Coal chemical industry carbon capture efficiency evaluation method and system based on deep learning

Abstract

The invention relates to the field of deep learning, and discloses a coal chemical industry carbon capture efficiency evaluation method and system based on deep learning, and the method comprises the steps: obtaining the process state data of a coal chemical industry carbon capture technology; a probabilistic neural network is used for executing multiple times of random forward inference on the process state data, and multiple efficiency predicted values and observation noise variances are obtained; calculating a first variance component characterizing a cognitive uncertainty based on the plurality of efficiency predictors; calculating a second variance component representing accidental uncertainty based on the plurality of observation noise variances; and generating an out-of-distribution working condition alarm signal based on the first variance component or generating a sensor noise abnormity alarm signal based on the second variance component. Through a dual-channel output structure and an uncertainty decoupling mechanism, accurate targeted identification of model cognitive defects and data quality defects is realized; the blind prediction problem under the non-stable working condition is solved.

Description

Technical Field

[0001] This invention relates to a method and system for evaluating the carbon capture efficiency in coal chemical industry based on deep learning, belonging to the field of deep learning technology. Background Technology

[0002] In current coal chemical carbon capture processes, evaluating capture efficiency is a prerequisite for achieving optimized process control. Traditional online gas analyzers suffer from limitations such as large measurement lag, high maintenance costs, and susceptibility to gas-liquid entrainment. Data-driven soft measurement technology is gradually becoming mainstream. Deep learning algorithms, with their nonlinear feature extraction capabilities, are widely used to establish mapping models between process state variables and key performance indicators. These deterministic neural network models are typically trained based on the independent and identically distributed (IOD) assumption, assuming that the statistical distribution of future inference data is consistent with the training dataset, and fitting the input-output function relationship by minimizing the prediction error. Simply eliminating hardware dependence cannot completely solve the reliability problem of the evaluation system. Existing mainstream deep learning evaluation strategy algorithm architectures still have logical shortcomings. Chinese invention patent CN117668481A discloses a non-invasive coal chemical carbon capture efficiency evaluation method and system, which uses convolutional neural networks to process image spectral data and long short-term memory networks to process time series data, attempting to improve detection accuracy through multimodal feature fusion. However, it essentially constructs a deterministic point estimation model. Such models are typically trained based on the independent and identically distributed (IOD) assumption, assuming that the statistical regularity of future inference data and the training dataset is consistent.

[0003] Coal chemical production is affected by changes in raw coal type, load fluctuations, absorbent decay, and equipment aging. The statistical characteristics of process data are non-stationary and time-varying. When operating conditions drift or extreme operating conditions are not covered by training data, deterministic neural networks lack an inherent self-diagnostic mechanism to perceive changes in the distribution of input data. The model outputs high-confidence scalar predictions in unknown areas outside the training domain. Overconfidence in unknown operating conditions makes it difficult for DCS systems to distinguish between effective predictions and blind guesses, leading to misleading operational risks. To address the reliability issues of models for non-stationary operating conditions, existing technologies often use ensemble learning methods to train multiple independent neural network models and weight the prediction results to evaluate uncertainty. Although this method can quantify confidence, it faces resource and efficiency bottlenecks in industrial edge computing scenarios: training and storing multiple independent models increases computational overhead and memory usage, making it difficult to deploy on embedded controllers with limited computing power; the response latency of multi-model parallel inference does not meet the millisecond-level response requirements of real-time control loops; the ensemble prediction variance only reflects the cognitive uncertainty of model parameters, making it difficult to distinguish the random uncertainty caused by data noise and failing to identify abnormal operating conditions or sensor failures.

[0004] Therefore, how to endow deep learning models with inherent self-diagnostic capabilities under limited computing resources, and to achieve decoupling and quantification of cognitive uncertainty and accidental uncertainty, has become the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of this invention is as follows: A method for evaluating the carbon capture efficiency in coal chemical industry based on deep learning, comprising the following steps:

[0006] Obtain the current process status data of the coal chemical carbon capture process;

[0007] The process state data is input into a pre-trained probabilistic deep neural network, which contains a random deactivation (Dropout) layer that remains on during the inference phase. The output layer is configured as an independent dual-channel structure to output the efficiency prediction and the observation noise variance, respectively.

[0008] Utilizing probabilistic deep neural networks to perform process state data execution Sub-random forward inference, due to the random blocking effect of the dropout layer, obtains A number of efficiency predictions with differences and The variance of observation noise, where It is an integer greater than 1;

[0009] based on The statistical variance of each efficiency prediction value is used as the first variance component. The first variance component is used to quantify the cognitive uncertainty caused by the distribution of the weight parameters of the probabilistic deep neural network.

[0010] based on The arithmetic mean of the variance of each observation noise is used as the second variance component, which is used to quantify the random uncertainty caused by the inherent noise of the process state data.

[0011] Call the preset first threshold and second threshold, compare the first variance component with the first preset threshold, and compare the second variance component with the second preset threshold;

[0012] When the first variance component is greater than the first preset threshold, an off-distribution alarm signal is generated to indicate the model's cognitive defects.

[0013] When the second variance component is greater than the second preset threshold, a sensor noise abnormality alarm signal is generated to indicate data quality defects.

[0014] Preferred, execute The steps of the next random forward inference follow a confidence-driven cascaded inference logic, specifically including: performing a first number of random forward inferences to obtain a first set of efficiency prediction values, the first number being less than... Calculate the local statistical dispersion of the first set of efficiency predictions and compare it with the rapid screening threshold. When the local statistical dispersion is less than the rapid screening threshold, terminate subsequent inference and output the arithmetic mean of the first set of efficiency predictions as the evaluation result. When the local statistical dispersion is greater than or equal to the rapid screening threshold, continue random forward inference of the remaining quantity, where the remaining quantity is... The difference between the first quantity and the second quantity.

[0015] Preferably, the training process of the probabilistic neural network uses the Gaussian log-likelihood loss function. The Gaussian log-likelihood loss function is used to constrain the noise distribution of the input data of the probabilistic neural network while minimizing the prediction error. Its calculation logic satisfies the following relationship: ,in, This represents the value of the loss function. Indicates the number of samples in the training batch. This represents the actual efficiency value. This represents the predicted efficiency value. This represents the variance of the observation noise.

[0016] Preferably, the steps of random forward inference include: in each inference process, randomly blocking the connection paths of some neurons in the probabilistic neural network according to a preset inactivation probability; mapping the process state data to a single efficiency prediction value and observation noise variance through the preserved neuron connection paths; aggregating the differential activation path outputs generated by multiple inferences to construct a set of probability distributions for efficiency prediction values ​​and observation noise variance.

[0017] Preferably, the calculation steps for the first variance component include: calculating The arithmetic mean of the predicted efficiency values ​​is used as the expected efficiency evaluation value, and the following calculations are performed: The degree of dispersion of each efficiency prediction value relative to the expected efficiency evaluation is used as an indicator to characterize the degree to which process state data deviates from the distribution of training data of the probability neural network.

[0018] Preferably, the method further includes a benchmark threshold calibration step: using a probabilistic neural network to perform multiple random forward inferences on the samples in the training dataset, statistically analyzing the cognitive uncertainty set and the accidental uncertainty set corresponding to the samples respectively, determining a first preset threshold based on the statistical quantile of the cognitive uncertainty set, and determining a second preset threshold based on the statistical quantile of the accidental uncertainty set.

[0019] Preferably, the process status data includes absorber temperature distribution data, lean liquor flow rate data, flue gas velocity data, and reboiler steam pressure data; the steps for obtaining the process status data of the coal chemical carbon capture process at the current moment include: normalizing the collected raw sensor data to map it to the input value range of the probabilistic neural network.

[0020] Preferably, the method further includes a closed-loop control step: when the first variance component is less than a first preset threshold and the second variance component is less than a second preset threshold, output... The arithmetic mean of the predicted efficiency values ​​is fed into the distributed control system; the distributed control system adjusts the lean liquid circulation rate or steam input rate based on the arithmetic mean to maintain the carbon capture efficiency within the preset target range.

[0021] Preferably, the step of generating an alarm signal for an out-of-distribution operating condition further includes: triggering an active learning process to mark the current process state data as a high-value sample; requesting human experts to annotate the high-value samples; using the annotated high-value samples to incrementally update and train the probabilistic neural network, expanding the known distribution domain of the probabilistic neural network; and generating a sensor noise anomaly alarm signal further includes: locating the key input variable channel that causes the second variance component to increase; cutting off the automatic control loop permission of the key input variable channel; and outputting maintenance instructions for the physical instruments corresponding to the key input variable channel.

[0022] A deep learning-based carbon capture efficiency evaluation system for coal chemical industry includes:

[0023] The data acquisition module is used to acquire the process status data of the coal chemical carbon capture process at the current moment;

[0024] The probabilistic inference module is used to input process state data into a pre-trained probabilistic neural network and to perform operations on the process state data using the probabilistic neural network. Sub-random forward inference, to obtain One efficiency prediction value and The observation noise variance is denoted by a probabilistic neural network containing a stochastically deactivated structure that remains active during the inference phase, and the output layer is configured as a dual-channel structure that outputs the efficiency prediction and the observation noise variance. It is an integer greater than 1;

[0025] Uncertainty decoupling calculation module, used for... The statistical variance of each efficiency prediction value is calculated as the first variance component, and based on... The arithmetic mean of the variances of the observation noise is used as the second variance component; the first variance component is used to characterize the cognitive uncertainty caused by the parameter distribution of the probabilistic neural network, and the second variance component is used to characterize the random uncertainty caused by the inherent noise of the process state data.

[0026] The fault diagnosis alarm module is used to compare the first variance component with the first preset threshold and the second variance component with the second preset threshold; when the first variance component is greater than the first preset threshold, an off-distribution working condition alarm signal indicating model cognitive defects is generated; when the second variance component is greater than the second preset threshold, a sensor noise abnormality alarm signal indicating data quality defects is generated.

[0027] Compared with the prior art, the beneficial effects of the present invention are:

[0028] 1. Based on a probabilistic inference architecture with a random deactivation mechanism, this approach addresses the problem of overconfidence in predictions by deterministic neural networks under non-stationary conditions at the algorithmic structure level. During the inference phase, the network's internal connections are randomly blocked, allowing multiple differentiated activation paths to be triggered for a single input of the same process state data. This yields an approximate Bayesian posterior distribution prediction set without increasing the training cost of independent models. It transforms a single scalar estimate into a probability distribution estimate, using the statistical dispersion of the prediction set to directly characterize the model's uncertainty regarding the current input data. This enables the evaluation system to autonomously identify abnormal conditions outside the training domain based on the consistency of the network's internal response, avoiding misleading high-confidence evaluation results when data distribution drifts.

[0029] 2. A confidence-driven cascaded inference logic is adopted to solve the problem of fixed computational resource consumption in industrial edge computing scenarios using traditional Monte Carlo approximation methods. By performing a small amount of initial random forward inference and calculating local statistical dispersion, the sample prediction difficulty is mapped to the execution depth of the computation graph. When the local dispersion is lower than the steady-state locking threshold, the current working condition is determined to be in a high confidence interval and the subsequent inference process is automatically truncated. When prediction divergence occurs within the model, full inference is activated to obtain complete distribution information. Based on the dynamic gating mechanism of intermediate statistical features, the system reduces the average inference computing load and response latency during long-term steady-state operation while ensuring the detection sensitivity of complex working conditions.

[0030] 3. By utilizing a dual-channel output structure in conjunction with a heteroscedastic regression loss function, orthogonal decoupling of cognitive uncertainty and accidental uncertainty is achieved. The neural network is configured to simultaneously output the mean of efficiency prediction and the variance of observation noise, enabling the model to learn the inherent noise manifold of the input data while minimizing the prediction error. By separately statistically analyzing the dispersion of the prediction mean and the expected value of the noise variance in multiple inferences, the cognitive blind spot caused by missing model training data and the data quality degradation caused by physical interference from sensors are clearly distinguished. This ability to identify the root cause of the fault provides a differentiated decision-making basis for subsequent control strategies, avoiding the misjudgment of sensor noise as operating condition drift that triggers unnecessary model retraining or parameter adjustments. Attached Figure Description

[0031] Figure 1This is a schematic diagram of the carbon capture efficiency evaluation and closed-loop control process that integrates an uncertainty decoupling mechanism according to the present invention.

[0032] Figure 2 This is a schematic diagram comparing the cumulative inference time of confidence-driven cascaded inference and fixed inference strategies according to the present invention;

[0033] Figure 3 This is a schematic diagram of the multi-dimensional components and logical architecture of the coal chemical carbon capture efficiency evaluation system of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in further detail below. Obviously, the described embodiments are only some embodiments of this invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0035] This invention provides a deep learning-based method and system for evaluating the efficiency of carbon capture in coal chemical industry. The system comprises a process state perception module, a probabilistic inference calculation engine, an uncertainty decoupling analysis module, and a closed-loop decision control module. These modules interact via industrial Ethernet or fieldbus. The process state perception module collects process state data from the coal chemical carbon capture process. The module uses a sensor array distributed at key nodes of the absorption tower and regeneration tower to collect data at a preset sampling frequency. to Between specific values, process state data are used as multidimensional feature vectors. The physical components include temperature measurement data distributed along the height of the absorber, lean liquor inlet flow rate data, raw material flue gas velocity data, and reboiler steam pressure data. Before being input into the neural network, the standardization preprocessing unit performs zero-mean unit variance normalization on the process state data based on the statistical characteristics of the training dataset, mapping physical quantities of different dimensions to the dimensionless standard normal distribution interval. The probabilistic inference calculation engine uses a pre-trained probabilistic deep neural network, which includes an input layer, several stacked hidden layers, and an output layer. Random deactivation layers are embedded between hidden layers or between hidden layers and the output layer. The random deactivation layers remain on during the inference phase, i.e., they are activated with a preset deactivation probability during each forward propagation. By randomly setting the activation values ​​of some neurons to zero according to a Bernoulli distribution, random perturbations are introduced into the network structure. The preset inactivation probability p follows a layer-differentiated setting strategy: for shallow hidden layers adjacent to the input layer, p is set to a low probability range, such as 0.05 to 0.15, to preserve the original feature integrity of key physical variables such as absorber temperature and lean liquid flow rate, and to prevent distortion of input physical information due to high-probability random blocking; for deep hidden layers adjacent to the output layer, p is set to a high probability range, such as 0.3 to 0.5, to introduce sufficient random perturbations into the abstract feature space, ensuring that the model has sufficient cognitive uncertainty estimation capability. The output layer of the probabilistic neural network is an independent dual-channel structure, outputting scalars representing the expected carbon capture efficiency prediction. and a scalar representing the variance of the inherent observation noise of the current input data. The network weight training process uses the heteroscedastic Gaussian log-likelihood loss function, calculated as follows: ,in, This represents the value of the loss function. Indicates the number of samples in the training batch. This represents the actual efficiency value. This represents the predicted efficiency value. This represents the variance of the observation noise, and the model fits the true efficiency value while minimizing the loss function. And learn the characteristics of the input data and the observation noise. Nonlinear mapping relationship between them.

[0036] The system executes a confidence-driven cascaded stochastic forward inference procedure. During real-time evaluation, the inference engine evaluates the same normalized input vector. Execute the first preset quantity stochastic forward inference, Set as Due to the effect of the random deactivation layer, this The next inference corresponds to Different subnetwork structures generate the first set of differentiated efficiency predictions. The system calculates the local statistical dispersion of the predicted values ​​and compares the local dispersion with a preset rapid screening threshold. In comparison, when the local dispersion is less than When the system determines that the model is in a steady-state condition in the current input region, it terminates subsequent inference and directly outputs the arithmetic mean of the predicted values ​​as the final evaluation result. This applies when the local dispersion is greater than or equal to... When this happens, the system activates the remaining inference process and continues to execute random forward inference until the total number of iterations reaches [a certain threshold]. N is set to 50 iterations, as determined by statistical convergence verification. This value is selected following the principle of balancing accuracy and time delay: under non-stationary, high-fluctuation conditions, when the number of inferences is less than 30, the statistical volatility of the variance estimate is significantly higher (>15%), leading to insufficient confidence; while when the number exceeds 50 (e.g., 100), the marginal gain of accuracy in uncertainty quantification decreases, and the accumulated inference time will cause the response delay to exceed the real-time control cycle; simultaneously, the initial number of inferences n... min The value is set to 5, based on the percentage of steady-state operation time (>80%) in historical operating conditions, thus achieving adaptive matching between inference computing power and operating condition complexity; subsequently, the system performs uncertainty decoupling analysis based on the complete N inference results; and obtains... Group inference results Subsequently, the uncertainty decoupling analysis module calculates the cognitive uncertainty index and the accidental uncertainty index, and the system calculates... One efficiency prediction value Statistical variance as the first variance component The calculation formula is as follows: ,in, for The mean of the predicted values, the first variance component quantifies the model weight distribution, leading to cognitive uncertainty. This characterizes the degree to which the current input data deviates from the training data distribution. The system calculates... Observation noise variance The arithmetic mean as the second variance component The calculation formula is as follows: The second variance component quantization data inherently possesses random uncertainty, characterizing the sensor noise or transient interference level. The system calls a pre-calibrated benchmark threshold for judgment. The benchmark threshold calibration procedure involves inputting samples from the training set into a trained probabilistic neural network, statistically analyzing the cognitive uncertainty and random uncertainty values ​​corresponding to all samples, and selecting the 95th percentile of the statistical distribution determined based on the sensitivity balance principle. This quantization standard aims to strictly control the theoretical false alarm rate of the system under normal steady-state operating conditions to within 5%, thereby establishing the effective boundary of the confidence interval. This avoids frequent false alarms caused by long-tailed noise while ensuring the detection sensitivity to small distribution shifts. The system selects this 95th percentile as the first preset threshold. Second preset threshold The closed-loop decision control module executes control or maintenance strategies based on the comparison results of the two variance components with a threshold. When the first variance component... Greater than When the system determines that the current operating condition is an out-of-distribution condition, it generates an out-of-distribution alarm signal to indicate a cognitive defect in the model, triggers the active learning process, marks the current process state data as high-value samples and stores them in the annotation database, and requests human experts to intervene in annotation for incremental model updates. During this process, the system disables automatic control permissions. When the second variance component... Greater than When the system determines that the current data quality is abnormal, it generates a sensor noise anomaly alarm signal and calls the gradient absolute value sorting algorithm to locate the cause of σ. 2 The key input variable channel is increased, and the specific execution logic is as follows: the variance value σ of the second channel of the output layer is calculated using the backpropagation mechanism. 2 The gradient vector ∇ is obtained by taking the partial derivative with respect to the input layer feature vector x. x σ 2 Calculate the absolute value of each element in the gradient vector and sort them in descending order. The input dimension corresponding to the absolute value of the first gradient element is then defined as σ. 2 The system raises the threshold of key input variable channels, automatically disconnects the corresponding automatic control loop, and outputs maintenance commands. When both variance components are below their respective thresholds, the system determines the current evaluation result is reliable. The arithmetic mean of the predicted sub-efficiency values The signal is sent to the distributed control system, which then calculates the adjustment command for the lean liquor circulation rate or the reboiler steam input rate. For example, when the evaluation efficiency is lower than the target set value, the output frequency of the lean liquor pump inverter is increased through the PID algorithm to maintain the carbon capture efficiency within the preset target range.

[0037] Example 1: In the carbon capture section of a coal-to-oil chemical industrial park with an annual output of 4 million tons, when the feed coal type of the upstream gasifier is switched from bituminous coal under design conditions to high-volatile lignite, the concentration gradient of flue gas components and velocity field data at the inlet of the capture tower exhibit non-stationary characteristics that deviate from the statistical distribution of the historical training dataset, resulting in a process state data vector. Drifting outside the known manifold of the probabilistic deep neural network, and facing off-distribution operating conditions, the process state perception module... The system continuously collects real-time data including absorber temperature distribution, lean liquor flow rate, and reboiler pressure. Zero-mean, unit-variance normalization mapping is performed in the standardized preprocessing unit. The inference engine then initiates a confidence-driven cascaded inference process based on the input vector at that moment. In the initial random forward inference, because the random deactivation layer in the probabilistic deep neural network remains on during the inference phase, and the network weights lack robust activation response consistency to input features from unseen lignite conditions, this... The set of predicted efficiency values ​​generated by the second inference exhibits high dispersion. The system calculates the local statistical dispersion of this set of predicted values ​​and finds that the current local statistical dispersion value is 0.085, which exceeds the preset rapid screening threshold of 0.02. This indicates that there is a discrepancy in the model's understanding of the current operating conditions.

[0038] Given that local dispersion triggers unsteady-state decision logic, the system automatically activates the remaining inference computing resources and continues to perform random forward inference until the total number of iterations is reached. Next, complete probability distribution information is obtained; the uncertainty decoupling analysis module is based on... Group inference results Expand the calculation, where This is a single-shot efficiency prediction. The system calculates the variance of noise for a single observation. One efficiency prediction value Statistical variance was analyzed to obtain the first variance component. The value is higher than the preset first threshold. The quantitative representation model suffers from cognitive uncertainty regarding the current lignite operating conditions due to the lack of training data; system calculation Observation noise variance The arithmetic mean is used to derive the second variance component. The value is still below the preset second threshold. This indicates that the current sensor physical link and signal quality are at a normal level, ruling out hardware fault interference. Based on the decoupling analysis results, the closed-loop decision control module determines that it is currently in the distributed operating condition outside the model's cognitive blind zone. The system generates an alarm signal for the distributed operating condition and automatically blocks the automatic adjustment authority for the frequency converter of the lean liquor circulation pump to prevent the lean liquor flow rate from fluctuating significantly or solvent flooding accidents due to the model's overconfidence in outputting incorrect efficiency evaluation values. The distributed control system switches the control loop to hold mode according to the alarm signal and simultaneously triggers the active learning process. The current process status data containing lignite operating condition characteristics is marked as high-value samples and stored in the database for human experts to annotate and use for incremental updates of neural network weights.

