Steam pipeline state monitoring method and system based on machine learning

Abstract

The invention discloses a steam pipeline state monitoring method and system based on machine learning, relates to the technical field of machine learning, and is used for accurately identifying an abnormal state of a steam pipeline. The method is technically characterized by comprising the steps of extracting time-frequency characteristics of signals through a wavelet convolution layer, capturing global frequency-domain characteristics through Fourier branches, extracting local time characteristics through convolution branches, performing anomaly detection in combination with an attention mechanism and an Isolation Forest model, and finally performing classification through a LightGBM model to realize abnormal state monitoring of the steam pipeline. According to the method, the problem of high false alarm rate in the traditional technology is solved, the vibration state of the steam pipeline can be accurately identified and classified, the accuracy, noise immunity and calculation efficiency of anomaly detection can be effectively improved, the robustness of the whole system is improved, and the method has a wide application prospect.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a steam pipeline status monitoring method and system based on machine learning. Background Art

[0002] Steam pipes are core components in thermal power generation, used to transport high-temperature and high-pressure steam generated by boilers to steam turbines to drive generator sets to generate electricity. The normal operation of steam pipes is essential to ensure system efficiency and reliability. With the development of high-temperature steam power generation technology, it is particularly important to monitor the operating status of steam pipes, especially to prevent problems such as condensate accumulation and water hammer, which may lead to energy loss or equipment damage, and in severe cases, cause safety accidents.

[0003] During the operation of the steam pipeline, its vibration state is closely related to different working conditions, especially after startup, shutdown, condensate accumulation and water hammer. After startup, due to the rapid increase in pressure and temperature, the pipeline will vibrate due to thermal expansion and contraction; after shutdown, the sharp drop in temperature and pressure will also cause vibration. Condensate accumulation will cause unstable steam flow and vibration, while water hammer is caused by condensate hitting the pipe wall at high speed, causing violent vibration.

[0004] However, steam pipes face many challenges in industrial operation, which affect the efficiency, reliability and safety of the system. For example, condensate will accumulate in the steam pipe. This condensate will form an insulating layer, reduce heat transfer efficiency, affect the heating process, cause energy loss, and may even cause water hammer, generate strong vibration and noise, and then damage pipes and equipment, affect system stability, and hinder normal industrial production.

[0005] In order to meet these challenges, it is particularly important to adopt advanced monitoring technology. By real-time monitoring of pipeline vibration status, it is possible to quickly identify and locate the problem areas that affect the operation of the unit, and understand the operating status of the pipeline in real time, including but not limited to the accumulation of condensate and the occurrence of water hammer.

[0006] Through these data, the crew can quickly monitor whether there is condensate accumulation or water hammer in the pipeline, determine whether the pipeline is operating normally under these conditions, and take corresponding measures in a timely manner, thereby effectively reducing property losses caused by these problems, ensuring production safety, and reducing the probability of safety accidents. In addition, the crew can use this technology to more effectively repair and maintain the production equipment itself, extending the service life of the equipment and reducing the loss of production efficiency caused by frequent maintenance and the high cost caused by frequent equipment replacement.

[0007] In traditional algorithms, the steam pipe vibration identification method is limited by external environmental disturbances, human disturbances and the algorithm itself. The existence of these factors leads to large errors and high false alarm rates when using traditional algorithms to monitor the vibration status of steam pipes in industrial production, which will affect the accuracy and timeliness of steam pipe monitoring.

[0008] Most of the traditional algorithms are based on time domain, frequency domain and time-frequency domain methods to analyze the vibration state of the steam pipeline, so as to judge the operating state of the steam pipeline. These parameters are difficult to adapt to the complex actual environment generated by the operation of the thermal power generator. Therefore, improving the robustness and adaptability of the algorithm to the environment is the key to enhancing the effect of steam pipeline vibration monitoring. Although the feature-based monitoring method can improve the recognition rate of abnormal steam pipeline conditions to a certain extent, the large amount of vibration signal data will lead to excessive workload. It not only requires the maintenance team to have mature experience in it, but also the result has a high false alarm rate, which requires a lot of manpower and material resources to be invested in it, limiting the application of the algorithm to abnormal steam pipeline state monitoring. Traditional monitoring methods often face the problems of high false alarm rate and low recognition accuracy in high noise environments. Summary of the invention

[0009] In view of the above problems, the present invention proposes a steam pipeline status monitoring method and system based on machine learning to more accurately identify the abnormal status of the steam pipeline, overcome the limitations of traditional methods, and greatly improve the monitoring accuracy and robustness.

[0010] According to one aspect of the present invention, a method for monitoring a steam pipeline state based on machine learning is proposed, the method comprising the following steps:

[0011] Collect vibration signals of steam pipes;

[0012] Preprocessing the vibration signal;

[0013] The preprocessed data is input into the trained dynamic wavelet Fourier attention network model for feature extraction;

[0014] The extracted features are input into a trained anomaly detection model for anomaly detection; if the anomaly detection result indicates that the vibration signal is an abnormal signal, the extracted features are input into a trained classification model for classification to obtain a classification result.

