Unconventional gas heat exchanger operation state prediction method based on thermodynamic entropy production

Abstract

This invention discloses an unconventional gas heat exchanger operating status prediction method based on thermodynamic entropy production, belonging to the fields of thermal energy engineering and artificial intelligence prediction technology. This invention collects and normalizes heat exchanger operating parameters; constructs heat transfer entropy production, viscous entropy production, and their multi-scale evolution characteristics based on the second law of thermodynamics, and concatenates them with the original parameters to form an enhanced feature matrix; constructs a cost-sensitive genetically optimized multilayer perceptron, uses the Morlet wavelet activation function to extract local mutation features, introduces an adaptive cost-sensitive loss function to enhance the sensitivity of early degradation identification, and globally optimizes network weights, regularization parameters, and wavelet scale factors through a genetic algorithm; constructs an adaptive probability confidence criterion based on the prediction residual, and combines it with comprehensive thermodynamic degradation indicators to form a four-level graded early warning system. This invention solves the problems of poor physical consistency, insensitive early degradation identification, and lack of confidence criteria in existing technologies, improving prediction accuracy and practicality.

Description

Technical Field

[0001] This invention relates to the fields of thermal energy engineering and artificial intelligence prediction technology, and in particular to an unconventional gas heat exchanger operation status prediction method based on thermodynamic entropy production. Background Technology

[0002] Heat exchangers are core thermal equipment in petroleum, chemical, power, and nuclear energy engineering. With the development of new energy technologies, unconventional gases (such as supercritical carbon dioxide, hydrogen, helium, syngas, and high-temperature flue gas) are increasingly used as heat transfer media. Unlike conventional air or water vapor, the thermophysical properties (density, specific heat capacity, viscosity, and thermal conductivity) of unconventional gases exhibit strong nonlinear changes with temperature and pressure when operating within heat exchangers, especially in the near-critical or high-temperature regions, where abrupt changes in properties lead to extremely complex flow and heat transfer mechanisms. Accurately predicting the operating status of such heat exchangers under complex conditions is crucial for predictive maintenance, ensuring production safety, and optimizing energy efficiency.

[0003] Currently, performance prediction for unconventional gas heat exchangers mainly relies on the following three types of methods: The first category is numerical simulation methods based on computational fluid dynamics (CFD). These methods obtain detailed information about the flow and temperature fields by solving the Navier-Stokes equations and energy equations. However, CFD methods are computationally extremely expensive, with a single simulation taking hours or even days, which cannot meet the needs of online real-time prediction and rapid response in industrial settings.

[0004] The second category is prediction methods based on traditional empirical correlations. These methods derive empirical formulas for heat transfer coefficients and pressure drops by fitting specific operating conditions and conventional working fluids. Under the strongly nonlinear physical properties of unconventional gases, the assumptions of these correlations are severely violated, leading to a significant decrease in prediction accuracy, with relative errors reaching over 30%, and extremely poor generalization ability across different operating conditions.

[0005] The third category is data-driven neural network prediction methods that have emerged in recent years. Existing technologies mostly employ BP neural networks or RBF neural networks, using parameters such as temperature, flow rate, and pressure as input to directly predict the heat transfer coefficient or outlet temperature. However, this type of method has the following significant drawbacks: (1) Lack of physical constraints: Pure data-driven black box models rely entirely on the statistical regularities of training data and ignore the basic laws of thermodynamics. Under extreme conditions where the training sample coverage is insufficient, the model prediction results often violate the second law of thermodynamics (for example, the predicted heat transfer entropy production is negative), resulting in poor physical interpretability and weak robustness.

[0006] (2) Insensitive to early degradation: The loss function of existing models usually adopts mean squared error (MSE), which treats the prediction error of all samples equally. In the early stage of heat exchanger performance degradation, the degradation characteristics are weak. Traditional MSE loss cannot drive the model to learn these key samples in a concentrated manner, resulting in the model's extremely poor ability to identify mild degradation state and delaying the best maintenance time.

[0007] (3) Lack of prediction confidence criteria: Most existing methods output deterministic single-point prediction values, which cannot quantify the credibility of the prediction results themselves. Operators cannot judge whether a prediction value is "highly credible" or "potentially inaccurate", making it difficult to form engineering-level hierarchical early warning and scientific decision support.

[0008] (4) Optimization algorithms are prone to local optima: Conventional gradient descent is sensitive to the initial weights of the network and is prone to local optima under complex non-convex loss functions, resulting in insufficient generalization ability of the model.

