Martensite heat-resistant steel service state intelligent evaluation method based on multi-source feature fusion and deep learning

Abstract

The invention relates to the technical field of martensitic heat-resistant steel service state evaluation, in particular to a martensitic heat-resistant steel service state intelligent evaluation method based on multi-source feature fusion and deep learning. The method comprises the following steps: firstly, acquiring temperature and stress data in a pipeline operation process, acquiring a microscopic structure image through an optical microscope and an electron microscope, and testing hardness and tensile property data through a hardness meter and a tensile testing machine; then, operation is carried out according to the steps of data preprocessing, feature extraction, feature fusion, model construction and training, service state evaluation and the like in sequence, and finally intelligent evaluation of the service state of the martensitic heat-resistant steel in the high-temperature pipeline is achieved. According to the method, the multi-source characteristic information of the martensitic heat-resistant steel can be fully utilized, and the service state evaluation precision and efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of martensitic heat-resistant steel service status assessment, and specifically to an intelligent assessment method for the service status of martensitic heat-resistant steel based on multi-source feature fusion and deep learning, which is suitable for safety monitoring of high-temperature equipment in the fields of electricity, energy, etc. Background Art

[0002] Martensitic heat-resistant steel, due to its excellent high-temperature strength and oxidation resistance, is widely used in high-temperature equipment in the power and energy sectors. However, over long-term service, exposed to complex environments such as high temperature and stress, martensitic heat-resistant steel can undergo microstructural evolution and performance degradation, seriously impacting the safe and reliable operation of the equipment. Therefore, accurately assessing the service condition of martensitic heat-resistant steel is of great engineering significance.

[0003] Patent publication number CN109308486A proposes a multi-source image fusion and feature extraction algorithm based on deep learning. Although deep learning is used to fuse and extract features from multi-source images, it mainly focuses on multi-view classification of three-dimensional models, and is concerned with how to select the optimal view from different views to improve target classification accuracy.

[0004] Patent publication number CN106295714A proposes a deep learning-based multi-source remote sensing image fusion method, utilizing a deep convolutional neural network to automatically select and fuse features from multi-source remote sensing images to improve remote sensing image classification accuracy. However, this method is primarily applicable to the remote sensing field, which differs from the actual needs of assessing the service condition of martensitic heat-resistant steel. Furthermore, its feature fusion approach may not fully utilize the multi-source feature information of martensitic heat-resistant steel.

[0005] Existing methods for assessing the service condition of martensitic heat-resistant steel primarily include traditional mechanical property testing, microstructural observation, and single-feature-based machine learning. These methods suffer from insufficient utilization of feature information, low assessment accuracy, and low efficiency, making them difficult to meet the practical engineering requirements for rapid and accurate assessment of the service condition of martensitic heat-resistant steel. Summary of the Invention

[0006] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide an intelligent evaluation method for the service status of martensitic heat-resistant steel based on multi-source feature fusion and deep learning, so as to make full use of the multi-source feature information of martensitic heat-resistant steel and improve the accuracy and efficiency of service status evaluation.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] An intelligent assessment method for the service condition of martensitic heat-resistant steel based on multi-source feature fusion and deep learning includes the following steps:

[0009] Step 1: Data collection: Collect multi-source data of martensitic heat-resistant steel during service, including temperature, stress, strain, microstructure image, hardness, and tensile property data;

[0010] Step 2: Data preprocessing: Preprocess the collected multi-source data, including data cleaning, noise reduction, and normalization operations to improve data quality and usability;

[0011] Step 3: Feature extraction: Extract features from the preprocessed multi-source data separately, as follows:

[0012] (1) For the time series data of temperature, stress, and strain, the long short-term memory network (LSTM) is used for feature extraction to obtain the temporal features in the time series data;

[0013] (2) For microstructure image data, convolutional neural network (CNN) is used for feature extraction to obtain spatial features in microstructure images;

[0014] (3) Static data of hardness and tensile properties are directly used as original features;

[0015] Step 4: Feature fusion: The extracted temporal features, spatial features and original features are fused to form a multi-source fusion feature vector; feature fusion adopts early fusion, late fusion or intermediate fusion;

[0016] Step 5: Model construction and training: Build a deep learning-based service condition assessment model, which includes an input layer, a hidden layer, and an output layer. The input layer is a multi-source fusion feature vector, the hidden layer uses a multi-layer fully connected neural network, and the output layer is the service condition category. The model is trained using labeled martensitic heat-resistant steel service condition data, and the model parameters are optimized to enable the model to accurately map the multi-source fusion feature vector to the corresponding service condition category.

