A method, medium and system for monitoring carbon emissions of industrial equipment

Abstract

The application discloses an industrial equipment carbon emission monitoring method, medium and system, and the method comprises the following steps: acquiring power time series data of the industrial equipment and dividing the power time series data into fixed-length time window data; inputting the time window data into a one-dimensional residual network to perform multi-level feature extraction on the power signal and obtaining a feature map; inputting the feature map into a convolution block attention module to obtain a double-weighted feature map; inputting the double-weighted feature map into a bidirectional gated recurrent unit to obtain comprehensive time sequence features; inputting the comprehensive time sequence features into a fully connected classification layer to output a probability distribution of the equipment running state, and taking the class with the maximum probability as the predicted state of the equipment; based on a pre-established equipment state-carbon emission intensity mapping relationship, the identified predicted state is converted into the corresponding carbon emission intensity, and the carbon emission amount is calculated. The application can realize high-precision identification of the running state of the industrial equipment and real-time calculation of the carbon emission amount.

Description

Technical Field

[0001] This invention relates to the field of carbon emission technology, specifically to a method, medium, and system for monitoring carbon emissions from industrial equipment. Background Technology

[0002] In recent years, deep learning technology has made significant progress in fields such as time series analysis and equipment condition recognition. Existing technologies include equipment condition monitoring methods based on combined models such as CNN-LSTM and CNN-GRU, but these methods have the following shortcomings: (1) Traditional CNN structures are prone to gradient vanishing in deep networks, which limits the depth and expressive power of feature extraction; (2) A single attention mechanism (such as channel attention only or spatial attention only) cannot fully capture the importance of features, resulting in the loss of key information; (3) The concatenated structure of CNN and RNN lacks an effective feature fusion mechanism, resulting in poor temporal modeling performance; (4) The model has a large number of parameters and high computational complexity, making it difficult to deploy in real time on edge devices; (5) The lack of a complete monitoring process from equipment status to carbon emissions makes it impossible to accurately quantify carbon emissions.

[0003] Therefore, there is an urgent need for a high-precision, lightweight, and real-time deployable method for monitoring carbon emissions from industrial equipment to meet the needs of modern industrial intelligentization and low-carbon development. Summary of the Invention

[0004] To address the technical problems existing in the prior art, the present invention provides a method, medium, and system for monitoring carbon emissions from industrial equipment, enabling high-precision identification of the operating status of industrial equipment and real-time calculation of carbon emissions.

[0005] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows: A method for monitoring carbon emissions from industrial equipment includes the following steps: S1. Acquire the power time series data of industrial equipment, and perform normalization and sliding window processing on the power time series data to divide the continuous power time series data into time window data of fixed length; S2. Input the time window data into a one-dimensional residual network to perform multi-level feature extraction on the power signal and obtain a feature map; S3. Input the feature map into the convolutional block attention module, and perform double feature weighting through the channel attention sub-module and spatial attention sub-module connected in sequence to obtain a double-weighted feature map; S4. Input the double-weighted feature map into the bidirectional gated recurrent unit BiGRU. The feature sequence is processed in forward time order by the forward GRU unit and in reverse time order by the backward GRU unit. The hidden states of each time step are then concatenated to obtain the comprehensive temporal features. S5. Input the comprehensive temporal features into the fully connected classification layer, output the probability distribution of the device's operating state through the softmax function, and take the category with the highest probability as the predicted state of the device; S6. Based on the pre-established equipment status-carbon emission intensity mapping relationship, the identified predicted status is converted into the corresponding carbon emission intensity, and the carbon emission amount is calculated by combining the continuous operating time of the equipment in the corresponding status.

[0006] Preferably, in step S2, the one-dimensional residual network includes multiple residual blocks, each residual block including two one-dimensional convolutional layers, a batch normalization layer and a ReLU activation function; the output of the residual block is added to its input through a skip connection, expressed as: Output = F(X) + X, where F(X) represents the convolution operation and X is the input.