[0039] Example 2: To objectively verify the effectiveness and engineering performance of the method of the present invention, a high-fidelity test platform based on the joint simulation of AspenPlusDynamics and the Python deep learning framework was constructed. The platform uses historical operating data of a 400,000-ton-per-year carbon dioxide capture unit in a chemical industrial park as a benchmark, and the data sampling frequency is set to [missing information]. The test covers key variables such as absorber temperature distribution, lean liquor flow rate, flue gas velocity, and reboiler steam pressure. To simulate electromagnetic interference in a real industrial environment, a signal-to-noise ratio of [value missing] was actively superimposed on the original sensor signals during the test. Gaussian white noise; regarding experimental parameter settings, the number of random forward inferences was... Selecting and executing a decision-making procedure based on the trade-off between inference stability and computational delay, experiments show that when Less than At that time, the first variance component Statistical volatility exceeds It cannot stably represent the cognitive uncertainty of the model; when Increase to When the variance-estimated volatility converges to The average reasoning time for a single evaluation is controlled within the following range. Within this range, meet the requirements of the DCS system. Real-time control cycle requirements; if Continue to increase to As shown above, the accuracy gain decreases while the computation time increases linearly. Based on this, this embodiment will... Set as The experiment was designed with four comparative sample groups. The sample groups of this invention adopted the dual-channel probabilistic neural network and cascaded inference strategy in the specific implementation. The control group 1 adopted a traditional deterministic deep neural network with only a single scalar output. The control group 2 adopted a deep ensemble network containing 5 independently trained models. The control group 3 adopted a network with a Dropout layer but only a single-channel probabilistic network as the output layer. The models of each group were trained on the same training set and the input included three types of test samples: normal steady state, out-of-distribution operating conditions and abnormal sensor noise.

[0040] Table 1 shows the performance comparison data of each sample group under different test conditions. Referring to Table 1, under out-of-distribution conditions, the root mean square error of prediction for control group 1 is as high as... However, due to the inability to output uncertainty indicators, the DCS system misjudged it as a valid prediction; although control group 2 identified out-of-distribution features, its average inference time reached [missing information]. This exceeds real-time control requirements; control group 3 detected increased uncertainty, but could not distinguish whether it stemmed from operating condition drift or data noise, resulting in a fault diagnosis accuracy of only [percentage missing]. The first variance component of the sample group in this invention increases to [a certain value] under out-of-distribution operating conditions. Exceeding the threshold Under sensor noise conditions, the second variance component responds independently and increases to Exceeding the threshold This achieves precise decoupling of the root cause of the fault, resulting in a fault diagnosis accuracy rate of [missing information]. The average reasoning time is .

[0041] Table 1: Performance Comparison Test Data for Each Sample Group

[0042]

[0043] Experimental data confirms that this invention, through the synergy of dual-channel output and heteroscedasticity loss function, decouples cognitive uncertainty from accidental uncertainty while ensuring prediction accuracy, thus solving the problem of the inability of control group 3 to distinguish fault types; and utilizes a cascaded inference mechanism to reduce inference time under steady-state conditions from Reduce to This avoids the drawback of excessive computational resource consumption in control group 2.

[0044] Example 3: This example combines Figures 1 to 3 This paper describes a deep learning-based method and system for evaluating carbon capture efficiency in coal chemical industry. Figure 1 As shown, this module is configured to collect temperature, flow rate, velocity, and pressure data from the absorption tower. The collected raw data undergoes zero-mean unit variance normalization mapping via a standardized preprocessing unit to generate normalized feature vectors, which are then input to the probabilistic inference calculation engine. This engine internally deploys a probabilistic deep neural network containing a random deactivation layer (Dropout) and a dual-channel output structure. Driven by confidence, it executes a cascaded stochastic forward inference process, generating an inference set containing the efficiency prediction value and the observation noise variance through N inferences. This set is then transmitted to the uncertainty decoupling analysis module, which calculates the first variance component and the second variance component in parallel, representing the cognitive uncertainty of the model and the random uncertainty of the data, respectively. The system performs a split based on a dual-threshold judgment logic: if the first variance component is greater than a first preset threshold, the system generates an out-of-distribution operating condition alarm signal to indicate a cognitive defect in the model and triggers an active learning and model update process. In the process, the screening criterion is defined as samples whose first variance component exceeds the first preset threshold as high-value samples. The system stores such samples in the unlabeled database and requests human experts to intervene in the labeling for incremental model updates. After the weight update is completed, the system automatically performs threshold dynamic drift calibration, re-statistically updates the distribution quantiles using the updated validation set, and synchronously updates the first and second preset thresholds to adapt to the new cognitive boundaries of the model. If the second variance component is greater than the second preset threshold, the system generates a sensor noise anomaly alarm signal to indicate data quality defects and triggers the fault diagnosis and maintenance process, locates the key channel and cuts off the loop to perform maintenance. If the two variance components are lower than their respective thresholds, the data enters the closed-loop decision control module and outputs the predicted mean to verify reliability. Finally, the efficiency predicted mean is sent to the distributed control system (DCS) to adjust the lean liquid circulation volume or steam input to achieve closed-loop control.