[0015] Furthermore, the preprocessing of the vibration signal includes standardization processing and dimensionality reduction processing.

[0016] Furthermore, the inputting of the preprocessed data into the trained dynamic wavelet Fourier attention network model for feature extraction includes: extracting wavelet features from the input data using a trainable wavelet autoencoder branch, extracting Fourier features from the input data using a Fourier feature extraction branch, and extracting convolution features from the input data using a convolution feature branch; performing weighted fusion of wavelet features, Fourier features, and convolution features, and outputting fused features.

[0017] Furthermore, the extracting wavelet features from the input data using the trainable wavelet autoencoder branch includes:

[0018] Define two trainable parameter tensors: low-pass filter parameters W l ∈R k and high-pass filter parameter W h ∈R k , where R is the real number domain, representing the value range of the filter parameters, and k represents the filter length;

[0019] The first level wavelet decomposition is as follows:

[0020]

[0021]

[0022] Where, L (l) [n] represents the low-frequency characteristics of the first-level wavelet decomposition, H (l) [n] represents the high-frequency features of the first-level wavelet decomposition, m and n represent the input index and output index of the signal respectively, and x represents the input data;

[0023] The multi-level decomposition recursion is as follows:

[0024]

[0025]

[0026] After j levels of decomposition, multiple sub-band features are obtained, that is, a set of feature sets is obtained in the multi-level decomposition process, including low-frequency features from the highest level to high-frequency features at each level: {L j ,H j ,H (j-1) ,...,H (1)};

[0027] Multiple sub-band features are further extracted and compressed through one-dimensional convolutional layers and pooling layers to obtain compact high-level representations;

[0028] Through the inverse wavelet filter, the low-frequency and high-frequency features of each level are combined step by step to gradually restore the time domain representation with higher resolution, and finally reconstruct the signal close to the original input signal zw ; Among them, the reconstruction formula is:

[0029]

[0030] in, is the filter used in the decoder.

[0031] Furthermore, the method of extracting Fourier features from input data using the Fourier feature extraction branch includes: performing fast Fourier transform on the input data to obtain a complex spectrum, and extracting its amplitude and phase features; splicing the amplitude and phase along the channel direction to form an input feature tensor; performing convolution processing on the input feature tensor using a one-dimensional convolution to obtain an output feature tensor, i.e., the Fourier feature z f ;

[0032] The step of extracting convolution features from input data using a convolution feature branch includes: extracting convolution features z using a one-dimensional convolutional neural network. c .

[0033] Furthermore, the wavelet feature, the Fourier feature, and the convolution feature are weightedly fused, and the fused feature output includes: the fused feature output is expressed as:

[0034] H attention =Attention(Q,K,V)

[0035] In the formula, Attention represents the attention mechanism; Q = w Q ·z w , K = w K ·z f , V = w V ·z c , w Q ,w K ,w V They are the weight matrices corresponding to wavelet features, Fourier features, and convolution features respectively.

[0036] Furthermore, the weight matrix corresponding to the wavelet feature is:

[0037]

[0038] In the formula, F(k) represents the spectrum amplitude after Fourier transform, T represents the total length of the signal spectrum; k c is the frequency cut point;

[0039] The weight matrix corresponding to the Fourier feature is:

[0040]

[0041] The weight matrix corresponding to the convolution feature is:

[0042] w V =1-|w Q -w K |.

[0043] Furthermore, the step of inputting the extracted features into the trained anomaly detection model for anomaly detection includes: inputting the fused features into the IsolationForest model, averaging the number of splits on all random trees to obtain the path length h(x) of the sample, and then normalizing it to obtain an anomaly score; when the anomaly score is higher than a set threshold, the vibration signal corresponding to the sample is determined to be an abnormal signal, otherwise it is a normal signal; wherein the calculation formula of the path length is:

[0044]

[0045] Where, T i represents the i-th random tree; split(x,T i ) indicates that sample x is in the i-th random tree T i The number of splits on ; p represents the number of random trees;

[0046] The calculation formula for the anomaly score is:

[0047]

[0048] Where E(h(x)) represents the average path length of sample x in the Isolation Forest model; c(N) is the normalization factor; and N represents the total number of input samples.

[0049] Furthermore, the classification model adopts the LightGBM model.

[0050] According to another aspect of the present invention, a steam pipeline condition monitoring system based on machine learning is proposed, the system comprising:

[0051] A signal acquisition module configured to acquire a vibration signal of the steam pipe;

[0052] A preprocessing module, configured to preprocess the vibration signal;

[0053] A feature extraction module configured to input the preprocessed data into a trained dynamic wavelet Fourier attention network model for feature extraction;

[0054] An anomaly detection module configured to input the extracted features into a trained anomaly detection model for anomaly detection;

[0055] The classification module is configured to input the extracted features into a trained classification model for classification to obtain a classification result if the abnormality detection result indicates that the vibration signal is an abnormal signal.

