A method and system for detecting leaks in boiler pipelines

Abstract

This invention relates to the field of boiler pipeline leakage detection technology, specifically to a boiler pipeline leakage detection method and system. The boiler pipeline leakage detection method includes the following steps: real-time acquisition and preprocessing of initial signal data from multiple sensors to generate standard signal data from multiple sensors; designing a separate dual encoder and a symmetric decoder based on the standard signal data from multiple sensors, and constructing a TCAE model; automatically iteratively optimizing the hyperparameters of the TCAE model using Bayesian optimization to obtain an Opt-TCAE model; confirming an adaptive leakage detection threshold; inputting the standard signal data from multiple sensors into the Opt-TCAE model, and calculating the reconstruction error of the standard signal data from multiple sensors in real time, wherein: if the reconstruction error is greater than the adaptive leakage detection threshold and the duration is ≥5 minutes, a leak is determined.

Description

Technical Field

[0001] This invention relates to the field of boiler pipeline leakage detection technology, specifically to a boiler pipeline leakage detection method and system. Background Technology

[0002] In the operation of thermal power plants, steam boilers, as core power equipment, have metal pipes that are constantly exposed to harsh conditions of high temperature and high pressure. These pipes are prone to overheating damage, thermal fatigue cracking, internal corrosion, and material creep, leading to continuous thinning of the pipe wall and ultimately leakage. Industry statistics show that pipe-related failures account for as much as 60% of all boiler failures, causing over $5 billion in economic losses to the global power industry annually. This not only results in unplanned shutdowns causing capacity losses but can also trigger catastrophic safety accidents such as explosions. Therefore, accurate and timely detection of boiler tube leaks is crucial for ensuring the stable operation of the power system.

[0003] Currently, anomaly detection technology has become the main technical approach for boiler tube leakage monitoring. Among them, unsupervised learning algorithms have become the mainstream choice in this field because they do not require the collection of exhaustive fault data (boiler pipes are densely distributed and leak locations are random, making it impossible to cover all fault scenarios). In existing technologies, methods based on principal component analysis (PCA) capture the linear correlation between sensors to achieve anomaly identification by performing dimensionality reduction modeling on multi-sensor data. However, this method cannot handle the nonlinear characteristics of sensor signals during boiler operation and is prone to missed detections or false detections under complex operating conditions. Although detection methods based on traditional autoencoders can model nonlinear relationships, they only focus on the spatial correlation between sensors and ignore the time dependence of signals formed by the dynamic changes of boiler operating status over time (such as the time-series trends of parameters such as pressure and temperature). This results in insufficient sensitivity for identifying gradual leaks and susceptibility to false alarms caused by transient interference signals. In addition, existing unsupervised detection algorithms generally lack automatic hyperparameter optimization mechanisms. The values ​​of hyperparameters (such as latent dimension, convolution kernel size, etc.) directly determine the model performance. Manual adjustment not only relies on experience but is also difficult to adapt to the changes in operating conditions of boilers in different operating cycles.

[0004] Meanwhile, acoustic emission (AE) sensors, as the core sensing device for leak detection, are susceptible to interference from external factors such as boiler soot blowing operations (a routine operation to remove ash buildup on the inner walls of pipes), generating short-duration, high-amplitude noise peaks. This noise, superimposed on the leak signal, further exacerbates the false alarm problem of existing detection methods. The deficiencies of the aforementioned existing technologies in feature capture completeness, noise suppression capabilities, and parameter optimization adaptability make it difficult to meet the actual needs of industrial sites in terms of accuracy and reliability of boiler tube leak detection. Therefore, there is an urgent need for a detection technology solution that can take into account both temporal and spatial characteristics, effectively suppress noise, and achieve adaptive parameter optimization. Summary of the Invention

[0005] The purpose of this invention is to provide a boiler pipe leakage detection method to solve the problem of low accuracy and reliability of boiler pipe leakage detection methods in the prior art.

[0006] To address the above problems, this invention proposes a method for detecting leaks in boiler pipelines. The technical solution adopted is as follows: A method for detecting leaks in boiler pipelines includes the following steps: Real-time acquisition and preprocessing of initial signal data from multiple sensors to generate standard signal data from multiple sensors; Based on standard signal data from multiple sensors, a separate dual encoder and a symmetric decoder are designed, and a temporal convolutional autoencoder (TCAE) model is constructed. With the goal of minimizing the reconstruction error entropy of multi-sensor standard signal data, Bayesian optimization is used to automatically iteratively optimize the hyperparameters of the temporal convolutional autoencoder (TCAE) model, determine the optimal hyperparameter combination, and obtain the optimal temporal convolutional autoencoder (Opt-TCAE) model. An adaptive threshold setting method based on kernel density estimation is used to fit the reconstruction error probability distribution of standard signal data from multiple sensors and confirm the adaptive leak detection threshold. Multi-sensor standard signal data is input into the optimal temporal convolutional autoencoder Opt-TCAE model, and the reconstruction error of the multi-sensor standard signal data is calculated in real time, where: If the reconstruction error of the multi-sensor standard signal data is greater than the adaptive leak detection threshold and the duration is ≥5 minutes, it is determined to be a leak, and the detection result and quantitative evaluation index are output. Otherwise, it is considered normal.

[0007] Furthermore, the real-time acquisition and preprocessing of initial signal data from multiple sensors to generate standard signal data from multiple sensors includes: Acoustic emission signals and operational signal data are acquired in real time to obtain initial signal data from multiple sensors and construct a time series dataset; Based on time series datasets, moving average filtering and sliding window segmentation and scaling are applied to the initial signal data of multiple sensors to obtain standard signal data of multiple sensors.

[0008] Furthermore, the process of performing moving average filtering for noise reduction and sliding window segmentation and scaling on the initial signal data from multiple sensors based on the time-series dataset to obtain standard signal data from multiple sensors includes: The short-term high-amplitude noise generated by external interference in the acoustic emission signal data is processed by moving average filtering to reduce noise and smooth the data, and multi-channel sensor noise-reduced signal data is obtained. A sliding window is used to convert the time series corresponding to the time series dataset into time segments, generate multi-sensor signal input samples, and perform min-max scaling on each signal in the multi-sensor noise-reduced signal data and the running signal data to generate multi-sensor standard signal data.

[0009] Furthermore, based on multi-sensor standard signal data, a discrete dual encoder and a symmetric decoder are designed, and a temporal convolutional autoencoder (TCAE) model is constructed, including: Design a split dual encoder and a symmetric decoder. Based on the temporal features of multi-sensor standard signal data extracted by the split dual encoder and the spatial correlation features between multiple sensors, realize the compression, reconstruction and reconstruction error minimization training of multi-sensor standard signal data, and construct a temporal convolutional autoencoder (TCAE) model.

[0010] Furthermore, the design of the discrete dual encoder and symmetric decoder, based on the temporal features of the multi-sensor standard signal data extracted by the discrete dual encoder and the spatial correlation features between the multiple sensors, realizes the compression, reconstruction, and reconstruction error minimization training of the multi-sensor standard signal data, and constructs a temporal convolutional autoencoder (TCAE) model, including: By extracting the temporal features and spatial correlation features between multiple sensors from multi-sensor standard signal data using a separate dual encoder, compression of multi-sensor standard signal data can be achieved. By transposing the temporal characteristics and spatial correlation characteristics between multiple sensors of multi-sensor standard signal data using a symmetric decoder, the reconstruction temporal characteristics and reconstruction samples of multi-sensor standard signal data are output, thereby realizing the reconstruction of multi-sensor standard signal data. Using multi-sensor standard signal data as input, and reconstructing samples based on the multi-sensor standard signal data, the model is trained with the goal of minimizing the reconstruction error of the multi-sensor standard signal data, resulting in the Temporal Convolutional Autoencoder (TCAE) model.

[0011] Furthermore, the separate dual encoder includes a time encoder and an inter-sensor encoder, and the symmetrical decoder includes an inter-sensor decoder and a time decoder; The time encoder is used to extract the time features of multi-sensor standard signal data by compressing the time axis dimension using 1D depth convolution, batch normalization and ReLU activation. The sensor encoder is used to compress the sensor axis dimension and extract the spatial correlation features between multiple sensors by employing 1×1 convolution, batch normalization and ReLU activation. The inter-sensor decoder, symmetrical to the inter-sensor encoder, is used to reconstruct the sensor axis dimension through 1×1 transposed convolution and output the reconstruction time characteristics of multi-sensor standard signal data. The time decoder, symmetrical to the time encoder, is used to reconstruct the time axis dimension through 1D transposed depth convolution and output multi-sensor standard signal data reconstruction samples. The calculation of the 1D depthwise convolution is shown in Equation 1: Formula 1 In Equation 1, Let f be the convolution output of the j-th sensor standard signal at time step t, and let f be the filter length of the time encoder. Here, b represents the weights of the time encoder filter, and b represents the bias. For the j-th sensor standard signal at time... The standard signal data, where p is a value from 0 to f-1; The calculation of the 1×1 convolution is shown in Equation 2: Formula 2 In Equation 2, The output of the k-th sensor channel at time step t. For 1×1 convolution weights, For bias, Let be the time characteristic of the j-th sensor standard signal at time step t.

[0012] Furthermore, with the goal of minimizing the reconstruction error entropy of multi-sensor standard signal data, Bayesian optimization is used to automatically iteratively optimize the hyperparameters of the temporal convolutional autoencoder (TCAE) model, determining the optimal hyperparameter combination to obtain the optimal temporal convolutional autoencoder (Opt-TCAE) model, including: Randomly generate multiple initial hyperparameter combinations for the split dual encoder, and calculate the entropy values ​​of the corresponding multiple initial hyperparameters; A surrogate model is constructed by fitting the mapping relationship between multiple initial hyperparameter combinations and entropy values ​​using a Gaussian process. Based on the entropy values ​​of multiple initial hyperparameters in the surrogate model and the minimum entropy value among them, the next experimental point is selected according to the expected improvement criterion: Repeat steps 2 and 3 several times to perform automated iterative optimization of the hyperparameters of the temporal convolutional autoencoder (TCAE) model. Finally, select the parameter combination that minimizes the entropy value as the optimal hyperparameter combination, and thus obtain the optimal temporal convolutional autoencoder (Opt-TCAE) model.

[0013] Furthermore, the adaptive threshold setting method based on kernel density estimation fits the reconstruction error probability distribution of multi-sensor standard signal data to confirm the adaptive leak detection threshold, including: An adaptive threshold setting method for kernel density estimation fits the reconstruction error probability distribution of multi-sensor standard signal data to obtain a probability density function, and sets the 99.99% upper quantile of the probability density function as the leakage detection threshold. The formula for calculating the probability density function is shown in Equation 3: Formula 3 In Equation 3, Let be the probability density function. h is the number of samples in the validation set, and h is the bandwidth (selected through cross-validation). Gaussian kernel function Let be the reconstruction error value of the i-th sample in the multi-sensor signal input samples. This refers to the reconstruction error of standard signal data from multiple sensors.

[0014] Furthermore, the boiler pipeline leakage detection method further includes: The advance detection time is calculated by comparing the detection time of the Opt-TCAE model (optimal temporal convolutional autoencoder) with the detection time of the expert model. The false alarm rate is calculated by counting the number of times the reconstruction error of the multi-sensor standard signal data exceeds the adaptive leak detection threshold.

[0015] The present invention also proposes a system for performing a boiler pipeline leakage detection method, comprising: The multi-sensor standard signal data generation module is used to collect and preprocess initial signal data from multiple sensors in real time to generate multi-sensor standard signal data. The Temporal Convolutional Autoencoder (TCAE) model building module is used to design a discrete dual encoder and a symmetric decoder based on multi-sensor standard signal data, and to build a Temporal Convolutional Autoencoder (TCAE) model. The Temporal Convolutional Autoencoder (TCAE) model optimization module is used to automatically iteratively optimize the hyperparameters of the TCAE model using Bayesian optimization with the goal of minimizing the reconstruction error entropy of multi-sensor standard signal data, and determine the optimal hyperparameter combination to obtain the optimal Temporal Convolutional Autoencoder Opt-TCAE model. An adaptive leak detection threshold acquisition module is used to fit the reconstruction error probability distribution of multi-sensor standard signal data based on an adaptive threshold setting method of kernel density estimation, and to confirm the adaptive leak detection threshold. The leakage detection module inputs multi-sensor standard signal data into the optimal temporal convolutional autoencoder Opt-TCAE model and calculates the reconstruction error of the multi-sensor standard signal data in real time, where: If the reconstruction error of the multi-sensor standard signal data is greater than the adaptive leak detection threshold and the duration is ≥5 minutes, it is determined to be a leak, and the detection result and quantitative evaluation index are output. Otherwise, it is considered normal.

[0016] Compared with the prior art, this application has the following beneficial effects: This invention is an improved version, proposing a boiler tube leakage detection method based on the Optimal Temporal Convolutional Autoencoder (Opt-TCAE). First, it innovatively designs a TCAE model with a separate dual encoder and a symmetric decoder architecture. The separate dual encoder allows one independent encoder to extract the temporal dynamic dependencies of multi-sensor standard signal data, while the other independent encoder captures the spatial correlation features between sensors, achieving accurate fusion of temporal and spatial features, thus constructing the TCAE model. Second, a Bayesian optimization hyperparameter mechanism is constructed with the goal of minimizing the reconstruction error entropy of multi-sensor standard signal data, automatically optimizing the hyperparameters of the TCAE model. Simultaneously, this method and system also propose an adaptive threshold setting method based on kernel density estimation, fitting the reconstruction error probability distribution of multi-sensor standard signal data to determine the adaptive leakage detection threshold. The boiler pipe leakage detection method and system of this application integrate the entire process of "data acquisition and preprocessing - model construction - parameter optimization - threshold setting - leakage inference," clearly defining the parameters and interaction logic of each stage to form a complete industrial-grade solution, ultimately achieving high-precision, low-false-alarm leakage detection. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart of the boiler pipeline leakage detection method of the present invention. Detailed Implementation

[0018] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0019] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0020] The boiler pipeline leakage detection method and system of this application are described below with reference to the accompanying drawings.

[0021] The following is combined with Figure 1 This application provides a detailed description of the boiler pipeline leakage detection method.

[0022] Step 1: Real-time acquisition and preprocessing of initial signal data from multiple sensors to generate standard signal data from multiple sensors.

[0023] Specifically, the system acquires and preprocesses initial signal data from multiple sensors in real time to generate standard multi-sensor signal data. This includes: acquiring acoustic emission and operational signal data in real time to obtain initial multi-sensor signal data and constructing a time-series dataset; and performing moving average filtering for noise reduction and sliding window segmentation and scaling on the initial multi-sensor signal data based on the time-series dataset to obtain the standard multi-sensor signal data. Here, the multi-sensor signal consists of 23 sensor signals, including 20 acoustic emission (AE) signals and 3 operational signals.

[0024] Specifically, based on the time series dataset, moving average filtering and sliding window segmentation and scaling are performed on the initial signal data of multiple sensors to obtain standard signal data of multiple sensors. This includes: performing moving average filtering and smoothing on short-term high-amplitude noise generated by external interference in the acoustic emission signal data to obtain multi-sensor denoised signal data; using a sliding window to convert the time series corresponding to the time series dataset into time segments to generate multi-sensor signal input samples; and performing min-max scaling on each signal in the multi-sensor denoised signal data and the running signal data to generate standard signal data of multiple sensors.

[0025] In one specific implementation, firstly, the aforementioned 20 AE signals are collected, covering key areas such as the front and rear walls, left and right side walls, and rear flue of the furnace, denoted as... The three operating signals are for the replenishment water flow rate. Steam flow rate Generator power The sampling frequency was set to 1 sample / minute to construct a time series dataset. ,in, This represents the total data length.

[0026] Secondly, for the short-term high-amplitude noise generated by external interference such as soot blowing in the 20-channel AE signals, average filter noise reduction (MAF) smoothing is used to obtain multi-channel sensor noise-reduced signal data. The specific formula is shown in Equation 4. Formula 4 In Equation 4, Let be the filtered value of the i-th AE signal at time t, and k be the size of the filtering window, which can be adjusted according to the noise characteristics. Here, we take . , For the i-th AE signal at time... The initial signal data, where j is a value between 0 and k-1.

[0027] The filtered values ​​of each signal obtained by the above MAF smoothing process at different times are used to characterize the noise-reduced multi-channel sensor signal. This operation can retain the continuous abnormal signals related to boiler pipeline leakage and suppress instantaneous noise.

[0028] Finally, a sliding window is used to convert the time series corresponding to the time series dataset into time segments, with a window length of [missing information]. Generate multi-sensor signal input samples (where n is the sample size), and the formula for calculating the sample size is shown in Equation 5: Formula 5 Subsequently, each signal in the multi-channel sensor noise reduction signal data and operation signal data is subjected to min-max scaling, and the data is mapped to... The range is adjusted to eliminate the influence of amplitude differences, resulting in multi-sensor standard signal data. The calculation of the scaled multi-sensor standard signal data is shown in Equation 6: Formula 6 In Equation 6, This is the scaled standard signal data from the i-th sensor. These are the minimum and maximum values ​​of the standard signal data of the i-th sensor, respectively.

[0029] Step 2: Based on multi-sensor standard signal data, design a separate dual encoder and a symmetric decoder, and construct a temporal convolutional autoencoder (TCAE) model.

[0030] Specifically, based on multi-sensor standard signal data, a separate dual encoder and a symmetric decoder are designed, and a temporal convolutional autoencoder (TCAE) model is constructed. This includes: designing a separate dual encoder and a symmetric decoder; using the temporal features of the multi-sensor standard signal data extracted by the separate dual encoder and the spatial correlation features between the multi-sensors to achieve compression, reconstruction, and reconstruction error minimization training of the multi-sensor standard signal data; and constructing a temporal convolutional autoencoder (TCAE) model.

[0031] Specifically, a separate dual encoder and a symmetric decoder are designed. Based on the temporal features and spatial correlation features between multiple sensors extracted by the separate dual encoder, the compression, reconstruction, and reconstruction error minimization training of the multi-sensor standard signal data are achieved. A temporal convolutional autoencoder (TCAE) model is constructed, including: extracting the temporal features and spatial correlation features between multiple sensors from the multi-sensor standard signal data through the separate dual encoder to achieve compression of the multi-sensor standard signal data; transposing the temporal features and spatial correlation features between multiple sensors through the symmetric decoder to output the reconstruction temporal features and reconstruction samples of the multi-sensor standard signal data, thereby achieving reconstruction of the multi-sensor standard signal data; and training the model with the multi-sensor standard signal data as input and the reconstruction samples as the target, aiming to minimize the reconstruction error of the multi-sensor standard signal data, to obtain the temporal convolutional autoencoder (TCAE) model. The above-mentioned method uses a separate dual encoder to extract the temporal dynamic dependencies of multi-sensor standard signal data. It is paired with another independent encoder to capture the spatial correlation features between sensors, achieving accurate fusion of temporal and spatial features. Furthermore, through the reconstruction of multi-sensor standard signal data and training with minimizing reconstruction errors, the Temporal Convolutional Autoencoder (TCAE) model is constructed.