[0017] Step 6: Service status evaluation: After preprocessing and feature extraction, the multi-source data of the martensitic heat-resistant steel to be evaluated are input into the trained model, and the model outputs the service status category of the martensitic heat-resistant steel to be evaluated.

[0018] The intelligent assessment method for the service condition of martensitic heat-resistant steel based on multi-source feature fusion and deep learning, in step 2, for temperature data, the Z-score normalization method is used to convert the temperature data into a value with a mean of 0 and a standard deviation of 1; for microstructure image data, grayscale processing and normalization processing are performed.

[0019] In the intelligent assessment method for the service condition of martensitic heat-resistant steel based on multi-source feature fusion and deep learning, in step four, it is preferred to adopt an intermediate fusion method, that is, feature fusion is performed in the intermediate layers of different feature extraction networks to fully utilize the intermediate layer information of each feature extraction network.

[0020] The intelligent assessment method for the service condition of martensitic heat-resistant steel based on multi-source feature fusion and deep learning, wherein the intermediate fusion is specifically as follows: the last hidden layer output of the LSTM network, the last convolutional layer output of the CNN network and the static data are spliced to form a fused feature vector.

[0021] In the intelligent assessment method for the service condition of martensitic heat-resistant steel based on multi-source feature fusion and deep learning, in step five, the hidden layer adopts a 3-layer fully connected neural network with 128, 64, and 32 nodes respectively.

[0022] The design idea of the present invention is:

[0023] Existing technologies often focus on a single data source (such as microstructure or mechanical properties) or are based on traditional models (such as physical formulas and single machine learning). This invention integrates three types of multi-source information: time series data (temperature, stress), image data (microstructure), and static data (hardness). It also uses the LSTM+CNN deep learning architecture for feature fusion, achieving collaborative innovation among multi-source data fusion, intermediate feature interaction, and deep learning models.

[0024] (1) Multi-source data fusion: Integrate time series data (temperature, stress, strain), microstructure image data, and static performance data (hardness, tensile strength) to cover the full-dimensional information of "environment-structure-performance" of the material's service status.

[0025] (2) Intermediate feature interaction: Feature splicing is performed in the intermediate layers of LSTM and CNN to fully utilize the collaborative information of temporal features (such as temperature change trends) and spatial features (such as microstructure evolution).

[0026] (3) Deep learning model: A full-process intelligent model from data collection to evaluation was constructed, which realized the automatic classification of service status and significantly improved the evaluation efficiency.

[0027] The present invention collects multi-source data of martensitic heat-resistant steel, uses long short-term memory network (LSTM) and convolutional neural network (CNN) to respectively extract time series data, microstructure image data, time series features, spatial features, etc., and combines the original features of static data. It adopts intermediate fusion to perform feature fusion and constructs a service status assessment model based on deep learning. This model can make more full use of the multi-source feature information of martensitic heat-resistant steel, effectively improve the accuracy and efficiency of service status assessment, and solve the problems of insufficient feature utilization and low assessment efficiency in the prior art.

[0028] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0029] 1. The present invention collects multi-source data of martensitic heat-resistant steel, adopts deep learning methods for feature extraction and fusion, and constructs a deep learning evaluation model based on multi-source feature fusion. It can fully utilize the multi-source feature information of martensitic heat-resistant steel, improve the accuracy and efficiency of service status evaluation, and provide a strong guarantee for the safe and reliable operation of martensitic heat-resistant steel.

[0030] 2. The present invention integrates and optimizes feature interaction in the middle layer of the feature extraction network, and achieves a more comprehensive state assessment through the complementarity of multi-source data (such as microstructure images reflecting long-term damage and time series data capturing dynamic changes). Compared with the single feature method, the assessment error is reduced, and service data can be processed in real time to meet the rapid assessment needs of engineering scenarios, which is superior to traditional destructive testing. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a flow chart of steps 1 to 4 of the intelligent evaluation method for the service condition of martensitic heat-resistant steel in an embodiment.

[0032] Figure 2 This is a flow chart of steps five and six of the intelligent evaluation method for the service condition of martensitic heat-resistant steel in an embodiment. DETAILED DESCRIPTION

[0033] The following is a clear and complete description of the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0034] Example

[0035] The high-temperature pipelines of thermal power plants are exposed to complex working conditions such as high temperature and high pressure for a long time. Martensitic heat-resistant steel is its main material, and its service status is crucial to the safe and stable operation of the power plant.