[0056] The beneficial technical effects of the present invention are:

[0057] The present invention proposes a steam pipeline state monitoring method and system based on machine learning, which extracts the time-frequency characteristics of the signal through the wavelet convolution layer, captures the global frequency domain characteristics through the Fourier branch, extracts the local time characteristics through the convolution branch, combines the attention mechanism and the Isolation Forest model for anomaly detection, and finally classifies through the LightGBM model, which can achieve more efficient and accurate anomaly detection in the field of steam pipeline monitoring. The present invention has the following advantages:

[0058] 1) Introducing a trainable wavelet convolutional autoencoder for automatic feature extraction: Traditional methods based on manual features or fixed transformations have limitations when facing complex vibration signals. To this end, the present invention uses a trainable wavelet convolutional autoencoder to perform adaptive multi-scale time-frequency decomposition and feature extraction on the input data. Through a trainable wavelet filter and a multi-layer convolution structure, not only can high-level features containing spatial structure and local pattern information be automatically extracted, but the feature distribution of normal and abnormal samples can also be effectively distinguished in the wavelet domain, making subsequent anomaly detection more accurate and efficient.

[0059] 2) Organic combination of multi-branch feature fusion and Isolation Forest model: After feature extraction by trainable wavelet convolution autoencoder, other feature branches (such as Fourier branch, time domain convolution branch) are used in parallel to strengthen feature representation from multiple dimensions (time-frequency domain, frequency domain, time domain), and form high-quality compact feature vectors under the weighted fusion of channel splicing and attention mechanism. Subsequently, the refined features are input into the Isolation Forest model for unsupervised anomaly detection. Thanks to the front-end feature extraction and dimensionality reduction processing, Isolation Forest no longer directly faces high-dimensional and redundant raw data, but quickly and accurately identifies anomalies in the streamlined feature space. The detected abnormal samples are then finely classified using the LightGBM model, which can clearly present the specific abnormal state, facilitating operation and maintenance personnel to make decisions accurately and efficiently.

[0060] 3) Structural flexibility and scalability optimization: Different vibration data sources may have significant differences. In order to meet the needs of a wide range of applications, the model proposed in this invention allows flexible regulation of the number of trainable wavelet layers and convolutional layers, the size of the convolution kernel, and the wavelet decomposition level, so as to adapt to a variety of complex signal environments. At the same time, the feature dimensionality reduction strategy can be adjusted as needed, and the high-dimensional raw data can be mapped to a lower-dimensional feature space through standardization (StandardScaler) and other methods, reducing the performance loss and computational burden of Isolation Forest in high dimensions and improving the overall operating efficiency.

[0061] The present invention solves the problem of high false alarm rate in traditional technologies, can accurately identify and classify the vibration state of steam pipes, can effectively improve the accuracy, noise resistance and computing efficiency of anomaly detection, enhances the robustness of the overall system, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood by reading the following detailed description with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, in which:

[0063] Figure 1 is a flow chart of a method for monitoring a steam pipeline state based on machine learning according to an embodiment of the present invention;

[0064] Figure 2 is another flow chart of a method for monitoring a steam pipeline state based on machine learning according to an embodiment of the present invention;

[0065] Figure 3 This is another flow chart of a steam pipeline state monitoring method based on machine learning described in an embodiment of the present invention; in the figure, orange squares represent data standardization and dimensionality reduction operations, black squares represent wavelet convolution operations, yellow squares represent convolution operations, green squares represent pooling operations, gray squares represent Fourier transform operations, dark blue squares represent branch weight calculation and feature fusion operations, cyan squares represent feature extraction operations, red squares represent anomaly detection operations, and white squares represent state classification operations.

[0066] Figure 4 It is a structural schematic diagram of a steam pipeline status monitoring system based on machine learning described in an embodiment of the present invention. DETAILED DESCRIPTION

[0067] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0068] Those skilled in the art know that the embodiments of the present invention can be implemented as a system, device, equipment, method or computer program product. Therefore, the present disclosure can be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software. In this article, it is to be understood that any number of elements in the drawings is for illustration and not limitation, and any naming is only for distinction and does not have any limiting meaning.

[0069] The present invention provides an efficient steam pipeline monitoring method, which combines the Φ-OTDR system, the dynamic wavelet-Fourier attention network (Dynamic Wavelet-FourierAttentionNetwork, DWFA-Net), the isolation forest model and the lightweight gradient boosting machine algorithm (LightGBM) model, uses the standard scaler to perform feature scaling on the input data, eliminates the dimensional differences between different features, and performs dimensionality reduction processing on the data to reduce the subsequent branch calculation pressure. The signal is then sent to the branch network at the same time to extract the characteristic signal, and then the isolation forest model is used for anomaly detection, and finally the LightGBM model is used for anomaly classification. The present invention significantly improves the detection accuracy and real-time performance, and effectively solves the false alarm problem of traditional methods in high-noise environments.

