts a segment of raw real-time deviation sequence data prior to the trigger. The length of this data segment is typically preset to twice the threshold number of data points used to determine the "continuous exceedance" condition, ensuring complete coverage of the abnormal phase leading to the determination and its preceding context. For example, if the determination threshold is 600 consecutive points, then 1200 data points before the trigger point might be extracted. This raw data is completely saved, with each point containing three elements: the deviation value itself, a millisecond-level acquisition timestamp, and a position status code. This code is a simple integer identifier indicating the relative position of the data point to the dynamic deviation reference interval at that time (i.e., the moment the point was generated). For example, code 0 can be defined as the data point being inside the interval, code 1 as the data point being greater than the upper boundary, and code -1 as the data point being less than the lower boundary. Saving the raw data and its status codes allows any third party to reproduce the triggering logic afterward.

[0095] After data capture, the program calls a pattern analysis module to automatically analyze the data. This module categorizes the triggered abnormal deviation sequence into several preset typical patterns. The specific analysis is based on the monotonicity and fluctuation characteristics of the sequence. First, it calculates the first-order difference (i.e., the difference between adjacent points) of the sequence over the entire time period and counts the durations of positive and negative differences. If more than 80% of the difference values ​​are positive and the sequence shows a clear upward trend, it is marked as a positive monotonic drift pattern, which usually corresponds to a positive linear or exponential drift in the sensor signal. Conversely, if more than 80% of the difference values ​​are negative and the overall trend is downward, it is marked as a negative monotonic drift. If the sequence's difference values ​​alternate frequently between positive and negative, without a clear unidirectional trend, or exhibit periodic up-and-down fluctuations, it is marked as a non-monotonic oscillation pattern, which may indicate intermittent sensor malfunctions or strong periodic interference. The resulting pattern label, such as "positive monotonic drift," is written to the report as a string.

[0096] The report then records the parameter version number of the compensation function network used in this calibration event, for example, "CalibVer_20230915". It also records the start and end times of the calibration event. The start time is the trigger moment, and the end time is the moment when the exhaust gas back-calculation channel resumes output after repair and verification. The difference between the two is the total duration of the event, for example, 3560 seconds.

[0097] During the calibration event, online sliding statistics were performed on the shadow carbon emission intensity sequence output in real time by the compensation function network. The statistics included the arithmetic mean of all data points in the sequence, the maximum and minimum values ​​in the sequence, and the standard deviation of the sequence. These statistical summary values ​​were calculated in real time and finally entered into a report at the end of the event to provide a general description of the overall level and volatility of the surrogate data throughout the event.

[0098] Finally, all the above information items are structurally encapsulated. Each piece of information is organized into a key-value pair. For example, "event_timestamp_utc" is the key, and its corresponding value is "2023-10-27T14:35:18.239156Z"; "trigger_data_segment" is the key, and its value is an array containing 1200 points, each point being a sub-object containing a value, timestamp, and status code; "fault_mode" is the key, and its value is "POSITIVE_MONOTONIC_DRIFT". Using this predefined key-value pair format, such as JSON or XML, organizing all the key-value pairs generates a complete calibration report file that is machine-readable and clearly structured.

[0099] To ensure the integrity and tamper-proof nature of the report, a cryptographic hash checksum of the entire text content is calculated after the file is packaged. The SHA256 algorithm is typically used. This algorithm is applied to all bytes of the report file, generating a 64-character hexadecimal string, such as "a1b2c3d4e5f6...". This hash string is then appended to the end of the report file as the final key-value pair, or stored separately and associated with the report. Any minor modification to the report content will cause a drastic change in the hash value, thus providing an efficient mechanism for verifying content integrity.

[0100] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A comprehensive management method for carbon emissions from internal combustion engines based on multi-source data fusion, characterized in that, Includes the following steps: S1. Obtain real-time operating data and environmental dynamic parameters of the target internal combustion engine. The operating data includes the engine operating condition sequence and the exhaust gas concentration sequence uploaded by the on-board emission monitor. The environmental dynamic parameters include the intake air temperature sequence and the atmospheric pressure sequence. S2. Perform carbon balance dual-channel parallel calculation. The working condition fuel consumption estimation channel queries the dynamically calibrated body fuel consumption mapping model based on the working condition sequence and outputs the theoretical fuel consumption estimation sequence. The exhaust gas back-calculation channel outputs the measured fuel consumption back-calculation sequence based on the exhaust gas concentration sequence and the exhaust flow rate corrected by environmental parameters. S3. Compare the theoretical fuel consumption estimation sequence with the measured fuel consumption back-calculation sequence to generate a real-time deviation sequence, and input the sequence into a dynamic baseline learner. The learner outputs a dynamic deviation reference interval based on recent historical deviation characteristics. S4. When the real-time deviation sequence continues to exceed the dynamic deviation reference range, it is determined that the on-board emission monitor has drifted, triggering an online calibration event and freezing the exhaust gas back-calculation channel output; S5. Input the theoretical fuel consumption estimation sequence and environmental dynamic parameters into the compensation function network. The parameters of the compensation function network are calibrated based on the historical operating data of the target internal combustion engine, and the shadow carbon emission intensity sequence is calculated and output. S6. During the calibration event, output the shadow carbon emission intensity sequence as continuous carbon accounting data and generate a calibration report that records the drift characteristics and the identifier of the compensation network used. In step S5, the calibration process for the network parameters of the compensation function is as follows: Select a segment of historical data from the target internal combustion engine historical database as the calibration set. The historical real-time deviation sequence of this segment of historical data must be within the historical dynamic deviation reference range of the corresponding period throughout the entire process. Historical theoretical fuel consumption estimates are extracted from the calibration set as the target value sequence for network training. Simultaneously, the intake air temperature, atmospheric pressure, and coolant temperature sequences matched with timestamps from the calibration set are extracted as feature input sequences. A compensation function network framework is constructed, which contains three independent sub-functions, each corresponding to one of the three feature inputs. Each sub-function is a polynomial with unknown coefficients. The target value sequence is used as the regression target, and the feature input sequence is used as the independent variable. The least squares method is used to solve for the unknown coefficients of all polynomials. The set of coefficients obtained by solving is associated with a version identifier and stored. In S5, the process of calculating and outputting the shadow carbon emission intensity sequence using the compensation function network is as follows: After the online calibration event is triggered, the system receives data points from the real-time theoretical fuel consumption estimation sequence. For each data point, the system simultaneously acquires the intake air temperature, atmospheric pressure, and engine coolant temperature at the same time. Input the intake air temperature value into the calibrated intake air temperature compensation sub-function to obtain the temperature compensation factor; input the atmospheric pressure value into the calibrated atmospheric pressure compensation sub-function to obtain the pressure compensation factor; input the coolant temperature value into the calibrated coolant temperature compensation sub-function to obtain the thermal state compensation factor; multiply the three compensation factors together to obtain the comprehensive compensation coefficient. Multiply the current theoretical fuel consumption estimate by the comprehensive compensation coefficient to obtain the shadow fuel consumption value, and then multiply it by the fuel carbon content conversion coefficient to obtain the shadow carbon emission intensity value. Repeat the above process for each data point and output the shadow carbon emission intensity sequence in chronological order.

2. The method for comprehensive management of carbon emissions from internal combustion engines based on multi-source data fusion according to claim 1, characterized in that, In S2, the process of establishing the dynamically calibrated volumetric fuel consumption mapping model is as follows: During an initialization period after the target internal combustion engine starts, the instantaneous fuel consumption sequence from the high-precision fuel flow meter, the speed sequence from the crankshaft position sensor, and the load rate sequence from the intake manifold pressure sensor are collected simultaneously, and the timestamps of the three sequences are aligned. A two-dimensional operating condition plane is constructed with engine speed as the horizontal axis and load rate as the vertical axis. This plane is then uniformly divided into multiple rectangular grid cells. Each data point within the initialization period is mapped to the corresponding operating condition grid cell based on its engine speed and load rate values. For each grid cell containing data points, the median of all instantaneous fuel consumption values ​​within it is calculated and used as the reference fuel consumption feature value for that cell. Using all grid cells containing the reference fuel consumption feature value as control points, a continuous fuel consumption mapping surface is generated using a surface fitting algorithm, serving as the dynamically calibrated ontological fuel consumption mapping model. In real-time calculations, each pair of engine speed and load rate data from the engine operating condition sequence is used as coordinate points, and interpolation calculations are performed on the fuel consumption mapping surface to obtain theoretical fuel consumption estimates and output the sequence.

3. The method for comprehensive management of carbon emissions from internal combustion engines based on multi-source data fusion according to claim 2, characterized in that, In step S2, the process of correcting the exhaust flow rate with environmental parameters and using it for exhaust gas back-calculation is as follows: Read the fixed displacement parameters of the engine from the electronic control unit of the target internal combustion engine, and calculate the uncorrected theoretical exhaust volume flow rate sequence based on the real-time engine speed sequence and the displacement parameters; The intake air temperature sensor reading sequence and atmospheric pressure sensor reading sequence, which are aligned with the engine speed sequence timestamp, are acquired synchronously. A gas density correction model is established, with real-time intake air temperature value, atmospheric pressure value and standard state reference value as inputs and gas density correction coefficient as output. Each data point of the theoretical exhaust volume flow rate sequence is multiplied by the real-time gas density correction coefficient calculated by the model at the corresponding time to obtain the corrected exhaust mass flow rate sequence. The corrected exhaust mass flow rate sequence is combined with the exhaust carbon dioxide concentration sequence and the carbon mass balance equation to calculate and back-calculate the measured fuel consumption back-calculation sequence.

4. The method for comprehensive management of carbon emissions from internal combustion engines based on multi-source data fusion according to claim 1, characterized in that, In step S3, the process by which the dynamic baseline learner outputs the dynamic deviation reference interval based on recent historical deviation characteristics is as follows: A fixed-capacity circular buffer is initialized as a historical deviation queue, storing the latest real-time deviation data points in chronological order. A statistical trigger mechanism is set up so that when a preset statistical interval is reached, the distribution characteristics of all data points in the queue are calculated. The arithmetic mean and standard deviation of all data points in the queue are calculated. The interval width coefficient bound to the target internal combustion engine model and emission level is retrieved from the vehicle attribute association table. The product of the interval width coefficient and the standard deviation is subtracted from the average value, and the result is assigned as the current lower boundary of the dynamic deviation reference interval. The product of the interval width coefficient and the standard deviation is added to the average value, and the result is assigned as the current upper boundary. When a new real-time deviation data point is generated, it is appended to the head of the queue, the old data points at the tail are removed, and a new round of feature calculation and boundary update is triggered.

5. The method for comprehensive management of carbon emissions from internal combustion engines based on multi-source data fusion according to claim 1, characterized in that, In step S4, the condition for determining that the real-time deviation sequence continuously exceeds the dynamic deviation reference interval is: Set a minimum sustained over-limit time threshold, which is determined based on the shortest time required for the engine to recover from transient fluctuations to steady state. Based on a fixed sampling frequency, convert the time threshold into a corresponding threshold for the number of consecutive data points. Scan the real-time deviation sequence in reverse from the current moment, check the latest consecutive data points, and determine whether the value of each data point is greater than the current upper boundary or less than the current lower boundary. Record the number of points that are continuously outside the interval. When the number of points continuously outside the interval reaches the threshold, and the numerical directions of these points are consistent, it is determined that the continuous exceedance condition is met.

6. The method for comprehensive management of carbon emissions from internal combustion engines based on multi-source data fusion according to claim 1, characterized in that, In step S6, the content and format of the generated calibration report record are as follows: When a calibration event is triggered, the Coordinated Universal Time (UTC) timestamp of the event is recorded, and the original data segment of the real-time deviation sequence before the trigger is extracted. This original data segment of the real-time deviation sequence contains the values ​​of each point, millisecond-level timestamps, and position status codes relative to the boundary at that time. The extracted data segment is subjected to pattern analysis, and it is marked as positive monotonic drift, negative monotonic drift, or non-monotonic oscillation according to its monotonicity and fluctuation characteristics. The pattern markings are written into the report. Record the network parameter version number of the compensation function called in this event, record the start and end times of the event, calculate the total duration, perform online statistics on the output shadow carbon emission intensity sequence during the event duration, calculate its average, maximum, minimum and standard deviation, and enter the statistical summary value into the report. Organize the timestamp, trigger data segment, mode tag, version number, time span and statistical summary according to the predefined key-value pair format, encapsulate and generate a calibration report file, calculate the hash check value of the file content and append it to the end of the report.

Images