[0045] like Figure 2As shown in the figure, the horizontal axis represents the number of inferences, ranging from 0 to 10; the vertical axis represents the cumulative inference time, in milliseconds, ranging from 0 to 25. The figure contains three trend lines to characterize the time characteristics of different strategies: the dashed line represents a strategy with a fixed number of inferences (50), exhibiting stability in the high-time range, serving as a baseline upper limit for system performance; the solid line represents a cascaded inference strategy for unsteady-state conditions, where time increases linearly with the number of inferences, indicating that the system needs to perform more inferences to obtain confidence under complex conditions; the dotted line represents a cascaded inference strategy for steady-state conditions, reaching a plateau in time consumption with fewer inferences, showing that this strategy can quickly truncate the inference process and reduce computational latency under steady-state conditions. Figure 3 As shown, this fishbone diagram details the multidimensional structure of the coal chemical carbon capture efficiency evaluation system. The fish head points to the core theme of the coal chemical carbon capture efficiency evaluation system, and the main trunk of the fishbone branches out into six key sub-modules: the process state perception branch covers the absorption tower temperature distribution, lean liquid flow data, flue gas velocity / pressure, and normalized preprocessing elements; the probabilistic network architecture branch includes random deactivation layer Dropout, dual-channel output structure, and heteroscedasticity loss function elements; the cascaded inference strategy branch involves confidence-driven logic, local statistical dispersion, fast screening threshold determination, and N-times random inference elements; the uncertainty decoupling branch corresponds to cognitive uncertainty parameters and the first variance component, and random uncertainty noise and the second variance component elements; the fault diagnosis and alarm branch includes off-distribution operating condition alarms, cognitive defect identification, sensor noise anomalies, and data quality defects elements; and the closed-loop decision control branch includes active learning processes, DCS parameter adjustment, and maintenance instructions elements.

[0046] Example 4: This example provides a robust training procedure for enhancing the robustness of a noise injection and adversarial perturbation model. In the initial state definition phase, the procedure targets the set of weight parameters of a probabilistic deep neural network that has already undergone basic training. The implementation environment is a deep learning computing workstation equipped with a high-performance graphics processing unit. The core of the training procedure lies in constructing a dynamically enhanced training dataset. In each training iteration, a batch of process state samples is randomly selected from the original historical database. A physically constrained noise injection operation is performed on this batch of samples, superimposing random noise following a specific distribution onto the data in each input channel. For temperature measurement point data, the superimposed noise has a mean of zero and a variance of... Gaussian white noise is used to simulate thermocouple measurement fluctuations; the amplitude of the superimposed flow and pressure data is the original value. Uniformly distributed noise is used to simulate transmitter quantization error. Samples with added noise are labeled as enhanced samples and used as model input along with the original samples.

[0047] Building upon this foundation, an adversarial perturbation generation mechanism is further introduced. Utilizing the fast gradient sign method, based on the current model's gradient information on the input data, a small perturbation that maximizes the model's prediction error is calculated. This perturbation is then applied to the augmented samples, generating adversarial examples. During training, the model must minimize not only the prediction error of the original samples but also the prediction error and uncertainty bias of the adversarial examples. Through this adversarial training strategy, the model learns more robust feature representations, maintaining prediction stability even when faced with small but malicious input perturbations and correctly outputting high uncertainty estimates as warnings. To quantitatively evaluate the model's convergence state during training, a weighted composite loss function is defined, and weight coefficients are assigned to each of the three components. , and In this embodiment, the following is set To establish the dominant role of fitting the original data, set To introduce robustness constraints without compromising convergence, the following is set: To prevent overfitting, the weighted loss function consists of three parts: a Gaussian log-likelihood loss based on the heteroscedasticity of the original samples, a prediction consistency loss based on adversarial examples, and a regularization term. The prediction consistency loss constrains the model to make its predictions as consistent as possible between the original and adversarial samples, enhancing model smoothness. During training, the trend of the composite loss function value on the validation set is monitored in real time. When the validation set loss value is continuous... If no decrease occurs within a certain number of cycles, the model is considered to have reached convergence, the training process is terminated, and the final model weight parameters are the optimized parameters after robustness enhancement.

[0048] Example 5: Before deploying the probabilistic deep neural network in an actual industrial control system, a standardized initial benchmark calibration procedure is executed. This procedure determines the cognitive uncertainty benchmark distribution and inputs a historical normal operating data set covering the entire operating condition range into the model to be calibrated. In a second random forward inference, the first variance component of each sample is calculated. Based on the obtained set of first variance components, a probability density function is fitted, and the result is determined according to a pre-set confidence level. The statistical upper bound of the first variance component is determined and solidified as the first preset threshold for judging out-of-distribution operating conditions. This process ensures that the model's cognitive boundary of known operating conditions is quantitatively defined, avoiding the arbitrariness of manually setting thresholds. For benchmark calibration for accidental uncertainties, a calibration method based on controlled noise injection is adopted. In an offline testing environment, a white noise sequence with known variance gradient is gradually superimposed on the benchmark test signal. The output second variance component of the model is recorded at different noise levels, and a mapping curve between the input noise variance and the output second variance component is constructed. The curve is linearly regressed using the least squares method to correct the model output variance proportionality coefficient, so that the model output second variance component can estimate the true noise level of the input data without bias. After calibration, the value corresponding to the maximum allowable measurement error variance of the sensor is selected as the second preset threshold to establish the benchmark for judging data quality anomalies.

[0049] Example 6: This example provides a standardized procedure for initial parameter calibration and model construction to ensure the reproducibility and robustness of the coal chemical carbon capture efficiency evaluation system in actual engineering deployment. Before the system is put into operation, the hyperparameter configuration of the probabilistic deep neural network is determined based on the statistical distribution characteristics of historical operating data. For the selection of the number of network layers, an optimization process based on the trade-off between information bottleneck and gradient vanishing is executed. When the input feature dimension exceeds 50 and the sample size is greater than 100,000, the network depth is set to 5 to 8 layers to balance feature extraction capability and training stability, and the random inactivation probability. The settings follow the principle of minimizing calibration error based on validation set uncertainty, through... to Intra-interval grid search to determine the optimal calibration error value. To ensure that the model's output prediction distribution truly reflects the level of cognitive uncertainty, a threshold for uncertainty judgment is calibrated. A statistical inference procedure based on extreme value theory is executed, and historical operating data of the system under known steady-state normal operating conditions is collected as a calibration set. A trained probabilistic neural network is then used to infer the first variance component from this calibration set. With the second variance component An empirical distribution is used, and a generalized Pareto distribution is applied to fit the tail data of the two variance components. Based on the preset false alarm rate, the extreme points of the corresponding cumulative probability density function are calculated and used as the first preset threshold. Second preset threshold .

[0050] Furthermore, to address the issue of missing or abruptly changing input data during actual operation, the system incorporates a pre-test and repair module based on spatiotemporal correlation. When data loss is detected in a sensor channel, it utilizes the current values ​​of other relevant sensor data and the channel's historical time-series data to generate a filler value through a multivariate interpolation algorithm or a long short-term memory network predictor. The confidence weight is then calculated based on the prediction consistency principle. Specifically, the algorithm calculates the normalized Euclidean distance between the filler value and the channel's historical short-period moving average (e.g., the last 10 sampling points). The smaller the distance, the higher the confidence weight. If the filler data is... If the reliability falls below a preset safety threshold, such as 0.6, a degraded operating mode is automatically triggered, and a fault-tolerant strategy is implemented: the output value of the actuator in the current control loop is forcibly locked to the state of the previous effective time step, the optimization instructions of the deep learning model are blocked, and a manual takeover request is simultaneously output. This ensures the inherent safety of the system under data quality impairment conditions. The rapid screening threshold calibration is performed based on a statistical analysis procedure using historical steady-state operating data. At least 1000 consecutive time-series samples with carbon capture efficiency fluctuations within ±1% of the target set value are selected from the historical database to form a calibration dataset. For a single sample in the dataset, the probabilistic neural network performs five random forward inferences in the first stage. The coefficient of variation (COP) of the five efficiency predictions, i.e., the ratio of the standard deviation to the arithmetic mean, is calculated as a local dispersion index. The COP eliminates the dimensional influence of the absolute efficiency value on the dispersion judgment. The cumulative distribution function of the local dispersion index of the full calibration sample is statistically analyzed, and the corresponding 95% quantile value of the cumulative probability is selected as the fast screening threshold. The value is fixed in the inference engine configuration parameters and serves as the unique quantification boundary for determining whether the current working condition triggers full inference during the online operation phase. The computing resource scheduling logic is anchored to the statistical distribution characteristics of the actual working condition. The determination of the first and second preset thresholds follows an offline calculation process based on extreme value theory. After the model completes basic training, the network is activated using a validation dataset covering 50% to 110% of the full load range. For each input feature vector, 50 complete random forward inferences are performed. The statistical variance of the 50 efficiency predictions is extracted to form a cognitive uncertainty sample set, and the arithmetic mean of the 50 observation noise variances is extracted to form a random uncertainty sample set. The generalized Pareto distribution function is used to fit the tail data of the two sets, and the corresponding 0.The upper bound of the probability density of the 01-level signal level is set as the first and second preset thresholds. This procedure ensures that the alarm threshold setting is directly traced back to the limit response boundary under known operating conditions of the model. The input data normalization processing parameters are strictly locked to the global statistical characteristics of the offline training database. The arithmetic mean and standard deviation of the data in each physical channel are calculated by traversing the training set. The statistical constants are written into the non-volatile storage space of the online inference module. During the real-time operation, for sensor readings that deviate from the mean by more than three times the standard deviation, a boundary clamping operation is forcibly performed to limit the values ​​to within the physical boundary of plus or minus three times the standard deviation. The aforementioned statistical constants are used to complete the zero-mean unit variance mapping. The preprocessing logic ensures that the distribution of online input data values ​​is always constrained within the activation domain of the neural network.

[0051] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A deep learning-based method for evaluating the efficiency of carbon capture in coal chemical industry, characterized in that, Includes the following steps: Obtain the current process status data of the coal chemical carbon capture process; The process state data is input into a pre-trained probabilistic deep neural network, which contains a randomly deactivated layer that remains on during the inference phase. The output layer is configured as an independent dual-channel structure to output the efficiency prediction and the observation noise variance, respectively. Utilizing probabilistic deep neural networks to perform process state data execution Sub-random forward inference, due to the random blocking effect of the random deactivation layer, obtains A number of efficiency predictions with differences and The variance of observation noise, where It is an integer greater than 1; based on The statistical variance of each efficiency prediction value is used as the first variance component. The first variance component is used to quantify the cognitive uncertainty caused by the distribution of the weight parameters of the probabilistic deep neural network. based on The arithmetic mean of the variance of each observation noise is used as the second variance component, which is used to quantify the random uncertainty caused by the inherent noise of the process state data. Call the preset first threshold and second threshold, compare the first variance component with the first preset threshold, and compare the second variance component with the second preset threshold; When the first variance component is greater than the first preset threshold, an off-distribution alarm signal is generated to indicate the model's cognitive defects. When the second variance component is greater than the second preset threshold, a sensor noise abnormality alarm signal is generated to indicate data quality defects.

