Cooler blockage prediction method based on deep learning coupling physical constraint

Abstract

The invention relates to the technical field of state prediction of hydroelectric power generation equipment, and discloses a cooler blockage prediction method based on deep learning coupling physical constraint, which comprises the following steps of: acquiring temperature, pressure difference and flow parameters during operation of a cooler, and generating time sequence data; preprocessing the time sequence data to generate a time sequence sample vector; constructing a deep learning hybrid model, extracting local features through the deep learning hybrid model, and capturing global time sequence dependence; combining a Navier-Stokes equation and a cross entropy to construct a comprehensive loss function; and inputting the processed data into the deep learning hybrid model, and outputting a cooler blockage state classification result. According to the method, early accurate recognition of the blockage state of the cooler can be achieved, the problem of response lag of a traditional method is solved, early warning can be given out in the early stage of blockage, the blockage type is recognized in advance, and non-planned shutdown is avoided.

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 the operation of the turbine, the pipes of the bearing cooler are very likely to be blocked due to the accumulation of mud, scale, or foreign matter. Once blocked, the cooling water flow rate will be greatly reduced, making it difficult to fully exert the heat dissipation effect, causing the bearing temperature to rise sharply. Continuous high temperature will not only accelerate bearing wear, but in severe cases it will even cause the bearing to burn, which will in turn cause abnormal vibration of the unit, and in severe cases cause the unit to shut down. Therefore, it is of great significance to predict cooler blockage in advance. To a certain extent, it can ensure equipment safety, improve production efficiency, help achieve refined equipment management, and support long-term and stable operation of equipment.

[0003] The current industry generally adopts an alarm mechanism based on a single temperature threshold or a manual regular inspection method, which has the following technical pain points: 1) Response hysteresis: Traditional methods rely on a single temperature threshold alarm, but temperature anomalies usually occur in the late stage of congestion, making it impossible to achieve early warning; 2) High false alarm rate: Single temperature parameter is easily affected by the environment and lacks multi-physics field collaborative analysis; 3) Complex characteristic coupling: There is a nonlinear dynamic coupling relationship between pressure, flow and temperature parameters, which is difficult to accurately model with traditional mathematical models; 4) High maintenance cost: Preventive maintenance relies on manual experience and lacks data-driven decision support. Sudden severe blockages may lead to unplanned downtime, high single maintenance costs, and loss of power generation. Summary of the invention

[0004] 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, to realize the whole process operation from data collection, processing to state prediction, to reduce manual intervention, to identify the blockage type in advance, to avoid unplanned downtime, and to reduce the false alarm rate.

[0005] In order to overcome the problems existing in the prior art, the technical solution adopted by the present invention is: a cooler blockage prediction method based on deep learning coupled with physical constraints, comprising the following steps: S1, collecting the temperature, pressure difference and flow parameters of the cooler during operation and generating time series data; S2, preprocess the time series data to generate a time series sample vector; S3, build a deep learning hybrid model, extract local features through the deep learning hybrid model, and capture global temporal dependencies; S4. Construct a comprehensive loss function by combining the Navier-Stokes equations and cross-entropy; S5. Input the processed data into the deep learning hybrid model to output the classification results of the cooler clogging status; where the classification of the cooler clogging status includes normal, slight, moderate, and severe clogging.

[0006] Specifically, S1 includes installing temperature sensors, pressure sensors, and electromagnetic flowmeters on the cooler pipeline; According to the historical maintenance records, mark the cooler status as multiple working states, and the working states include normal state, slight clogging, moderate clogging, and severe clogging; Obtain the temperature, pressure difference, and flow rate data of the cooler during the full cycle of cold start, steady-state operation, load change, and shutdown of the unit, and generate the time-series data of the cooler under different working states.

[0007] In S2, the methods for preprocessing the time-series data include outlier processing, sliding window expansion, and data standardization.

