Cooler blockage prediction method based on deep learning coupled with physical constraints

Abstract

The present invention relates to the technical field of hydropower equipment state prediction. The present invention discloses a cooler blockage prediction method based on deep learning coupled with physical constraints, comprising the following steps: collecting the temperature, pressure difference, and flow parameters of the cooler during operation to generate time series data; preprocessing the time series data to generate time series sample vectors; constructing a deep learning hybrid model to extract local features and capture global time series dependencies through the deep learning hybrid model; combining the Navier-Stokes equations and cross entropy to construct a comprehensive loss function; inputting the processed data into the deep learning hybrid model to output a cooler blockage state classification result. The present invention can achieve early and accurate identification of the cooler blockage state, solving the problem of delayed response of traditional methods. It can issue an early warning of blockage, identify the blockage type in advance, and avoid unplanned downtime.

Description

Technical Field

[0001] The present invention relates to the technical field of hydropower equipment state prediction, and in particular to a cooler blockage prediction method based on deep learning coupled with physical constraints. Background Art

[0002] During turbine operation, the bearing cooler pipes are prone to clogging due to the accumulation of silt, scale, or foreign matter. Once clogged, the cooling water flow rate is significantly reduced, making it difficult to fully dissipate heat, causing the bearing temperature to rise sharply. Sustained high temperatures not only accelerate bearing wear but, in severe cases, can even cause bearing burnout, leading to abnormal vibration and, in severe cases, shutdown. Therefore, predicting cooler blockage in advance is of great significance. It can, to a certain extent, ensure equipment safety, improve production efficiency, facilitate refined equipment management, and support long-term stable operation.

[0003] The current industry generally adopts an alarm mechanism based on a single temperature threshold or manual regular inspections, which have the following technical pain points:

[0004] 1) Response hysteresis: Traditional methods rely on a single temperature threshold alarm, but temperature anomalies usually occur in the late stages of congestion, making early warning impossible.

[0005] 2) High false alarm rate: Single temperature parameter is easily affected by environmental interference and lacks multi-physics field collaborative analysis;

[0006] 3) Complex characteristic coupling: There is a nonlinear dynamic coupling relationship between pressure, flow and temperature parameters, which makes it difficult for traditional mathematical models to accurately model;

[0007] 4) High maintenance costs: Preventive maintenance relies on manual experience and lacks data-driven decision support. Sudden severe blockages can lead to unplanned downtime, resulting in high single-time repair costs and associated power generation losses. Summary of the Invention

[0008] In order to solve the current technical problems, the main purpose of the present invention is to provide a cooler blockage prediction method based on deep learning coupled with physical constraints, realize the whole process operation from data collection, processing to state prediction, reduce manual intervention, identify the blockage type in advance, avoid unplanned shutdowns, and reduce the false alarm rate.

[0009] To overcome the problems of the prior art, the present invention adopts a technical solution: a cooler blockage prediction method based on deep learning coupled with physical constraints, comprising the following steps:

[0010] S1. Collect the temperature, pressure difference and flow parameters of the cooler during operation and generate time series data;

[0011] S2. Preprocess the time series data to generate a time series sample vector;

[0012] S3. Build a deep learning hybrid model to extract local features and capture global temporal dependencies.

[0013] S4. Combining Navier-Stokes equation and cross entropy to construct a comprehensive loss function;

[0014] S5. Input the processed data into a deep learning hybrid model and output a cooler blockage status classification result; wherein the cooler blockage status classification includes normal, mild, moderate, and severe blockage.

[0015] S1 specifically includes installing a temperature sensor, a pressure sensor, and an electromagnetic flowmeter on the cooler pipeline;

[0016] Based on historical maintenance records, the cooler status is marked as multiple working states, including normal state, slight blockage, moderate blockage and severe blockage;

[0017] The temperature, pressure difference and flow rate data of the cooler are obtained during the entire cycle of the unit's cold start, steady-state operation, variable load and shutdown stages, and time series data of the cooler under different working conditions are generated.

[0018] In S2, the methods for preprocessing time series data include outlier processing, sliding window expansion, and data normalization.

[0019] The method for handling outliers is as follows:

[0020] The z-score algorithm is used to detect and eliminate transient interference points that exceed the threshold range. For the time series data of each parameter, the mean and standard deviation within the window length are calculated to eliminate abnormal points.

[0021] Then, the window slides one time unit, and the calculation and outlier detection are continued for the data in the new window until the entire sequence is traversed;

[0022] After the temperature, pressure, and flow data are processed by the z-score algorithm, the abnormal sensor data is identified through pressure-flow cross validation; for the outliers removed, the linear interpolation method is used to replace them, assuming that the outliers are located at x i , the adjacent data points are x i-1 、 x i+1 , the estimated value after replacement is:

[0023] .