2. The method for evaluating the efficiency of carbon capture in coal chemical industry based on deep learning according to claim 1, characterized in that, implement The steps of the next random forward inference follow a confidence-driven cascaded inference logic, specifically including: performing a first number of random forward inferences to obtain a first set of efficiency prediction values, the first number being less than... Calculate the local statistical dispersion of the first set of efficiency predictions and compare it with the rapid screening threshold. When the local statistical dispersion is less than the rapid screening threshold, terminate subsequent inference and output the arithmetic mean of the first set of efficiency predictions as the evaluation result. When the local statistical dispersion is greater than or equal to the rapid screening threshold, continue random forward inference of the remaining quantity, where the remaining quantity is... The difference between the first quantity and the second quantity.

3. The method for evaluating the efficiency of carbon capture in coal chemical industry based on deep learning according to claim 1, characterized in that, The training process of a probabilistic neural network uses the Gaussian log-likelihood loss function. This function is used to constrain the noise distribution of the input data while minimizing the prediction error. Its calculation logic satisfies the following relationship: ,in, This represents the value of the loss function. Indicates the number of samples in the training batch. This represents the actual efficiency value. This represents the predicted efficiency value. This represents the variance of the observation noise.

4. The method for evaluating the efficiency of carbon capture in coal chemical industry based on deep learning according to claim 1, characterized in that, The steps of stochastic forward inference include: in each inference process, randomly blocking the connection paths of some neurons in the probabilistic neural network according to the preset inactivation probability; mapping the process state data to a single efficiency prediction value and observation noise variance through the preserved neuron connection paths; aggregating the differential activation path outputs generated by multiple inferences to construct a set of probability distributions of efficiency prediction value and observation noise variance.

5. The method for evaluating the efficiency of carbon capture in coal chemical industry based on deep learning according to claim 1, characterized in that, The steps for calculating the first variance component include: calculating... The arithmetic mean of the predicted efficiency values ​​is used as the expected efficiency evaluation value, and the following calculations are performed: The degree of dispersion of each efficiency prediction value relative to the expected efficiency evaluation is used as an indicator to characterize the degree to which process state data deviates from the distribution of training data of the probability neural network.

6. The method for evaluating the efficiency of carbon capture in coal chemical industry based on deep learning according to claim 1, characterized in that, It also includes a baseline threshold calibration step: using a probabilistic neural network to perform multiple random forward inferences on the samples in the training dataset, statistically analyzing the cognitive uncertainty set and the accidental uncertainty set corresponding to the samples respectively, determining the first preset threshold based on the statistical quantile of the cognitive uncertainty set, and determining the second preset threshold based on the statistical quantile of the accidental uncertainty set.

7. The method for evaluating the efficiency of carbon capture in coal chemical industry based on deep learning according to claim 1, characterized in that, The process status data includes the temperature distribution data of the absorption tower, the lean liquor flow rate data, the flue gas velocity data, and the reboiler steam pressure data. The steps to obtain the process status data of the coal chemical carbon capture process at the current moment include: normalizing the collected raw sensor data to map it to the input value range of the probabilistic neural network.

8. The method for evaluating the efficiency of carbon capture in coal chemical industry based on deep learning according to claim 1, characterized in that, The method also includes a closed-loop control step: when the first variance component is less than a first preset threshold and the second variance component is less than a second preset threshold, the output is... The arithmetic mean of the predicted efficiency values ​​is fed into the distributed control system; the distributed control system adjusts the lean liquid circulation rate or steam input rate based on the arithmetic mean to maintain the carbon capture efficiency within the preset target range.

9. The method for evaluating the efficiency of carbon capture in coal chemical industry based on deep learning according to claim 1, characterized in that, The steps for generating off-distribution operating condition alarm signals also include: triggering an active learning process to mark the current process state data as high-value samples; requesting human experts to annotate the high-value samples; using the annotated high-value samples to incrementally update and train the probabilistic neural network, expanding the known distribution domain of the probabilistic neural network; and generating sensor noise anomaly alarm signals. The steps also include: locating the key input variable channel that causes the second variance component to increase; cutting off the automatic control loop permission of the key input variable channel; and outputting maintenance instructions for the physical instruments corresponding to the key input variable channel.

10. A deep learning-based coal chemical carbon capture efficiency evaluation system, used to implement the deep learning-based coal chemical carbon capture efficiency evaluation method described in claim 1, characterized in that, include: The data acquisition module is used to acquire the process status data of the coal chemical carbon capture process at the current moment; The probabilistic inference module is used to input process state data into a pre-trained probabilistic neural network and to perform operations on the process state data using the probabilistic neural network. Sub-random forward inference, to obtain One efficiency prediction value and Individual observation noise variance; The probabilistic neural network includes a stochastically deactivated structure that remains active during the inference phase, and the output layer is configured as a dual-channel structure that outputs the predicted efficiency value and the variance of the observation noise. It is an integer greater than 1; Uncertainty decoupling calculation module, used for... The statistical variance of each efficiency prediction value is calculated as the first variance component, and based on... The arithmetic mean of the variances of each observation noise is used as the second variance component. The first variance component is used to characterize the cognitive uncertainty caused by the parameter distribution of the probabilistic neural network, and the second variance component is used to characterize the random uncertainty caused by the inherent noise of the process state data. The fault diagnosis alarm module is used to compare the first variance component with the first preset threshold and to compare the second variance component with the second preset threshold. When the first variance component is greater than the first preset threshold, an off-distribution alarm signal indicating cognitive defects in the model is generated. When the second variance component is greater than the second preset threshold, a sensor noise anomaly alarm signal indicating a data quality defect is generated.

Images