[0008] The method of the outlier processing is as follows: Adopt the z-score algorithm to detect and remove the transient interference points beyond the threshold range. For the time-series data of each parameter, calculate the mean and standard deviation within the window length, and remove the outlier points; Subsequently, slide the window by one time unit, and continue to calculate and detect outlier points for the data within the new window until the entire sequence is traversed; After processing the temperature, pressure, and flow rate data with the z-score algorithm, identify the abnormal sensor data through pressure-flow cross-validation; for the removed outliers, use the linear interpolation method to replace them. Suppose the outlier is located at x i , and the adjacent data points before and after are x i-1 , x i+1 , and the estimated value after replacement is: .

[0009] The method of the sliding window expansion is as follows: For the clogging categories with a small sample size, use sampling and sliding windows to balance the dataset distribution; for the cooler operation data in the form of time series, determine the sliding window size. If the data is collected at intervals of one minute, the window size is set to m, that is, each window contains the data of consecutive m time points. Starting from the beginning position of the data, slide the window according to the set step size, and generate a new window of data each time it slides.

[0010] The method of the data standardization is as follows: After denoising the temperature, pressure, and flow rate data, to eliminate the differences brought about by different physical dimensions and improve the comparability of the data and the effectiveness of model processing, these data are respectively standardized, and the calculation formula is as follows: ; In the formula, X is the sequence data after denoising, X scaled is the data after standardization, median( X ) is the median of this sequence, and IQR( X ) is the interquartile range.

[0011] In S3, based on the deep learning hybrid model, in the transformer module, the encoder part is used for feature extraction; an output branch for pressure prediction is added to the output layer. After the data is feature-extracted by the transformer module, it is respectively connected to two fully connected layers. One fully connected layer is used to perform a classification task to determine the predicted category of cooler blockage, and the other fully connected layer outputs the pressure prediction value, which participates in the calculation process of the physical loss.

[0012] In S4, combining the Navier-Stokes equation and cross-entropy, construct a comprehensive loss function, including the following steps: S4.1. Introduce the Navier-Stokes equation constraint in the loss function to make the "flow rate - pressure" relationship of the prediction intermediate quantity of the deep learning hybrid model conform to the laws of fluid mechanics. Use the mean square error to measure the difference between the predicted pressure and the theoretical pressure, and 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, and construct the 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.

[0013] In S4.1, the simplified Navier-Stokes equation is: ; Assume that the cross-sectional area of the pipeline is uniform, the flow rate Q =v·A , A is the cross-sectional area, and further discretize to obtain the "flow rate - pressure" relationship constraint: ; In the formula: μ is the dynamic viscosity; L is the pipeline length; A is the cross-sectional area; Qwhere \(Q\) is the flow rate; \(t\) is the time variable; ρ \(\rho\) is the density of the fluid; v \(\vec{v}\) is the velocity field; \(P_{th}\) is the theoretical pressure; p \(P\) is the pressure of the fluid; The mean square error is used to measure the difference between the predicted pressure \(P\) pred and the theoretical pressure \(P_{th}\) theo to construct the first loss term : ; where \(N\) is the number of sample points, and \(i\) represents the \(i\)-th sample point; In S4.2, the mean square error is used to calculate the loss between the predicted pressure \(P\) pred and the actual measured pressure P to construct the second loss term : ; where \(N\) is the number of sample points, \(i\) represents the \(i\)-th sample point, P i \(P_{i}\) 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 the cross-entropy loss and the physical loss: ; where L total \(L\) is the total loss function, L physics \(L_{p}\) is the physical loss function, L CE \(L_{ce}\) is the cross-entropy loss function, α and β \(\alpha\) and \(\beta\) are weights.

[0014] In S5, the training and prediction of the deep learning hybrid model include the following steps: The training set, validation set, and test set are divided in a ratio of 7:2:1. The adaptive moment estimation weight decay optimizer is adopted, combined with the cosine annealing learning rate scheduler. During the model training stage, the parameters are optimized by minimizing the comprehensive loss function. After the 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 categories are output.

