Evaluation method and system for shell and tube heat exchangers in high pressure applications based on industrial big data

Abstract

The present invention relates to the technical field of industrial data processing, and specifically to a method and system for evaluating high-pressure shell and tube heat exchangers based on industrial big data, including: collecting industrial data information and performing normalization processing, combining the processing results into a multidimensional matrix, and mapping them to the pixel value range of an image, and obtaining a visualized image through the mapping results; determining a comprehensive intensity label, taking the visualized image as input, constructing a convolutional neural network model, and training the model; and evaluating high-pressure shell and tube heat exchangers through the convolutional neural network model. The present invention realizes a comprehensive, accurate and efficient evaluation of high-pressure shell and tube heat exchangers through systematic data collection, standardized data processing, advanced machine learning technology and automated evaluation processes, which not only improves the efficiency and accuracy of the evaluation process, but also enhances the reliability of the evaluation results, providing strong support for the optimization and safety management of industrial equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial data processing, and in particular to a method and system for evaluating a high-pressure shell and tube heat exchanger based on industrial big data. Background Art

[0002] In high-pressure application environments, shell and tube heat exchangers are widely used in the fields of petroleum, chemical industry, electric power, etc. As key equipment, their performance and quality directly affect the operating efficiency and safety of the entire system.

[0003] In the industrial production process, the data and information generated by the heat exchanger are multifaceted, often including industrial data and information such as structural form, material selection, safety devices and manufacturing process. Specifically, the industrial data of the structural form includes: wall thickness design data. Under high pressure environment, the wall thickness of the tube body and shell needs to be appropriately increased to improve the pressure resistance. The wall thickness design must comply with the relevant pressure vessel design standards, such as ASME standards; tube sheet design data. The tube sheet must have sufficient thickness and strength under high pressure conditions to prevent deformation and leakage, such as using forged tube sheets to increase strength; special connection method data, such as the connection between the tube sheet and the tube body usually adopts the method of expansion and welding to ensure the sealing performance under high pressure; partition design data, it is necessary to ensure that the partition has sufficient rigidity and strength under high pressure environment to avoid deformation, and the connection between the partition and the shell should be firm to prevent falling off.

[0004] Regarding material selection, industrial data information may specifically include: in order to withstand high pressure, select materials with high strength and high pressure resistance, such as stainless steel, alloy steel or titanium alloy; and, according to the properties of the fluid, select materials with excellent corrosion resistance to extend the service life of the equipment.

[0005] Industrial data on safety devices mainly rely on safety valves and pressure gauges installed in high-pressure systems to monitor and control pressure and prevent overpressure accidents; as well as well-designed pressure relief devices to ensure that pressure can be released in time when the pressure exceeds the limit to protect the safety of equipment and operators.

[0006] The industrial data of the manufacturing process is multifaceted. During the implementation process, some key points of the data include: high-precision processing flow, each component of the high-pressure heat exchanger needs to be processed with high precision to ensure close fit and reduce stress concentration; welding process control requirements, strict control of the welding process, selection of high-strength welding materials, to ensure welding quality and pressure resistance; non-destructive testing schemes, such as X-ray inspection of welds and key parts to ensure the absence of cracks and defects, or ultrasonic testing to evaluate internal defects of materials to ensure structural integrity; pressure testing schemes, such as water pressure tests higher than the working pressure before leaving the factory to test the pressure resistance and sealing performance of the heat exchanger, or air pressure tests as needed to further verify the sealing and pressure resistance of the equipment.

[0007] However, the existing high-pressure shell and tube heat exchangers often have the following problems in the production process due to the lack of effective application of industrial big data:

[0008] Various types of industrial data collected during the production process often exist in isolated forms, making it difficult to make effective and comprehensive use of them. Traditional methods are inefficient and error-prone when dealing with large-scale data and complex relationships. Traditional evaluation methods for high-pressure shell and tube heat exchangers usually rely on a single or a few parameters for evaluation. This method cannot fully reflect the overall performance and quality of the equipment, and may cause some important issues to be ignored. For example, evaluation based only on a single structural strength or material performance may underestimate the impact of safety devices or manufacturing processes on the overall performance of the equipment.

[0009] In addition, the current evaluation method mainly relies on expert experience and manual analysis. The whole process is time-consuming, inefficient, and easily affected by human factors. The objectivity and consistency of the evaluation results are difficult to guarantee. Summary of the invention

[0010] The present invention provides a method and system for evaluating a high-pressure shell and tube heat exchanger based on industrial big data, thereby effectively solving the problems pointed out in the background technology.