[0007] Preferably, the plurality of residual blocks include a first residual block, a second residual block, and a third residual block connected in sequence; The first residual block is used to expand the input feature dimension to 64 dimensions; The second residual block is used to expand the 64-dimensional features to 128 dimensions; The third residual block is used to expand the 128-dimensional feature to 256 dimensions; All convolutional layers have a kernel size of 3, a stride of 1, and use same padding to keep the time step length constant.

[0008] Preferably, in step S3, the channel attention submodule is used for: The input feature map is simultaneously subjected to global average pooling and global max pooling to obtain two one-dimensional channel descriptors. The two one-dimensional channel descriptors are respectively input into a shared multilayer perceptron (MLP); The outputs of the two MLPs are summed, and channel attention weights are generated using a sigmoid activation function. ; The channel attention weights The feature map is multiplied element by element to obtain the channel-weighted feature map F'.

[0009] Preferably, in step S3, the spatial attention submodule is used for: The channel-weighted feature map F' is subjected to average pooling and max pooling along the channel dimension to obtain two two-dimensional feature maps; The two two-dimensional feature maps are concatenated along the channel dimension and then subjected to a one-dimensional convolution operation. Spatial attention weights are generated after the sigmoid activation function. ; The spatial attention weight The feature map F' is multiplied element by element to obtain the final double-weighted feature map F''.

[0010] Preferably, in step S4, the bidirectional gated loop unit BiGRU consists of a forward GRU and a backward GRU, and the calculation of the GRU unit includes: Update Gate: ; in, The input vector at the current time step, The hidden state vector from the previous time step. The weight matrix is ​​input to the update gate. Represents standard matrix multiplication. It is the sigmoid activation function. The output of the update gate of the GRU unit at time t; Reset Door: ; in, The weight matrix is ​​input to the reset gate. To reset the gate's output at time t; Candidate hidden state: ; in, For activation function, The weight matrix is ​​input to the candidate hidden gate. This represents element-wise multiplication between matrices. The output of the candidate hidden gate at time t; Final hidden state: ; in, This represents the output of the final hidden state at time t.

[0011] Preferably, in step S4, the bidirectional gated recurrent unit (BiGRU) is divided into two layers, with the hidden dimension of the first layer set to 128 and the hidden dimension of the second layer set to 64; a dropout layer is added between the two layers to prevent overfitting.

[0012] Preferably, in step S6, the equipment status-carbon emission intensity mapping relationship is a pre-constructed lookup table, which contains the unit-time carbon emission intensity of each piece of equipment under different operating states in different industrial scenarios; the formula for calculating the carbon emission amount is: ,in, This indicates the direct carbon emissions of the equipment. Indicates device status The corresponding carbon emission intensity, This indicates the duration of continuous operation of the device in this state.

[0013] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, the computer program performing the steps of the method described above when run by a processor.

[0014] The present invention further discloses an industrial equipment carbon emission monitoring system, including a memory and a processor connected to each other, wherein the memory stores a computer program, and the computer program executes the steps of the method described above when run by the processor.

[0015] Compared with the prior art, the advantages of the present invention are as follows: This invention achieves effective fusion of multi-scale features through a 1D ResNet-CBAM-BiGRU hybrid architecture, achieving an average recognition accuracy of 90.56% in multiple industrial scenarios, significantly outperforming traditional methods. Employing sliding window technology and a lightweight network design, this invention enables real-time carbon emission monitoring at the 5-minute level, meeting the demands of intelligent carbon management in the Industry 4.0 era. Using non-intrusive load monitoring technology, this invention requires only one monitoring device installed on the main power line to achieve multi-device status identification, significantly reducing deployment costs compared to traditional invasive methods. This invention achieves high-precision prediction in five different industrial sectors: chemical, glass, metal, plastic, and steel, demonstrating its strong adaptability to varying industrial processes and equipment complexity. This invention establishes a complete technology chain from equipment status identification to carbon emission calculation, providing enterprises with a directly usable carbon emission monitoring solution. This invention boasts advantages such as strong feature extraction capabilities, efficient attention mechanism, accurate temporal modeling, lightweight model, and convenient deployment, enabling high-precision identification of industrial equipment operating status and real-time calculation of carbon emissions, providing technical support for industrial enterprises' carbon asset management and energy conservation and emission reduction. Attached Figure Description

[0016] Figure 1 This is a flowchart of the industrial equipment carbon emission monitoring method based on 1D ResNet-CBAM-BiGRU according to an embodiment of the present invention.