[0015] The present invention has the following beneficial effects: 1. The present invention collects time-series data of temperature, pressure, and flow rate by deploying a multi-source sensor network, and constructs a deep learning hybrid model to achieve multi-scale feature extraction. By fusing the hydrodynamic equation and the cross-entropy loss function, a comprehensive loss function under physical constraints is formed, enabling the model to optimize both the classification accuracy and the compliance with physical laws during the training process, and achieving a balance between data-driven and knowledge-guided. Through outlier filtering, sliding window data augmentation, and robust normalization preprocessing, combined with the collaborative feature extraction of bidirectional temporal convolution and attention mechanism, a four-level classification of the blockage degree is achieved, namely normal, slight, moderate, and severe blockage. This method can accurately identify the early stage of the cooler blockage state in advance, providing data-driven decision support for the intelligent operation and maintenance of the water turbine unit.

[0016] 2. The present invention combines the Navier-Stokes equation and cross-entropy to construct a comprehensive loss function. By integrating data-driven and physical prior knowledge, the interpretability of the model is enhanced and the risk of overfitting is reduced, constraining the model output to conform to the hydrodynamic laws and avoiding misjudgments of pure data-driven methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0018] Figure 1 It is a flowchart of the present invention.

[0019] Figure 2 It is a schematic diagram of the sliding window process of the few-shot data of the present invention.

[0020] Figure 3 It is a structural diagram of the deep learning model of the present invention.

[0021] Figure 4 It is a schematic diagram of the bidirectional temporal convolution network module of the present invention.

[0022] Figure 5 It is a schematic diagram of the transformer module of the present invention.

[0023] Figure 6 It is a flowchart of the construction of the comprehensive loss function of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] In order to make the above objects, features, and advantages of the present invention more obvious and understandable, the following will provide a detailed description of the specific embodiments of the present invention with reference to the drawings in the specification.

[0025] In the following description, many specific details are set forth in order to provide a thorough understanding of the present invention. However, the present invention may be practiced in other ways different from those described herein. Those skilled in the art can make similar generalizations without departing from the spirit of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0026] Secondly, as used herein, the term "this embodiment" or "embodiment" refers to specific features, structures, or characteristics that may be included in at least one implementation manner of the present invention.

[0027] Furthermore, the present invention is described in detail in conjunction with schematic diagrams. When detailing the embodiments of the present invention, for the convenience of explanation, the cross-sectional views showing the device structure will be enlarged locally in a non-general proportion, and the schematic diagrams are only examples and should not limit the scope of protection of the present invention herein. In addition, in actual production, three-dimensional spatial dimensions including length, width, and depth should be included.

[0028] Embodiment 1: See Figure 1 , this embodiment provides a cooler clogging prediction method based on deep learning coupled with physical constraints, including the following steps: S1. Collect the temperature, pressure difference, and flow rate parameters during the operation of the cooler to generate time series data.

[0029] In S1, install temperature sensors, pressure sensors, and electromagnetic flowmeters on the cooler pipeline.

[0030] According to the historical maintenance records, mark the cooler status as multiple working states, and the working states include: normal, slight, moderate, and severe clogging.

[0031] Obtain the temperature, pressure difference, and flow rate data of the cooler during the full cycle of cold start, steady-state operation, load change, and shutdown of the unit to ensure that the model training covers the entire operating condition field.

[0032] Synchronously collect the temperature, pressure, and flow rate parameters to form a time series data stream. For each complete sample, the time series data stream corresponding to each clogging level contains three-dimensional time series points, and at the same time, strictly ensure the time alignment of the temperature, pressure, and flow rate data.

[0033] S2. Preprocess the time series data to generate time series sample vectors.

[0034] In S2, the steps for preprocessing the time series data include outlier processing, sliding window expansion, and data standardization.

[0035] Specifically, the method for outlier processing is as follows: See Figure 2, the z-score algorithm (sliding window Z-score algorithm) is adopted, with the window length set to 60s. Transient interference points beyond the range of ±3σ are detected and removed, where σ is the sliding standard deviation. For each parameter time series data x t , calculate the mean value μ W within the window length w = 60 and the standard deviation σ W , and remove the abnormal points that satisfy the following formula: ; Subsequently, the window slides by one time unit, and the data within the new window is continuously calculated and abnormal points are detected until the entire sequence is traversed.