[0011] In order to achieve the above object, the technical solution adopted by the present invention is:

[0012] Evaluation method for shell and tube heat exchangers for high pressure applications based on industrial big data, including:

[0013] Collect industrial data information on the structure, material selection, safety devices and manufacturing process of shell and tube heat exchangers for high pressure applications;

[0014] Normalizing the industrial data information to map the value ranges of different features to the same interval;

[0015] Combining the normalized industrial data information into a multi-dimensional matrix, and mapping the multi-dimensional matrix into a pixel value range of an image to obtain a mapping result;

[0016] Creating a blank image matrix, and filling the mapping results into corresponding positions of the image matrix to obtain a visualized image;

[0017] Determine a comprehensive strength label for each industrial data information, wherein the comprehensive strength label is used to measure the overall performance and quality of the heat exchanger;

[0018] Taking the visualized image as input, constructing a convolutional neural network model, and using the comprehensive intensity label as the target label of the model to train the model;

[0019] The high-pressure shell and tube heat exchanger is evaluated by using the trained convolutional neural network model, and the evaluation result is output.

[0020] Furthermore, the comprehensive strength labels of each industrial data information are determined, including:

[0021] Providing separate scores for each of the industrial data information;

[0022] Each score is normalized to obtain a numerical range of 0 to 1;

[0023] A weight is assigned to each of the industrial data information, and the comprehensive strength label is the sum of the normalized results of each of the scores multiplied by the corresponding weight.

[0024] On the basis of the above embodiment, a convolutional neural network model is constructed, including:

[0025] Dividing the generated visualization images and the corresponding comprehensive intensity labels into a training set, a validation set and a test set;

[0026] Selecting a pre-trained model, the pre-trained model comprising at least three convolutional layers, freezing at least 50% of the convolutional layers, and adding a custom fully connected layer on top of the pre-trained model;

[0027] Compiling the pre-trained model is achieved by selecting a loss function, an optimizer, and an evaluation metric;

[0028] Using the training set data to train the pre-trained model, and verifying it on the verification set;

[0029] Unfreeze some convolutional layers of the pre-trained model and adjust the model;

[0030] The model performance is evaluated on the test set, and the optimization of the pre-trained model is continued based on the test results.

[0031] Furthermore, the part of the convolutional layers that are unfrozen are convolutional layers close to the top.

[0032] As a preferred embodiment of this invention, the convolutional neural network model structure constructed includes:

[0033] An input layer, receiving the visualization image;

[0034] A convolution layer, used for extracting local features from the visualized image and outputting a feature map, performing a filtering operation through a plurality of convolution kernels, each of which learns a different feature response mode;

[0035] An activation layer that introduces nonlinearity after the convolutional layer;

[0036] A maximum pooling layer to reduce the spatial dimension of the feature map and extract key features;

[0037] A flattening layer flattens the multi-dimensional feature map into a one-dimensional vector;

[0038] Several fully connected layers, where neurons in the fully connected layers are connected to all neurons in the previous layer for high-level feature learning;

[0039] The output layer is a fully connected layer that outputs the evaluation result through a single neuron.

[0040] Furthermore, the constructed convolutional neural network model structure includes:

[0041] An input layer, receiving the visualization image;

[0042] A convolution layer, used for extracting local features from the visualized image and outputting a feature map, performing a filtering operation through a plurality of convolution kernels, each of which learns a different feature response mode;

[0043] A residual block having several of the convolutional layers, introducing skip connections for the convolutional neural network;

[0044] An activation layer that introduces nonlinearity after the convolutional layer;

[0045] A global average pooling layer that converts the feature maps into a single output tensor;

[0046] Several fully connected layers, where neurons in the fully connected layers are connected to all neurons in the previous layer for high-level feature learning;

[0047] The output layer is a fully connected layer that outputs the evaluation result through a single neuron.

[0048] Furthermore, the constructed convolutional neural network model structure also includes adding a dropout layer between the fully connected layers to randomly shut down a number of neurons.

[0049] Further, the pre-trained model is trained using the training set data and verified on the verification set; including:

[0050] Loading the weight and bias parameters of the pre-trained model;

[0051] Setting training rounds, and performing cyclic training on the convolutional neural network model through a back propagation algorithm;

[0052] Monitor the cyclic training process, and after each training round, use the verification set to verify the convolutional neural network model, and adjust the model parameters according to the performance of the convolutional neural network model on the verification set until the model training termination condition is reached.

[0053] Furthermore, it also includes setting an early stopping strategy, and determining whether to start the early stopping strategy based on the performance of the convolutional neural network model on the verification set.

[0054] High-pressure shell and tube heat exchanger evaluation system based on industrial big data, including:

[0055] Industrial data collection module, which collects industrial data information on the structure, material selection, safety devices and manufacturing process of high-pressure shell and tube heat exchangers;

[0056] An industrial data processing module normalizes the industrial data information to map the value ranges of different features to the same interval;

[0057] An industrial data mapping module combines the normalized industrial data information into a multi-dimensional matrix, and maps the multi-dimensional matrix into a pixel value range of an image to obtain a mapping result;

[0058] A visualization image generation module creates a blank image matrix and fills the mapping results into corresponding positions of the image matrix to obtain a visualization image;

[0059] A comprehensive strength label determination module determines the comprehensive strength label of each industrial data information, wherein the comprehensive strength label is used to measure the overall performance and quality of the heat exchanger;

[0060] A convolutional neural network building module takes the visualization image as input to build a convolutional neural network model, and uses the comprehensive intensity label as a target label of the model to train the model;

[0061] The convolutional neural network application module evaluates the high-pressure shell and tube heat exchanger through the trained convolutional neural network model and outputs the evaluation result.