[0017] Figure 2 This is a diagram of the BiGRU structure in an embodiment of the present invention.

[0018] Figure 3 This is a distribution chart of F1 scores for different methods in the embodiments of the present invention.

[0019] Figure 4This is a graph showing the 24-hour direct carbon emission intensity monitoring results of industrial equipment in an embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0021] like Figure 1 As shown, the industrial equipment carbon emission monitoring method provided in this embodiment of the invention includes the following steps: S1, Data Acquisition and Preprocessing: The total power time series data of industrial equipment is collected by the power monitoring equipment installed on the main power supply line at a frequency of 1Hz; the collected power data is normalized to eliminate the influence of power level differences between different equipment; the sliding window technology is used to divide the continuous power time series data into fixed-length time windows, with the window length set to 300 time steps (corresponding to 5 minutes of monitoring time).

[0022] S2, 1D ResNet Feature Extraction: A one-dimensional residual network (1D ResNet) is used to extract features from the power signal at multiple levels to obtain a feature map. The gradient vanishing problem of deep networks is solved by the residual connection structure. Multiple residual blocks are constructed, each containing two one-dimensional convolutional layers, a batch normalization layer, and a ReLU activation function; The jump connection structure of the residual block is as follows: (1) Where F(X) represents the convolution operation and X is the input; when the input and output dimensions do not match, the dimension of X is adjusted by 1×1 convolution; The first residual block expands the input features from F dimensions to 64 dimensions, with a kernel size of 3 and a stride of 1. The second residual block expands the 64-dimensional features to 128-dimensional features, with a kernel size of 3 and a stride of 1. The third residual block expands the 128-dimensional features to 256-dimensional features, with a kernel size of 3 and a stride of 1. All convolutional layers use the same padding method to keep the time step length constant, and batch normalization layers accelerate convergence and prevent overfitting.

[0023] S3, Dual Attention Mechanism: The Convolutional Block Attention Module (CBAM) includes two attention modules: channel and spatial. The channel attention submodule and the spatial attention submodule are concatenated, and the feature map is weighted by channel and location features in turn to obtain a dual-weighted feature map; the CBAM structure is as follows. Figure 2 As shown.

[0024] Specifically, the channel attention submodule: a) Perform global average pooling and global max pooling on the input feature map simultaneously to obtain two one-dimensional channel descriptors; b) Input the two one-dimensional channel descriptors into a shared multilayer perceptron (MLP), which employs a bottleneck structure and has a dimensionality reduction ratio of 16. c) After summing the outputs of the two MLPs, channel attention weights are generated using the sigmoid activation function. ; d) Adjust channel attention weights The input feature map is multiplied element by element to obtain the channel-weighted feature map F'.

[0025] Spatial Attention Submodule: a) Perform average pooling and max pooling along the channel dimension on the channel-weighted feature map F' to obtain two two-dimensional feature maps; b) Concatenate the two two-dimensional feature maps along the channel dimension and perform a one-dimensional convolution operation with a kernel size of 7. c) Spatial attention weights are generated using the sigmoid activation function. ; d) Spatial attention weights Multiplying the feature map F' element by element yields the final double-weighted feature map F'' (corresponding to the second feature map).

[0026] The channel attention and spatial attention work together to dynamically weight the features one channel at a time and one position at a time.