[0036] After processing the temperature, pressure, and flow data using the z-score algorithm (sliding window Z-score algorithm), sensor abnormal data is further identified through pressure-flow cross-validation.

[0037] In the cooling pipe system, when the pressure rises significantly but the flow rate is extremely low, such as when 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 outliers, the method of linear interpolation is used for replacement. Suppose the outlier is located at x i , and the adjacent data points before and after are x i-1 , x i+1 , and the estimated value after replacement is: .

[0038] The method of sliding window expansion is as follows: Refer to Figure 2 , for the clogging category with a small sample size, sampling and sliding windows are used to balance the dataset distribution. For the cooler operation data in the form of time series, determine the appropriate sliding window size. Assume that the data is collected at one-minute intervals, and the window size can be set to m, that is, each window contains data of consecutive m time points. Starting from the beginning position of the data, slide the window according to the set step size, and a new window of data is generated each time.

[0039] The method of data standardization is as follows: After denoising the temperature, pressure, and flow data, to eliminate the differences brought by different physical dimensions and improve the comparability of the data and the effectiveness of model processing, these data are respectively standardized using RobustScaler. The calculation formula is: ; In the formula, X is the denoised sequence data, Xscaled is the standardized data, median( X ) is the median of this sequence, and IQR( X ) is the interquartile range.

[0040] S3. Construct a deep learning hybrid model (Bi-TCN-Transformer) to extract local features and capture global temporal dependencies through the deep learning hybrid model.

[0041] See Figure 3 , and the steps to construct the deep learning hybrid model (Bi-TCN-Transformer) include: S3.1. Construction of the Bi-TCN module: See Figure 4 , the Bidirectional Temporal Convolutional Network (Bi-TCN) introduces a bidirectional structure on the basis of the traditional temporal convolutional network and models sequence data from both the forward and backward directions simultaneously.

[0042] TCN expands the receptive field through dilated convolution to extract local detailed features of pressure pulsation and temperature gradient, and can effectively model the progressive change pattern of cooler parameters without stacking multiple layers.

[0043] In the Bi-TCN module, the forward and backward temporal convolutions are set the same, and the operations of dilated convolution - max pooling - batch normalization are completed in sequence. The difference is that when performing the backward temporal convolution, the input vector is reversed, and finally the vectors of the two branches are concatenated and output.

[0044] S3.2: Construction of the Transformer module: See Figure 5 , the Transformer module captures the dynamic coupling relationship between long-term trends and parameters through the multi-head attention mechanism. The multi-head attention mechanism dynamically focuses on key periods by calculating the correlation weights between any two time points, suppressing noise or irrelevant information.

[0045] The multi-head attention mechanism first performs a linear mapping on the input sequence to obtain three groups of representations: query (Q), key (K), and value (V). The matrix calculation is: ; X is the input sequence; W Q , W K , W V is the weight matrix.

[0046] Then, by performing dot product operations on these representations and through scaling and the softmax function, the attention weights for each position are calculated as follows: ; where, d k is the dimension of the key; QK T is the dot product operation.

[0047] The values at each position are weighted and summed according to the attention weights to form the output of the multi-head attention. Finally, the outputs of multiple attention heads are concatenated together and, after passing through a linear mapping, the final output of the multi-head attention is obtained, as shown in the following equation: ; where head i = Attention( XW i Q , XW i K , XW i V ); h is the number of attention heads; W O is the linear mapping matrix of the concatenated attention heads.

[0048] In the Transformer module, the encoder part is used for feature extraction.

[0049] S3.3. Output layer construction: Different from traditional single-task classification models, on top of the existing classification model architecture, an output branch dedicated to pressure prediction is added.