[0062] The technical solution of the present invention can achieve the following technical effects:

[0063] The present invention realizes a comprehensive, accurate and efficient evaluation of shell and tube heat exchangers for high-pressure applications through systematic data collection, standardized data processing, advanced machine learning technology and automated evaluation process. It not only improves the efficiency and accuracy of the evaluation process, but also enhances the reliability of the evaluation results, providing strong support for the optimization and safety management of industrial equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0065] Figure 1 This is a schematic diagram of the sources of industrial data information;

[0066] Figure 2 It is a flow chart of the evaluation method of shell and tube heat exchanger for high pressure application based on industrial big data;

[0067] Figure 3 A schematic diagram of the process of building a convolutional neural network model;

[0068] Figure 4 A schematic diagram comparing the two convolutional neural network model structures constructed;

[0069] Figure 5 Schematic diagram of the structure of the shell and tube heat exchanger evaluation system for high-pressure applications based on industrial big data. DETAILED DESCRIPTION

[0070] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0071] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0072] like Figure 1 and 2 As shown in the figure, the evaluation method of high-pressure shell and tube heat exchanger based on industrial big data includes:

[0073] S01: Collect industrial data information on the structure, material selection, safety devices and manufacturing process of shell and tube heat exchangers for high pressure applications;

[0074] S02: normalize the industrial data information to map the value ranges of different features to the same interval;

[0075] Because different features may have different data ranges and units, in this step, all information is mapped to the same numerical range, usually between 0 and 1, through normalization to ensure that they have similar weights and comparability in subsequent processing;

[0076] S03: combining the normalized industrial data information into a multi-dimensional matrix, and mapping the multi-dimensional matrix into a pixel value range of the image to obtain a mapping result;

[0077] Taking a specific high-pressure shell and tube heat exchanger as an example, assume that the specific data of its structure, material selection, safety device and manufacturing process are as follows:

[0078] (1) Structural form:

[0079] Wall thickness design: The wall thickness of the tube body and shell are 10mm and 20mm respectively, which conforms to ASME standards.

[0080] Tube sheet design: Forged tube sheet with a thickness of 30mm.

[0081] Special connection method: The connection between the tube sheet and the tube body is achieved by expansion and welding.

[0082] Partition design: The partition thickness is 5mm and the material is high-strength steel to ensure rigidity and strength.

[0083] (2) Material selection:

[0084] Tube material: Stainless steel 304, with high strength and corrosion resistance.

[0085] Shell material: Alloy steel P11, suitable for high temperature and high pressure environment.

[0086] Tube sheet material: titanium alloy, with excellent corrosion resistance.

[0087] (3) Safety devices:

[0088] Safety valve: installed on the shell side, with the set pressure at 20MPa.

[0089] Pressure gauge: Installed in the pipe, monitoring pressure range is 0-25MPa.

[0090] Pressure relief device: The set pressure is 22MPa to ensure timely pressure relief in case of overpressure.

[0091] (4) Manufacturing process:

[0092] High-precision processing: The processing accuracy of the tube body and shell reaches ±0.1mm.

[0093] Welding process: TIG welding is adopted, with high welding strength and no cracks in the weld.

[0094] Non-destructive testing: X-ray testing of welds and ultrasonic testing of key parts.

[0095] Pressure test: Conduct a 30MPa water pressure test and a 25MPa air pressure test.

[0096] The normalization and mapping results of the above data are as follows:

[0097] Serial number Industrial data information Normalized value Mapping Data 1 Wall thickness design (tube body) 0.2 51 2 Wall thickness design (shell) 0.4 102 3 Tube sheet design 0.6 153 4 Partition design 0.1 25 5 Pipe body material 0.7 178 6 Shell material 0.5 127 7 Tube sheet material 0.9 229 8 Safety valve set pressure 0.8 204 9 Pressure gauge monitoring range 1.0 255 10 Pressure relief device set pressure 0.88 224 11 High-precision machining 0.1 25 12 Welding process 0.8 204 13 Nondestructive Testing 0.9 229 14 Pressure test (water pressure) 1.0 255 15 Pressure test (air pressure) 0.83 211

[0098] The mapping result of the multi-dimensional matrix after mapping is as follows. According to the amount of industrial data information, blank positions can be filled:

[0099] | 51 | 102 | 153 | 25 |

[0100] | 178 | 127 | 229 | 204 |

[0101] | 255 | 224 | 25 | 204 |

[0102] | 229 | 255 | 211 | 0 |

[0103] S04: Create a blank image matrix, and fill the mapping results into the corresponding positions of the image matrix to obtain a visualized image;

[0104] In this step, the information mapping result obtained in step S03 is filled into the corresponding position of the image matrix, thereby generating a visual image, which contains a variety of information of the high-pressure shell and tube heat exchanger; the multi-dimensional matrix mapping and image generation process retains the spatial relationship and feature association of the data, and provides rich context information for subsequent convolutional neural network processing. Specifically, the spatial relationship and feature association specifically include:

[0105] (1) Relative positions between data features

[0106] When the industrial data information is converted into a multidimensional matrix and mapped to the pixel value range of the image, each data feature is placed at a specific position in the image matrix, and these positions reflect the relative relationship between the data features, such as:

[0107] Neighboring features: Placing related data features in adjacent positions in the image matrix helps CNN capture the association between them during convolution operations; Partitioning features: Different types of data features can be distributed in different areas of the image, helping the model distinguish different categories of information.

[0108] (2) Local and global patterns of data

[0109] When processing images, CNN can extract local and global patterns by scanning the images with convolution kernels. After mapping industrial data to images, the convolution kernel can capture the local relationship between adjacent pixels, such as the relationship between wall thickness design and material selection. Through multi-layer convolution and pooling operations, CNN can gradually learn higher-level global patterns, such as the overall trend of comprehensive performance and safety.

[0110] (3) Spatial distribution and correlation of data

[0111] After mapping the data features to the image, each pixel not only represents a single value, but also forms a certain association with the surrounding pixels, that is, other data features. This association is the "spatial relationship". In the actual application of industrial equipment, some data features have an intrinsic correlation. For example, there may be a correlation between wall thickness design and material selection, and this relationship can be represented by adjacent pixels in the image; the distribution pattern of data with different features in the image can also reveal some valuable information, such as the potential correlation between high-voltage test results and welding processes.

[0112] S05: Determine the comprehensive strength label of each industrial data information, which is used to measure the overall performance and quality of the heat exchanger;

[0113] In the above process, steps S04 and S05 can be performed simultaneously, or the order can be switched and completed before building the model;

[0114] S06: Use the visualized image as input to build a convolutional neural network model, use the comprehensive intensity label as the target label of the model, and train the model;

[0115] The convolutional neural network model can be specially designed to process data with a grid structure, such as images. The model extracts features layer by layer through convolutional layers and pooling layers to capture local and global patterns in the data. In this embodiment, the industrial data information of the structural form, material selection, safety device and manufacturing process of the high-pressure shell and tube heat exchanger is encoded into an image form, and then processed by the convolutional neural network model to cope with the influence of different information. The normalized intensity label, such as 0.8875, will be used as the target label of the convolutional neural network model for training the model. In the implementation process, encoding information into an image form and processing it using a convolutional neural network model can realize an automated evaluation process, thereby reducing the need for manual intervention and enabling faster evaluation.

[0116] S07: Evaluate the high-pressure shell and tube heat exchanger through the trained convolutional neural network model and output the evaluation results.

[0117] The present invention realizes a comprehensive, accurate and efficient evaluation of shell and tube heat exchangers for high-pressure applications through systematic data collection, standardized data processing, advanced machine learning technology and automated evaluation process. It not only improves the efficiency and accuracy of the evaluation process, but also enhances the reliability of the evaluation results, providing strong support for the optimization and safety management of industrial equipment.

[0118] Furthermore, the comprehensive strength labels of each industrial data information are determined, including:

[0119] A01: Provide separate scores for each industrial data information;

[0120] A02: Each score is normalized to obtain a value range of 0 to 1;

[0121] A03: Assign weights to each piece of industrial data information, and the comprehensive strength label is the sum of the normalized results of each score multiplied by the corresponding weight.

[0122] During the implementation process, the strength label provides a standardized evaluation method to facilitate comparison and analysis between different heat exchangers. During the implementation process, when scoring each industrial data information separately, assuming that the full score of the following aspects is 100, the scores of each industrial data information are as follows: structural strength: 85, material performance: 90, safety: 88, manufacturing process: 92.

[0123] Each score is normalized to obtain a numerical range of 0~1 in the following specific way: take each score and divide it by 100: structural strength normalization: 85 / 100 = 0.85, material performance normalization: 90 / 100 = 0.90, safety normalization: 88 / 100 = 0.88, manufacturing process normalization: 92 / 100 = 0.92.

[0124] Assuming that the weights of each aspect are equal, that is, each aspect has a weight of 0.25, the calculation of the comprehensive strength label is as follows:

[0125] Combined strength label = (0.85 * 0.25) + (0.90 * 0.25) + (0.88 * 0.25) + (0.92* 0.25) = 0.2125 + 0.225 + 0.22 + 0.23 = 0.8875.