[0070] The embodiment of the present invention proposes a steam pipeline state monitoring method based on machine learning, such as Figures 1 to 3 As shown, the method comprises the following steps:

[0071] S1, collecting vibration signals of steam pipes;

[0072] S2, preprocessing the vibration signal;

[0073] S3, input the preprocessed data into the trained dynamic wavelet Fourier attention network model for feature extraction;

[0074] S4, input the extracted features into the trained anomaly detection model for anomaly detection;

[0075] S5. If the abnormality detection result indicates that the vibration signal is an abnormal signal, the extracted features are input into a trained classification model for classification to obtain a classification result.

[0076] The method starts at S1. In S1, a vibration signal of a steam pipeline is collected.

[0077] According to an embodiment of the present invention, a Φ-OTDR monitoring system is used to collect time domain signals, and Rayleigh backscattering is used to obtain the optical signal intensity at each position in the optical cable of the steam pipeline. The Rayleigh backscattering signal generated by each pulse signal emitted by the light source is taken as a row, thereby forming a data space-time matrix A. Each row in this data matrix A represents the spatial domain, and each column represents the intensity of the current component, that is, the signal intensity generated by the optical cable at this position, which reflects the time domain. According to the system parameter setting, each data space-time matrix A represents a certain spatial distance and time span.

[0078] Then, S2 is executed, in which the vibration signal is preprocessed; the preprocessing includes standardization processing and dimensionality reduction processing.

[0079] According to an embodiment of the present invention, a section of normal operating condition data collected is used as a benchmark data set. The normal data is loaded in batches, and the collected spatiotemporal data is standardized through standardization (StandardScaler) to eliminate environmental interference, ensure that the data mean is 0, and the variance is 1, thereby improving the availability of the data. Secondly, the standardized data is incrementally fitted and reduced in dimension through the incremental principal component analysis (IncrementalPCA) method, that is, the data is processed in blocks, which alleviates the memory and computing pressure. This preprocessing helps to improve the convergence speed of the model, reduce the negative impact of outliers on the model training process, and ultimately improve the detection accuracy and stability of the model.

[0080] Each input data sample in the training data set needs to be labeled before preprocessing, that is, the collected data is labeled according to its corresponding status. In the operation and maintenance data, the startup and shutdown time periods are all normal operating states except for the time periods when condensate accumulation and water hammer phenomena occur, which cause severe vibration of the pipeline. The startup sample is marked as normal sample one, and the shutdown sample is marked as normal sample two; when condensate accumulation and water hammer phenomena occur, the pipeline vibrates violently due to the impact, which may cause pipeline damage and reduce the life of the equipment. The two obtained samples are marked as abnormal sample one and abnormal sample two. The corresponding labels can improve the accuracy of subsequent model detection and classification.

[0081] By achieving moderate dimensionality reduction through standardization and incremental PCA, tree-based models such as the Isolation Forest model and the LightGBM model can operate in a feature space with more concentrated information, thereby effectively reducing the false alarm rate and improving the accuracy and stability of fault classification. This design significantly reduces the computational complexity while maintaining recognition accuracy, and enhances robustness to environmental changes and noise interference.

[0082] Then execute S3, in which the preprocessed data is input into the trained dynamic wavelet Fourier attention network model for feature extraction; specifically, the trainable wavelet autoencoder branch is used to extract wavelet features from the input data, the Fourier feature extraction branch is used to extract Fourier features from the input data, and the convolution feature branch is used to extract convolution features from the input data; the wavelet features, Fourier features, and convolution features are weightedly fused and the fused features are output.

[0083] According to an embodiment of the present invention, the traditional Isolation Forest model divides the data set by randomly selecting features and feature values, and identifies abnormal points based on the path length required for the data point to be isolated to a single node. Its unsupervised and efficient characteristics give it certain advantages in large-scale data monitoring. However, when it is applied to steam pipeline status monitoring, it still faces the following challenges:

[0084] 1) Difficulty in processing high-dimensional data: As the data dimension increases, the data becomes extremely sparse in high-dimensional space, and distance measurement and feature discrimination are no longer reliable. At this time, Isolation Forest requires more trees and greater depth to accurately capture the characteristic distribution of anomalies, which not only significantly increases the computational overhead, but also easily increases the false alarm and missed alarm rates, thereby affecting the efficiency and accuracy of real-time monitoring.

[0085] 2) Unclear definition of anomalies leads to false alarms: In actual steam pipeline operation, the normal state itself may have large fluctuations. Isolation Forest only relies on isolation to judge anomalies, which may judge normal fluctuations as anomalies, resulting in frequent false alarms. This not only reduces the credibility of the system, but also increases the judgment burden and subsequent screening costs of operation and maintenance personnel.