[0050] After the Transformer module has completed feature extraction on the data, it is connected to two fully connected layers respectively. One fully connected layer is used to perform the classification task to accurately determine the predicted category of cooler blockage; the other fully connected layer outputs the pressure prediction value. This pressure prediction value participates in the calculation process of the physical loss to construct a comprehensive loss function. The model is optimized by the constructed comprehensive loss, and finally the predicted category of cooler blockage is output.

[0051] S4. Refer to Figure 6 , and combine the Navier-Stokes equation and cross-entropy to construct a comprehensive loss function.

[0052] S4.1. Physical constraint design: Introduce the constraint of the simplified Navier-Stokes equation into the loss function to ensure that the predicted flow-pressure relationship conforms to the laws of fluid mechanics and reduce false alarms caused by sensor noise. The Navier-Stokes equation is a classical equation describing fluid motion, and its complete form is: ; In the formula, is the fluid density; is the velocity field, , representing the components of the velocity in the x, y, and z directions; t is the time variable; is the gradient operator, ; p is the pressure; μ is the dynamic viscosity; is the external force.

[0053] In the cooler scenario of the present invention, the following simplifications are made: 1) Steady flow assumption: The flow of cooling water is steady, and the time term is ignored = 0; 2) One-dimensional flow approximation: The water flow is mainly along the axial direction of the pipeline, and the radial and tangential components are ignored, that is ; 3) Incompressible fluid: is a constant, ; 4) Volume force is ignored: .

[0054] The simplified one-dimensional Navier-Stokes equation is: ; Assume that the cross-sectional area of the pipeline is uniform, the pipeline length is L , the flow rate Q =v·A , A is the cross-sectional area, and further discretization gives the theoretical pressure-flow relationship constraint: ; In the formula: μ is the dynamic viscosity; L is the pipeline 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; is the theoretical pressure; p is the pressure of the fluid; S4.2: Physical loss construction: Use the mean square error to measure the predicted pressure P pred and the theoretical pressureP theo Construct the first loss term based on the difference between : ; In the formula, N is the number of sample points, and i represents the i-th sample point.

[0055] This loss term reflects the deviation between the model prediction result and the result calculated based on the physical equation, and prompts the model to learn the pressure distribution that conforms to the physical law.

[0056] Use the mean square error to calculate the loss between the predicted pressure P pred and the actually measured pressure P, and construct the second loss term L 2 : ; In the formula: N is the number of sample points, i represents the i-th sample point, P i is the i-th actually measured pressure value; This loss term reflects the fitting degree between the model prediction and the actual observed data, and ensures that the model can capture the pressure change characteristics in the actual physical system.

[0057] S4.3: Construction of the total loss function: Combine the above two loss terms to form the final physical loss function: ; On this basis, combine the cross-entropy loss function commonly used in deep learning classification tasks to construct the total loss function with physical constraints. The total loss function is the weighted sum of the cross-entropy loss and the physical loss: ; In the formula, 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 are the weights. In this embodiment, α = 0.7, β = 0.3.

[0058] S5. Input the processed data into the deep learning hybrid model and output the classification result of the cooler clogging state; among them, the classification of the cooler clogging state includes normal, slight, moderate, and severe clogging.

[0059] In S5, the training and prediction of the deep learning hybrid model (Bi-TCN-Transformer) include the following steps: S5.1. Dataset Division: Divide the training set, validation set, and test set in a ratio of 7:2:1.

[0060] S5.2: Deep Learning Framework and Parameter Settings: Build a model using the PyTorch deep learning framework. Set the batch size to 16, which determines the number of data samples input to the model during each training session and affects the stability and efficiency of model training. At the same time, determine the number of iteration steps to be 500 based on the dataset size and model complexity. The number of iteration steps represents the number of times the model trains on the entire training dataset, and sufficient iteration steps help the model fully learn the data features.

[0061] S5.3: Optimizer and Learning Rate Settings: Use the Adaptive Moment Estimation with Weight Decay optimizer (AdamW optimizer), with a learning rate of 0.0001 and a weight decay of 0.0001, in conjunction with a cosine annealing learning rate scheduler. As the number of training epochs increases, the learning rate is dynamically adjusted to exhibit a cosine function - like decay.