[0009] In summary, there is an urgent need for a heat exchanger condition prediction method that integrates thermodynamic and physical mechanisms, is sensitive to early degradation, can output confidence criteria, and has global optimization capabilities, in order to overcome the shortcomings of existing technologies. Summary of the Invention

[0010] The purpose of this invention is to provide an unconventional gas heat exchanger operating status prediction method based on thermodynamic entropy production, in order to solve the technical problems of poor physical consistency, insensitive early degradation identification, lack of confidence criteria, and weak generalization ability in the prior art.

[0011] To achieve the above objectives, this invention provides a method for predicting the operating status of unconventional gas heat exchangers based on thermodynamic entropy production, comprising the following steps: Step S1: Collect the thermal operating parameters and structural design parameters of the unconventional gas heat exchanger, construct the original sample dataset, and perform data cleaning and normalization preprocessing on the original sample dataset. Step S2: Based on the second law of thermodynamics, construct heat transfer entropy production, viscous entropy production, and multi-scale evolution characteristics to form a set of mechanistic characteristic quantities for characterizing the irreversible loss state evolution of heat exchangers; Step S3: The mechanism feature quantity obtained in step S2 is fused with the parameters after preprocessing in step S1 to construct an enhanced feature matrix containing prior knowledge of thermodynamic physics. Step S4: Construct a cost-sensitive genetically optimized multilayer perceptron prediction model. This model is based on a multilayer perceptron, with its hidden layer using a multi-resolution wavelet activation function. The loss function is a weighted mean square error loss function with an embedded adaptive cost-sensitive factor. A genetic algorithm is used to jointly optimize the initial weights, regularization parameters, and scale parameters of the multilayer perceptron and the multi-resolution wavelet activation function globally. Step S5: Use the enhanced feature matrix as the input of the prediction model, and the expected operating status index of the heat exchanger as the output to predict the operating status of the heat exchanger at future times. Based on the dynamic distribution of the prediction residuals, construct an adaptive probability confidence criterion. Step S6: Based on the adaptive probability confidence criterion and the comprehensive thermodynamic degradation index in the mechanism characteristic quantity, generate a four-level graded early warning criterion that includes normal state, mild degradation state, moderate degradation state and severe degradation state.

[0012] Preferably, the thermal operating parameters in step S1 include at least the inlet temperatures of the cold and hot streams of the heat exchanger. outlet temperature mass flow rate and the pressure drop between the tube side and the shell side. The structural design parameters include at least the inner diameter of the heat exchange tubes. , outer diameter Pipe length Fin spacing and effective heat exchange volume The normalization preprocessing uses the range normalization method to map each parameter to... The interval, the formula is: ; in, These are the original parameter values. and These are the maximum and minimum values ​​of the parameter in the dataset, respectively. This is the normalized value.

[0013] Preferably, in step S2, the heat transfer entropy production is constructed. With this viscous entropy production The mathematical model is as follows: ; ; in, For local heat flux density, This refers to the local mainstream temperature of unconventional gases. For local thermal conductivity, For local dynamic viscosity, Let be the dissipation function, which is approximated for turbulent flow inside the pipe as follows: , For flow rate, The coordinates are radial; this multi-scale evolution characteristic includes at least the first-order difference between the heat transfer entropy production and the viscous entropy production. Second-order difference and variance within the sliding window This forms a set of 8-dimensional mechanistic feature quantities: ; ; Preferably, the method for constructing the enhanced feature matrix in step S3 is as follows: [The method involves] using the mechanism feature quantity... The normalized thermal operating parameters and structural design parameters are then horizontally combined to form an enhanced feature matrix with expanded dimensions. ,in This is the normalized original parameter vector.

[0014] Preferably, in step S4, the multi-resolution wavelet activation function adopts the Morlet wavelet basis function, the mathematical expression of which is: ; in, For input signal, As a scale factor, The translation factor is used; the multilayer perceptron contains an input layer, at least two hidden layers and an output layer. The output of each hidden layer is mapped by the multi-resolution wavelet activation function, and a batch normalization layer is added after each hidden layer.

[0015] Preferably, in step S4, the adaptive cost-sensitivity factor The dynamic adjustment mechanism is as follows: ; in, For the first The true label value of each sample and These are the mean and variance of the labels for the normal state samples, respectively. This is a preset sensitivity intensity coefficient, with a value range of [value range missing]. The weighted mean square error loss function The expression is: ; in, This refers to the batch sample size. These are the model's predicted values.