[0126] Based on the above embodiments, Figure 3 As shown, a convolutional neural network model is constructed, including:

[0127] B01: Divide the generated visualization images and the corresponding comprehensive intensity labels into training set, validation set and test set;

[0128] B02: Select a pre-trained model that includes at least three convolutional layers, freeze at least 50% of the convolutional layers, and add a custom fully connected layer on top of the pre-trained model;

[0129] The design of at least three convolutional layers allows shallow, mid-level, and deep features to be extracted. Specifically, shallow features extract low-level features, such as edges and textures, which are very important for capturing the basic structure of an image. Mid-level features extract mid-level features, such as shapes and simple patterns, which help the model understand the content of the image.

[0130] Deep feature extraction is high-level features, such as complex patterns and global information, which help the model make higher-level judgments and classifications. The model fully extracts multi-level features from low-level to high-level, giving the model higher expressiveness and understanding capabilities.

[0131] B03: Compile the pre-trained model by selecting the loss function, optimizer and evaluation metric;

[0132] B04: Use the training set data to train the pre-trained model and verify it on the validation set;

[0133] B05: Unfreeze some convolutional layers of the pre-trained model and adjust the model;

[0134] B06: Evaluate the model performance on the test set and continue to optimize the pre-trained model based on the test results.

[0135] In the implementation process, as a specific method, the data can be divided into training sets (70-80%), validation sets (10-15%) and test sets (10-15%) to ensure that the model has independent data for evaluation during the training, validation and testing stages. For the model, select models pre-trained on large-scale datasets (such as ImageNet), such as VGG16, ResNet50, etc. These models have good feature extraction capabilities, freeze the convolutional layers of the pre-trained model to keep its features learned on large-scale datasets; add a custom fully connected layer on top of the pre-trained model to adapt to specific tasks in a targeted manner.

[0136] In step B03, the loss function may be the mean square error (MSE). As a specific choice, the Adam optimizer may be used, which performs well in most cases due to its adaptive learning rate adjustment capability. Finally, the mean absolute error (MAE) may be selected as an evaluation indicator to measure the difference between the predicted value and the actual value.

[0137] Use the training set data to train the model, and verify it on the validation set to observe the performance of the model on unseen data, so as to adjust the model parameters and prevent overfitting. Specifically, you can set parameters such as the number of training rounds (epochs) and batch size (batch size). After the initial training, unfreeze some convolutional layers of the pre-trained model to allow it to be fine-tuned on the new data set, thereby improving the adaptability and accuracy of the model. After unfreezing, the model needs to be recompiled and fine-tuned using a lower learning rate to prevent drastic adjustments to the learned weights.

[0138] Furthermore, some of the convolutional layers that are unfrozen are the convolutional layers near the top.

[0139] This part of the convolutional layer extracts high-level features, such as complex patterns and global information, and different tasks have different requirements for these high-level features. Unfreezing the convolutional layers near the top allows the model to adjust according to the data characteristics of the new task, so as to better adapt to the specific task. For example, in the evaluation of high-pressure shell and tube heat exchangers, unfreezing the top convolutional layer enables the model to better understand and capture the high-level features of specific industrial data.

[0140] As a preferred embodiment of this embodiment, Figure 4 As shown in the figure, two convolutional neural network model structures are constructed. The one on the left side of the figure includes:

[0141] Input layer, receiving multi-dimensional data of visualization images;

[0142] The convolution layer is used to extract local features from the visualized image and output a feature map. It performs filtering operations through several convolution kernels, and each convolution kernel learns a different feature response mode. The convolution operation can effectively capture the spatial relationship in the visualized image through the convolution kernel sliding on the input data, thereby extracting important features such as edges and textures. The setting of the convolution layer can select the appropriate convolution kernel size and number according to the characteristics of the visualized image data.

[0143] The activation layer introduces nonlinearity after the convolution layer, which increases the expressive power of the convolutional neural network and enables it to learn more complex feature maps. The activation function can help the convolutional neural network learn nonlinear relationships, so that it can better fit complex data. For example, activation functions such as ReLU, Sigmoid, and Tanh can usually be followed by an activation function layer after each convolution layer to perform nonlinear transformation on the results of the convolution operation.

[0144] The maximum pooling layer reduces the computational complexity by reducing the spatial dimension of the feature map, while extracting the main features and enhancing the receptive field and invariance of the convolutional neural network. The main function of the maximum pooling is to downsample the feature map. By taking the maximum value of each local area to reduce the spatial dimension of the feature map, the amount of data and the amount of computation can be reduced, while retaining important features and enhancing the receptive field and invariance of the convolutional neural network. Specifically, the receptive field refers to the size of the response area of ​​a specific neuron in a certain layer of the neural network to the input data, including the local receptive field and the global receptive field. Expanding the receptive field can help the model understand contextual information in a larger range, which is especially important for processing input data with large size or global features. Invariance refers to the robustness and stability of the neural network to certain changes in the input data, including translation invariance, scale invariance and rotation invariance. The enhancement of invariance enables the neural network to more effectively process real-world data with diversity and variability, and enhances the generalization ability and robustness of the model. It should be noted that the setting of the maximum pooling layer can reasonably select the pooling kernel size and step size to ensure the effectiveness of feature extraction and the appropriate size of the feature map.