[0062] S5.4: Model Training: During the model training stage, optimize the parameters by minimizing the comprehensive loss function. The cross - entropy loss drives the model to continuously adjust the parameters to improve the accuracy of predicting the cooler blockage category, making the prediction results as close as possible to the true category; while the physical loss ensures that the pressure prediction results follow the laws of fluid physics, ensuring the reliability of the model in pressure prediction. The two work together to prompt the model to accurately predict the cooler blockage category while conforming to the laws of fluid physics, significantly improving the prediction accuracy and generalization performance of the model.

[0063] S5.5: Model Prediction: After completing the training of the model through the above - mentioned training process, use the test set for the final evaluation of the model's prediction performance and output the prediction results for the cooler blockage category.

[0064] The present invention can achieve early and accurate identification of the cooler blockage state, solve the problem of lagging response in traditional methods, and can issue early warnings at the early stage of blockage. The present invention reduces the false alarm rate through multi - source sensing data and multi - physical field collaborative analysis, combining physical constraints and data - driven methods. The deep - learning hybrid model of the present invention can effectively handle the non - linear dynamic coupling relationship between pressure, flow rate, and temperature parameters to achieve accurate modeling. The present invention realizes full - process automation, reduces manual intervention, identifies the blockage type in advance, avoids unplanned shutdowns, and reduces maintenance costs.

[0065] Example 2: I. Fault Level Description When the cooler is blocked, the temperature change is an important indicator, but relying solely on temperature has limitations. The cooler temperature will vary under different operating conditions and loads of the unit. At this time, a single indicator cannot measure the degree of blockage. For example, a decrease in flow rate may be caused by blockage or pump failure. Therefore, using the method of threshold alarm to deal with cooler blockage has certain limitations. In fact, the blockage level is a non-linear classification of the multi-parameter coupling state, rather than a simple threshold matching.

[0066] The degree of cooler blockage is equivalent to the cooling efficiency. Refer to IEC 60953 "Code for Thermal Performance Tests of Hydraulic Turbines" to define the blockage level. In practical applications, it is necessary to consider factors such as the type of hydraulic turbine, the specifications of the cooler, and the operating requirements for definition. That is, based on the multi-parameter comprehensive criterion of industry standards, considering the calculation of cooler efficiency, combined with maintenance records and expert experience to give the blockage level division, and it is necessary to repeatedly iterate and correct to determine the division standard.

[0067] In the present invention, the consideration of the fault level can be described from two directions: If simply summarized, each blockage level is determined based on industry standards, historical maintenance records, and expert experience; And the detailed description is as follows: Normal: Parameters such as temperature, pressure, and flow rate fluctuate within the specified normal range, and no abnormal vibration is detected; Slight blockage: The flow area is reduced by 10% - 20%, the pressure drop increase rate ≤ 15%, the flow rate decrease rate ≤ 10%, and the temperature fluctuation is within ±5% of the normal operating condition; Moderate blockage: The flow area is reduced by 20% - 40%, the pressure drop increase rate is 15% - 30%, the flow rate decrease rate is 10% - 20%, and the temperature rises abnormally by 5% - 10%.

[0068] Severe blockage: The flow area is reduced by ≥ 40%, the pressure drop increase rate ≥ 30%, the flow rate decrease rate ≥ 20%, and the temperature rises abnormally by more than 10%.

[0069] It seems that the blockage level of the cooler can be judged by the changes in parameters such as temperature, pressure, and flow rate, but the judgments of these changes are all based on the situation where the blockage has occurred. The purpose of the present invention is to discover the change trend of the cooler operating parameters through learning historical data and issue a warning before the fault develops to significantly affect parameters such as temperature, pressure, and flow rate.