[0086] 3) Lack of explainability and difficulty in tracing the source of the problem: As a black box model, Isolation Forest can only output anomaly scores and cannot clearly point out the key features or specific causes of the anomaly. For steam pipeline operation scenarios, this limits the operation and maintenance engineers from quickly locating problems and taking precise and effective intervention measures.

[0087] 4) Unable to capture complex time series patterns: Isolation Forest assumes that data are independent of each other and does not consider time order and dynamic evolution characteristics. In typical time series data such as vibration signals, important time series dependencies and non-stationary changes may be ignored, resulting in insufficient sensitivity to the detection of dynamic abnormal patterns, missing reports or misjudging normal time series changes as abnormalities.

[0088] 5) Lack of labeled samples makes it difficult to evaluate performance: The frequency of abnormal events in actual steam pipelines is low, and labeled data is scarce, making it difficult to objectively evaluate the performance of unsupervised models. It is impossible to tune model parameters and evaluate indicators based on known abnormal samples, which increases deployment risks and uncertainties.

[0089] Based on the above problems, the present invention proposes a multi-branch feature extraction scheme with a dynamic wavelet-Fourier attention network (Dynamic Wavelet-Fourier Attention Network, DWFA-Net) as the core. The model can extract features through three feature extraction branches, namely, training wavelet autoencoder branch, Fourier feature extraction branch and convolution feature branch, and dynamically adjust the weights of wavelet, Fourier and convolution branches according to signal characteristics. According to the spectral characteristics or time domain characteristics of the input signal, the most relevant branch is dynamically activated to ensure the optimal feature extraction ratio, and then the wavelet-Fourier joint transform is used to combine local and global features and integrate them into a multi-component feature vector, which enhances the model's ability to express complex signals and improves training speed and classification accuracy. In this model, accurate analysis of complex vibration signals can be achieved on the basis of multi-domain fusion without tedious hyperparameter adjustment or excessive data and computing resources. By introducing a trainable wavelet convolutional autoencoder (WCAE), the original data is adaptively decomposed into multi-scale time-frequency decomposition, and the global frequency domain and time domain features are extracted by combining Fourier and convolution feature branches. The attention mechanism is then used to perform weighted fusion of wavelet and Fourier features, adaptively select key feature areas, significantly reduce feature redundancy, improve the stability and generalization ability of the model, and reduce the risk of model overfitting.

[0090] In this model, the dynamic wavelet Fourier attention network (DWFA-Net) uses a trainable wavelet convolutional autoencoder to automatically extract the multi-scale time-frequency features of the signal, accurately capturing the changes in the equipment's operating status and the characteristics of fault signals. Combining the deep pattern mining of the time domain signal by the convolutional feature branch and the modeling ability of the global spectrum information by the Fourier branch, adaptive feature extraction is achieved through a dynamic branch selection mechanism. At the same time, the cross-transformation attention mechanism performs weighted fusion of multi-branch features, which can efficiently focus on key feature areas under high-dimensional input.

[0091] Feature extraction branch 1: Trainable wavelet convolutional autoencoder;

[0092] 1) Two trainable parameter tensors can be defined in the process of building the model: low-pass filter parameter W l ∈R k And the high-pass filter parameter W h ∈R k , where R is the real number domain, representing the value range of the filter parameters, and k represents the filter length; the initial parameters of the two convolution kernels can be set as Morlet coefficients. x is the one-dimensional data after standardization and dimensionality reduction.

[0093] 2) Multi-level wavelet decomposition and encoder implementation:

[0094] 21) First level wavelet decomposition:

[0095]

[0096] Where, L (1) [n] represents the low-frequency characteristics of the first-level wavelet decomposition, H (1) [n] represents the high-frequency characteristics of the first-level wavelet decomposition, m and n represent the input index and output index of the signal respectively; at the same time, the output continues to be sampled to achieve multi-scale analysis.

[0097] 22) Multi-level decomposition is performed recursively:

[0098]

[0099] After j levels of decomposition, multiple sub-band features are obtained, that is, a set of features obtained in the multi-level decomposition process, including low-frequency features from the highest level to high-frequency features at each level: {L j ,H j ,H (j-1) ,...,H (1)}; These sub-band features can be regarded as the time-frequency domain decomposition results at multiple scales.

[0100] 23) Integration of convolution and pooling layers: After obtaining the wavelet subbands, these features are further extracted and compressed through one-dimensional convolution and pooling layers. wavelet :

[0101] z wavelet =f conv-pool ([L (j) ;H (j) ;...;H (1) ])

[0102] Among them, [;] represents the splicing operation, f conv-poolRepresents the pooling operation. After multiple convolutions and pooling, a compact high-level representation z is obtained. w .

[0103] 3) Reconstruction loss and training of autoencoder: The decoder converts the low-frequency features L( j ) and high-frequency features H( j ) are gradually merged until the final signal z is reconstructed w .

[0104] The reconstruction process formula is as follows:

[0105]

[0106] in is the filter used in the decoder.