[0024] The method for sliding window expansion is as follows:

[0025] For congestion categories with small sample sizes, sampling and sliding windows are used to balance the distribution of the data set. For cooler operation data in time series form, the sliding window size is determined. If the data is collected at intervals of one minute, the window size is set to m, that is, each window contains data from m consecutive time points. Starting from the starting position of the data, the window is slid according to the set step size, and each sliding generates a new window of data.

[0026] The method of data standardization is as follows:

[0027] After denoising the temperature, pressure, and flow data, these data are standardized to eliminate the differences caused by different physical dimensions and improve the comparability of the data and the effectiveness of the model processing. The calculation formula is:

[0028] ;

[0029] Where, X is the denoised sequence data, X scaled is the standardized data, median( X ) is the median of the sequence, IQR( X ) is the interquartile range.

[0030] In S3, based on a deep learning hybrid model, the encoder part is used for feature extraction in the converter module. An output branch for pressure prediction is added to the output layer. After the converter module completes feature extraction of the data, it is connected to two fully connected layers. One of the fully connected layers is used to perform classification tasks to determine the predicted category of cooler blockage, and the other fully connected layer outputs the pressure prediction value, which is involved in the calculation of physical loss.

[0031] In S4, the Navier-Stokes equation and cross entropy are combined to construct a comprehensive loss function, which includes the following steps:

[0032] S4.1. Introduce the Navier-Stokes equation constraint into the loss function to ensure that the relationship between the intermediate quantity "flow-pressure" predicted by the deep learning hybrid model conforms to the laws of fluid mechanics. Use the mean square error to measure the difference between the predicted pressure and the theoretical pressure to construct the first loss term.

[0033] S4.2. Use the mean square error to calculate the loss between the predicted pressure and the actual measured pressure to construct a second loss term;

[0034] S4.3. Combine the above two loss terms to form the final physical loss function. On this basis, combine the cross-entropy loss function in the deep learning classification task to construct a total loss function with physical constraints. The total loss function is the weighted sum of the cross-entropy loss function and the physical loss function.

[0035] In S4.1, the simplified Navier-Stokes equations are:

[0036] ;

[0037] Assuming the pipe cross-sectional area is uniform, the flow rate Q =v·A , A is the cross-sectional area, and further discretization yields the “flow-pressure” relationship constraint:

[0038] ;

[0039] Where: μ is the dynamic viscosity; L is the pipe length; A is the cross-sectional area; Q is the flow rate; t is the time variable; ρ is the density of the fluid; v is the velocity field; For theoretical pressure; p is the pressure of the fluid;

[0040] The mean square error is used to measure the predicted pressure P pred With the theoretical pressure P theo The difference between them constructs the first loss term :

[0041] ;

[0042] Where: N is the number of sample points, i represents the i-th sample point;

[0043] In S4.2, the mean square error is used to calculate the predicted pressure P pred The actual measured pressure P The loss between them constructs the second loss term :

[0044] ;

[0045] Where: N is the number of sample points, i represents the i-th sample point, P i is the i-th measured pressure value;

[0046] In S4.3, the above two loss terms are combined to form the final physical loss function:

[0047] ;

[0048] The total loss function is the weighted sum of cross entropy loss and physical loss:

[0049] ;

[0050] Where, L total is the total loss function, L physics is the physical loss function, L CE is the cross entropy loss function, α and β is the weight.

[0051] In S5, the training and prediction of the deep learning hybrid model includes the following steps:

[0052] The training set, validation set, and test set are divided into a ratio of 7:2:1. An adaptive moment estimation weight decay optimizer is used in conjunction with cosine annealing learning rate scheduling. During the model training phase, parameter optimization is performed by minimizing the comprehensive loss function. After model training is completed, the test set is used for the final evaluation of the model prediction performance, and the prediction results for the cooler blockage category are output.

[0053] The present invention has the following beneficial effects:

[0054] 1. The present invention collects temperature, pressure, and flow time series data by deploying a multi-source sensor network, and constructs a deep learning hybrid model to realize multi-scale feature extraction. The fluid mechanics equation is integrated with the cross entropy loss function to form a comprehensive loss function under physical constraints, so that the model can optimize the classification accuracy and compliance with physical laws at the same time during the training process, and achieve a balance between data-driven and knowledge-guided. Through outlier filtering, sliding window data enhancement and robust normalization preprocessing, combined with the collaborative feature extraction of bidirectional time series convolution and attention mechanism, a four-level classification of blockage degree is realized, namely normal, mild, moderate and severe blockage. This method can identify the early and accurate identification of the cooler blockage status in advance, and provide data-driven decision support for the intelligent operation and maintenance of the turbine unit.