[0036] like Figure 1-Figure 2As shown, this embodiment provides an intelligent evaluation method for the service condition of martensitic heat-resistant steel based on multi-source feature fusion and deep learning, comprising the following steps:

[0037] Step 1: Data collection: Through sensors, microscopes, material testing machines and other equipment, collect time series data such as temperature, stress, strain, microstructure image data, and static data such as hardness and tensile properties of martensitic heat-resistant steel during service.

[0038] For example, temperature sensors and stress sensors are installed on the high-temperature pipelines of a thermal power plant to regularly collect temperature and stress data during pipeline operation. At the same time, small samples of the pipelines are regularly cut, and microstructure images are obtained using optical microscopes and electron microscopes. The hardness and tensile performance data are tested using a hardness tester and a tensile testing machine.

[0039] (1) Sensor installation and data acquisition: Thermocouple temperature sensors and strain gauge stress sensors are installed at key locations on the high-temperature pipeline. Temperature and stress data are collected every 10 minutes. The signals are transmitted to a computer via a data acquisition card and stored as time series data files. At the same time, a fiber optic strain sensor installed on the pipeline monitors the strain in real time, and the data is also transmitted to the computer.

[0040] (2) Microstructure image acquisition: Small samples of the pipe with a thickness of approximately 10 mm were regularly taken (every 3000 hours of operation). The microstructure morphology was observed using an optical microscope at a magnification of 500 times, and microstructure images were taken. The sample surface was then scanned using a scanning electron microscope to obtain a higher-resolution microstructure image with a resolution of 1024 × 1024 pixels.

[0041] (3) Hardness and tensile properties test: The cut samples were processed into standard hardness test blocks and tensile specimens. Hardness was tested using a hardness tester according to national standards. Each sample was tested at 5 points and the average value was taken as the hardness value of the sample. Tensile tests were performed on the tensile specimens using a universal material testing machine to obtain performance indicators such as tensile strength.

[0042] Step 2: Data preprocessing: Clean the collected multi-source data to remove outliers and missing values. Use filtering algorithms, such as Gaussian filtering, to reduce noise in time series data. Normalize all data to convert them to values within the range [0, 1] to improve data comparability and model training efficiency. For example, for temperature data, use the Z-score normalization method to convert the temperature data to values with a mean of 0 and a standard deviation of 1. For microstructure image data, grayscale processing and normalization are performed.

[0043] Step 3: Feature extraction:

[0044] 1. For time series data such as temperature, stress, and strain, construct an LSTM network model. Input the time series data into the LSTM network, and extract time series features through the hidden layer of the LSTM network. The LSTM network structure consists of an input layer, several LSTM layers, and an output layer. The number of nodes in the input layer corresponds to the characteristic dimension of the time series data. The number of nodes in the LSTM layer is set based on the complexity of the data. The number of nodes in the output layer corresponds to the dimension of the extracted time series features. For example, setting the number of input layer nodes to 3 (temperature, stress, strain), the number of LSTM layer nodes to 64, and the number of output layer nodes to 32 will extract 32-dimensional time series features through the LSTM network.

[0045] 2. For microstructure image data, a CNN network model is constructed. The preprocessed microstructure image is input into the CNN network, and spatial features are extracted through the CNN's convolutional and pooling layers. The CNN network structure consists of an input layer, several convolutional layers, a pooling layer, and a fully connected layer. The number of nodes in the input layer is equal to the number of pixels in the image. The convolutional layers use convolution kernels of different sizes to extract different image features. The pooling layer is used to reduce the feature dimensionality. The number of nodes in the fully connected layer is equal to the dimension of the extracted spatial features. For example, a microstructure image is resized to 224×224 pixels and input into a CNN network consisting of three convolutional layers and two pooling layers. Finally, a fully connected layer is used to extract 64-dimensional spatial features.

[0046] 3. For static data such as hardness and tensile properties, they are directly used as original features, such as hardness values and tensile strength values.

[0047] Step 4: Feature fusion:

[0048] Using an intermediate fusion approach, the temporal features extracted by the LSTM network, the spatial features extracted by the CNN network, and the static raw features are fused in the fully connected layer. Specifically, the temporal features, spatial features, and raw features are input into different fully connected layers, and then the outputs of each fully connected layer are concatenated to form a multi-source fused feature vector. For example, 32-dimensional temporal features, 64-dimensional spatial features, and 2-dimensional raw features (hardness and tensile strength) are concatenated into a 98-dimensional multi-source fused feature vector.

