A method for detecting leakage of a tube wall of a boiler heating surface

Abstract

The application provides a boiler heating surface pipe wall leakage detection method, and belongs to the technical field of industrial intelligent detection based on machine learning; the method adopts a heterogeneous perception architecture combining acoustic array positioning of self-adaptive signal decomposition and physical field correction optimized by an improved sparrow search algorithm, extracts complementary features from time domain statistical features and frequency energy spectrum through a specially designed double-flow space-time attention fusion neural network, captures and identifies weak leakage signals in a strong noise environment. Based on the identification result, the sound wave propagation path is grid corrected using a three-dimensional temperature field distribution model of the furnace, a nonlinear positioning equation set is constructed, and three-dimensional spatial sub-m accurate positioning of the leakage sound source is realized without human intervention, completely solving the false positives and false negatives caused by background noise interference in traditional acoustic monitoring and the positioning deviation caused by temperature gradient.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of machine learning-based industrial intelligent detection, and particularly relates to a boiler heating surface pipe wall leakage detection method. BACKGROUND

[0002] The heating surface pipe system of the boiler equipment of a thermal power plant has been running in an extremely harsh working condition of high temperature and high pressure and flue gas scouring for a long time. The heating surface pipe wall leakage is the primary cause of the unscheduled shutdown of the boiler. According to industry statistics, the shutdown accidents caused by four-pipe leakage account for more than 40%. Once a small leakage occurs, if it cannot be discovered and handled in time within the golden window period of 10 to 30 minutes, the high-pressure jet flow of up to 15 MPa to 25 MPa will quickly blow away the adjacent pipe row, causing a chain reaction of pipe burst accidents, causing huge economic losses and seriously threatening the safety of on-site personnel.

[0003] The detection methods commonly used in the industrial field at present have significant defects. The traditional manual listening method seriously depends on the personal experience of the inspection personnel, cannot realize 24-hour all-weather real-time monitoring, and often cannot distinguish the weak leakage sound with a sound intensity lower than 10 dB in the furnace. Although the conventional acoustic monitoring technology realizes non-contact measurement, during the operation of the boiler, the furnace is filled with combustion noise with a frequency of 50 Hz to 500 Hz, blow ash noise of up to 110 dB, and fan airflow noise. The sound pressure level of these strong background noise is often much higher than the jet flow sound signal generated by early small leakage, causing the leakage characteristics to be completely submerged, and the signal-to-noise ratio of the system is usually lower than 5 dB, which is prone to false positives or false negatives. In addition, most of the existing positioning methods are based on the assumption of ideal normal temperature sound speed (about 340 m / s), ignoring the nonlinear influence of the extremely uneven temperature field distribution inside the furnace on the sound wave propagation speed, resulting in a sound source positioning error of usually more than 2 m, which cannot provide accurate fault point coordinates for maintenance personnel.

[0004] Therefore, there is an urgent need for an intelligent detection method that can adaptively filter out strong background colored noise, automatically mine weak leakage characteristics from massive data, and realize sub-meter high-precision positioning combined with temperature field distribution. SUMMARY

[0005] In view of the above problems, the present application proposes a boiler heating surface pipe wall leakage detection method based on adaptive modal decomposition and spatio-temporal attention fusion network. This method is not only a simple threshold discrimination of acoustic monitoring data, but also a complete closed loop of adaptive anti-noise perception, heterogeneous feature deep reasoning, dual-flow spatio-temporal attention fusion, and multi-physical field coupling positioning, thereby solving the problem of the blind area of weak signal perception of traditional acoustic methods in dealing with strong background noise interference, and the spatial positioning deviation problem of the traditional normal temperature sound speed model in the non-uniform temperature field environment.

[0006] The application provides a boiler heating surface pipe wall leakage detection method, comprising the following processes:

[0007] S1, analog voltage signals are collected based on omnidirectional deployment of an acoustic sensor array and a piezoelectric ceramic sensor array, and one-dimensional acoustic time history signals are formed after preprocessing;

[0008] S2, a parameter optimization model minimizing average envelope entropy is constructed with the acoustic time history signals as input, a sparrow search algorithm introducing chaos mapping and Cauchy variation is adopted to search for optimal parameters of variational mode decomposition, intrinsic mode components are obtained by performing variational mode decomposition on the acoustic time history signals with the optimal parameters, mode screening and leakage signal reconstruction are performed, and an enhanced leakage signal is output;

[0009] S3, the enhanced leakage signal is subjected to double-branch processing, feature vectors extracted by a frequency domain feature extraction branch and a time domain feature extraction branch are fused through a cross-attention feature fusion module, and then a probability distribution vector of a leakage state is output through a full connection layer;

[0010] S4, when a total leakage probability exceeds a threshold value, an alarm is given and data are locked; sensor nodes are selected to participate in positioning, a generalized cross-correlation phase transform algorithm is used to extract a signal time difference; a non-uniform three-dimensional sound velocity grid model is constructed in combination with a furnace temperature field; a ray tracing method is used to calculate a predicted time difference; a nonlinear residual error objective function is jointly constructed, and a particle swarm algorithm is used for iterative solving to inverse a three-dimensional position coordinate of a leakage source.