[0027] S4, BiGRU Temporal Modeling: In order to effectively capture the temporal dependencies and state transition patterns of industrial equipment operating states, a bidirectional gated cyclic unit (BiGRU) is used to process the CBAM-weighted feature sequence to capture the long-term temporal dependencies of equipment operating states. BiGRU consists of a forward GRU and a backward GRU. The forward GRU processes the feature sequence in ascending time order, while the backward GRU processes the feature sequence in descending time order. The update gate calculation formula for the GRU unit is as follows: (2) In the formula, The input vector at the current time step, The hidden state vector from the previous time step. The weight matrix is ​​input to the update gate. Represents standard matrix multiplication. It is the sigmoid activation function. The output of the update gate of the GRU unit at time t, with a value range of [0, 1].

[0028] The formula for resetting the door is: (3) In the formula, The weight matrix is ​​input to the reset gate. This is to reset the gate's output at time t.

[0029] The formula for calculating candidate hidden states is: (4) In the formula, As the activation function, select The activation function is used to ensure that the output of the candidate hidden states includes both positive and negative values, thereby enhancing the model's ability to represent information. The weight matrix is ​​input to the candidate hidden gate. This represents element-wise multiplication between matrices. Let be the output of the candidate hidden door at time t.

[0030] The final hidden state calculation formula is: (5) In the formula, This represents the output of the final hidden state at time t.

[0031] A two-layer BiGRU network was constructed, with the first layer having a hidden dimension of 128 and the second layer having a hidden dimension of 64. A dropout layer was added between the two layers with a dropout rate of 0.3 to prevent overfitting. At each time step, the forward and backward hidden states were concatenated to form a feature representation containing bidirectional temporal information. The bidirectional hidden states at the last time step were extracted as the comprehensive temporal feature representation of the entire time window for subsequent classification.

[0032] S5, Device Status Classification: Input the time-series features output by BiGRU into the fully connected classification layer, perform nonlinear transformation through a multilayer perceptron, and output the probability distribution of the device's operating status. Specifically, a two-layer fully connected classification network is constructed. The first layer maps the features output by the BiGRU from the input dimension to a 64-dimensional hidden space using the ReLU activation function. A dropout layer is added after the first fully connected layer with a dropout rate of 0.5 to enhance the model's generalization ability. The second fully connected layer maps the 64-dimensional hidden features to an output dimension equal to the number of device state categories, K. One-hot encoding is used to represent the device state labels, transforming the multi-state recognition task into a K-classification problem. The output vector is normalized to a probability distribution using the softmax function. (6) Where z is the output of the fully connected layer. Indicates exponentiation. For summation operations, Predict the probability that the sample belongs to class i for the model.

[0033] Select the category with the highest probability as the predicted state of the device: (7) in, An index for the predicted state of the device. This indicates a parameter search operation, which can find the parameter function value that maximizes the objective function.

[0034] The model training uses the cross-entropy loss function: (8) in, This represents element-wise multiplication between matrices. Represents a logarithmic function.

[0035] S6, In the calculation of the industrial equipment carbon emission monitoring method based on 1D ResNet-CBAM-BiGRU: Establish an equipment status-carbon emission intensity mapping table, which contains the unit time carbon emission intensity corresponding to each operating status (no load, light load, rated load, heavy load, start-up, shutdown, etc.), and convert the identified equipment status into the corresponding carbon emission parameters. S7. This invention employs multiple classic metrics to comprehensively evaluate the classification performance of the equipment condition recognition model. Considering the characteristics of multi-classification tasks such as industrial equipment condition recognition, the following metrics are selected for model performance evaluation. Specifically: Accuracy: Measures the overall prediction accuracy of the model, calculated as the proportion of all correctly predicted samples out of the total sample. The formula is: (9) in, , , , These represent true positives, true negatives, false positives, and false negatives, respectively.