[0107]

[0108] By minimizing the reconstruction error filter And the convolution layer parameters will be adaptively adjusted so that WCAE can best represent the characteristics of normal signals.

[0109] Feature extraction branch 2: Fourier branch;

[0110] Given an input time domain signal x∈R, first perform a fast Fourier transform (FFT) on it as follows:

[0111]

[0112] The complex spectrum X[K] is obtained, and the amplitude and phase characteristics are extracted as follows:

[0113] A[k]=|X[k]|,φ[k]=arg(X[k]].

[0114] The amplitude A and phase φ are concatenated along the channel direction to form the input feature tensor as follows:

[0115]

[0116] Use a one-dimensional convolution module to process the above frequency domain features. Assume that the parameters of the convolution layer are as follows:

[0117] Convolution kernel weight: w f ; Bias term: k f ;

[0118] For output channel j and feature position t, the convolution result is:

[0119]

[0120] Among them, z f is the output feature tensor, and K is the convolution kernel size.

[0121] Feature extraction branch three: convolution feature branch;

[0122] Use a one-dimensional convolutional neural network to extract features from the input x∈R. Suppose there is a layer of convolution:

[0123]

[0124] Among them, σ(.) is a nonlinear activation function. If M layers of convolution and pooling are used:

[0125]

[0126] Finally, the convolution branch feature representation is obtained:

[0127]

[0128] After the above three feature extraction branches are processed, the three branch features are expressed as z w ,z f ,z c .

[0129] Then, the joint feature extraction and attention mechanism is as follows.

[0130] The fused feature output is:

[0131] H attention =Attention(Q,K,V)

[0132] Among them, H attention is the dynamic branch output, i.e. the fused feature representation; Q = w Q ·z W , K = w K ·z f , V = w V ·z c , z w ,z f ,z c They are wavelet feature, Fourier feature and convolution feature respectively; w Q ,w K ,w V They are the weight matrices corresponding to wavelet features, Fourier features, and convolution features respectively.

[0133] The attention mechanism is calculated as follows:

[0134]

[0135] Where, d k is the feature dimension.

[0136] In an embodiment of the present invention, a dynamic branch selection mechanism is introduced to calculate the weight matrices corresponding to wavelet features, Fourier features and convolution features. According to the energy distribution and feature importance of different feature branches, the weight of each branch is adaptively adjusted to achieve a fusion effect of dynamically optimizing the output features.

[0137] The branch weight is calculated as follows:

[0138]

[0139]

[0140] w V =1-|w Q -w K |

[0141] In the above formula, the frequency energy distribution is calculated as follows: Fast Fourier Transform output F = FFT (X), k c is the frequency cut-off point.

[0142] Then, S4 is executed, in which the extracted features are input into the trained anomaly detection model for anomaly detection.

[0143] According to an embodiment of the present invention, the anomaly detection model (IsolationForest) is trained as follows: the feature signal after calculation and fusion by the weight matrix is ​​input into the IsolationForest model, and the anomaly is identified based on the isolated path length of the sample. The path length calculation is used to measure the ease with which the sample is isolated. The isolation forest randomly splits the feature space of the sample. The short sample path length usually indicates that the sample is sparsely distributed, while the long path indicates that the sample is densely distributed. By averaging the number of splits on all trees, the path length h(x) of the sample is obtained, and then the anomaly score is obtained by normalization. The anomaly score reflects the degree of deviation of the sample from the normal distribution. The score value close to 1 indicates that the distribution of the sample is significantly different from that of most samples and may be an outlier. When the anomaly score is higher than the set threshold, the model determines that the sample is an abnormal signal; otherwise, the sample is considered to be a normal signal.

[0144] The path length calculation formula is:

[0145]

[0146] Where, T i represents the i-th random tree; split(x,T i ) indicates that sample x is in the i-th random tree T i The number of splits on ; p represents the total number of trees in the isolation forest (the number of random trees);

[0147] The anomaly score calculation formula is as follows:

[0148]

[0149] Where E(h(x)) represents the average path length of sample x in the Isolation Forest model; c(N) is the normalization factor used to normalize the path length, and N represents the total number of input samples, which is used to standardize the path length.

[0150] Then, S5 is executed. In S5, if the abnormality detection result indicates that the vibration signal is an abnormal signal, the extracted features are input into a trained classification model for classification to obtain a classification result.

[0151] According to an embodiment of the present invention, the classification model adopts a lightweight gradient boosting machine algorithm - LightGBM model. After the labeled water hammer phenomenon samples and condensate accumulation samples undergo the same preprocessing and feature extraction steps, they are used to train the LightGBM model to achieve classification.

[0152] In the actual monitoring stage, new data fragments are continuously acquired, and normalization and PCA dimensionality reduction are performed on them. Signal features are dynamically acquired through DWFA-Net, and the features are predicted using Isolation Forest. If the result is normal, it means that the sample is consistent with the normal distribution during training, and no alarm is required. If the result is abnormal, and multiple samples are abnormal in succession, the features of the abnormal sample are input into the trained LightGBM model, the abnormal type category label is obtained, and an alarm is issued.