[0055] 2. This paper combines the Navier-Stokes equations and cross entropy to construct a comprehensive loss function. This integrates data-driven and physical prior knowledge to enhance model interpretability and reduce the risk of overfitting. It also constrains model output to conform to the laws of fluid mechanics, avoiding misjudgments that would be driven purely by data. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0057] Figure 1 Flowchart of the present invention.

[0058] Figure 2 Schematic diagram of the sliding window process of small sample data in the present invention.

[0059] Figure 3 This is a structural diagram of the deep learning model of the present invention.

[0060] Figure 4 Schematic diagram of the bidirectional temporal convolutional network module of the present invention.

[0061] Figure 5 Schematic diagram of the converter module of the present invention.

[0062] Figure 6 A flow chart is constructed for the comprehensive loss function of the present invention. DETAILED DESCRIPTION

[0063] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0064] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0065] Secondly, the "present embodiment" or "embodiment" referred to herein refers to a specific feature, structure or characteristic that may be included in at least one implementation of the present invention.

[0066] Furthermore, the present invention is described in detail with reference to schematic diagrams. For ease of illustration, when describing the embodiments of the present invention, cross-sectional views illustrating device structures may be partially enlarged and not to scale. Furthermore, the schematic diagrams are merely illustrative and should not limit the scope of protection of the present invention. Furthermore, in actual production, the three-dimensional dimensions of length, width, and depth should be included.

[0067] Example 1:

[0068] See also Figure 1This embodiment provides a cooler blockage prediction method based on deep learning coupled with physical constraints, including the following steps:

[0069] S1. Collect the temperature, pressure difference and flow parameters of the cooler during operation and generate time series data.

[0070] In S1, temperature sensors, pressure sensors and electromagnetic flow meters are installed on the cooler pipeline.

[0071] Based on historical maintenance records, the cooler status is marked as a variety of working states, including normal, slight, moderate, and severe blockage.

[0072] Obtain the temperature, pressure difference and flow data of the cooler during the entire cycle of the unit's cold start, steady-state operation, variable load and shutdown stages to ensure that model training covers the entire operating field.

[0073] Temperature, pressure, and flow parameters are collected synchronously to form a time-series data stream. Each complete sample, corresponding to each congestion level, contains three-dimensional time points, while strictly ensuring the time alignment of temperature, pressure, and flow data.

[0074] S2. Preprocess the time series data to generate a time series sample vector.

[0075] In S2, the steps of preprocessing time series data include outlier processing, sliding window expansion, and data normalization.

[0076] Specifically, the method for handling outliers is as follows:

[0077] See also Figure 2 , using the z-score algorithm (sliding window Z-score algorithm), setting the window length to 60s, detecting and eliminating transient interference points that exceed the range of ±3σ, where σ is the sliding standard deviation. x t , calculate the window length W = mean μ within 60 w and standard deviation σ W , remove the outliers that satisfy the following formula:

[0078] ;

[0079] Then, the window slides one time unit and continues to calculate and detect outliers on the data in the new window until the entire sequence is traversed.

[0080] After processing the temperature, pressure, and flow data using the z-score algorithm (sliding window Z-score algorithm), abnormal sensor data is further identified through pressure-flow cross-validation.

[0081] In the cooling pipe system, when the pressure rises sharply but the flow rate is extremely low, such as ΔP>0.2MPa and Q<50% of the rated flow rate, it is determined that the sensor is faulty and needs to be removed. For the removed abnormal values, the linear interpolation method is used to replace them. Assume that the abnormal value is located at x i , the adjacent data points are x i-1 、 x i+1 , the estimated value after replacement is:

[0082] .

[0083] The method of sliding window expansion is as follows:

[0084] See also Figure 2 For congestion categories with smaller sample sizes, we use sampling and sliding windows to balance the dataset distribution. For time series chiller operating data, we determine an appropriate sliding window size. Assuming data is collected every minute, the window size can be set to m, meaning each window contains data from m consecutive time points. Starting from the starting point of the data, we slide the window at the specified step size, generating a new window of data with each slide.

[0085] The data normalization method is as follows:

[0086] After denoising the temperature, pressure, and flow data, RobustScaler normalization is performed on these data to eliminate the differences caused by different physical dimensions and improve data comparability and the effectiveness of model processing. The calculation formula is:

[0087] ;

[0088] Where, X is the denoised sequence data, X scaled is the standardized data, median( X ) is the median of the sequence, IQR( X ) is the interquartile range.

[0089] S3. Build a deep learning hybrid model (Bi-TCN-Transformer) to extract local features and capture global temporal dependencies through the deep learning hybrid model.

[0090] See also Figure 3 , the steps to build a deep learning hybrid model (Bi-TCN-Transformer) include:

[0091] S3.1. Bi-TCN module construction:

[0092] See also Figure 4 ,Bidirectional Temporal Convolutional Network (Bi-TCN), based on the traditional temporal convolutional network, introduces a bidirectional structure and models sequence data from both forward and reverse directions.

[0093] TCN expands the receptive field through dilated convolution, extracts local detailed features of pressure pulsation and temperature gradient, and effectively models the gradual change pattern of cooler parameters without stacking multiple layers.

[0094] In the Bi-TCN module, the forward and reverse temporal convolutions are configured identically, performing the dilated convolution, maximum pooling, and batch normalization operations in sequence. The difference is that the reverse temporal convolution reverses the input vector, and the vectors from the two branches are concatenated and output.

[0095] S3.2: Transformer module construction:

[0096] See also Figure 5 The Transformer module captures the dynamic coupling relationship between long-term trends and parameters through a multi-head attention mechanism. The multi-head attention mechanism dynamically focuses on key periods by calculating the correlation weight between any two time points and suppressing noise or irrelevant information.

[0097] The multi-head attention mechanism first performs a linear mapping on the input sequence to obtain three sets of representations: query (Q), key (K), and value (V). The matrix calculation is:

[0098] ;

[0099] X is the input sequence; W Q , W K , W V is the weight matrix.

[0100] Then, the attention weights for each position are calculated by performing dot product operations on these representations and passing them through scaling and softmax functions:

[0101] ;

[0102] Where, d k is the dimension of the key; QK T is the dot product operation.

[0103] The value of each position is weighted and summed according to the attention weight to form the output of the multi-head attention. Finally, the outputs of multiple attention heads are connected in series and linearly mapped to obtain the final multi-head attention output, as shown in the following formula:

[0104] ;

[0105] Where, head i = Attention( XW i Q , XW i K , XW i V ); h is the number of attention heads; W O It is the linear mapping matrix of the concatenated attention head.

[0106] In the Transformer module, the encoder part is used for feature extraction.

[0107] S3.3. Output layer construction:

[0108] Different from the traditional single-task classification model, an output branch dedicated to pressure prediction is added on top of the existing classification model architecture.

[0109] After the Transformer module extracts features from the data, it is fed into two fully connected layers. One fully connected layer performs classification, accurately determining the predicted cooler blockage category; the other fully connected layer outputs a pressure prediction. This pressure prediction is incorporated into the calculation of physical loss to construct a comprehensive loss function. This comprehensive loss function is used to optimize the model and ultimately output the predicted cooler blockage category.

[0110] S4, see Figure 6 , combining the Navier-Stokes equation and cross entropy to construct a comprehensive loss function.

[0111] S4.1. Physical Constraint Design:

[0112] The simplified Navier-Stokes equations (Navier-Stokes equations) constraints are introduced into the loss function to ensure that the predicted flow-pressure relationship conforms to the laws of fluid mechanics and reduce false positives caused by sensor noise. The Navier-Stokes equations are classic equations that describe fluid motion. Their complete form is:

[0113] ;

[0114] Where, is the fluid density; is the velocity field, , represents the components of velocity in the x, y, and z directions; t is the time variable; is the gradient operator, ; p For pressure; μ is the dynamic viscosity; For external force.

[0115] In the cooler scenario of the present invention, the following simplifications are made:

[0116] 1) Steady-state flow assumption: cooling water flow is steady-state, and the time term is ignored =0;

[0117] 2) One-dimensional flow approximation: The water flow is mainly along the axial direction of the pipe, and the radial and tangential components are ignored, that is, ;

[0118] 3) Incompressible fluid: is a constant, ;

[0119] 4) Ignore body forces: .

[0120] The simplified one-dimensional Navier-Stokes equations are:

[0121] ;

[0122] Assuming the pipe cross-sectional area is uniform, the pipe length is L , flow Q =v·A , A is the cross-sectional area, and further discretization gives the theoretical pressure-flow relationship constraint:

[0123] ;

[0124] Where: μ is the dynamic viscosity; L is the pipe length; A is the cross-sectional area; Q is the flow rate; t is the time variable; ρ is the density of the fluid; v is the velocity field; For theoretical pressure; p is the pressure of the fluid;

[0125] S4.2: Physical Loss Construction:

[0126] The mean square error is used to measure the predicted pressure P pred and theoretical pressure Ptheo The difference between them constructs the first loss term :

[0127] ;

[0128] Where N is the number of sample points, and i represents the i-th sample point.

[0129] This loss term reflects the deviation between the model prediction results and the calculation results based on physical equations, prompting the model to learn the pressure distribution that conforms to physical laws.

[0130] Use the mean square error to calculate the predicted pressure P pred The loss between the actual measured pressure P and the second loss term L2 is constructed:

[0131] ;

[0132] Where: N is the number of sample points, i represents the i-th sample point, P i is the ith actual measured pressure value;

[0133] This loss reflects the degree of fit between the model prediction and the actual observation data, ensuring that the model can capture the pressure change characteristics of the actual physical system.

[0134] S4.3: Construction of total loss function:

[0135] Combining the above two loss terms, we can form the final physical loss function:

[0136] ;

[0137] On this basis, we combine the cross entropy loss function commonly used in deep learning classification tasks to construct a total loss function with physical constraints. The total loss function is the weighted sum of the cross entropy loss and the physical loss:

[0138] ;

[0139] Where, L total is the total loss function, L physics is the physical loss function, L CE is the cross entropy loss function, α and β is the weight. In this embodiment, α =0.7, β =0.3.

[0140] S5. Input the processed data into a deep learning hybrid model and output a cooler blockage status classification result; wherein the cooler blockage status classification includes normal, mild, moderate, and severe blockage.

[0141] In S5, the training and prediction of the deep learning hybrid model (Bi-TCN-Transformer) includes the following steps:

[0142] S5.1. Dataset division: The training set, validation set, and test set are divided into two parts in a ratio of 7:2:1.

[0143] S5.2: Deep Learning Framework and Parameter Settings: The model was built using the PyTorch deep learning framework. The batch size was set to 16. This value determines the number of data samples fed into the model during each training session, affecting the stability and efficiency of model training. Furthermore, the number of iterations was set to 500 based on the dataset size and model complexity. The iteration number represents the number of times the model trains on the entire training dataset. A sufficient number of iterations helps the model fully learn the data characteristics.

[0144] S5.3: Optimizer and learning rate settings: Adaptive moment estimation weight decay optimizer (AdamW optimizer) is used, with a learning rate of 0.0001 and a weight decay of 0.0001. In combination with cosine annealing learning rate scheduling, the learning rate is dynamically adjusted as the number of training rounds increases, so that it exhibits a cosine function-like decay change.

[0145] S5.4: Model Training: During the model training phase, parameter optimization is performed by minimizing a comprehensive loss function. The cross-entropy loss drives the model to continuously adjust parameters, improving the accuracy of the cooler blockage prediction and ensuring that the prediction results are as close to the actual category as possible. The physical loss ensures that the pressure prediction results adhere to the laws of fluid physics, ensuring the reliability of the model in pressure prediction. The synergistic effect of these two factors enables the model to accurately predict the cooler blockage category while adhering to the laws of fluid physics, significantly improving the model's prediction accuracy and generalization performance.

[0146] S5.5: Model prediction: After completing the model training through the above training process, use the test set to perform a final evaluation of the model prediction performance and output the prediction results for the cooler blockage category.

[0147] The present invention can achieve early and accurate identification of the cooler blockage state, solve the problem of delayed response of traditional methods, and issue an early warning at the blockage stage.

[0148] The present invention reduces the false alarm rate through collaborative analysis of multi-source sensor data and multi-physical fields, combining physical constraints with data driving.

[0149] The present invention is based on a deep learning hybrid model that can effectively handle the nonlinear dynamic coupling relationship between pressure, flow and temperature parameters to achieve accurate modeling.

[0150] The present invention realizes full process automation, reduces manual intervention, identifies the type of blockage in advance, avoids unplanned downtime, and reduces maintenance costs.

[0151] Example 2:

[0152] 1. Fault level description

[0153] When a cooler is clogged, temperature changes are a key indicator, but relying solely on temperature has limitations. Cooler temperatures vary under different operating conditions and loads, so a single indicator cannot accurately measure the degree of clog. For example, a drop in flow rate may be caused by a clog or pump failure. Therefore, using threshold alarms to address cooler clogs has certain limitations. In reality, the clog level is a nonlinear classification of multiple coupled parameters, not a simple threshold matching.

[0154] The degree of cooler blockage is equivalent to cooling efficiency. Blockage levels are defined based on IEC 60953, "Test Procedure for Thermal Performance of Hydraulic Turbines." In practice, this definition requires a combination of factors such as turbine model, cooler specifications, and operational requirements. This classification is based on industry-standard, multi-parameter comprehensive criteria, taking into account cooler efficiency calculations, combined with maintenance records and expert experience. This classification requires repeated iteration and refinement.

[0155] In the present invention, the fault level can be described from two aspects:

[0156] To put it simply, each blockage level is determined based on industry standards, historical inspection and maintenance records, and expert experience;

[0157] The detailed description is as follows:

[0158] Normal: Parameters such as temperature, pressure, and flow rate fluctuate within the specified normal range, and no abnormal vibration is detected;

[0159] Slight blockage: Flow area is reduced by 10%-20%, pressure drop increase is ≤15%, flow rate decrease is ≤10%, and temperature fluctuation is within ±5% of normal operating conditions;

[0160] Moderate blockage: Flow area decreases by 20%-40%, pressure drop increases by 15%-30%, flow rate decreases by 10%-20%, and temperature abnormally increases by 5%-10%.