[0070] II. Case Explanation Taking the prediction of the blockage situation during the operation of the bearing cooler of a hydraulic turbine unit in a certain hydropower station as an example, the steps are as follows: Multi-source sensor data acquisition: Temperature sensors, pressure sensors, and flow meters are arranged on the inner wall of the cooler pipeline to collect the temperature, flow rate, and pressure data of the cooler during operation. The three-parameter time-series data is continuously collected at a frequency of 1 Hz through a synchronous acquisition system to achieve time alignment of the three-parameter data. The acquisition period covers all operating conditions of the unit equipment startup, steady-state operation, and shutdown. A set of data samples is generated within 10 minutes, and each blockage level is determined based on industry standards, historical maintenance records, and expert experience. The collected data is labeled to create a dataset of cooler blockage conditions.

[0071] Sample example: Temperature data: ; Pressure data: ; Flow rate data: 。

[0072] For data preprocessing, refer to Figure 2 , and the sliding window Z-score algorithm is used to detect transient interference. The window length is 60 seconds, and the mean μ and standard deviation σ are calculated. Data points outside the range of ±3σ are removed. Then, further cross-validation of pressure and flow rate is used to identify abnormal data. When ΔP > 0.2 MPa and Q < 50% of the rated flow rate, it is determined that the sensor fails and the data needs to be removed. For the removed outliers, linear interpolation is used for replacement; Among the collected data of various situations, the number of normal data is more than that of other categories under normal circumstances. Direct use will result in data imbalance, making the deep learning model tend to learn more features of normal data and insufficiently learn the features of other minority categories. This causes the model to perform well in identifying normal situations but poorly in detecting abnormal situations or other rare categories, and the generalization ability decreases. Therefore, for samples with fewer categories (such as severe blockage), sliding window expansion is performed. Starting from the starting position of the time-series data, a segment of data with a length of 600 is intercepted as the first sliding window. The data within the current sliding window is used as a sample and labeled with the blockage category. The window is slid to the right by a step size of 200, and a new segment of data with a length of 600 is intercepted as the next sliding window until the window slides to the end of the time-series data. Through the sliding window method, multiple samples containing blockage features are generated to expand the number of samples in the blockage category. Finally, the temperature, pressure, and flow rate data are standardized respectively to eliminate the dimensional differences between different features.

[0073] Deep learning model construction: According to the characteristics of the sensor time-series data, a model is constructed by combining the characteristics of the bidirectional time-series convolutional module (Bi-TCN) and the transformer module.

[0074] (1) Bidirectional time-series convolutional module (Bi-TCN) Design a bidirectional temporal convolutional module, which includes two independent branches, forward and backward, to capture the temporal dynamic features in sensor data respectively.

[0075] Forward branch: Scan the input temporal data sequentially through multiple one-dimensional convolutional kernels to extract the cumulative change rules of temperature, pressure, and flow parameters within the historical time window. After each layer of convolution, introduce a non-linear activation function to enhance the model's ability to express complex patterns, and gradually compress the temporal length through max-pooling operations to reduce the computational complexity.

[0076] Backward branch: After reversing the input temporal data, adopt the same convolution and pooling operations as the forward branch to extract the sudden fluctuation features in the reverse time dimension.

[0077] Feature fusion: Concatenate the output features of the forward and backward branches in the channel dimension to form a fusion feature matrix containing bidirectional spatio-temporal information, providing a global analysis basis for subsequent modules.

[0078] (2)Transformer module Sensor data belongs to time series data. The Transformer module can capture long-range dependencies in the sequence, mine complex temporal patterns in the data, and build a global temporal modeling module based on the Transformer encoder. Its core is the multi-head self-attention mechanism, which splits the fusion feature matrix into multiple subspaces, calculates the interaction weights of the query Q, key K, and value V vectors respectively, captures different parts of the sequence without being limited to the local range, and understands the global complex relationships.

[0079] (3)Output layer In the present invention, on the basis of the existing classification model architecture, an output branch dedicated to pressure prediction is added. After the data is feature-extracted by the Transformer module, it is connected to two fully-connected layers respectively. One fully-connected layer is used to perform the classification task to determine the predicted category of cooler blockage; the other fully-connected layer outputs the pressure prediction value. This pressure prediction value participates in the calculation process of the physical loss to construct a comprehensive loss function. The model is optimized by the constructed comprehensive loss, and finally the predicted category of cooler blockage is output.