[0153] It should be noted that in order to ensure that this model maintains optimal performance in a variety of application scenarios, the parameters of each branch structure can be flexibly adjusted according to task requirements, such as adjusting the level of wavelet decomposition, the number of layers and channels of the convolution branch, and the spectrum feature extraction method of the Fourier branch. In addition, the hyperparameters of Isolation Forest and LightGBM (such as the number and depth of trees) can also be optimized to adapt to different environmental conditions, signal types, and noise levels. Through these strategies, the model can efficiently respond to diverse application requirements and achieve accurate fault detection and status assessment.

[0154] The present invention combines the Φ-OTDR monitoring system, dynamic branch selection mechanism, wavelet-Fourier joint transform, multi-domain feature fusion strategy (wavelet, Fourier, convolution) and efficient tree model (Isolation Forest and LightGBM) to achieve real-time monitoring and identification of steam pipeline status under the deep learning framework. As an unsupervised anomaly detection algorithm, Isolation Forest, combined with multi-branch feature representation, can quickly locate abnormal data points, significantly improve the accuracy and real-time performance of pipeline status monitoring in complex industrial scenarios, and thus provide higher reliability and stability for industrial applications. Through this method, on the basis of multi-scale time-frequency domain feature extraction and dimensionality reduction of data, the optimized features are input into IsolationForest for efficient anomaly detection, and LightGBM is used to classify and finely distinguish the detected anomalies, thereby effectively reducing the false alarm rate, improving the model calculation efficiency and adaptability to time series data, and ultimately enhancing the accuracy, reliability and interpretability of steam pipeline status monitoring.

[0155] The present invention organically combines the dynamic wavelet Fourier attention network, unsupervised anomaly detection (Isolation Forest) and the lightweight gradient boosting machine algorithm (LightGBM) to form an efficient and scalable anomaly detection framework. Through this integrated solution, the advantages of Isolation Forest are fully utilized in the field of anomaly detection, and the performance is significantly improved with the support of feature extraction and dimensionality reduction. After the above improvements, the method of the present invention not only achieves more efficient training and recognition under limited computing and data resource conditions, but also significantly reduces the false alarm rate and enhances the model's adaptability to environmental noise and non-stationary changes. Due to the organic combination of the dynamic wavelet Fourier attention network and Isolation Forest, the analysis depth and breadth of time series data are improved, and the features of various scales can be fully mined, thereby significantly improving the detection accuracy and robustness of the model.

[0156] Another embodiment of the present invention provides a steam pipeline status monitoring system based on machine learning, such as Figure 4 As shown, the system includes:

[0157] A signal acquisition module 410 configured to acquire a vibration signal of the steam pipe;

[0158] A preprocessing module 420, configured to preprocess the vibration signal;

[0159] A feature extraction module 430, which is configured to input the preprocessed data into a trained dynamic wavelet Fourier attention network model for feature extraction;

[0160] Anomaly detection module 440, which is configured to input the extracted features into the trained anomaly detection model for anomaly detection;

[0161] The classification module 450 is configured to input the extracted features into a trained classification model for classification to obtain a classification result if the abnormality detection result indicates that the vibration signal is an abnormal signal.

[0162] The function of the steam pipe state monitoring system based on machine learning described in this embodiment can be explained by the aforementioned steam pipe state monitoring method based on machine learning. Therefore, for the parts not described in detail in the system embodiment, please refer to the above method embodiment.

[0163] It should be noted that, although several units, modules or submodules are mentioned in the above detailed description, this division is merely exemplary and not mandatory. In fact, according to an embodiment of the present invention, the features and functions of two or more modules described above can be embodied in one module. Conversely, the features and functions of one module described above can be further divided into being embodied by multiple modules.

[0164] In addition, although the operations of the method of the present invention are described in a specific order in the drawings, this does not require or imply that the operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0165] Although the spirit and principle of the present invention have been described with reference to several specific embodiments, it should be understood that the present invention is not limited to the disclosed specific embodiments, and the division of various aspects does not mean that the features in these aspects cannot be combined to benefit, and such division is only for the convenience of expression. The present invention is intended to cover various modifications and equivalent arrangements included in the spirit and scope of the attached claims.

Claims

1. A steam pipeline status monitoring method based on machine learning, characterized in that: The following steps are involved: Collect vibration signals of steam pipes; Preprocessing the vibration signal; The preprocessed data is input into the trained dynamic wavelet Fourier attention network model for feature extraction; The extracted features are input into a trained anomaly detection model for anomaly detection; if the anomaly detection result indicates that the vibration signal is an abnormal signal, the extracted features are input into a trained classification model for classification to obtain a classification result.

2. A method for monitoring steam pipeline status based on machine learning according to claim 1, characterized in that: The preprocessing of the vibration signal includes standardization processing and dimension reduction processing.