[0145] The flattening layer flattens the multi-dimensional feature map into a one-dimensional vector. Since the output of the convolutional layer is a multi-dimensional feature map, and the input of the fully connected layer requires a one-dimensional vector, a flattening layer can be added between the convolutional part and the fully connected part. The flattening layer is a transition layer connecting the convolutional part and the fully connected part. The features extracted by the convolutional layer can be converted into a one-dimensional vector to provide a flattened input for the fully connected layer. However, it is necessary to ensure that the size of the flattened vector matches the input size of the fully connected layer.

[0146] Fully connected layer, the neurons in the fully connected layer are connected to all the neurons in the previous layer, which is used for high-level feature learning; the fully connected layer is also called a densely connected layer, whose neurons are connected to all the neurons in the previous layer, which can learn the complex relationship between different features and further improve the expressive power of the model. It may increase the number of parameters and computational complexity of the model, and can be properly regularized and optimized.

[0147] Output layer,The output layer is a fully connected layer that outputs the evaluation results through a single neuron.

[0148] like Figure 4 As shown on the right side of the figure, another convolutional neural network model structure is constructed. Compared with the above convolutional neural network model structure, the following changes are made:

[0149] Keep the input layer, convolution layer, activation layer, fully connected layer and output layer in the convolutional neural network model structure;

[0150] A residual block is added after the convolutional layer, the flattening layer is removed, and the maximum pooling layer is modified to a global average pooling layer.

[0151] Through the above changes, the final model structure is:

[0152] Input layer, receiving visualization images;

[0153] The convolution layer is used to extract local features from the visualized image and output a feature map. The filter operation is performed through several convolution kernels, and each convolution kernel learns a different feature response mode.

[0154] The residual block has several convolutional layers and introduces skip connections for convolutional neural networks;

[0155] Activation layer, which introduces nonlinearity after the convolutional layer;

[0156] A global average pooling layer that converts the feature map into a single output tensor;

[0157] Several fully connected layers, where neurons are connected to all neurons in the previous layer for high-level feature learning;

[0158] Output layer,The output layer is a fully connected layer that outputs the evaluation results through a single neuron.

[0159] On the basis of the above-mentioned embodiment, other implementation methods are obtained by changing the convolutional neural network model structure constructed by the above-mentioned changes, retaining the input layer, convolution layer, activation layer, fully connected layer and output layer in the convolutional neural network model structure, adding a residual block after the convolution layer, and each residual block contains multiple convolution layers and jump connections for extracting and propagating features. The gradient vanishing problem in deep networks can be solved by jump connections, allowing information to flow more freely in the network, and adding the input directly to the output of the last convolution layer, passing the original input information through identity mapping, which helps to reduce the gradient vanishing problem. Multiple residual blocks are reused in the network, and each residual block can learn a specific level of feature representation. By modifying the maximum pooling layer to a global average pooling layer and performing global average pooling on the spatial dimension of each feature map, the number of parameters can be reduced while maintaining the global information of the feature map. The function of the global average pooling layer is to perform average pooling on the entire feature map, taking the average of all elements of the feature map of each channel to obtain a feature vector. The dimension of this feature vector is usually the same as the number of channels, retaining the important feature information in each channel and reducing the complexity of the model by reducing the spatial dimension. At the same time, the global average pooling layer can convert the convolutional feature map into a one-dimensional vector so that it can be connected to the fully connected layer. Due to the existence of the global average pooling layer, the flattening layer is no longer needed to map the multi-dimensional features into a one-dimensional vector, and the flattening layer can be removed.

[0160] Furthermore, the constructed convolutional neural network model structure also includes adding a dropout layer between the fully connected layers to randomly shut down a number of neurons.

[0161] Specifically, the dropout layer is a regularization technique used to reduce overfitting. It is applied after the fully connected layer of a deep neural network. The dropout layer randomly turns off some neurons, forcing the neural network to learn more robust features, thereby reducing the degree of overfitting of the model on the training set, making the network more flexible, and reducing the dependence between neurons, thereby improving the generalization ability of the model on new data. During model training, the dropout layer will randomly select a part of neurons with a certain probability and set their outputs to zero. These ignored neurons will not participate in the calculation of the current iteration; in order to keep the expected value of the output unchanged, the outputs of the remaining neurons that are not ignored will be appropriately scaled during the training process; it should be noted that the dropout layer only works in the training phase of the model, and will be disabled in the test phase. No dropout operation will be performed, but the outputs of all neurons will be involved in the calculation, but without scaling.