[0032] In one specific implementation, the split dual encoder includes a time encoder and an inter-sensor encoder, and the symmetrical decoder includes an inter-sensor decoder and a time decoder.

[0033] A time encoder is used to compress the time axis dimension and extract the temporal features of multi-sensor standard signal data by employing 1D depthwise convolution, batch normalization (BN), and ReLU activation. Specifically, it uses two layers of 1D depthwise convolution (DV), batch normalization (BN), and ReLU activation to compress only the time axis dimension while retaining the sensor axis dimension. The 1D depthwise convolution calculation formula is shown in Equation 1: Formula 1 In Equation 1, Let f be the convolution output of the j-th sensor standard signal at time step t, and let f be the filter length of the time encoder. Here, b represents the weights of the time encoder filter, and b represents the bias. For the j-th sensor standard signal at time... The standard signal data, where p is a value from 0 to f-1; Here, after the first convolutional layer, the time axis dimension changes from the above-mentioned... Compress to ( (The length of the first layer filter), and then further compressed after the second convolution to... ( (Ensuring the first layer is twice the length of the second layer), extract the temporal features of the multi-sensor standard signal data, and finally output the temporal feature map. (T is the compressed time axis dimension).

[0034] An inter-sensor encoder is used to compress the sensor axis dimension and extract spatial correlation features between multiple sensors by employing 1×1 convolution, batch normalization, and ReLU activation. Two layers of 1×1 convolution, BN, and ReLU activation are used to compress only the sensor axis dimension without changing the time axis dimension. The 1×1 convolution calculation formula is shown in Equation 2: Formula 2 In Equation 2, The output of the k-th sensor channel at time step t. For 1×1 convolution weights, For bias, Let be the time characteristic of the j-th sensor standard signal at time step t.

[0035] Here, the first convolutional layer compresses the number of sensor channels from 23 to... (Optimize variables), second-level compression to Extract spatial correlation features among multiple sensors, and finally output latent vector. .

[0036] The inter-sensor decoder, symmetrical to the inter-sensor encoder, is used to reconstruct the sensor axis dimensions through 1×1 transposed convolutions, outputting the reconstruction time characteristics of multi-sensor standard signal data. Specifically, it uses two layers of 1×1 convolutions to convert the number of channels from... Restored to 23, output .

[0037] The temporal decoder, symmetrical to the temporal encoder, is used to reconstruct the temporal axis dimension through 1D transposed depthwise convolutions, outputting reconstructed samples of standard multi-sensor signal data. Specifically, it restores the temporal axis dimension from T to 60 using two layers of 1D transposed depthwise convolutions (Transposed Depth-wise Conv), outputting reconstructed samples. .

[0038] Using multi-sensor standard signal data as input, and reconstructing samples based on the multi-sensor standard signal data, the model is trained with the goal of minimizing the reconstruction error (RE) of the multi-sensor standard signal data. Here, the reconstruction error is calculated using the root mean square error (RMSE), as shown in Equation 7: Formula 7 In Equation 7, N is the total number of samples of the multi-sensor standard signal data. Let be the input value of the standard signal from the i-th sensor for the n-th sample at time step t. This is the corresponding reconstructed value.

[0039] Step 3: With the goal of minimizing the reconstruction error entropy of multi-sensor standard signal data, Bayesian optimization is used to automatically iteratively optimize the hyperparameters of the temporal convolutional autoencoder (TCAE) model, determine the optimal hyperparameter combination, and obtain the optimal temporal convolutional autoencoder (Opt-TCAE) model.

[0040] Specifically, aiming to minimize the reconstruction error entropy of multi-sensor standard signal data, Bayesian optimization is used to automatically iteratively optimize the hyperparameters of the temporal convolutional autoencoder (TCAE) model, determining the optimal hyperparameter combination and obtaining the optimal Opt-TCAE model. This process includes: randomly generating multiple initial hyperparameter combinations for the split dual encoder and calculating the entropy values ​​of the corresponding initial hyperparameters; using a Gaussian process to fit the mapping relationship between the multiple initial hyperparameter combinations and the entropy values ​​to construct a surrogate model; based on the entropy values ​​of the multiple initial hyperparameters in the surrogate model and the minimum entropy value, selecting the next experimental point using the expected improvement criterion; repeating steps 2 and 3 several times to automatically iteratively optimize the hyperparameters of the TCAE model, ultimately selecting the parameter combination that minimizes the entropy value as the optimal hyperparameter combination, thus obtaining the optimal Opt-TCAE model. Here, the number of times steps 2 and 3 are repeated can be determined based on the actual situation.

[0041] In one specific implementation, the REs (Entropy Representations) of the multi-sensor standard signal data should be centrally distributed. The lower the RE entropy value of the multi-sensor standard signal data, the more concentrated the distribution, and the more accurate the model's representation of the normal state. The entropy calculation formula is shown in Equation 8: Formula 8 In Equation 8, H is the entropy, and m is the number of RE intervals. Let RE be the value of the i-th interval. The probability of.