[0016] Preferably, the joint global optimization process of the genetic algorithm in step S4 includes: using real number encoding to encode the initial weight matrix of the multilayer perceptron. L2 regularization parameters and the scaling factor of the wavelet activation functions of each hidden layer. Encoded as a chromosome; the root mean square error of the prediction model on the independent validation set. The reciprocal of the value is used as the fitness function; tournament selection, simulated binary crossover operator (SBX) and polynomial mutation operator are used to generate offspring population; the evolution is repeated until the maximum number of iterations or the fitness function value converges, and the optimal chromosome decoding result is output and assigned to the multilayer perceptron.

[0017] Preferably, the method for constructing the adaptive probability confidence criterion in step S5 includes: based on the predicted residual sequence within the recent time window. The mean of the residuals is dynamically estimated using the exponentially weighted moving average method. and standard deviation Assuming the residuals follow a normal distribution, then for the new predicted value... Its confidence interval is Confidence level Defined as: ; in, The cumulative distribution function of the standard normal distribution. This is a dynamically adjusted confidence coefficient. Similarity to the enhanced feature matrix There is a negative correlation; as the deviation between the distribution of the input sample and the training sample increases, As the value increases, the confidence interval widens.

[0018] Preferably, the comprehensive thermodynamic degradation index in step S6 The definition of is: ; The specific rules for this four-level early warning criterion are as follows: when When this occurs, it is considered a normal state; when or At that time, it was determined to be in a state of mild degradation, and a yellow alert was issued; when and At that time, the system was judged to be in a moderately deteriorated state, and an orange alert was issued. when At that time, regardless of confidence level The value indicates a severely deteriorated state, triggering a red alert. in, , , The threshold is defined based on historical data and satisfies... .

[0019] Preferably, the method further includes a model adaptive update step S7: when the newly acquired real-time data is determined to have entered a stable operating condition, the cost-sensitive genetically optimized multilayer perceptron prediction model is incrementally fine-tuned using the new data, the elastic weight consolidation method is used to prevent catastrophic forgetting, and the model parameters are updated to adapt to the long-term performance degradation of the heat exchanger.

[0020] Therefore, the unconventional gas heat exchanger operation state prediction method based on thermodynamic entropy production using the above structure, as described in this invention, has the following beneficial effects: (1) This invention embeds the entropy production theory of the second law of thermodynamics into a neural network prediction model through feature-level fusion. The constructed heat transfer entropy production, viscous entropy production and their multi-scale evolution characteristics are used as inputs, providing the model with prior knowledge about the irreversible loss of the system from a mechanistic perspective. This enables the model to output prediction results that conform to thermodynamic laws even under extreme conditions with few training samples, significantly improving generalization ability and physical interpretability.

[0021] (2) This invention proposes an adaptive cost-sensitive loss function, which dynamically adjusts the weights according to the degree to which the sample deviates from the normal state, forcing the model to focus on learning the subtle signs of early degradation during the training process. Compared with the traditional mean squared error loss, this method can significantly improve the recognition rate of mild degradation state, thus gaining valuable advance time for predictive maintenance.

[0022] (3) Based on the exponentially weighted moving average statistics of the prediction residuals, this invention constructs a dynamically adaptive probability confidence criterion. This criterion not only quantifies the credibility of the prediction results, but also automatically adjusts the confidence interval width according to the similarity between the input sample and the training distribution, realizing the leap from "single-point prediction" to "credible interval prediction", and providing a scientific basis for hierarchical early warning.

[0023] (4) This invention uses a genetic algorithm to jointly optimize the initial weights, regularization parameters, and scale factor of the wavelet activation function of the multilayer perceptron, effectively avoiding the problem of traditional BP algorithm getting trapped in local optima. At the same time, the introduction of multi-resolution wavelet activation function enables the model to capture both the long-term trend and local mutations in the heat exchanger performance evolution, and the prediction accuracy is significantly improved compared with traditional neural networks.

[0024] (5) This invention combines the probability confidence criterion with the physical index of entropy production ratio to form a logic decision tree for redundancy verification. By setting four levels—normal, mild, moderate, and severe—and providing clear warning signals for different levels, operators can directly arrange maintenance plans based on the warning results, which has extremely high engineering practical value.

[0025] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0026] Figure 1 This is an overall flowchart of the present invention; Figure 2 A schematic diagram of the dynamic adjustment curve of the adaptive cost-sensitive factor; Figure 3 The comprehensive thermodynamic degradation index in Example 1 Evolution curves and schematic diagram of the four-level early warning results; Figure 4 This is a schematic diagram showing the comparison of predicted syngas heat exchanger outlet temperature and the adaptive confidence interval in Example 2. Detailed Implementation

[0027] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0028] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0029] Example 1 This embodiment uses a printed circuit board heat exchanger (PCHE) applied in a supercritical CO2 Brayton cycle as an example, as shown below: Step S1, Data Acquisition and Preprocessing: like Figure 1As shown, the first step of this invention is to collect and preprocess data. Specifically, in this embodiment, the operating data and structural parameters of the heat exchanger are collected over 40 days of continuous operation. The main structural parameters of the heat exchanger are: hydraulic diameter of the flow channel 2.0 mm, core length 0.5 m, and effective heat exchange volume 0.02 m³. 3 The heat exchanger has 120 plates. Thermal operating parameters include: cold stream (CO2) inlet temperature 32-45℃, outlet temperature 65-85℃, mass flow rate 0.5-2.0 kg / s; hot stream (another CO2 stream) inlet temperature 120-150℃, outlet temperature 70-95℃, mass flow rate 0.5-2.0 kg / s; tube-side pressure drop and shell-side pressure drop. These parameters are first entered... Figure 1 The "range standardization and normalization processing" module uses the formula Map all parameters to the [0,1] interval to form the original normalized feature vector. The purpose of this step is to eliminate the influence of different units of measurement on subsequent model training, while preserving the relative change information of the original data.

[0030] Step S2: Construct mechanistic feature quantities: This step introduces the second law of thermodynamics into the feature construction process. Specifically, it calculates the heat transfer entropy production at each sampling time. and viscous entropy production Because the physical properties of CO2 change drastically near the critical region, its thermal conductivity... Dynamic viscosity All data are dynamically obtained from the NIST REFPROP database based on local temperature and pressure. The formula for calculating heat transfer entropy production is as follows: The formula for calculating viscous entropy production is: Furthermore, to characterize the trend of entropy production over time, this invention also calculates multi-scale evolutionary features: first-order difference. Second-order difference and variance within the sliding window Thus forming 8-dimensional mechanistic characteristic quantities These characteristic quantities are not simply statistical data, but physical information extracted from the thermodynamic irreversible loss mechanism, providing prior knowledge about the essential operating state of the heat exchanger for subsequent models.

[0031] Step S3: Construct the enhanced feature matrix: This step will normalize the original parameters obtained in step S1. (This embodiment has a total of 11 parameters) and the 8-dimensional mechanism feature quantity obtained in step S2. Horizontal concatenation is performed to form a 19-dimensional enhanced feature matrix. Each row of this matrix corresponds to a sampling time point, and each column represents a feature dimension. This concatenation operation achieves the fusion of physical mechanisms and original data at the feature level, enabling the subsequent model to receive thermodynamic constraints at the input end. This is a key step in achieving high generalization ability and physical consistency in this invention.

[0032] Step S4: Construct a cost-sensitive genetically optimized multilayer perceptron prediction model: This step constructs the core prediction model, which is based on a multilayer perceptron and includes an input layer (19 nodes), two hidden layers (32 nodes and 16 nodes respectively), and an output layer (3 nodes, which predict the overall heat transfer coefficient). outlet temperature and pressure drop The hidden layer uses the Morlet wavelet activation function, whose expression is: , where the scale factor Translation factor The network adapts and adjusts itself during training, thereby giving it multi-resolution analysis capabilities, enabling it to simultaneously capture both long-term trends and local mutations in heat exchanger performance evolution.

[0033] In model training, this invention introduces a dual optimization mechanism: First, global optimization using genetic algorithms (GA). For example... Figure 1 As shown in the GA optimization loop, this invention optimizes the initial weight matrix of the network. L2 regularization parameters and wavelet scaling factors of each hidden layer A chromosome is encoded using real numbers, and the reciprocal of the validation set RMSE is used as the fitness function. Through tournament selection, simulated binary crossover, and polynomial mutation, the network iterates and evolves, ultimately decoding the optimal chromosome and assigning it to the network. This global optimization process effectively avoids the problems of traditional gradient descent methods, such as sensitivity to initial values ​​and susceptibility to local optima.

[0034] Second, an adaptive cost-sensitive loss function. For example... Figure 2 As shown, the cost-sensitive factor designed in this invention The rate of change increases exponentially with the degree to which the sample deviates from the normal state. Figure 2 The horizontal axis represents the degree to which the sample deviates from the normal state. The vertical axis represents the weights. .from Figure 2 It can be clearly seen that when the sample is in the normal range (deviation close to 0), the weight is approximately 1; as the deviation increases (entering the early deterioration stage), the weight rapidly rises to 5-6. The corresponding mathematical formula for this curve is: ,in The weighted loss function is: This loss function forces the network to pay more attention to early degraded samples during backpropagation, thereby significantly improving its sensitivity to subtle signs of failure.

[0035] Step S5, Adaptive probability confidence criterion: While outputting predicted values, the model also executes a confidence assessment module in parallel. This step uses an exponentially weighted moving average method (with a decay factor of 0.9) to dynamically estimate the mean of recent predicted residuals. and standard deviation For the new predicted value Its confidence interval is Confidence level ,in The standard normal cumulative distribution function is... The system dynamically adjusts based on the similarity between the input sample and the training distribution.

[0036] Figure 4 The actual effect of this confidence criterion is visually demonstrated through specific data. Figure 4 In this study, the predicted value curve of this invention closely matches the measured value curve, while the predicted value of traditional BP shows significant deviation in the component abrupt change region. Meanwhile, Figure 4 The shaded area represents the first half of the time series (stable operating conditions), where the 95% confidence interval is relatively narrow (approximately ±2.2℃), indicating high prediction reliability. After the 6th sampling point (the region of abrupt component change), the confidence interval automatically widens to approximately ±3.8℃, corresponding to a higher confidence level. The value decreased from 0.93 to 0.71. This change can be clearly observed in the graph, allowing operators to intuitively judge the reliability of the prediction results, thus achieving an upgrade from "single-point prediction" to "reliable prediction".

[0037] Step S6, Level 4 Early Warning: like Figure 1 As shown in the four-level early warning module of "Normal / Yellow / Orange / Red", this invention will determine the confidence level. With comprehensive thermodynamic degradation index Combined, these elements form a four-level early warning mechanism for redundancy checks. Figure 3 This demonstrates the operation of this embodiment over 40 days. The evolution curve and early warning results are shown in the figure. The three horizontal dashed lines in the figure represent three preset thresholds: (Normal upper limit) (Boundary between mild and moderate) (The dividing line between moderate and severe).

[0038] from Figure 3 It can be clearly seen that: Day 28 The curve dropped to 2.06 for the first time, below... And confidence level The value dropped to 0.89, triggering a yellow alert. Figure 3 The yellow marker indicates "slight degradation, attention recommended"; Day 33 It further decreased to 1.62, below And confidence level Triggering an orange alert ( Figure 3 The orange marker indicates "moderate degradation; inspection recommended." Day 38 It dropped to 0.97, below Triggering a red alert ( Figure 3 The red marker indicates "severe degradation, requiring emergency shutdown and inspection".

[0039] Figure 3 The early warning timing sequence in the data matches the comparison data in Table 1 perfectly, proving that the present invention can provide maintenance personnel with clear, tiered early warning signals with reliability verification.

[0040] Step S7: Model adaptive update (see...) Figure 1 Feedback loop "S7 (optional): Incremental model update" like Figure 1 As shown by the dashed feedback arrow pointing from S7 to S4, once the newly acquired real-time data is determined to have entered a stable operating condition, the model enters the incremental update phase. This phase employs the Elastic Weight Consolidation (EWC) method, applying additional secondary penalty terms to important parameters in the network to prevent the model from forgetting previously acquired knowledge when learning new data. This allows the model to adapt online to the long-term performance degradation and operating condition drift of the heat exchanger, maintaining good predictive performance.

[0041] Predictive performance comparison and verification: Table 1 provides a comprehensive comparison of the prediction accuracy and early warning timeliness of the method of this invention with two comparative methods (traditional BP neural network and cost-insensitive genetic optimization MLP). Experimental data show that the overall heat transfer coefficient of the method of this invention is... The predicted RMSE is 0.071, and the outlet temperature is... The predicted MAE is 1.23℃, and the voltage drop is... The predicted RMSE was 1.08 kPa, significantly better than the comparative method. Regarding early warning timeliness, this invention issued a yellow warning on day 28, 6 days earlier than the traditional BP (day 34).

[0042] Table 1

[0043] Figure 4The visualization confirms the above data: in the component mutation region, the predicted value of this invention (orange curve) closely follows the measured value (blue curve), while the traditional BP predicted value (gray curve) deviates significantly. In addition, the confidence interval (orange shaded band) output by this invention automatically widens in the mutation region, providing additional reliability information for engineering decision-making, which is completely lacking in traditional methods.

[0044] Example 2 This embodiment uses a shell-and-tube heat exchanger for cooling syngas from coal gasification as an example to verify the effectiveness of the invention under mixed gas conditions. The main components of the syngas are H2 (25-35%), CO (35-45%), CO2 (10-20%), H2O (5-15%), and a small amount of N2. The heat exchanger's structural parameters are: heat exchange tube inner diameter 25 mm, tube length 6.0 m, number of tube bundles 320, shell inner diameter 800 mm, and baffle spacing 400 mm.

[0045] Strictly follow during implementation Figure 1 The entire process is shown. In step S2, the local thermal conductivity and viscosity of the mixed gas are calculated using Wilke's mixing rule. In step S4, considering the frequent fluctuations in the composition of the syngas, refer to... Figure 2 The curve shape was analyzed, and the sensitivity coefficient α was set to 6 to enhance the model's response sensitivity to sudden changes in component composition. Under the sudden drop in CO content from 40% to 25%, the method of this invention achieved an outlet temperature prediction MAE of 4.2℃ and a pressure drop prediction RMSE of 1.3 kPa, while the traditional BP neural network achieved an MAE of 12.6℃ and an RMSE of 3.2 kPa. Figure 4 As shown, the predicted values ​​of this invention are in high agreement with the measured values ​​in the mutation region, and the confidence interval automatically widens (the shaded area expands). The confidence level drops from 0.93 to 0.71, and the system simultaneously issues a yellow warning (see reference). Figure 3 The system employs an early warning logic to alert operators to the uncertainty of the forecast results, achieving a dual output of accurate prediction and reliability measurement.

[0046] Example 3 This embodiment uses a finned tube flue gas heat exchanger applied to waste heat recovery in the steel industry as an example to verify the dust accumulation early warning capability of the present invention under dusty and high-temperature conditions. The flue gas temperature reaches 650-800℃, and the dust concentration is 10-50 g / Nm³. 3 The main parameters of the heat exchanger are: finned tube outer diameter 50 mm, fin spacing 8 mm, and tube bundle length 4.5 m.

[0047] according to Figure 1 The complete process shown demonstrates the execution of this method. In step S4, the sensitivity intensity coefficient of the adaptive cost sensitivity factor is... Set to 8 (i.e.) Figure 2 The curve is steeper to enhance the identification of weak signals of heat transfer attenuation in the early stage of ash accumulation. In a 60-day continuous ash accumulation experiment, this method issued a yellow warning when the heat transfer coefficient decreased by only 4.8% (on the 33rd day of operation) (the warning triggering process is similar to...). Figure 3 The yellow warning point logic is consistent with the traditional temperature difference method, which only issues a warning when the heat transfer coefficient decreases by 12.5% ​​(on day 42), while the BP neural network only issues a warning when the coefficient decreases by 15.3% (on day 45). This invention provides an approximately 9-day advance window for scheduling online dust removal on-site, effectively avoiding unplanned downtime and fully verifying the effectiveness of the method. Figure 1 The complete technical solution shown demonstrates its engineering applicability in harsh industrial environments.

[0048] The above three embodiments are based on different unconventional gases (supercritical CO2, syngas, high-temperature flue gas) and different heat exchanger types (PCHE, shell and tube, finned tube), and strictly follow the principles of... Figure 1 The complete process shown, and make full use of Figures 2 to 4 The cost-sensitivity curve, early warning sequence, and confidence interval visualization method shown systematically illustrate all the technical details of this invention, from data acquisition, mechanism feature extraction, enhancement matrix construction, genetically optimized wavelet MLP training, confidence assessment to hierarchical early warning and model updating. Those skilled in the art, based on the guidance of the above embodiments, can extend the method of this invention to other types of heat exchangers and working fluids without any inventive effort.

[0049] Example 4 This embodiment uses a helium water shell-and-tube heat exchanger applied in the main helium blower system of a high-temperature gas-cooled reactor as an example to verify the effectiveness of the present invention under high-temperature and high-pressure inert gas conditions.

[0050] Step S1, Data Acquisition and Preprocessing: The heat exchanger has the following structural parameters: heat exchange tube inner diameter 20 mm, wall thickness 2.5 mm, tube length 4.0 m, number of tubes 280, shell inner diameter 600 mm, baffle spacing 350 mm, and effective heat exchange volume 0.38 m³. 3 The hot stream is high-temperature, high-pressure helium, with an inlet temperature of 250-350℃, an outlet temperature of 160-220℃, a mass flow rate of 3.0-8.0 kg / s, and a pressure of 7.0-9.0 MPa; the cold stream is water, with an inlet temperature of 30-50℃, an outlet temperature of 90-130℃, and a mass flow rate of 5.0-12.0 kg / s. 60 consecutive days of operational data were collected and pre-processed using the aforementioned method.

[0051] Step S2: Construct mechanistic feature quantities: Helium is a typical unconventional gas, and its thermal conductivity and viscosity increase significantly with increasing temperature (unlike conventional gases, whose thermal conductivity increases at varying rates with temperature). These physical properties are dynamically obtained based on local temperature and pressure by accessing the Helium-specific property database. Figure 1 The process shown involves integrating the local entropy production rate piecewise along the pipe length to obtain the global entropy production characteristics, and then calculating the multi-scale evolution characteristics.

[0052] Steps S3-S5: Model Prediction and Confidence Output: according to Figure 1 The complete process of this method was executed. During 60 days of continuous monitoring, the model's predicted MAE for helium outlet temperature was 1.8℃, the predicted RMSE for pressure drop was 2.1 kPa, and the predicted RMSE for overall heat transfer coefficient was 0.065. The confidence level of the model output remained between 0.92 and 0.95 under steady-state conditions, verifying the good generalization ability of this invention in different unconventional gas working fluids.

[0053] Steps S6-S7: Early Warning Verification and Model Update On day 35, the model issued a yellow alert, indicating "mild performance degradation, monitoring recommended." On day 42, the alert was upgraded to orange. On day 48, a planned shutdown for maintenance revealed slight scaling on the inner wall of the tube bundle (average fouling thermal resistance approximately 0.0003m). 2 The heat transfer coefficient decreased by approximately 8% (·K / W) compared to the initial value, which is highly consistent with the model's early warning timeline. After incrementally updating the model using the EWC method in step S7, the subsequent prediction accuracy remained stable, further verifying the engineering applicability of the present invention.

[0054] This embodiment fully demonstrates that the present invention is not only applicable to supercritical CO2, syngas, and high-temperature flue gas, but also to inert unconventional gas working fluids such as high-temperature and high-pressure helium, covering the real application scenarios of the main helium blower helium-water heat exchanger.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. 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 still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for predicting the operating status of unconventional gas heat exchangers based on thermodynamic entropy production, characterized in that, Includes the following steps: Step S1: Collect the thermal operating parameters and structural design parameters of the unconventional gas heat exchanger, construct the original sample dataset, and perform normalization preprocessing on the original sample dataset; Step S2: Based on the second law of thermodynamics, construct heat transfer entropy production and viscous entropy production to form mechanistic characteristic quantities for characterizing the irreversible loss state evolution of the heat exchanger; Step S3: The mechanism feature quantity obtained in step S2 is fused with the parameters after preprocessing in step S1 to construct an enhanced feature matrix containing prior knowledge of thermodynamic physics. Step S4: Construct a cost-sensitive genetically optimized multilayer perceptron prediction model, in which the hidden layer adopts the wavelet activation function, the loss function is the weighted mean square error loss function with embedded adaptive cost-sensitive factors, and the genetic algorithm is used to jointly optimize the initial weights, regularization parameters and scale parameters of the wavelet activation function of the multilayer perceptron. Step S5: Use the enhanced feature matrix as input to the prediction model to predict the operating state of the heat exchanger at future times, and construct an adaptive probability confidence criterion based on the dynamic distribution of the prediction residuals. Step S6: Generate a four-level graded early warning criterion based on the adaptive probability confidence criterion and the comprehensive thermodynamic degradation index in the mechanism characteristic quantity.