[0162] Furthermore, the pre-trained model is trained using the training set data and verified on the validation set; including:

[0163] C01: Load the weights and bias parameters of the pre-trained model; the parameters of the pre-trained model have been adjusted and optimized through a large amount of data in the previous training process, have good initial values, and can effectively extract image features. Since the newly added layer structure is designed to adapt to new tasks, it can be randomly initialized.

[0164] C02: Set the training rounds and perform cyclic training on the convolutional neural network model through the back propagation algorithm;

[0165] Specifically, when training a model, you can first define hyperparameters such as the batch size, training rounds, and learning rate during training. These parameters will affect the convergence speed of the model during training and the final performance. The training data set is divided into small batches, each batch contains a certain number of sample data. In each training round, the training algorithm will loop through all batches, and before each training round, the training data can be randomly rearranged to ensure that the model does not learn unnecessary rules due to deviations in the data order. For each training round, forward propagation is performed in the model to calculate the model prediction output and calculate the loss value using the selected loss function; then backpropagation is performed to calculate the gradient of the loss function with respect to the model parameters based on the loss value, and the selected optimizer is used to update the model weight parameters according to the gradient to reduce the loss value.

[0166] C03: Monitor the cyclic training process. After each training round, use the validation set to validate the convolutional neural network model. Adjust the model parameters according to the performance of the convolutional neural network model on the validation set until the model training termination conditions are met.

[0167] This step monitors the training set loss and possible evaluation indicators to evaluate the performance of the model on the training data, and after each training round, uses the validation data set to evaluate the performance of the model, including calculating the loss and evaluation indicators on the validation set, such as accuracy, precision, etc. Finally, based on the performance of the validation set and the monitoring results during the training process, the model's hyperparameters, such as learning rate, regularization parameters, etc., or the model structure, such as adding or removing layers, can be adjusted.

[0168] Furthermore, it also includes setting an early stopping strategy and determining whether to start the early stopping strategy based on the performance of the convolutional neural network model on the validation set.

[0169] In this embodiment, the early stopping strategy refers to a strategy in the training process of the machine learning model, which is used to avoid overfitting of the model and improve the generalization ability. First, an evaluation indicator that performs well on the validation set can be selected as the monitoring indicator for early stopping, and an early stopping condition can be defined. For example, if the monitoring indicator has not improved in several consecutive training rounds, early stopping is triggered. At the end of each training round, the monitoring indicators on the validation set are calculated and recorded to check whether the early stopping conditions are met. Once the early stopping conditions are triggered, that is, the performance on the validation set has not improved further, the training process is stopped, which means that the performance of the model on the validation set has reached a relatively good level, avoiding the risk of overfitting caused by continued training.

[0170] Evaluation system for shell and tube heat exchangers for high pressure applications based on industrial big data, such as Figure 5 As shown, including:

[0171] Industrial data collection module, which collects industrial data information on the structure, material selection, safety devices and manufacturing process of high-pressure shell and tube heat exchangers;

[0172] The industrial data processing module normalizes the industrial data information to map the value ranges of different features to the same interval;

[0173] The industrial data mapping module combines the normalized industrial data information into a multi-dimensional matrix, and maps the multi-dimensional matrix into the pixel value range of the image to obtain a mapping result;

[0174] A visualization image generation module creates a blank image matrix and fills the mapping results into the corresponding positions of the image matrix to obtain a visualization image;

[0175] Comprehensive strength label determination module, which determines the comprehensive strength label of each industrial data information. The comprehensive strength label is used to measure the overall performance and quality of the heat exchanger;

[0176] The convolutional neural network building module takes the visualization image as input and builds the convolutional neural network model. It uses the comprehensive intensity label as the target label of the model to train the model.

[0177] The convolutional neural network application module evaluates the high-pressure shell and tube heat exchanger through the trained convolutional neural network model and outputs the evaluation results.

[0178] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. An evaluation method for high-pressure shell and tube heat exchangers based on industrial big data, characterized in that: include: Collect industrial data information on the structure, material selection, safety devices and manufacturing process of shell and tube heat exchangers for high pressure applications; Normalizing the industrial data information to map the value ranges of different features to the same interval; Combining the normalized industrial data information into a multi-dimensional matrix, and mapping the multi-dimensional matrix into a pixel value range of an image to obtain a mapping result; Creating a blank image matrix, and filling the mapping results into corresponding positions of the image matrix to obtain a visualized image; The multi-dimensional matrix mapping and image generation process preserves the spatial relationship and feature association of the data, which includes: the relative position between data features, the local and global patterns of the data, and the spatial distribution and association of the data; Determine the comprehensive strength label of each industrial data information, which is used to measure the overall performance and quality of the heat exchanger, including: Providing separate scores for each of the industrial data information; Each score is normalized to obtain a numerical range of 0 to 1; Assigning a weight to each of the industrial data information, the comprehensive strength label being the sum of the normalized result of each of the scores multiplied by the corresponding weight; Taking the visualized image as input, constructing a convolutional neural network model, and using the comprehensive intensity label as the target label of the model to train the model; The high-pressure shell and tube heat exchanger is evaluated by using the trained convolutional neural network model, and the evaluation result is output.