[0080] (4)Loss function design In order to improve the physical rationality of the model decision, the constraints of fluid mechanics laws are introduced in the training process. Based on the simplified Navier-Stokes equation, the theoretical dynamic relationship between flow parameters and pressure parameters is established. The mean square error between the pressure change characteristics implicitly learned by the model and the theoretical value and the actual value is calculated as an additional loss term. This loss forces the model to optimize the classification accuracy and physical law compliance at the same time during the training process, avoiding non-physical predictions caused by data noise or abnormal samples.

[0081] The total loss function is composed of the cross entropy loss (dominant classification accuracy) and the 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.

[0082] 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. Set the batch size (batchsize) to 16 and the number of iterations (epoch) to 500. The model is trained and validated on the training set and validation set. During the training process, the hyperparameters of the model are adjusted according to the results of the validation set, and the performance indicators of the model (accuracy, loss value) are monitored to prevent the model from overfitting. The test set is used to perform a final performance evaluation on the model parameters that have been trained and performed well in the validation set, and finally output the probability distribution of four types of congestion states. Execute the following warnings based on the prediction situation: Minor blockage: triggers a yellow warning, prompting maintenance within 72 hours; Moderate congestion: triggers an orange warning, requiring shutdown and inspection within 48 hours; Severe blockage: Red warning and automatic initiation of emergency shutdown protocol.

[0083] As can be seen from the above fault level classification, cooler blockage is affected by multiple factors, and the factors are not simply linearly related. It is difficult to measure the blockage of the cooler by the change of a single indicator. In actual production, the blockage of the cooler can only be known after the accident occurs, causing a large amount of economic losses. Threshold alarms usually sound an alarm only after indicators such as temperature have exceeded the normal range. At this time, the blockage may have caused a certain degree of damage to the equipment. The deep learning model can learn the fault evolution mode that is not explicitly expressed by traditional criteria, and achieve 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 collects the temperature, pressure difference and flow parameters of the cooler during operation and generates time series data; S2, preprocess the time series data to generate a time series sample vector; S3, build a deep learning hybrid model, extract local features and capture global temporal dependencies through the deep learning hybrid model; 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 the cooler blockage status classification result; wherein the cooler blockage status classification includes normal, slight, moderate and severe blockage.

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 pipeline; Based on historical maintenance records, the cooler status is marked into multiple working states, including normal, slight, moderate and severe blockage; The temperature, pressure difference and flow data of the cooler are obtained during the whole cycle of the unit's cold start, steady-state operation, variable load and shutdown stages, and the 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 standardization.

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 excluded outliers, 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 of sliding window expansion is as follows: For the blockage category with a small sample size, sampling and sliding windows are used to balance the distribution of the data set; for the cooler operation data in the form of time series, 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 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: ； In the formula, 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 S3, based on the 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 respectively. One of the fully connected layers is used to perform classification tasks to determine the prediction category of cooler blockage, and the other fully connected layer outputs the pressure prediction value, which is involved in the calculation process of physical loss.

8. 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, including the following steps: S4.

1. Introduce Navier-Stokes equation constraints into the loss function so that the relationship between the predicted intermediate quantity "flow-pressure" of the deep learning hybrid model conforms to the law 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, using the mean square error to calculate the loss between the predicted pressure and the actual measured pressure, and constructing 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.

9. The cooler blockage prediction method based on deep learning coupled with physical constraints according to claim 8 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 gives the "flow-pressure" relationship constraint: ； Where: μ is the dynamic viscosity; L is the length of the pipeline; 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 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: ； In the formula, 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.

10. The cooler blockage prediction method based on deep learning coupled with physical constraints according to claim 1, 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 7:2:1 ratios. The adaptive moment estimation weight decay optimizer is used in conjunction with the cosine annealing learning rate scheduling. During the model training phase, parameter optimization is performed by minimizing the comprehensive loss function. After the 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