[0161] Severe blockage: flow area decreases by ≥40%, pressure drop increases by ≥30%, flow rate decreases by ≥20%, and temperature increases abnormally by more than 10%.

[0162] It seems that the blockage level of the cooler can be judged by changes in parameters such as temperature, pressure, and flow, but the judgment of these changes is based on the situation where blockage has already occurred. The present invention aims to discover the changing trend of the cooler operating parameters by learning historical data, and issue an early warning before the fault develops to a point that significantly affects parameters such as temperature, pressure, and flow.

[0163] 2. Case Description

[0164] Taking the prediction of the blockage of the bearing cooler of a turbine unit of a hydropower station during operation as an example, the following steps are taken:

[0165] Multi-source sensor data acquisition involves placing temperature sensors, pressure sensors, and flow meters on the inner walls of the cooler pipes to collect temperature, flow, and pressure data during operation. A synchronous acquisition system continuously collects time series data of these three parameters at a 1Hz frequency, achieving time alignment of the three-parameter data. The acquisition cycle covers all operating conditions, including unit startup, steady-state operation, and shutdown. A set of data samples is generated within 10 minutes. The blockage levels are determined based on industry standards, historical maintenance records, and expert experience. The collected data is annotated to create a cooler blockage data set.

[0166] Sample Example: Temperature Data: ;

[0167] Pressure data: ;

[0168] Traffic data: .

[0169] Data preprocessing see Figure 2 A sliding window Z-score algorithm is used to detect transient interference with a 60-second window length. The mean μ and standard deviation σ are calculated, and data points outside the ±3σ range are removed. Further pressure-flow cross-validation is used to identify abnormal data. When ΔP > 0.2 MPa and Q < 50% of the rated flow, the sensor is considered faulty and needs to be removed. Linear interpolation is used to replace the removed outliers.

[0170] Among the various types of data collected, normal conditions are more numerous than other categories. Directly using this data will result in data imbalance, causing the deep learning model to learn more features of normal data and insufficient features of other minority categories. This results in the model performing well in identifying normal conditions, but performing poorly in detecting abnormal conditions or other rare categories, resulting in reduced generalization. Therefore, a sliding window expansion is performed on samples with fewer categories (such as severe congestion). Starting from the beginning of the time series data, a 600-byte segment of data is selected as the first sliding window. The data within the current sliding window is used as a sample and labeled as the congestion category. The window is then slid to the right with a step size of 200, and a new 600-byte segment of data is selected as the next sliding window. This process continues until the window reaches the end of the time series data. This sliding window method generates multiple samples containing congestion features, thereby increasing the number of samples in the congestion category. Finally, the temperature, pressure, and flow data are standardized to eliminate dimensional differences between the different features.

[0171] The deep learning model is constructed based on the characteristics of sensor time series data, combining the characteristics of the bidirectional time series convolution module (Bi-TCN) and the transformer module (transformer).

[0172] (1) Bidirectional Temporal Convolution Module (Bi-TCN)

[0173] A bidirectional temporal convolution module is designed, which consists of two independent branches, forward and reverse, to capture the temporal dynamic features in sensor data respectively.

[0174] Forward branch:

[0175] The model sequentially scans the input time series data using multiple layers of one-dimensional convolution kernels to extract the cumulative changes in temperature, pressure, and flow parameters within the historical time window. A nonlinear activation function is introduced after each convolution layer to enhance the model's ability to represent complex patterns. Max pooling operations are then used to gradually compress the time series length, reducing computational complexity.

[0176] Reverse branch:

[0177] After reversing the input time series data, the same convolution and pooling operations as the forward branch are used to extract the sudden fluctuation features of the reverse time dimension.

[0178] Feature fusion:

[0179] The output features of the forward and reverse branches are spliced ​​in the channel dimension to form a fusion feature matrix containing bidirectional spatiotemporal information, providing a global analysis basis for subsequent modules.

[0180] (2) Transformer module

[0181] Sensor data is time series data. The transformer module can capture long-range dependencies in the sequence and mine complex time series patterns in the data. A global time series modeling module is built based on the Transformer encoder. Its core is the multi-head self-attention mechanism, which splits the fused feature matrix into multiple subspaces and calculates the interaction weights of the query Q, key K, and value V vectors respectively. This captures different parts of the sequence without being limited to the local scope and understands the complex global relationships.