[0011] Preferably, S1 specifically comprises: acoustic sensor arrays and piezoelectric ceramic sensors are deployed in a distributed topology layout scheme, analog voltage signals containing high-frequency jet noise generated by pipe wall rupture are collected, and the signals are transmitted to a high-speed synchronous data acquisition card after being conditioned by a preamplifier; secondly, the data acquisition card performs analog-to-digital conversion on the signals to obtain discrete voltage sequences, and converts the sequences into discrete sound pressure sequences based on a sensitivity coefficient, and then splices the sequences to generate original acoustic observation data in channel order; finally, an infinite impulse response band-pass filter is used to filter out power frequency and low frequency interference, and one-dimensional preprocessed acoustic time history signals used for subsequent algorithm processing are formed within a locked time window.

[0012] Preferably, in S2, a parameter optimization model minimizing average envelope entropy is constructed, a sparrow search algorithm introducing chaos mapping and Cauchy variation is adopted to search for optimal parameters of variational mode decomposition, and the specific process is as follows:

[0013] The acoustic time history signals are decomposed by variational mode decomposition VMD Two key parameters, the number of modes And the penalty factor Are determined; the envelope entropy of each modal component obtained by VMD decomposition is taken as an evaluation index: the first The envelope entropy of each modal component is denoted as... and with average envelope entropy As a fitness index for the sparrow search algorithm;

[0014] In the initialization phase, a Logistic chaotic mapping is introduced to obtain the initial sparrow positions. ,in , ;

[0015] Using the first dimension component For the number of modes Encode it and map it to a preset integer range. ; Utilizing the second dimension component For the penalty factor Encode the values ​​and map them to a preset real number range. ;

[0016] Sparrow position The corresponding meaning is, to A set of candidate solutions for performing VMD decomposition Introducing Cauchy variation during individual location update, i.e., in the updated parameter combination... Apply random perturbations to enable the search to escape local minima; iterative search ultimately yields a fitness index that satisfies the fitness metric. The minimum optimal parameters are denoted as follows: and .

[0017] Preferably, in step S2, variational mode decomposition is performed on the acoustic time history signal using optimal parameters to obtain intrinsic mode components, followed by mode screening and leakage signal reconstruction. The specific process is as follows:

[0018] Use optimal parameters and right Perform VMD decomposition, Decomposition The intrinsic mode components; the first Each modal component is denoted as... And at the same time, its corresponding center frequency is obtained, denoted as . ,Right now For the first Modal center angular frequency;

[0019] Then calculate the first one respectively. Modal components and correlation coefficient With the Modal components cliff ; select the modal that satisfies both and as the effective leakage modal set, denoted as S, if S is empty, adopt the bottom-up strategy: calculate the comprehensive score for each modal, select the top modals with the highest score to form set S; wherein the correlation coefficient threshold is denoted as , set ; the kurtosis threshold is denoted as , set ;

[0020] Finally, linearly superimpose the modal components in set S to reconstruct the enhanced leakage signal .

[0021] Preferably, the frequency domain feature extraction branch is specifically:

[0022] The enhanced leakage signal output by S2 is subjected to short-time Fourier transform to generate a two-dimensional time-frequency spectrum reflecting the energy distribution of the signal with time and frequency, which is defined as an input tensor ; input to a feature extraction network comprising 4 cascaded convolution blocks, each convolution block comprising a convolution layer, a batch normalization layer and a ReLU activation layer, and the convolution operation extracts local texture features through a set of learnable filters, which is mathematically expressed as:

[0023] ;

[0024] wherein is a weight matrix, is a bias term, represents the feature map tensor output by the layer convolution operation, represents the input feature map of the previous layer, represents a linear rectification activation function; after feature extraction by the fourth convolution block, a feature map containing high-level semantic information is obtained; a global average pooling operation is performed on to calculate the average response value of each channel in the spatial dimension, generating a frequency domain feature vector .

[0025] Preferably, the time domain feature extraction branch is specifically:

[0026] The enhanced leakage signal output by S2 is subjected to full-dimensional time domain statistical analysis to sequentially calculate 12 physical statistical feature quantities including mean, root mean square, variance, standard deviation, peak value, peak-to-peak value, rectified mean value, skewness, kurtosis, waveform factor, pulse factor and margin factor;

[0027] The calculated 12-dimensional feature data are stacked in chronological order to construct a time-domain feature sequence matrix with a time step of 50. The sequence matrix is ​​input into a bidirectional long short-term memory network (BiLSTM) for each time step. The forward layer calculates the hidden state. Backward layer calculates hidden states The hidden states in the two directions are concatenated to obtain the combined state vector at the current time step. Extract the output of the last time step as the temporal feature vector. .

[0028] Preferably, S3 outputs a four-class probability vector. ,in This represents the probability of the normal state. , , These represent the probabilities of early leakage, moderate leakage, and pipe rupture, respectively; the probabilities of the three leakage-related categories are combined to form the total leakage probability. ,in ;when If the acoustic data exceeds the preset safety threshold of 0.85 for three consecutive seconds, a leakage event is determined to have occurred. An audible and visual alarm is immediately triggered, and the acoustic data within that time window is automatically locked, initiating the location procedure.

[0029] Preferably, in step S4, sensor node pairs are selected to participate in the positioning process, specifically as follows:

[0030] