A carbon emission collection method based on big data collection

Abstract

The present application relates to the technical field of big data, and specifically discloses a carbon emission collection method based on big data collection, which synchronously collects the power load, equipment vibration and emission port infrared thermal imaging data of a carbon emission source, extracts composite features reflecting production load, equipment state and thermal plume change in a sliding time window, and inputs the features into a pre-trained long short-term memory network to obtain a carbon emission estimate value corresponding to the time window; the estimate value is compared with a reported value of an emission monitoring device in a dynamic deviation manner, an abnormal timestamp is identified and an abnormal record is generated, so that the reliability and traceability of the carbon emission collection result are improved.

Description

Technical Field

[0001] This invention relates to the field of big data technology, and specifically to a method for collecting carbon emissions based on big data collection. Background Technology

[0002] Current carbon emission data collection typically relies on continuous emission monitoring systems (CEMS) installed at emission outlets. These systems calculate emissions by collecting data such as carbon dioxide concentration, flue gas velocity, and outlet cross-sectional area. While this method directly reflects the monitoring results at the emission outlet, its data source is relatively singular. When sensor probes accumulate dust, experience zero-point drift, sampling pipeline blockage, or localized malfunctions in the emission monitoring equipment, the system may still continuously output seemingly continuous data. This makes it difficult for maintenance personnel to promptly determine whether the reported carbon emissions for a given period have deviated from the actual emission levels. Furthermore, the electrical load generated by the target carbon emission source during production, the vibration status of key production equipment, and thermal plume images from the exhaust outlet are all correlated with production intensity and emission status. However, current collection methods typically do not integrate this multi-source heterogeneous data for indirect emission estimation and anomaly verification. This results in a lack of independent reference data when the emission monitoring equipment itself malfunctions, affecting the reliability of carbon emission data collection and subsequent verification and analysis. Summary of the Invention

[0003] The purpose of this invention is to provide a carbon emission collection method based on big data collection, thereby solving the above-mentioned technical problems.

[0004] The objective of this invention can be achieved through the following technical solutions: A method for collecting carbon emissions based on big data includes the following steps: Acquire multimodal data of the target carbon emission source within a selected time period. The multimodal data includes electricity load data, production equipment vibration data, and infrared thermal imaging data of exhaust gas emission outlets. Set a sliding time window, extract multimodal data within the sliding time window and filter it to obtain a sub-multimodal dataset; Extract composite feature vectors from the sub-multimodal datasets, input each composite feature vector into a pre-trained long short-term memory network, and output carbon emission estimates. The estimated carbon emissions are compared with the carbon emissions reported by the emission monitoring equipment. Abnormal timestamps are identified based on the comparison results, and abnormal collection records are generated based on the abnormal timestamps.

[0005] As a further aspect of the present invention, the process of acquiring multimodal data is as follows: Instantaneous voltage and current are collected by non-intrusive load monitoring terminals deployed on the main power supply circuit of the target carbon emission source, and the active power sequence is calculated as the electricity load data. Acceleration waveforms are collected as vibration data of the production equipment by vibration sensors installed on the bearing housings of drive motors in key production equipment. The infrared thermal imager, positioned directly opposite the exhaust outlet, continuously captures images of thermal radiation intensity, which are then used as infrared thermal imaging data.

[0006] As a further aspect of the present invention: the process of obtaining the sub-multimodal dataset is as follows: Set the sliding step size of the sliding time window. The sliding step size matches the time interval between two adjacent frames in the infrared thermal imaging data. Move the sliding time window according to the sliding step size within the selected time period. For each sliding position, the center time of the sliding time window is used as the time reference. Data segments of the power load data before and after each preset time period are extracted as the initial first type of segment. Data segments of the production equipment vibration data before and after each preset time period are extracted as the initial second type of segment. Image sequences of the infrared thermal imaging data before and after each preset time period are extracted as the initial sequence. Median filtering is performed on all original sampling points in the initial first and second class sub-segments to obtain the first and second class sub-segments. For each frame of the image in the initial sequence, non-uniformity correction is first performed based on the two-point calibration method, and then Gaussian smoothing is performed on the corrected image to obtain the target sequence. The first and second class segments, together with the target sequence, constitute a sub-multimodal dataset within a single sliding time window.

[0007] As a further aspect of the present invention, the process of extracting the composite feature vector is as follows: Wavelet packet decomposition of a class of sub-segments is performed to the third level to obtain multiple sub-bands. The proportion of signal energy in each sub-band to the total energy before decomposition is calculated. The active power in a class of sub-segments is integrated over time to obtain the window power consumption. The proportions corresponding to each sub-band and the window power consumption are combined to form the first feature vector. Empirical mode decomposition is performed on the second type of sub-segment to screen out the intrinsic mode functions containing the dominant frequency component. The Hilbert marginal spectrum of the intrinsic mode function is calculated, and the difference between the frequency centroid of the marginal spectrum and the frequency centroid of the equipment no-load state reference is extracted as the second feature. Differentiate between every two adjacent frames in the target sequence to obtain a differential image sequence. Using the emission port location as the initial search starting point, extract the set of pixels whose differential amplitude is greater than the preset differential threshold and are connected to the emission port location. Use the set of pixels as the gas plume region. Calculate the orientation histogram of pixel intensity gradient within the gas plume region, extract the principal orientation angles from the orientation histogram, convert the principal orientation angles into sine and cosine components, calculate the mean of the sine and cosine components of all principal orientation angles, and calculate the circumferential average direction and circumferential dispersion from the mean of the sine and cosine components. The pixel area, circumferential average direction, and circumferential dispersion of the gas plume region are used as the third feature vector. The first, second, and third feature vectors are concatenated in sequence to obtain the composite feature vector corresponding to the sliding time window.

[0008] As a further aspect of the present invention: the process of obtaining an estimate of carbon emissions: Obtain multiple historical composite feature vectors, and calculate the carbon emission reference amount for the time period corresponding to the historical composite feature vector based on fuel consumption and fuel carbon content; Construct a network consisting of an input layer, two long short-term memory layers, and a fully connected output layer. Set the number of output nodes in the fully connected output layer to one, and set the number of nodes in the input layer to the dimension of the composite feature vector. Randomly initialize the weights of the input layer, the two long short-term memory layers, and the fully connected output layer. The historical composite feature vectors are input into the network one by one. After each composite feature vector is input, the hidden state is updated by two layers of long short-term memory, and the fully connected output layer outputs a carbon emission estimate. The deviation between the output carbon emission estimate and the corresponding carbon emission reference is calculated, and the weights of the network are updated through backpropagation of the error to obtain the pre-trained long short-term memory network. All current composite feature vectors are input into a pre-trained long short-term memory network in chronological order. For each composite feature vector input, the fully connected output layer outputs an estimated carbon emission value.

[0009] As a further aspect of the present invention: the process of identifying points where the deviation exceeds the dynamic tolerance range is as follows: The carbon emission concentration sequence, flue gas velocity sequence, and emission outlet cross-sectional area recorded by the emission monitoring equipment are periodically acquired. The carbon emission concentration value, flue gas velocity value, and emission outlet cross-sectional area at the same time point are multiplied by the duration of the time interval to obtain the reported carbon emission amount. Get the sum of all reported carbon emissions A1 within the current sliding time window, and get the carbon emission estimate A2 for the current sliding time window. Calculate the deviation between A1 and A2. The deviation values ​​corresponding to the current sliding time window and the previous preset number of sliding time windows are sorted in ascending order. The deviation values ​​in the first and second order proportions are used as the lower and upper tolerance limits, respectively. The first and second order proportions are preset and the first order proportion is < the second order proportion. If the deviation value of the current sliding time window is less than the lower tolerance limit or greater than the upper tolerance limit, then the timestamp T1 corresponding to the center time of the current sliding time window is marked as an abnormal timestamp.

[0010] As a further aspect of the present invention, the process of generating abnormal acquisition records is as follows: Using the abnormal timestamp as the time center, extract multimodal data within a preset length before and after it; The abnormal timestamp, the corresponding composite feature vector, the reported carbon emissions, and the extracted multimodal data are packaged into a single data collection anomaly record; The abnormal records are added to the pre-set abnormal record database, and the abnormal timestamp and the corresponding reported carbon emissions are transmitted to the carbon verification terminal.

[0011] The beneficial effects of this invention compared to the prior art are: This invention provides a verification basis independent of the monitoring equipment itself, even when emission monitoring equipment continues to output data despite anomalies such as dust accumulation, drift, or blockage. This eliminates the reliance on manual experience or a single monitoring curve for judging abnormal data points. By comparing the reported results from emission outlets with emission trends reflected in the production process, it can more promptly identify periods where reported carbon emissions do not match actual production emissions, reducing the risk of abnormal data being directly used for carbon accounting and regulatory reporting. Furthermore, once the abnormal timestamp and its corresponding data are recorded, maintenance personnel can conduct investigations around specific abnormal periods, reducing the scope of blind inspections, improving the efficiency of anomaly location, and providing traceable data for subsequent carbon verification. This enhances the reliability, verifiability, and anomaly traceability of carbon emission data collection results. Attached Figure Description

[0012] The invention will now be further described with reference to the accompanying drawings.

[0013] Figure 1 This is a flowchart illustrating a carbon emission collection method based on big data collection according to the present invention. Figure 2 This is a schematic diagram of the process for extracting composite feature vectors according to the present invention. Detailed Implementation

[0014] 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.

[0015] Please see Figures 1-2 As shown, this invention is a method for collecting carbon emissions based on big data collection, comprising the following steps: Acquire multimodal data of the target carbon emission source within a selected time period. The multimodal data includes electricity load data, production equipment vibration data, and infrared thermal imaging data of exhaust gas emission outlets. In a preferred embodiment of the present invention, the process of acquiring multimodal data is as follows: First, determine the main power supply circuit of the target carbon emission source. Install a non-intrusive load monitoring terminal on the incoming side of this main power supply circuit. Connect the voltage sampling terminal to the voltage measurement terminal of the main power supply circuit, and obtain the current change in the main power supply circuit through current inductance. The non-intrusive load monitoring terminal collects instantaneous voltage and instantaneous current according to a unified sampling clock. Calculate the active power for the instantaneous voltage and current values ​​corresponding to the same sampling time, and arrange them in the order of sampling time to form an active power sequence. This active power sequence is used as the electricity load data.

[0016] Vibration sampling points are selected at the drive motor bearing housing on key production equipment directly related to the production process of the target carbon emission source. Vibration sensors are fixed to the outer surface of the bearing housing, ensuring the sensor's sensitive direction aligns with the primary vibration transmission direction of the bearing housing. The vibration sensor continuously collects the acceleration changes of the bearing housing as the equipment operates and outputs acceleration waveforms according to the sampling time. These acceleration waveforms are used as the vibration data of the production equipment. For example, the key production equipment could be a drive motor that powers a fan, pump, or conveyor mechanism; the vibration at its bearing housing can form a continuous waveform depending on the equipment's operating state.

[0017] The infrared thermal imager is then positioned directly facing the exhaust outlet, ensuring its imaging field of view covers the outlet and the surrounding airflow diffusion area. The imager's position and orientation remain constant throughout the acquisition process. The imager acquires images of the thermal radiation intensity at the outlet in a continuous frame acquisition manner. Each frame carries the corresponding acquisition time, and these frames are arranged according to their acquisition times to form infrared thermal imaging data. For example, after exhaust gas exits the outlet, it appears as a thermal radiation area that changes relative to the surrounding background in the image; continuous frame images record this change over time.

[0018] Set a sliding time window, extract multimodal data within the sliding time window and filter it to obtain a sub-multimodal dataset; In another preferred embodiment of the present invention, the process of obtaining the sub-multimodal dataset is as follows: First, the acquisition time of each frame in the infrared thermal imaging data is read, and the time interval between the acquisition times of two adjacent frames is calculated. This time interval is then set as the sliding step size of the sliding time window. The start time of a pre-selected time period is used as the initial position of the sliding time window, and the sliding time window moves sequentially backward according to the sliding step size until it reaches the end position of the selected time period. For each sliding position, the start and end times of the sliding time window are read, and the midpoint between them is taken as the center time. Using this center time as the time base, a preset time is extended forward and backward to form the current intercepted time range.

[0019] Based on the sampling time carried by each active power sampling point in the power load data, all active power sampling points whose sampling time falls within the current intercept time range are selected and arranged in order of sampling time as the initial first-class sub-segment; based on the sampling time carried by each acceleration sampling point in the production equipment vibration data, all acceleration sampling points whose sampling time falls within the current intercept time range are selected and arranged in order of sampling time as the initial second-class sub-segment; based on the acquisition time carried by each frame of infrared thermal imaging data, all thermal radiation intensity images whose acquisition time falls within the current intercept time range are selected and arranged in order of acquisition time as the initial sequence.

[0020] For each active power sampling point in the initial first-class sub-segment, a preset number of adjacent sampling points before and after the sampling point are read. The read sampling values ​​are sorted according to their numerical values, and the sampling value in the middle position after sorting is recorded as the filtered sampling value of the sampling point. After all sampling points are processed, the first-class sub-segment is obtained. For each acceleration sampling point in the initial second-class sub-segment, the same median filtering process is applied, and the sampling value in the middle position after sorting is recorded as the filtered sampling value of the sampling point. After all sampling points are processed, the second-class sub-segment is obtained.

[0021] Before performing two-point calibration for non-uniformity correction on each frame of the initial sequence, two-point reference acquisition of the infrared thermal imager is completed. The infrared thermal imager is aligned with a uniform radiation source, ensuring the uniform radiation source fills the imaging field of view of the infrared thermal imager. The uniform radiation source is adjusted to a first radiation state. After the output of the uniform radiation source stabilizes, the calibrated radiation intensity value corresponding to the first radiation state is read and recorded as the low radiation reference value. Multiple frames of the first reference image are continuously acquired. For any pixel location, the response value of that pixel location in multiple frames of the first reference image is read. The average of the read response values ​​is calculated, and the average value is taken as the low response value of that pixel location. After traversing all pixel locations, all low response values ​​are arranged according to their original pixel locations to form a low radiation reference image. Subsequently, the uniform radiation source is adjusted to a second radiation state, where the radiation intensity is higher than that of the first radiation state. After the output of the uniform radiation source stabilizes, the calibrated radiation intensity value corresponding to the second radiation state is read and recorded as the high radiation reference value. Multiple frames of the second reference image are continuously acquired. For any pixel location, read the response value of that pixel location in multiple frames of the second reference image, average the read response values, and use the average value as the high response value of that pixel location; after traversing all pixel locations, arrange all high response values ​​according to their original pixel locations to form a high radiometric reference image.

[0022] When correcting the current image in the initial sequence, the original thermal radiation intensity value of any pixel in the current image is read, along with the corresponding low response value, high response value, low radiation reference value, and high radiation reference value. The difference between the high radiation reference value and the low radiation reference value is calculated, as is the difference between the high response value and the low response value. Dividing the former difference by the latter yields the gain correction for that pixel. The product of the gain correction and the low response value is then subtracted from the low radiation reference value to obtain the bias correction for that pixel. The original thermal radiation intensity value of that pixel in the current image is multiplied by the gain correction, and then the bias correction is added to obtain the corrected thermal radiation intensity value for that pixel. This process is repeated for all pixel positions in the current image, arranging all corrected thermal radiation intensity values ​​according to their original pixel positions to form a non-uniformly corrected image.

[0023] When performing Gaussian smoothing on non-uniformity-corrected images, a Gaussian convolution kernel of a preset size is selected. Neighboring pixels are read sequentially, centered on each pixel. The thermal radiation intensity values ​​of these neighboring pixels are then weighted and summed according to the Gaussian convolution kernel weights. The weighted sum is recorded as the smoothed thermal radiation intensity value of that pixel. After processing all pixels, a smoothed image is obtained. After performing non-uniformity correction and Gaussian smoothing on all images in the initial sequence sequentially, they are arranged according to their original acquisition times to form the target sequence.

[0024] The first-class sub-segment, the second-class sub-segment, and the target sequence corresponding to the current sliding time window are bound into a data combination under the same time window, and this data combination is recorded as the sub-multimodal dataset of the current sliding time window.

[0025] Extract composite feature vectors from the sub-multimodal datasets, input each composite feature vector into a pre-trained long short-term memory network, and output carbon emission estimates. In a preferred embodiment of the present invention, the process of extracting the composite feature vector is as follows: When performing wavelet packet decomposition on a class of sub-segments, a preset wavelet basis is selected, and the class of sub-segments is used as the input signal to be decomposed layer by layer to the third layer, obtaining all sub-band signals of the third layer. The sum of squares of each sample value in each sub-band signal is calculated and denoted as the energy of that sub-band signal; the sum of squares of each original sample value in the class of sub-segments is calculated and denoted as the total energy before decomposition; the energy of each sub-band signal is divided by the total energy before decomposition to obtain the proportion corresponding to that sub-band. Then, the sampling time interval of adjacent active power sampling points in the class of sub-segments is read, and the average value of two adjacent active power sampling values ​​is multiplied by the corresponding time interval to obtain the electricity consumption within that interval. The electricity consumption within all intervals is accumulated to obtain the window electricity consumption. The proportions corresponding to each sub-band are arranged in order of sub-band number, and the window electricity consumption is added to the last position to form the first feature vector. For example, if the wavelet packet decomposition to the third layer yields 8 sub-bands, then the 8 proportions are arranged first, and then the window electricity consumption is arranged.

[0026] When performing empirical mode decomposition (EMD) on the type II sub-segments, the type II sub-segments are used as input signals, and EMD is used to obtain several intrinsic mode functions (EMFs) and residual terms. First, a Fast Fourier Transform (FFT) is performed on the type II sub-segments to obtain the frequency point with the largest amplitude, which is denoted as the dominant frequency. Then, a FFT is performed on each EMF to obtain its principal frequency point. The principal frequency point is compared with the dominant frequency, and EMFs whose principal frequency point matches the dominant frequency or falls within a preset proximity range are selected and denoted as EMFs containing the dominant frequency component. A Hilbert Transform is performed on the selected EMFs to obtain the instantaneous amplitude and instantaneous frequency corresponding to each sampling time. Frequency intervals are divided according to a preset frequency interval, and the instantaneous amplitudes falling within the same frequency interval are accumulated to obtain the marginal spectral amplitude of that frequency interval. The marginal spectral amplitudes of all frequency intervals are arranged in frequency order to form the Hilbert marginal spectrum. Multiply the center frequency of each frequency interval by the marginal spectral amplitude of that interval, sum the results, and then divide by the sum of all marginal spectral amplitudes to obtain the frequency centroid of the marginal spectrum. Read the frequency centroid of the equipment's no-load reference frequency, and subtract the reference frequency centroid from the current frequency centroid to obtain the second feature.

[0027] For two adjacent frames in the target sequence, subtract each other point by point according to the same pixel position, and take the absolute value of the difference to form a difference image. After processing all adjacent frames in sequence, a difference image sequence is obtained. For each difference image, a connectivity search is performed starting from the pre-marked emission port position. Pixels adjacent to the emission port position are read. If the pixel difference magnitude is greater than a preset difference threshold, it is added to the set to be expanded, and the neighboring pixels around that pixel are read. The above process is repeated until no new neighboring pixels meet the condition. The final pixel set is recorded as the gas plume region corresponding to the difference image, and the number of pixels in the set is counted to obtain the pixel area of ​​the gas plume region.

[0028] For each gas plume region, the intensity value of each pixel within that region is read from the corresponding image in the target sequence. The Sobel operator is used to calculate the horizontal and vertical gradients respectively. The gradient angle of the pixel is determined based on the vertical and horizontal gradients, and the gradient magnitude is obtained by taking the square root of the sum of the squares of the two gradients. The range from 0 degrees to 360 degrees is divided into several angle intervals, for example, each interval is 10 degrees. The gradient angle of each pixel is assigned to the corresponding angle interval, and the gradient magnitude of the pixel is accumulated into that angle interval to form a direction histogram. The angle interval with the largest accumulated value is selected, and the center angle of that angle interval is recorded as the principal direction angle.

[0029] After obtaining the principal direction angles for all differential images, each principal direction angle is first converted from an angle value to a radian value. Then, the sine and cosine values ​​of the radian value are calculated to obtain the corresponding sine and cosine components. The mean of the sine components is obtained by averaging all the sine components, and the mean of the cosine components is obtained by averaging all the cosine components. Using the mean of the cosine components as the abscissa and the mean of the sine components as the ordinate, the polar angle of the coordinate point relative to the origin is calculated. This polar angle is converted to an angle value and recorded as the circumferential mean direction; if the angle value is less than 0 degrees, 360 degrees is added. The distance from the coordinate point to the origin is calculated, and 1 is subtracted from this distance to obtain the circumferential dispersion. The average pixel area of ​​all gas plume regions, the circumferential mean direction, and the circumferential dispersion are arranged sequentially to form the third feature vector.

[0030] Finally, the first feature vector, the second feature vector, and the third feature vector are connected in the above order to form a numerical sequence, which is denoted as the composite feature vector of the current sliding time window.

[0031] Understandably, the second feature selects the intrinsic mode functions (EMFs) containing the dominant frequency component in the vibration signal and extracts the difference between the frequency centroid of its Hilbert marginal spectrum and the equipment's no-load reference value. This is because when key production equipment operates under different load conditions, the mechanical coupling relationship between the drive motor, bearing housing, and driven mechanism changes with the stress state, and the dominant frequency component and its energy distribution in the vibration signal also shift accordingly. Empirical mode decomposition (EMD) can break down complex acceleration waveforms into several EMFs with different time scales. Selecting EMFs containing the dominant frequency component can reduce the interference of other non-primary vibration components on the characterization of the operating state. The Hilbert marginal spectrum can reflect the energy distribution of the dominant frequency-related vibration component on the frequency axis, while the frequency centroid compresses this distribution into a comparable value. By subtracting the current frequency centroid from the no-load reference frequency centroid, the vibration frequency shift caused by the equipment transitioning from no-load to production load can be expressed. This allows the second feature to reflect the actual operating intensity changes of key production equipment and provides input data related to the production process state for subsequent carbon emission estimation.

[0032] The gas plume region is determined by differentiating adjacent infrared images and combining the connectivity of the emission port location. This is because after exhaust gas is continuously discharged from the emission port, it typically appears as a thermal radiation region that changes relative to the surrounding background in the infrared image, while the static background shows little change between adjacent frames. Differentiating between two adjacent frames highlights the thermal radiation region that moves or diffuses over time from the static background. Then, using the emission port location as the initial search starting point, only the set of pixels with a difference amplitude exceeding a preset difference threshold and connected to the emission port location is retained. This filters out regions with spatial continuity related to the emission process, reducing the possibility of other independent heat sources or random noise in the image being mistakenly identified as emission plumes. This set of pixels corresponds to the spatial coverage of the emission port thermal plume on the image plane, and its pixel area can reflect the scale of plume expansion within the current time window. After participating in the composite feature vector, it can supplement the spatial change information on the emission port side, allowing carbon emission estimation to not only rely on electricity consumption and vibration signals but also to consider the actual emission state of the emission port.

[0033] The circumferential mean direction and circumferential dispersion are used to process the main direction angles of gas plumes because angle data has a circular characteristic, with 0 degrees and 360 degrees being adjacent in physical direction. Directly averaging angle values ​​would miscalculate adjacent directions as being far apart. Converting each main direction angle into sine and cosine components is equivalent to mapping the direction angle to a point on the unit circle. Calculating the mean of the sine and cosine components separately yields a composite direction in circular space. The polar angle determined by these two means represents the average diffusion direction of the plume's overall bias within the target time window. The vector length formed by these two means reflects the concentration of the main direction angles, further yielding the circumferential dispersion. The resulting circumferential mean direction characterizes the main direction of plume diffusion, while the circumferential dispersion characterizes the stability of the diffusion direction. Together, they describe the morphological changes of the thermal plume at the emission outlet within the time window, providing a dynamic image-based reference for identifying whether reported carbon emissions match actual emission conditions.

[0034] In this embodiment, infrared thermal imaging data is used to characterize the motion state and spatial diffusion scale of the thermal plume at the exhaust outlet, and is not used to directly measure carbon dioxide concentration. After the exhaust gas is discharged from the outlet, there is a difference in thermal radiation intensity between it and the surrounding background, or a region of motion change relative to the static background is formed in consecutive image frames; therefore, after differential processing of two adjacent frames in the target sequence, the set of pixels with a difference amplitude greater than a preset difference threshold and connected to the outlet location can characterize the thermal plume region escaping from the outlet. The pixel area of ​​the gas plume region is used to represent the spatial expansion scale of the thermal plume at the outlet in the image plane within the sliding time window, and the circumferential average direction and circumferential dispersion are used to represent the main diffusion direction and directional stability of the thermal plume; the above infrared features are used as auxiliary features for indirect estimation of carbon emissions in the construction of composite feature vectors, rather than as direct measurement results of emission concentration.

[0035] A preferred embodiment of the present invention describes the process of obtaining an estimated carbon emission value: First, obtain the historical composite feature vectors corresponding to multiple completed sliding time windows, and read the start and end times of each sliding time window. Based on this time range, query the fuel metering records to obtain the fuel consumption within that sliding time window, and read the corresponding fuel carbon content. Multiply the fuel consumption by the fuel carbon content to obtain the carbon mass involved in emission calculations within that sliding time window. Since a carbon dioxide molecule consists of one carbon atom and two oxygen atoms, and the relative atomic mass of a carbon atom is 12, while the relative molecular mass of a carbon dioxide molecule is 44, multiply the carbon mass by 44 and divide by 12 to obtain the carbon dioxide mass completely converted from fuel carbon within that sliding time window. Record this carbon dioxide mass as the carbon emission reference amount within that sliding time window and use it as the training label corresponding to the historical composite feature vector.

[0036] Subsequently, a Long Short-Term Memory (LSTM) network is constructed. The input layer receives the values ​​from the composite feature vector, and the number of nodes in the input layer is the same as the dimension of the composite feature vector. The input layer is then connected to the first LSM layer. The output sequence of the first LSM layer is input to the second LSM layer. The current hidden state output by the second LSM layer is input to the fully connected output layer. The fully connected output layer has one output node, whose output value represents the carbon emission estimate for the current sliding time window. The weight parameters in the input layer, the two LSM layers, and the fully connected output layer are randomly initialized, and preset learning rates, training epochs, and batch lengths are set.

[0037] During training, historical composite feature vectors are arranged into a historical feature sequence according to the chronological order of their corresponding sliding time windows, and then input into the network one by one in this order. For each historical composite feature vector input, the first long short-term memory (LSM) layer receives the vector and updates its first hidden state; the second LSM layer receives the output of the first LSM layer and updates its second hidden state; and the fully connected output layer outputs a carbon emission estimate based on the second hidden state. The difference between this carbon emission estimate and the carbon emission reference value for the same sliding time window is calculated to obtain the training bias. This training bias is then propagated back along the connection directions of the fully connected output layer, the second LSM layer, the first LSM layer, and the input layer to update the weights of each layer. The historical feature sequence is repeatedly input and the weights are updated according to a preset number of training rounds. After training, the updated network weights are retained, resulting in a pre-trained long short-term memory network.

[0038] When making the current estimate, all composite feature vectors arranged in chronological order within the currently selected time period are read and input into the pre-trained long short-term memory network one by one. For each current composite feature vector input, the two long short-term memory layers update the current hidden state according to the current input and the hidden state of the previous sliding time window. The fully connected output layer outputs a value, which is recorded as the carbon emission estimate for the current sliding time window.

[0039] It's important to note that the model architecture uses an input layer, two long short-term memory layers, and a fully connected output layer connected sequentially. This is because the composite feature vector is a set of values ​​arranged in a fixed order. The input layer first receives these values ​​according to their feature dimensions and feeds them into the temporal processing structure. Then, the long short-term memory layers update the hidden states according to the order of the sliding time windows. Finally, the fully connected output layer maps the hidden states to a specific carbon emission estimate. The number of nodes in the input layer is set to the number of dimensions of the composite feature vector, ensuring that each window's features, such as electricity consumption, frequency band ratio, vibration frequency shift, plume area, circumferential mean direction, and circumferential dispersion, have a corresponding input location. The two long short-term memory layers handle the transmission relationship of these features between consecutive time windows. The fully connected output layer has only one output node because the current task only needs to output the carbon emission estimate corresponding to the current sliding time window. The underlying logic of this architecture is to first fully receive multimodal features, then extract time dependencies, and finally form a scalar output of emissions. There is a clear domain correspondence between the input and output, so it can transform the comprehensive changes on the production process side and the emission outlet side into carbon emission estimates that can be used for anomaly verification.

[0040] Two long short-term memory (LSM) layers are employed because carbon emissions are not solely determined by instantaneous features within the current sliding time window; they are also influenced by changes in production load, equipment operating inertia, and the continuity of airflow at the emission outlet. The first LSM layer receives the composite feature vector input in chronological order and converts the electricity consumption, vibration, and infrared features within each sliding time window into latent states with temporal memory, preserving the relationship between adjacent time windows. The second LSM layer continues to receive the latent state sequence output from the first layer, further extracting higher-level continuous change patterns based on the existing temporal memory. Essentially, this process transforms the static features of a single time window into context-dependent temporal features, allowing the network to not only see the current production state but also utilize the continuous trends of load fluctuations, equipment vibration changes, and plume changes in previous time windows. Since carbon emission sources typically exhibit a temporal continuity from load changes to emission changes during production, two LSM layers can more fully express this continuous state, making the output carbon emission estimate more closely match the actual emission state corresponding to the current sliding time window.

[0041] In this embodiment, each historical composite feature vector corresponds to a defined sliding time window, and the carbon emission reference is statistically calculated according to the start and end times of this sliding time window. Specifically, the fuel consumption of the target carbon emission source within the sliding time window is obtained, and the carbon content of the corresponding fuel is read. Based on the fuel consumption and carbon content within the sliding time window, the carbon emission reference within the sliding time window is calculated, ensuring a one-to-one correspondence between the historical composite feature vector and the carbon emission reference in terms of time range. Multiple historical composite feature vectors are arranged in chronological order according to their corresponding sliding time windows to form a historical feature sequence, which is then input into the Long Short-Term Memory (LSTM) network sequentially. When the LSM layer receives the current historical composite feature vector, it inherits the hidden state formed after the input of the previous sliding time window and updates the current hidden state in conjunction with the current historical composite feature vector. The fully connected output layer outputs the carbon emission estimate corresponding to the current sliding time window based on the current hidden state. During training, the deviation between each carbon emission estimate and the carbon emission reference within the same sliding time window is calculated, and the error is propagated backward along the time expansion direction of the historical feature sequence to update the network weights.

[0042] The estimated carbon emissions are compared with the carbon emissions reported by the emission monitoring equipment. Abnormal timestamps are identified based on the comparison results, and abnormal collection records are generated based on the abnormal timestamps.

[0043] In a preferred embodiment of the present invention, the process of identifying points whose deviations exceed the dynamic tolerance range is as follows: The carbon emission concentration sequence, flue gas velocity sequence, and emission outlet cross-sectional area recorded by the emission monitoring equipment are periodically acquired. The carbon emission concentration value, flue gas velocity value, and emission outlet cross-sectional area at the same time point are multiplied by the duration of the time interval to obtain the reported carbon emission amount. Get the sum of all reported carbon emissions A1 within the current sliding time window, and get the carbon emission estimate A2 for the current sliding time window. Calculate the deviation between A1 and A2. The deviation values ​​corresponding to the current sliding time window and the previous preset number of sliding time windows are sorted in ascending order. The deviation values ​​in the first and second order proportions are used as the lower and upper tolerance limits, respectively. The first and second order proportions are preset and the first order proportion is < the second order proportion. If the deviation value of the current sliding time window is less than the lower tolerance limit or greater than the upper tolerance limit, then the timestamp T1 corresponding to the center time of the current sliding time window is marked as an abnormal timestamp.

[0044] In this embodiment, the deviation between A1 and A2 is represented by a signed deviation, which is the difference between the sum of all reported carbon emissions (A1) within the current sliding time window and the estimated carbon emissions (A2) for the current sliding time window. A positive signed deviation indicates that the reported carbon emissions from the emission monitoring device are higher than the multimodal data estimation result; a negative signed deviation indicates that the reported carbon emissions from the emission monitoring device are lower than the multimodal data estimation result. After sorting the signed deviations corresponding to the current sliding time window and a preset number of previous sliding time windows in ascending order, the deviation values ​​at the first order proportion are used as the lower tolerance bound, and the deviation values ​​at the second order proportion are used as the upper tolerance bound. If the signed deviation of the current sliding time window is less than the lower tolerance bound or greater than the upper tolerance bound, it indicates that the currently reported carbon emissions have deviated abnormally from the recent deviation distribution, and the timestamp corresponding to the center time of the current sliding time window is marked as an abnormal timestamp.

[0045] In another preferred embodiment of the present invention, the process of generating abnormal acquisition records is as follows: Using the abnormal timestamp as the time center, extract multimodal data within a preset length before and after it; The abnormal timestamp, the corresponding composite feature vector, the reported carbon emissions, and the extracted multimodal data are packaged into a single data collection anomaly record; The abnormal records are added to the pre-set abnormal record database, and the abnormal timestamp and the corresponding reported carbon emissions are transmitted to the carbon verification terminal.

[0046] The overall idea of ​​this invention is to introduce multimodal data that reflects the actual production status of carbon emission sources and the outflow status of emission outlets, without relying entirely on the output results of emission monitoring equipment. First, electricity load data characterizes the strength of production load; vibration data of production equipment characterizes the operating status of key equipment; and infrared thermal imaging data of exhaust gas outlets characterizes the diffusion scale and direction of thermal plumes. Then, these data are processed uniformly within the same sliding time window to extract composite features. A pre-trained long short-term memory network is used to obtain an estimated value of carbon emissions within this time window. Since this estimated value originates from the combined changes on the production process side and the emission outlet image side, rather than from the same sampling link of the emission monitoring equipment, when the emission monitoring equipment reports abnormal carbon emissions due to probe dust accumulation, zero-point drift, or sampling pipeline blockage, the multimodal estimation result can serve as an independent reference, comparing the deviation with the reported carbon emissions, and identifying abnormal timestamps based on the recent deviation distribution. This approach can identify abnormal data collection points that are not easily detected by a single monitoring curve, and record the abnormal timestamps and corresponding data to provide a basis for subsequent maintenance, investigation, and carbon verification. This solves the problem of existing carbon emission collection methods having a single data source and difficulty in timely identifying abnormal data collection points.

[0047] 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 present invention should still fall within the scope of the present invention.

Claims

1. A method for collecting carbon emissions based on big data collection, characterized in that, Includes the following steps: Acquire multimodal data of the target carbon emission source within a selected time period. The multimodal data includes electricity load data, production equipment vibration data, and infrared thermal imaging data of exhaust gas emission outlets. Set a sliding time window, extract multimodal data within the sliding time window and filter it to obtain a sub-multimodal dataset; Extract composite feature vectors from the sub-multimodal datasets, input each composite feature vector into a pre-trained long short-term memory network, and output carbon emission estimates. The estimated carbon emissions are compared with the carbon emissions reported by the emission monitoring equipment. Abnormal timestamps are identified based on the comparison results, and abnormal collection records are generated based on the abnormal timestamps.

2. The carbon emission collection method based on big data collection according to claim 1, characterized in that, The process of acquiring multimodal data is as follows: Instantaneous voltage and current are collected by non-intrusive load monitoring terminals deployed on the main power supply circuit of the target carbon emission source, and the active power sequence is calculated as the electricity load data. Acceleration waveforms are collected as vibration data of the production equipment by vibration sensors installed on the bearing housings of drive motors in key production equipment. The infrared thermal imager, positioned directly opposite the exhaust outlet, continuously captures images of thermal radiation intensity, which are then used as infrared thermal imaging data.

3. The carbon emission collection method based on big data collection according to claim 1, characterized in that, The process of obtaining the sub-multimodal dataset is as follows: Set the sliding step size of the sliding time window. The sliding step size matches the time interval between two adjacent frames in the infrared thermal imaging data. Move the sliding time window within the selected time period according to the sliding step size. For each sliding position, the center time of the sliding time window is used as the time reference. Data segments of the power load data before and after each preset time period are extracted as the initial first type of segment. Data segments of the production equipment vibration data before and after each preset time period are extracted as the initial second type of segment. Image sequences of the infrared thermal imaging data before and after each preset time period are extracted as the initial sequence. Median filtering is performed on all original sampling points in the initial first and second class sub-segments to obtain the first and second class sub-segments. For each frame of the image in the initial sequence, non-uniformity correction is first performed based on the two-point calibration method, and then Gaussian smoothing is performed on the corrected image to obtain the target sequence. The first and second class segments, together with the target sequence, constitute a sub-multimodal dataset within a single sliding time window.

4. The carbon emission collection method based on big data collection according to claim 1, characterized in that, The process of extracting composite feature vectors is as follows: Wavelet packet decomposition of a class of sub-segments is performed to the third level to obtain multiple sub-bands. The proportion of signal energy in each sub-band to the total energy before decomposition is calculated. The active power in a class of sub-segments is integrated over time to obtain the window power consumption. The proportions corresponding to each sub-band and the window power consumption are combined to form the first feature vector. Empirical mode decomposition is performed on the second type of sub-segment to screen out the intrinsic mode functions containing the dominant frequency component. The Hilbert marginal spectrum of the intrinsic mode function is calculated, and the difference between the frequency centroid of the marginal spectrum and the frequency centroid of the equipment no-load state reference is extracted as the second feature. Differentiate between every two adjacent frames in the target sequence to obtain a differential image sequence. Using the emission port location as the initial search starting point, extract the set of pixels whose differential amplitude is greater than the preset differential threshold and are connected to the emission port location. Use the set of pixels as the gas plume region. Calculate the orientation histogram of pixel intensity gradient within the gas plume region, extract the principal orientation angles from the orientation histogram, convert the principal orientation angles into sine and cosine components, calculate the mean of the sine and cosine components of all principal orientation angles, and calculate the circumferential average direction and circumferential dispersion from the mean of the sine and cosine components. The pixel area, circumferential average direction, and circumferential dispersion of the gas plume region are used as the third feature vector. The first, second, and third feature vectors are concatenated in sequence to obtain the composite feature vector corresponding to the sliding time window.

5. The carbon emission collection method based on big data collection according to claim 1, characterized in that, The process of obtaining carbon emission estimates: Obtain multiple historical composite feature vectors, and calculate the carbon emission reference amount for the time period corresponding to the historical composite feature vector based on fuel consumption and fuel carbon content; Construct a network consisting of an input layer, two long short-term memory layers, and a fully connected output layer. Set the number of output nodes in the fully connected output layer to one, and set the number of nodes in the input layer to the dimension of the composite feature vector. Randomly initialize the weights of the input layer, the two long short-term memory layers, and the fully connected output layer. The historical composite feature vectors are input into the network one by one. After each composite feature vector is input, the hidden state is updated by two layers of long short-term memory, and the fully connected output layer outputs a carbon emission estimate. The deviation between the output carbon emission estimate and the corresponding carbon emission reference is calculated, and the weights of the network are updated through backpropagation of the error to obtain the pre-trained long short-term memory network. All current composite feature vectors are input into a pre-trained long short-term memory network in chronological order. For each composite feature vector input, the fully connected output layer outputs an estimated carbon emission value.

6. The carbon emission collection method based on big data collection according to claim 1, characterized in that, The process of identifying points whose deviations exceed the dynamic tolerance range is as follows: The carbon emission concentration sequence, flue gas velocity sequence, and emission outlet cross-sectional area recorded by the emission monitoring equipment are periodically acquired. The carbon emission concentration value, flue gas velocity value, and emission outlet cross-sectional area at the same time point are multiplied by the duration of the time interval to obtain the reported carbon emission amount. Get the sum of all reported carbon emissions A1 within the current sliding time window, and get the carbon emission estimate A2 for the current sliding time window. Calculate the deviation between A1 and A2. The deviation values ​​corresponding to the current sliding time window and the previous preset number of sliding time windows are sorted in ascending order. The deviation values ​​in the first and second order proportions are used as the lower and upper tolerance limits, respectively. The first and second order proportions are preset and the first order proportion is < the second order proportion. If the deviation value of the current sliding time window is less than the lower tolerance limit or greater than the upper tolerance limit, then the timestamp T1 corresponding to the center time of the current sliding time window is marked as an abnormal timestamp.

7. The carbon emission collection method based on big data collection according to claim 1, characterized in that, The process of generating collection anomaly records is as follows: Using the abnormal timestamp as the time center, extract multimodal data within a preset length before and after it; The abnormal timestamp, the corresponding composite feature vector, the reported carbon emissions, and the extracted multimodal data are packaged into a single data collection anomaly record; The abnormal records are added to the pre-set abnormal record database, and the abnormal timestamp and the corresponding reported carbon emissions are transmitted to the carbon verification terminal.

Images