[0182] (3) Output layer

[0183] This invention builds upon the existing classification model architecture by adding an output branch specifically for pressure prediction. After the Transformer module extracts features from the data, it is fed into two fully connected layers: one performs the classification task, determining the predicted cooler blockage category; the other outputs the predicted pressure value. This pressure prediction is then incorporated into the calculation of physical losses to construct a comprehensive loss function. This comprehensive loss function is then used to optimize the model, ultimately outputting the predicted cooler blockage category.

[0184] (4) Loss function design

[0185] To enhance the physical plausibility of the model's decisions, fluid dynamics constraints are introduced during training. Based on the simplified Navier-Stokes equations, a theoretical dynamic relationship between flow and pressure parameters is established. The mean squared error (MSE) between the implicitly learned pressure variation characteristics and the theoretical and actual values ​​is calculated as an additional loss term. This loss forces the model to simultaneously optimize classification accuracy and physical compliance during training, avoiding non-physical predictions caused by data noise or anomalous samples.

[0186] The total loss function is composed of cross entropy loss (dominant classification accuracy) and physical loss (dominant law fitting) weighted at a ratio of 7:3. The model parameters are adjusted through back propagation to achieve a balance between data-driven and knowledge-guided approaches.

[0187] Model training and congestion status prediction: The cooler congestion data set is divided into training set, validation set, and test set in a ratio of 7:2:1. The AdamW optimizer (learning rate 1e-4, weight decay 1e-4) is used with cosine annealing learning rate scheduling to balance convergence speed and generalization performance. The batch size (batchsize) is set to 16 and the number of iterations (epochs) is set to 500. The model is trained and validated on the training and validation sets. During training, the model hyperparameters are adjusted based on the results of the validation set. The model performance indicators (accuracy, loss value) are monitored to prevent model overfitting. The test set is used to conduct a final performance evaluation of the model parameters that have been trained and performed well in the validation set, and the probability distribution of the four types of congestion states is finally output. The following warnings are executed based on the prediction results:

[0188] Minor blockage: triggers a yellow warning, prompting maintenance within 72 hours;

[0189] Moderate congestion: triggers an orange alert, requiring shutdown and inspection within 48 hours;

[0190] Severe congestion: Red warning and automatic activation of emergency shutdown protocol.

[0191] As shown in the fault classification above, cooler blockage is influenced by multiple factors, and the relationships between these factors are not simply linear. It is difficult to measure cooler blockage status based on changes in a single indicator. Furthermore, in actual production, cooler blockage status is often only known after an incident occurs, resulting in significant economic losses. Threshold alarms typically only sound after indicators such as temperature have exceeded normal ranges, at which point the blockage may have already caused significant damage to the equipment. Deep learning models, however, can learn fault evolution patterns not explicitly expressed by traditional criteria, enabling early warning.

Claims

1. A cooler blockage prediction method based on deep learning coupled with physical constraints, characterized in that: The following steps are involved: S1. Collect the temperature, pressure difference and flow parameters of the cooler during operation and generate time series data; S2. Preprocess the time series data to generate a time series sample vector; S3. Build a deep learning hybrid model to extract local features and capture global temporal dependencies; S4. Combining Navier-Stokes equation and cross entropy to construct a comprehensive loss function; S5. Input the processed data into the deep learning hybrid model and output a cooler blockage status classification result; wherein the cooler blockage status classification includes normal, mild, moderate, and severe blockage; In S3, the steps to build a deep learning hybrid model include: S3.

1. Construction of bidirectional temporal convolution module: Based on the temporal convolutional network, the sequential data is modeled from both the forward and reverse directions. In the bidirectional temporal convolution module, the forward and reverse temporal convolution settings are consistent, and the dilated convolution, maximum pooling, and batch normalization operations are completed in sequence. S3.2, converter module construction: Based on the features output by the bidirectional temporal convolution module, the multi-head attention mechanism is used to capture the dynamic coupling relationship between long-term trends and parameters, generating global temporal dependency features. S3.

3. Output layer construction: The output layer adds an output branch for pressure prediction. After the transformer module completes feature extraction on the data, it is connected to two fully connected layers. One fully connected layer is used to perform classification tasks and determine the predicted category of cooler blockage. The other fully connected layer outputs the pressure prediction value, which is used in the calculation of physical loss. Based on the deep learning hybrid model, in the transformer module, the encoder part is used for feature extraction.

2. The cooler blockage prediction method based on deep learning coupled with physical constraints according to claim 1 is characterized in that: S1 includes: Install temperature sensors, pressure sensors and electromagnetic flow meters on the cooler pipelines; Based on historical maintenance records, the cooler status is marked as a variety of working states, including normal, slight, moderate and severe blockage; The temperature, pressure difference and flow rate data of the cooler are obtained during the entire cycle of the unit's cold start, steady-state operation, variable load and shutdown stages, and time series data of the cooler under different working conditions are generated.

3. The cooler blockage prediction method based on deep learning coupled with physical constraints according to claim 1 is characterized in that: In S2, the methods for preprocessing time series data include outlier processing, sliding window expansion, and data normalization.

4. The cooler blockage prediction method based on deep learning coupled with physical constraints according to claim 3 is characterized in that: The method for handling outliers is as follows: The z-score algorithm is used to detect and eliminate transient interference points that exceed the threshold range. For the time series data of each parameter, the mean and standard deviation within the window length are calculated to eliminate abnormal points. Then, the window slides one time unit, and the calculation and outlier detection are continued for the data in the new window until the entire sequence is traversed; After the temperature, pressure, and flow data are processed by the z-score algorithm, the abnormal sensor data is identified through pressure-flow cross validation; for the outliers removed, the linear interpolation method is used to replace them, assuming that the outliers are located at x i , the adjacent data points are x i-1 、 x i+1 , the estimated value after replacement is: 。 5. The cooler blockage prediction method based on deep learning coupled with physical constraints according to claim 3 is characterized in that: The method for sliding window expansion is as follows: For congestion categories with small sample sizes, sampling and sliding windows are used to balance the distribution of the data set. For cooler operation data in time series form, the sliding window size is determined. If the data is collected at intervals of one minute, the window size is set to m, that is, each window contains data from m consecutive time points. Starting from the starting position of the data, the window is slid according to the set step size, and each sliding generates a new window of data.

6. The cooler blockage prediction method based on deep learning coupled with physical constraints according to claim 3 is characterized in that: The method of data standardization is as follows: After denoising the temperature, pressure, and flow data, these data are standardized to eliminate the differences caused by different physical dimensions and improve the comparability of the data and the effectiveness of the model processing. The calculation formula is: ； Where, X is the denoised sequence data, X scaled is the standardized data, median( X ) is the median of the sequence, IQR( X ) is the interquartile range.

7. The cooler blockage prediction method based on deep learning coupled with physical constraints according to claim 1 is characterized in that: In S4, the Navier-Stokes equation and cross entropy are combined to construct a comprehensive loss function, which includes the following steps: S4.

1. Introduce the Navier-Stokes equation constraint into the loss function to ensure that the relationship between the predicted intermediate quantity "flow-pressure" in the deep learning hybrid model conforms to the laws of fluid mechanics. Use the mean square error to measure the difference between the predicted pressure and the theoretical pressure to construct the first loss term. S4.

2. Use the mean square error to calculate the loss between the predicted pressure and the actual measured pressure to construct a second loss term; S4.

3. Combine the above two loss terms to form the final physical loss function. On this basis, combine the cross-entropy loss function in the deep learning classification task to construct a total loss function with physical constraints. The total loss function is the weighted sum of the cross-entropy loss function and the physical loss function.

8. The cooler blockage prediction method based on deep learning coupled with physical constraints according to claim 7 is characterized in that: In S4.1, the simplified Navier-Stokes equations are: ； Assuming the pipe cross-sectional area is uniform, the flow rate Q =v·A , A is the cross-sectional area, and further discretization yields the "flow-pressure" relationship constraint: ； Where: μ is the dynamic viscosity; L is the pipe length; A is the cross-sectional area; Q is the flow rate; t is the time variable; ρ is the density of the fluid; v is the velocity field; For theoretical pressure; p is the pressure of the fluid; The mean square error is used to measure the predicted pressure P pred With the theoretical pressure P theo The difference between them constructs the first loss term : ； Where: N is the number of sample points, i represents the i-th sample point; In S4.2, the mean square error is used to calculate the predicted pressure P pred The actual measured pressure P The loss between them constructs the second loss term : ； Where: N is the number of sample points, i represents the i-th sample point, P i is the i-th measured pressure value; In S4.3, the above two loss terms are combined to form the final physical loss function: ； The total loss function is the weighted sum of cross entropy loss and physical loss: ； Where, L total is the total loss function, L physics is the physical loss function, L CE is the cross entropy loss function, α and β is the weight.

9. The cooler blockage prediction method based on deep learning coupled with physical constraints according to claim 1 is characterized in that: In S5, the training and prediction of the deep learning hybrid model includes the following steps: The training set, validation set, and test set are divided into a ratio of 7:2:

1. An adaptive moment estimation weight decay optimizer is used in conjunction with cosine annealing learning rate scheduling. During the model training phase, parameter optimization is performed by minimizing the comprehensive loss function. After model training is completed, the test set is used for the final evaluation of the model prediction performance, and the prediction results for the cooler blockage category are output.

Images