3. The method for monitoring the state of a steam pipeline based on machine learning according to claim 1, characterized in that: The method of inputting the preprocessed data into the trained dynamic wavelet Fourier attention network model for feature extraction includes: extracting wavelet features from the input data using a trainable wavelet autoencoder branch, extracting Fourier features from the input data using a Fourier feature extraction branch, and extracting convolution features from the input data using a convolution feature branch; performing weighted fusion of wavelet features, Fourier features, and convolution features, and outputting fused features.

4. The method for monitoring the state of a steam pipeline based on machine learning according to claim 3, characterized in that: The extracting wavelet features from input data using a trainable wavelet autoencoder branch comprises: Define two trainable parameter tensors: low-pass filter parameters W l ∈R k and high-pass filter parameter W h ∈R k , where R is the real number domain, representing the value range of the filter parameters, and k represents the filter length; The first level wavelet decomposition is as follows: Where, L (l) [n] represents the low-frequency characteristics of the first-level wavelet decomposition, H (l) [n] represents the high-frequency features of the first-level wavelet decomposition, m and n represent the input index and output index of the signal respectively, and x represents the input data; The multi-level decomposition recursion is as follows: After j levels of decomposition, multiple sub-band features are obtained, that is, a set of feature sets is obtained in the multi-level decomposition process, including low-frequency features from the highest level to high-frequency features at each level: {L j ,H j ,H (j-1) ,...,H (1) }; Multiple sub-band features are further extracted and compressed through one-dimensional convolutional layers and pooling layers to obtain compact high-level representations; Through the inverse wavelet filter, the low-frequency features L of each level are reconstructed step by step. (j) and high frequency features H (j) Step by step, restore to a higher resolution time domain representation until the final signal z is reconstructed w ; Among them, the reconstruction formula is: Among them, W l rev , is the filter used in the decoder.

5. The method for monitoring the state of a steam pipeline based on machine learning according to claim 4, characterized in that: The method of extracting Fourier features from input data using the Fourier feature extraction branch includes: performing fast Fourier transform on the input data to obtain a complex spectrum, and extracting its amplitude and phase features; splicing the amplitude and phase along the channel direction to form an input feature tensor; performing convolution processing on the input feature tensor using a one-dimensional convolution to obtain an output feature tensor, i.e., the Fourier feature z f ; The step of extracting convolution features from input data using a convolution feature branch includes: extracting convolution features z using a one-dimensional convolutional neural network. c .

6. The method for monitoring the state of a steam pipeline based on machine learning according to claim 5, characterized in that: The wavelet feature, Fourier feature, and convolution feature are weightedly fused, and the output fusion feature includes: the fusion feature is expressed as: H attention =Attention(Q,K,V) In the formula, Attention represents the attention mechanism; Q = w Q ·z W , K = w K ·z f , V = w V ·z c , w Q ,w K ,w V They are the weight matrices corresponding to wavelet features, Fourier features, and convolution features respectively.

7. The method for monitoring the state of a steam pipeline based on machine learning according to claim 6, characterized in that: The weight matrix corresponding to the wavelet feature is: In the formula, F(k) represents the spectrum amplitude after Fourier transform, T represents the total length of the signal spectrum; k c is the frequency cut point; The weight matrix corresponding to the Fourier feature is: The weight matrix corresponding to the convolution feature is: In V =1-|in Q -In K |。 8. The method for monitoring the state of a steam pipeline based on machine learning according to claim 7, characterized in that: The step of inputting the extracted features into the trained anomaly detection model for anomaly detection includes: inputting the fused features into the IsolationForest model, averaging the number of splits on all random trees to obtain the path length h(x) of the sample, and then normalizing it to obtain an anomaly score; when the anomaly score is higher than a set threshold, the vibration signal corresponding to the sample is determined to be an abnormal signal, otherwise it is a normal signal; wherein the calculation formula of the path length is: Where, T i represents the i-th random tree; split(x,T i ) indicates that sample x is in the i-th random tree T i The number of splits on ; p represents the number of random trees; The calculation formula for the anomaly score is: Where E(h(x)) represents the average path length of sample x in the Isolation Forest model; c(N) is the normalization factor; and N represents the total number of input samples.

9. The method for monitoring the state of a steam pipeline based on machine learning according to claim 8, characterized in that: The classification model adopts the LightGBM model.

10. A steam pipeline status monitoring system based on machine learning, characterized in that: include: A signal acquisition module configured to acquire a vibration signal of the steam pipe; A preprocessing module, configured to preprocess the vibration signal; A feature extraction module configured to input the preprocessed data into a trained dynamic wavelet Fourier attention network model for feature extraction; An anomaly detection module configured to input the extracted features into a trained anomaly detection model for anomaly detection; The classification module is configured to input the extracted features into a trained classification model for classification to obtain a classification result if the abnormality detection result indicates that the vibration signal is an abnormal signal.

Images