[0042] With "time encoder filter length" (Second layer) and "Number of 1×1 convolution channels of the encoder between sensors" "(Second layer)" is the optimization variable, and the optimization range is as follows: (Since the number of sensors / 2 ≈ 11). The Bayesian optimization process is as follows: Initialization: Randomly generate 10 initial parameter combinations and calculate the corresponding entropy value H; Constructing a proxy model: Using a Gaussian process (GP) to fit the mapping relationship between the parameters and H; Acquisition function: Selects the next test point based on the expected improvement (EI) criterion.

[0043] In the formula, This is the current minimum entropy value. For expectations; Iterative update: Repeat steps 2-3 a total of 20 times, and finally select the parameter combination that minimizes H as the optimal hyperparameter.

[0044] Step 4: Fit the reconstruction error probability distribution of multi-sensor standard signal data using an adaptive threshold setting method based on kernel density estimation, and confirm the adaptive leak detection threshold.

[0045] Specifically, the adaptive threshold setting method based on kernel density estimation fits the reconstruction error probability distribution of multi-sensor standard signal data to determine the adaptive leak detection threshold, including: fitting the reconstruction error probability distribution of multi-sensor standard signal data based on kernel density estimation to obtain the probability density function, and setting the 99.99% upper quantile of the probability density function as the leak detection threshold. The formula for calculating the probability density function is shown in Equation 3: Formula 3 In Equation 3, Let be the probability density function. h is the number of samples in the validation set, and h is the bandwidth (selected through cross-validation). Gaussian kernel function Let be the reconstruction error value of the i-th sample in the multi-sensor signal input samples. This refers to the reconstruction error of standard signal data from multiple sensors.

[0046] To ensure an extremely low false negative rate, the 99.99% upper quantile of the PDF was set as the leakage detection threshold. ,Right now: .

[0047] Step 5: Input the multi-sensor standard signal data into the optimal temporal convolutional autoencoder Opt-TCAE model, and calculate the reconstruction error of the multi-sensor standard signal data in real time, where: If the reconstruction error of the multi-sensor standard signal data is greater than the adaptive leak detection threshold and the duration is ≥5 minutes, it is determined to be a leak, and the detection result and quantitative evaluation index are output. Otherwise, it is considered normal.

[0048] In one specific implementation, multi-sensor standard signal data is input into the Opt-TCAE model of the optimal temporal convolutional autoencoder, and the reconstruction error RE value of the multi-sensor standard signal data is calculated. .

[0049] Among them, if If the leakage continues to exceed the threshold and the duration is ≥5 minutes (the duration is set to 5 minutes to avoid instantaneous interference), it is judged as "leakage", and the detection results and quantitative evaluation indicators are output. Otherwise, it is judged as "normal".

[0050] The boiler pipeline leakage detection method of this application also includes: calculating the advance detection time by comparing the detection time of the optimal time convolutional autoencoder Opt-TCAE model with the expert detection time; and calculating the false alarm rate by counting the number of times the reconstruction error of the multi-sensor standard signal data is greater than the adaptive leakage detection threshold.

[0051] In one specific implementation, the model detection time is recorded. With the expert testing time Compare and calculate the advance detection time:

[0052] In the reconstruction error of statistical multi-sensor standard signal data Calculate the false alarm rate:

[0053] In the formula, The total normal operating time (hours) is multiplied by 60 and converted to minutes (to match the sampling frequency).

[0054] This application also provides a system for performing the above-described boiler pipeline leakage detection method, comprising: The multi-sensor standard signal data generation module is used to collect and preprocess initial signal data from multiple sensors in real time to generate multi-sensor standard signal data. The Temporal Convolutional Autoencoder (TCAE) model building module is used to design a discrete dual encoder and a symmetric decoder based on multi-sensor standard signal data, and to build a Temporal Convolutional Autoencoder (TCAE) model. The Temporal Convolutional Autoencoder (TCAE) model optimization module is used to automatically iteratively optimize the hyperparameters of the TCAE model using Bayesian optimization with the goal of minimizing the reconstruction error entropy of multi-sensor standard signal data, and determine the optimal hyperparameter combination to obtain the optimal Temporal Convolutional Autoencoder Opt-TCAE model. An adaptive leak detection threshold acquisition module is used to fit the reconstruction error probability distribution of multi-sensor standard signal data based on an adaptive threshold setting method of kernel density estimation, and to confirm the adaptive leak detection threshold. The leakage detection module inputs multi-sensor standard signal data into the optimal temporal convolutional autoencoder Opt-TCAE model and calculates the reconstruction error of the multi-sensor standard signal data in real time, where: If the reconstruction error of the multi-sensor standard signal data is greater than the adaptive leak detection threshold and the duration is ≥5 minutes, it is determined to be a leak, and the detection result and quantitative evaluation index are output. Otherwise, it is considered normal.