[0049] Step 5: Model construction and training:

[0050] A deep learning-based service condition assessment model was constructed. The model's input layer consists of a 98-dimensional multi-source fusion feature vector. The hidden layer utilizes a three-layer fully connected neural network with 128, 64, and 32 nodes, respectively. The output layer consists of three nodes, corresponding to the three service conditions of martensitic heat-resistant steel: good, fair, and poor. The model was trained using the cross-entropy loss function and the Adam optimization algorithm. The training data consisted of labeled martensitic heat-resistant steel service condition data, with 80% of the training set and 20% of the validation set. During training, the model's accuracy on the validation set was regularly evaluated. Training was terminated when accuracy stopped improving, and the optimal model was saved.

[0051] Step 6: Service Status Assessment:

[0052] The multi-source data for the martensitic heat-resistant steel to be evaluated is preprocessed and feature extracted according to the above steps to obtain a multi-source fusion feature vector. This vector is then fed into the trained model, which then outputs the service condition category of the martensitic heat-resistant steel to be evaluated. For example, for a martensitic heat-resistant steel sample to be evaluated, temperature, stress, and strain data, microstructure image data, and hardness and tensile properties data are collected. After preprocessing and feature extraction, a multi-source fusion feature vector is obtained. After input into the model, the model outputs a service condition category of "good," "fair," or "poor."

[0053] Implementation results show that the present invention first collects temperature and stress data during pipeline operation, obtains microstructural images using optical and electron microscopes, and measures hardness and tensile properties using a hardness tester and a tensile testing machine. It then sequentially follows the aforementioned steps of data preprocessing, feature extraction, feature fusion, model construction and training, and service condition assessment, ultimately achieving intelligent assessment of the service condition of martensitic heat-resistant steel in high-temperature pipelines.

[0054] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent assessment method for the service condition of martensitic heat-resistant steel based on multi-source feature fusion and deep learning, characterized in that: The following steps are involved: Step 1: Data collection: Collect multi-source data of martensitic heat-resistant steel during service, including temperature, stress, strain, microstructure image, hardness, and tensile property data; Step 2: Data preprocessing: Preprocess the collected multi-source data, including data cleaning, noise reduction, and normalization operations to improve data quality and usability; Step 3: Feature extraction: Extract features from the preprocessed multi-source data separately, as follows: (1) For the time series data of temperature, stress, and strain, the long short-term memory network (LSTM) is used for feature extraction to obtain the temporal features in the time series data; (2) For microstructure image data, convolutional neural network (CNN) is used for feature extraction to obtain spatial features in microstructure images; (3) Static data of hardness and tensile properties are directly used as original features; Step 4: Feature fusion: Fuse the extracted temporal features, spatial features and original features to form a multi-source fusion feature vector; Fusion can be done in the form of early fusion, late fusion or intermediate fusion; Step 5: Model construction and training: Build a deep learning-based service condition assessment model, which includes an input layer, a hidden layer, and an output layer. The input layer is a multi-source fusion feature vector, the hidden layer uses a multi-layer fully connected neural network, and the output layer is the service condition category. The model is trained using labeled martensitic heat-resistant steel service condition data, and the model parameters are optimized to enable the model to accurately map the multi-source fusion feature vector to the corresponding service condition category. Step 6: Service status evaluation: After preprocessing and feature extraction, the multi-source data of the martensitic heat-resistant steel to be evaluated are input into the trained model, and the model outputs the service status category of the martensitic heat-resistant steel to be evaluated.

2. The intelligent assessment method for the service condition of martensitic heat-resistant steel based on multi-source feature fusion and deep learning according to claim 1, characterized in that: In step 2, the temperature data is converted into a value with a mean of 0 and a standard deviation of 1 using the Z-score normalization method; the microstructure image data is grayscaled and normalized.

3. The intelligent assessment method for the service condition of martensitic heat-resistant steel based on multi-source feature fusion and deep learning according to claim 1, characterized in that: In step 4, it is preferred to adopt an intermediate fusion method, that is, to perform feature fusion in the intermediate layers of different feature extraction networks to fully utilize the intermediate layer information of each feature extraction network.

4. The intelligent assessment method for the service condition of martensitic heat-resistant steel based on multi-source feature fusion and deep learning according to claim 3, characterized in that: The specific intermediate fusion is: concatenating the last hidden layer output of the LSTM network, the last convolutional layer output of the CNN network and the static data to form a fused feature vector.

5. The intelligent assessment method for the service condition of martensitic heat-resistant steel based on multi-source feature fusion and deep learning according to claim 1, characterized in that: In step 5, the hidden layer uses a 3-layer fully connected neural network with 128, 64, and 32 nodes respectively.

Images