2. The high-pressure shell and tube heat exchanger evaluation method based on industrial big data according to claim 1 is characterized in that: Build a convolutional neural network model, including: Dividing the generated visualization images and the corresponding comprehensive intensity labels into a training set, a validation set and a test set; Selecting a pre-trained model, the pre-trained model comprising at least three convolutional layers, freezing at least 50% of the convolutional layers, and adding a custom fully connected layer on top of the pre-trained model; Compiling the pre-trained model is achieved by selecting a loss function, an optimizer, and an evaluation metric; Using the training set to train the pre-trained model, and verifying it on the verification set; Unfreeze some convolutional layers of the pre-trained model and adjust the model; The model performance is evaluated on the test set, and the optimization of the pre-trained model is continued based on the test results.

3. The high-pressure application shell and tube heat exchanger evaluation method based on industrial big data according to claim 2 is characterized in that: The unfrozen part of the convolutional layers is the convolutional layer close to the top.

4. The method for evaluating high-pressure shell and tube heat exchangers based on industrial big data according to claim 2 is characterized in that: The structure of the convolutional neural network model constructed includes: An input layer, receiving the visualization image; A convolution layer, used for extracting local features from the visualized image and outputting a feature map, performing a filtering operation through a plurality of convolution kernels, each of which learns a different feature response mode; An activation layer that introduces nonlinearity after the convolutional layer; A maximum pooling layer to reduce the spatial dimension of the feature map and extract key features; A flattening layer flattens the multi-dimensional feature map into a one-dimensional vector; Several fully connected layers, where neurons in the fully connected layers are connected to all neurons in the previous layer for high-level feature learning; The output layer is a fully connected layer that outputs the evaluation result through a single neuron.

5. The high-pressure application shell and tube heat exchanger evaluation method based on industrial big data according to claim 2 is characterized in that: The constructed convolutional neural network model structure includes: An input layer, receiving the visualization image; A convolution layer, used for extracting local features from the visualized image and outputting a feature map, performing a filtering operation through a plurality of convolution kernels, each of which learns a different feature response mode; A residual block having several of the convolutional layers, introducing skip connections for the convolutional neural network; An activation layer that introduces nonlinearity after the convolutional layer; A global average pooling layer that converts the feature maps into a single output tensor; Several fully connected layers, where neurons in the fully connected layers are connected to all neurons in the previous layer for high-level feature learning; The output layer is a fully connected layer that outputs the evaluation result through a single neuron.

6. The method for evaluating high-pressure shell and tube heat exchangers based on industrial big data according to claim 4 or 5, characterized in that: The constructed convolutional neural network model structure also includes adding a dropout layer between the fully connected layers to randomly shut down a number of neurons.

7. The method for evaluating high-pressure shell and tube heat exchangers based on industrial big data according to claim 2, characterized in that: The pre-trained model is trained using the training set and verified on the verification set; including: Loading the weight and bias parameters of the pre-trained model; Setting training rounds, and performing cyclic training on the convolutional neural network model through a back propagation algorithm; Monitor the process of the cyclic training, and after each training round, use the verification set to verify the convolutional neural network model, and adjust the model parameters according to the performance of the convolutional neural network model on the verification set until the model training termination condition is reached.

8. The method for evaluating high-pressure shell and tube heat exchangers based on industrial big data according to claim 7 is characterized in that: It also includes setting an early stopping strategy, and determining whether to start the early stopping strategy based on the performance of the convolutional neural network model on the validation set.

9. A high-pressure shell and tube heat exchanger evaluation system based on industrial big data, using a high-pressure shell and tube heat exchanger evaluation method based on industrial big data as claimed in claim 1, characterized in that: include: Industrial data collection module, which collects industrial data information on the structure, material selection, safety devices and manufacturing process of high-pressure shell and tube heat exchangers; An industrial data processing module normalizes the industrial data information to map the value ranges of different features to the same interval; An industrial data mapping module combines the normalized industrial data information into a multi-dimensional matrix, and maps the multi-dimensional matrix into a pixel value range of an image to obtain a mapping result; A visualization image generation module creates a blank image matrix and fills the mapping results into corresponding positions of the image matrix to obtain a visualization image; A comprehensive strength label determination module determines the comprehensive strength label of each industrial data information, wherein the comprehensive strength label is used to measure the overall performance and quality of the heat exchanger; A convolutional neural network building module takes the visualization image as input to build a convolutional neural network model, and uses the comprehensive intensity label as a target label of the model to train the model; The convolutional neural network application module evaluates the high-pressure shell and tube heat exchanger through the trained convolutional neural network model and outputs the evaluation result.

Images