[0055] Here, those skilled in the art will understand that the specific methods of the above-described boiler pipeline leak detection system have been referenced above. Figure 1 The description of the boiler pipeline leakage detection method is detailed here, so its repeated description will be omitted.

[0056] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. The scope of patent protection of the present invention shall be determined by the claims. Similarly, any equivalent structural changes made based on the description and drawings of the present invention shall also be included within the scope of protection of the present invention.

Claims

1. A method for detecting leaks in boiler pipelines, characterized in that, Includes the following steps: Real-time acquisition and preprocessing of initial signal data from multiple sensors to generate standard signal data from multiple sensors; Based on standard signal data from multiple sensors, a separate dual encoder and a symmetric decoder are designed, and a temporal convolutional autoencoder (TCAE) model is constructed. With the goal of minimizing the reconstruction error entropy of multi-sensor standard signal data, Bayesian optimization is used to automatically iteratively optimize the hyperparameters of the temporal convolutional autoencoder (TCAE) model, determine the optimal hyperparameter combination, and obtain the optimal temporal convolutional autoencoder (Opt-TCAE) model. An adaptive threshold setting method based on kernel density estimation is used to fit the reconstruction error probability distribution of standard signal data from multiple sensors and confirm the adaptive leak detection threshold. Multi-sensor standard signal data is input into the optimal temporal convolutional autoencoder Opt-TCAE model, and the reconstruction error of the multi-sensor standard signal data is calculated in real time, where: If the reconstruction error of the multi-sensor standard signal data is greater than the adaptive leak detection threshold and the duration is ≥5 minutes, it is determined to be a leak, and the detection result and quantitative evaluation index are output. Otherwise, it is considered normal.

2. The boiler pipeline leakage detection method according to claim 1, characterized in that, The real-time acquisition and preprocessing of initial signal data from multiple sensors to generate standard signal data from multiple sensors includes: Acoustic emission signals and operational signal data are acquired in real time to obtain initial signal data from multiple sensors and construct a time series dataset; Based on time series datasets, moving average filtering and sliding window segmentation and scaling are applied to the initial signal data of multiple sensors to obtain standard signal data of multiple sensors.

3. The boiler pipeline leakage detection method according to claim 2, characterized in that, The process involves applying moving average filtering and noise reduction, along with sliding window segmentation and scaling, to the initial signal data from multiple sensors based on a time-series dataset to obtain standard multi-sensor signal data, including: The short-term high-amplitude noise generated by external interference in the acoustic emission signal data is processed by moving average filtering to reduce noise and smooth the data, and multi-channel sensor noise-reduced signal data is obtained. A sliding window is used to convert the time series corresponding to the time series dataset into time segments, generate multi-sensor signal input samples, and perform min-max scaling on each signal in the multi-sensor noise-reduced signal data and the running signal data to generate multi-sensor standard signal data.

4. The boiler pipeline leakage detection method according to claim 1, characterized in that, The design of a discrete dual encoder and a symmetric decoder based on multi-sensor standard signal data, and the construction of a temporal convolutional autoencoder (TCAE) model, includes: Design a split dual encoder and a symmetric decoder. Based on the temporal features of multi-sensor standard signal data extracted by the split dual encoder and the spatial correlation features between multiple sensors, realize the compression, reconstruction and reconstruction error minimization training of multi-sensor standard signal data, and construct a temporal convolutional autoencoder (TCAE) model.

5. The boiler pipeline leakage detection method according to claim 4, characterized in that, The design employs a separate dual encoder and a symmetric decoder. Based on the temporal features of the multi-sensor standard signal data extracted by the separate dual encoder and the spatial correlation features between the multiple sensors, it achieves compression, reconstruction, and reconstruction error minimization training of the multi-sensor standard signal data. A temporal convolutional autoencoder (TCAE) model is constructed, including: By extracting the temporal features and spatial correlation features between multiple sensors from multi-sensor standard signal data using a separate dual encoder, compression of multi-sensor standard signal data can be achieved. By transposing the temporal characteristics and spatial correlation characteristics between multiple sensors of multi-sensor standard signal data using a symmetric decoder, the reconstruction temporal characteristics and reconstruction samples of multi-sensor standard signal data are output, thereby realizing the reconstruction of multi-sensor standard signal data. Using multi-sensor standard signal data as input, and reconstructing samples based on the multi-sensor standard signal data, the model is trained with the goal of minimizing the reconstruction error of the multi-sensor standard signal data, resulting in the Temporal Convolutional Autoencoder (TCAE) model.