2. The method for predicting the operating status of an unconventional gas heat exchanger based on thermodynamic entropy production according to claim 1, characterized in that, The thermal operating parameters in step S1 include at least the inlet temperatures of the cold and hot streams of the heat exchanger. outlet temperature mass flow rate and the pressure drop between the tube side and the shell side. The structural design parameters include at least the inner diameter of the heat exchange tubes. , outer diameter Pipe length Fin spacing and effective heat exchange volume The normalization preprocessing uses the range normalization method to map each parameter to... The interval, the formula is: ； in, These are the original parameter values. and These are the maximum and minimum values ​​of the parameter in the dataset, respectively. This is the normalized value.

3. The method for predicting the operating status of an unconventional gas heat exchanger based on thermodynamic entropy production according to claim 1, characterized in that, In step S2, the mechanism feature quantity also includes multi-scale evolution features. The multi-scale evolution features include at least the first-order difference and second-order difference between the heat transfer entropy production and the viscous entropy production, as well as the variance within the sliding window, forming a set of 8-dimensional mechanism feature quantities.

4. The method for predicting the operating status of an unconventional gas heat exchanger based on thermodynamic entropy production according to claim 1, characterized in that, In step S2, the heat transfer entropy production is constructed. The mathematical model is as follows: ； This viscous entropy production The mathematical model is as follows: ； in, For local heat flux density, This refers to the local mainstream temperature of unconventional gases. For local thermal conductivity, For local dynamic viscosity, Let be the dissipation function. For effective heat exchange volume.

5. The method for predicting the operating status of an unconventional gas heat exchanger based on thermodynamic entropy production according to claim 1, characterized in that, The method for constructing the enhanced feature matrix in step S3 is as follows: the mechanism feature quantity is horizontally spliced ​​with the normalized thermal operating parameters and structural design parameters to form an enhanced feature matrix with expanded dimensions.

6. The method for predicting the operating status of an unconventional gas heat exchanger based on thermodynamic entropy production according to claim 1, characterized in that, In step S4, the wavelet activation function adopts the Morlet wavelet basis function; the multilayer perceptron includes an input layer, at least two hidden layers and an output layer, the output of each hidden layer is mapped by the wavelet activation function, and a batch normalization layer is added after each hidden layer.

7. The method for predicting the operating status of an unconventional gas heat exchanger based on thermodynamic entropy production according to claim 1, characterized in that, The adaptive cost sensitivity factor in step S4 The dynamic adjustment mechanism is as follows: ； in, For the first The true label value of each sample and These are the mean and variance of the labels for the normal state samples, respectively. This is a preset sensitivity intensity coefficient; the weighted mean square error loss function The expression is: ； in, This refers to the batch sample size. These are the model's predicted values.

8. The method for predicting the operating status of an unconventional gas heat exchanger based on thermodynamic entropy production according to claim 1, characterized in that, Step S4, the joint global optimization process of the genetic algorithm, includes: encoding the initial weight matrix, L2 regularization parameters, and scale factors of the wavelet activation functions of each hidden layer of the multilayer perceptron into a chromosome using real-number encoding; using the reciprocal of the root mean square error of the prediction model on the independent validation set as the fitness function; generating offspring population using tournament selection, simulated binary crossover operator, and polynomial mutation operator; repeating the evolution until the maximum number of iterations or the fitness function value converges, outputting the optimal chromosome decoding result, and assigning it to the multilayer perceptron.

9. The method for predicting the operating status of an unconventional gas heat exchanger based on thermodynamic entropy production according to claim 1, characterized in that, The method for constructing the adaptive probability confidence criterion in step S5 includes: dynamically estimating the mean of the residuals based on the predicted residual sequence within a recent time window using the exponentially weighted moving average method. and standard deviation Assuming the residuals follow a normal distribution, then for the new predicted value... Its confidence interval is Confidence level Defined as: ； in, The cumulative distribution function of the standard normal distribution. The confidence coefficient is dynamically adjusted.

10. The method for predicting the operating status of an unconventional gas heat exchanger based on thermodynamic entropy production according to claim 1, characterized in that, In step S6, the comprehensive thermodynamic degradation index The definition of is: ； The specific rules for this four-level early warning criterion are as follows: when and When this occurs, it is considered a normal state; when or At that time, it was determined to be in a state of mild degradation, and a yellow alert was issued; when and At that time, the system was determined to be in a moderately deteriorated state, and an orange alert was issued. when At that time, it was determined to be in a severely deteriorated state, and a red alert was issued; in, The threshold is determined based on historical data.

Images