[0036] Precision: Evaluates the proportion of samples that a model predicts to belong to a certain class, but which actually belong to that class. It reflects the reliability of the model's predictions. The formula is: Recall: Assesses the proportion of samples that actually belong to a certain category but are correctly identified by the model, reflecting the model's ability to identify each category. The calculation formula is: (10) F1 score: The harmonic mean of precision and recall, which comprehensively considers the model's predictive accuracy and completeness. The calculation formula is: (11) The F1 score balances the trade-off between precision and recall, providing a comprehensive evaluation of model performance. (12).

[0037] As can be seen from the above technical solution of the present invention, the present invention has the following technical effects: This invention achieves effective fusion of multi-scale features through a 1D ResNet-CBAM-BiGRU hybrid architecture, achieving an average recognition accuracy of 90.56% in multiple industrial scenarios, which is significantly better than traditional methods.

[0038] This invention employs sliding window technology and lightweight network design to achieve real-time carbon emission monitoring at the 5-minute level, meeting the demand for intelligent carbon management in the Industry 4.0 era.

[0039] This invention employs non-intrusive load monitoring technology, which only requires the installation of one monitoring device on the main power supply line to achieve the status identification of multiple devices, significantly reducing deployment costs compared to traditional intrusive methods.

[0040] This invention has achieved high-precision prediction in five different industrial fields, including chemical, glass, metal, plastic and steel, proving the method's strong adaptability to different industrial processes and equipment complexities.

[0041] This invention establishes a complete technology chain from equipment status identification to carbon emission calculation, providing enterprises with a directly usable carbon emission monitoring solution.

[0042] This invention has the advantages of strong feature extraction capability, efficient attention mechanism, accurate temporal modeling, lightweight model, and convenient deployment. It can realize high-precision identification of the operating status of industrial equipment and real-time calculation of carbon emissions, providing technical support for carbon asset management and energy conservation and emission reduction of industrial enterprises.

[0043] Example 1 This invention provides a method for monitoring carbon emissions from industrial equipment based on 1D ResNet-CBAM-BiGRU. It utilizes deep learning technology to accurately identify equipment status and, combined with the equipment-carbon emission mapping relationship, achieves real-time and accurate monitoring of carbon emissions from industrial equipment. The specific steps are as follows: 1) Before implementing this invention, a problem modeling for industrial equipment status identification is first performed: Industrial equipment typically has well-defined operating states, which are closely related to specific production processes. Taking a glass factory as an example, its production process includes raw material processing using crushers, high-temperature melting in electric furnaces, forming and processing, and heat treatment in electric quenching furnaces. In this complete process, electric furnaces and electric quenching furnaces are the main carbon-emitting devices.

[0044] The four operating states of the electric furnace are coded: Shutdown status: [1,0,0,0], carbon emission intensity is 0 kg CO2 / h; High-emission operation mode: [0,1,0,0], carbon emission intensity is 415.25 kg CO2 / h; Operating status: [0,0,1,0], carbon emission intensity is 207.43 kg CO2 / h; Second highest emission operating state: [0,0,0,1], carbon emission intensity is 345.04 kg CO2 / h; 2) Secondly, construct a 1D ResNet feature extraction module, a CBAM dual attention mechanism module, and a bidirectional GRU temporal modeling module; 3) Next, data preprocessing and feature extraction The sliding window technique is used to process the input power timing data of industrial equipment: Data normalization: The input power data is Z-score normalized to ensure the numerical stability of network training.

[0045] Sliding window settings: The window length is set to 300 time steps, corresponding to a 5-minute monitoring time window, with a step size of 1, to achieve continuous sliding sampling.

[0046] 1D ResNet Feature Extraction: The preprocessed power time series data is input into the 1D ResNet module, and local features and spatial patterns are extracted through the constructed convolutional layers and residual blocks.

[0047] 4) Temporal feature modeling and attention weighting: BiGRU temporal modeling: The features extracted by 1D ResNet are input into the BiGRU module to process temporal dependencies and output bidirectional temporal features.

[0048] Dynamic attention weighting: The attention mechanism is constructed by inputting the output features of BiGRU, generating dynamic weights and performing feature weighting to obtain the enhanced feature representation.