6. The boiler pipeline leakage detection method according to claim 5, characterized in that, The separate dual encoder includes a time encoder and an inter-sensor encoder, and the symmetrical decoder includes an inter-sensor decoder and a time decoder. The time encoder is used to extract the time features of multi-sensor standard signal data by compressing the time axis dimension using 1D depth convolution, batch normalization and ReLU activation. The sensor encoder is used to compress the sensor axis dimension and extract the spatial correlation features between multiple sensors by employing 1×1 convolution, batch normalization and ReLU activation. The inter-sensor decoder, symmetrical to the inter-sensor encoder, is used to reconstruct the sensor axis dimension through 1×1 transposed convolution and output the reconstruction time characteristics of multi-sensor standard signal data. The time decoder, symmetrical to the time encoder, is used to reconstruct the time axis dimension through 1D transposed depth convolution and output multi-sensor standard signal data reconstruction samples. The calculation of the 1D depthwise convolution is shown in Equation 1: Formula 1 In Equation 1, Let f be the convolution output of the j-th sensor standard signal at time step t, and let f be the filter length of the time encoder. Here, b represents the weights of the time encoder filter, and b represents the bias. For the j-th sensor standard signal at time... The standard signal data, where p is a value from 0 to f-1; The calculation of the 1×1 convolution is shown in Equation 2: Formula 2 In Equation 2, The output of the k-th sensor channel at time step t. For 1×1 convolution weights, For bias, Let be the time characteristic of the j-th sensor standard signal at time step t.

7. The boiler pipeline leakage detection method according to claim 1, characterized in that, The method aims to minimize the reconstruction error entropy of multi-sensor standard signal data. Bayesian optimization is used to automatically iteratively optimize the hyperparameters of the temporal convolutional autoencoder (TCAE) model, determining the optimal hyperparameter combination and obtaining the optimal Opt-TCAE model, including: Randomly generate multiple initial hyperparameter combinations for the split dual encoder, and calculate the entropy values ​​of the corresponding multiple initial hyperparameters; A surrogate model is constructed by fitting the mapping relationship between multiple initial hyperparameter combinations and entropy values ​​using a Gaussian process. Based on the entropy values ​​of multiple initial hyperparameters in the surrogate model and the minimum entropy value among them, the next experimental point is selected according to the expected improvement criterion: Repeat steps 2 and 3 several times to perform automated iterative optimization of the hyperparameters of the temporal convolutional autoencoder (TCAE) model. Finally, select the parameter combination that minimizes the entropy value as the optimal hyperparameter combination, and thus obtain the optimal temporal convolutional autoencoder (Opt-TCAE) model.

8. The boiler pipeline leakage detection method according to claim 3, characterized in that, The adaptive threshold setting method based on kernel density estimation fits the reconstruction error probability distribution of multi-sensor standard signal data to determine the adaptive leak detection threshold, including: An adaptive threshold setting method based on kernel density estimation fits the reconstruction error probability distribution of multi-sensor standard signal data to obtain a probability density function, and sets the 99.99% upper quantile of the probability density function as the leakage detection threshold. The formula for calculating the probability density function is shown in Equation 3: Formula 3 In Equation 3, Let be the probability density function. h is the number of samples in the validation set, and h is the bandwidth (selected through cross-validation). Gaussian kernel function Let be the reconstruction error value of the i-th sample in the multi-sensor signal input samples. This refers to the reconstruction error of standard signal data from multiple sensors.

9. The boiler pipeline leakage detection method according to claim 1, characterized in that, The boiler pipeline leakage detection method further includes: The advance detection time is calculated by comparing the detection time of the optimal temporal convolutional autoencoder Opt-TCAE model with the expert detection time. The false alarm rate is calculated by counting the number of times the reconstruction error of the multi-sensor standard signal data exceeds the adaptive leak detection threshold.

10. A system for performing a boiler pipe leakage detection method, characterized in that, include: The multi-sensor standard signal data generation module is used to collect and preprocess initial signal data from multiple sensors in real time to generate multi-sensor standard signal data. The Temporal Convolutional Autoencoder (TCAE) model building module is used to design a discrete dual encoder and a symmetric decoder based on multi-sensor standard signal data, and to build a Temporal Convolutional Autoencoder (TCAE) model. The Temporal Convolutional Autoencoder (TCAE) model optimization module is used to automatically iteratively optimize the hyperparameters of the TCAE model using Bayesian optimization with the goal of minimizing the reconstruction error entropy of multi-sensor standard signal data, and determine the optimal hyperparameter combination to obtain the optimal Temporal Convolutional Autoencoder Opt-TCAE model. An adaptive leak detection threshold acquisition module is used to fit the reconstruction error probability distribution of multi-sensor standard signal data based on an adaptive threshold setting method of kernel density estimation, and to confirm the adaptive leak detection threshold. The leakage detection module inputs multi-sensor standard signal data into the optimal temporal convolutional autoencoder Opt-TCAE model and calculates the reconstruction error of the multi-sensor standard signal data in real time, where: If the reconstruction error of the multi-sensor standard signal data is greater than the adaptive leak detection threshold and the duration is ≥5 minutes, it is determined to be a leak, and the detection result and quantitative evaluation index are output. Otherwise, it is considered normal.

Images