[0049] Residual connection: The attention-weighted features are fused with the original features through residual connection to prevent information loss.

[0050] 5) Equipment status classification and identification Classification layer construction: A fully connected classification layer is added after the feature extraction network, and the number of neurons in the output layer corresponds to the number of device states.

[0051] Softmax activation: The softmax activation function is used to convert the output into a probability distribution; State determination: Select the category with the highest probability as the predicted state. 6) Establish a state-carbon emission intensity mapping relationship Based on the operating characteristics of industrial equipment, a state-emission intensity lookup table is established, as shown in Table 1: Table 1. Direct Carbon Emission Intensity of Equipment Status in Typical Industrial Scenarios (Unit: kg CO2 / h)

[0052] 7) Real-time carbon emission calculation Carbon emissions are calculated in real time based on the identified device status and mapping table. Instantaneous emission intensity acquisition: based on the identified equipment status i Obtain the corresponding carbon emission intensity from the mapping table. E i ; Time-based emission calculation: Calculate the carbon emissions of the equipment during a specific period of continuous operation under certain conditions.

[0053] in, This indicates the direct carbon emissions of the equipment. Indicates device status The corresponding carbon emission intensity, This indicates the duration of continuous operation of the device in this state.

[0054] Cumulative emissions statistics: By monitoring the equipment state transition process in real time and dynamically recording the duration of each operating state, the total carbon emissions are calculated cumulatively.

[0055] Example 2 For complex industrial scenarios involving the collaborative operation of multiple devices, this invention provides an extended implementation scheme: Multi-device feature fusion: Equipment power decomposition: Non-intrusive load decomposition technology is used to extract the power contribution of each device from the total power signal.

[0056] Feature concatenation: Concatenate 1D ResNet-CBAM-BiGRU features from multiple devices to form a joint feature representation.

[0057] Interactive attention: Introducing an interactive attention mechanism to learn the mutual influence relationships between devices.

[0058] Joint state recognition Multi-task learning: Simultaneously predict the operating status of multiple devices and share underlying feature representations.

[0059] Constraint optimization: Physical constraints are added to ensure the rationality of the combination of multiple device states.

[0060] Joint decoding: Conditional random field (CRF) is used to jointly decode the state sequences of multiple devices.

[0061] Experimental verification results: Experiments were conducted using the Industrial Equipment Identification Dataset (IAID) for verification, combined with... Figure 3 and Figure 4 Analysis of experimental results: Comparative analysis of method performance like Figure 3 The box plots of F1 score distributions for different methods clearly demonstrate the superior performance of the method of this invention. From... Figure 3 It can be observed intuitively that: SVM method performance analysis: The traditional SVM method has the largest F1 score distribution range and the lowest average score, which is about 70.59%. It also exhibits significant variability, reflecting the instability of traditional machine learning methods in complex industrial scenarios.

[0062] CNN method performance analysis: The CNN method performed moderately, with an average of about 80.78%, but it also had some performance fluctuations, indicating that simple convolutional feature extraction has limitations in temporal modeling.

[0063] Performance analysis of the LSTM method: The F1 score distribution of the LSTM method is relatively concentrated, but the overall performance level is limited, with an average of about 80.78%, indicating the limitations of traditional time-series methods in the identification of complex industrial equipment.

[0064] Performance analysis of the 1D ResNet-BiGRU method: The 1D ResNet-BiGRU method exhibits good performance stability. Its F1 score distribution is relatively high and compact, with an average of 88.23%, which proves the effectiveness of combining temporal modeling and convolutional feature extraction.

[0065] Performance analysis of the 1D ResNet-CBAM-BiGRU method: The 1D ResNet-CBAM-BiGRU method proposed in this invention performs best among all the comparison methods. It not only has the highest average F1 score of 90.56%, but also the most compact distribution, showing excellent cross-device generalization ability and stability.

[0066] Table 2 shows a detailed performance comparison of the five methods on seven industrial devices.

[0067] Table 2 Performance Comparison of Five Methods

[0068] Verification of the effectiveness of carbon emission monitoring: like Figure 4 The 24-hour direct carbon emission intensity monitoring results of the industrial equipment shown verify the effectiveness of the method of the present invention in practical applications: Chemical plant reactor monitoring results: From Figure 4 As can be seen from the first sub-figure, the reactor exhibited a relatively stable carbon emission pattern during the monitoring period. The method of this invention accurately captured the equipment's operating state transitions, and the predicted curve highly overlapped with the actual curve, demonstrating good state identification capability.

[0069] Glass factory equipment monitoring results: Figure 4 The second sub-figure shows that, as a high-energy-consuming industrial setting, the combined operation of the glass factory's electric melting furnace and electric quenching furnace generated the highest carbon emission intensity, exceeding 400 kg CO2 / h at its peak. The method of this invention successfully identified the complex multi-equipment collaborative operation mode and accurately predicted the dynamic changes in emission intensity.

[0070] Monitoring results of electric heat treatment furnaces in metal plants: Figure 4 The third sub-figure illustrates the intermittent operation characteristics of the electrothermal treatment furnace in the metal plant during the day, exhibiting a clear start-up and shutdown cycle. The method of this invention accurately identifies this periodic state change pattern.

[0071] Monitoring results of injection molding machines in plastic factories: From Figure 4 The fourth subplot shows that the injection molding machine exhibits a typical cyclical production pattern, maintaining a relatively stable carbon emission intensity during operation. The predicted results almost perfectly match the actual results.

[0072] Monitoring results of electric arc furnaces in steel plants: Figure 4 The fifth sub-figure shows that the electric arc furnace combination in the steel plant mainly operated at low power during the monitoring period, resulting in relatively low overall emission intensity. The method of this invention effectively distinguishes the operating states of different electric arc furnaces, accurately reflecting the actual production situation.

[0073] Quantitative analysis results verification: Table 3 summarizes the calculated daily direct carbon emissions for the five factories. The daily direct carbon emission data shows significant differences in carbon emission levels across different industrial scenarios. The model demonstrated good accuracy within the monitored time period, validating the effectiveness of the 1D ResNet-CBAM-BiGRU model in industrial carbon emission monitoring.

[0074] Table 3 Calculation results of direct carbon emissions from industrial equipment over 24 hours

[0075] This invention utilizes a 1D ResNet-CBAM-BiGRU deep learning architecture to achieve high-precision identification of industrial equipment status and real-time carbon emission monitoring, providing an effective solution for intelligent carbon management in industrial enterprises. This method is non-invasive, low-cost, and high-precision, making it suitable for carbon emission monitoring needs in various industrial scenarios.

[0076] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for monitoring carbon emissions from industrial equipment, characterized in that, Includes the following steps: S1. Acquire the power time series data of industrial equipment, and perform normalization and sliding window processing on the power time series data to divide the continuous power time series data into time window data of fixed length; S2. Input the time window data into a one-dimensional residual network to perform multi-level feature extraction on the power signal and obtain a feature map; S3. Input the feature map into the convolutional block attention module, and perform double feature weighting through the channel attention sub-module and spatial attention sub-module connected in sequence to obtain a double-weighted feature map; S4. Input the double-weighted feature map into the bidirectional gated recurrent unit BiGRU. The feature sequence is processed in forward time order by the forward GRU unit and in reverse time order by the backward GRU unit. The hidden states of each time step are then concatenated to obtain the comprehensive temporal features. S5. Input the comprehensive temporal features into the fully connected classification layer, output the probability distribution of the device's operating state through the softmax function, and take the category with the highest probability as the predicted state of the device; S6. Based on the pre-established equipment status-carbon emission intensity mapping relationship, the identified predicted status is converted into the corresponding carbon emission intensity, and the carbon emission amount is calculated by combining the continuous operating time of the equipment in the corresponding status.

2. The method for monitoring carbon emissions from industrial equipment according to claim 1, characterized in that, In step S2, the one-dimensional residual network contains multiple residual blocks, each residual block including two one-dimensional convolutional layers, a batch normalization layer and a ReLU activation function; the output of the residual block is added to its input through a skip connection, expressed as: Output = F(X) + X, where F(X) represents the convolution operation and X is the input.

3. The method for monitoring carbon emissions from industrial equipment according to claim 2, characterized in that, Multiple residual blocks include a first residual block, a second residual block, and a third residual block connected in sequence; The first residual block is used to expand the input feature dimension to 64 dimensions; The second residual block is used to expand the 64-dimensional features to 128 dimensions; The third residual block is used to expand the 128-dimensional feature to 256 dimensions; All convolutional layers have a kernel size of 3, a stride of 1, and use same padding to keep the time step length constant.

4. The method for monitoring carbon emissions from industrial equipment according to claim 1, 2, or 3, characterized in that, In step S3, the channel attention submodule is used for: The input feature map is simultaneously subjected to global average pooling and global max pooling to obtain two one-dimensional channel descriptors. The two one-dimensional channel descriptors are respectively input into a shared multilayer perceptron (MLP); The outputs of the two MLPs are summed, and channel attention weights are generated using a sigmoid activation function. ; The channel attention weights The feature map is multiplied element by element to obtain the channel-weighted feature map F'.

5. The method for monitoring carbon emissions from industrial equipment according to claim 4, characterized in that, In step S3, the spatial attention submodule is used to: The channel-weighted feature map F' is subjected to average pooling and max pooling along the channel dimension to obtain two two-dimensional feature maps; The two two-dimensional feature maps are concatenated along the channel dimension and then subjected to a one-dimensional convolution operation. Spatial attention weights are generated after the sigmoid activation function. ; The spatial attention weight The feature map F' is multiplied element by element to obtain the final double-weighted feature map F''.

6. The method for monitoring carbon emissions from industrial equipment according to claim 1, 2, or 3, characterized in that, In step S4, the bidirectional gated recurrent unit BiGRU consists of a forward GRU and a backward GRU, and the calculation of the GRU unit includes: Update Gate: ; in, The input vector at the current time step, The hidden state vector from the previous time step. The weight matrix is ​​input to the update gate. Represents standard matrix multiplication. It is the sigmoid activation function. The output of the update gate of the GRU unit at time t; Reset Door: ; in, The weight matrix is ​​input to the reset gate. To reset the gate's output at time t; Candidate hidden state: ; in, For activation function, The weight matrix is ​​input to the candidate hidden gate. This represents element-wise multiplication between matrices. The output of the candidate hidden gate at time t; Final hidden state: ; in, This represents the output of the final hidden state at time t.

7. The method for monitoring carbon emissions from industrial equipment according to claim 1, 2, or 3, characterized in that, In step S4, the bidirectional gated recurrent unit (BiGRU) is divided into two layers. The hidden dimension of the first layer is set to 128, and the hidden dimension of the second layer is set to 64. A dropout layer is added between the two layers to prevent overfitting.

8. The method for monitoring carbon emissions from industrial equipment according to claim 1, 2, or 3, characterized in that, In step S6, the equipment status-carbon emission intensity mapping relationship is a pre-constructed lookup table, which contains the unit-time carbon emission intensity of each piece of equipment under different operating states in different industrial scenarios; the formula for calculating the carbon emission amount is: ,in, This indicates the direct carbon emissions of the equipment. Indicates device status The corresponding carbon emission intensity, This indicates the duration of continuous operation of the device in this state.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program, when run by a processor, performs the steps of the method as described in any one of claims 1-8.

10. An industrial equipment carbon emission monitoring system, comprising an interconnected memory and a processor, wherein the memory stores a computer program, characterized in that, The computer program, when run by a processor, performs the steps of the method as described in any one of claims 1-8.

Images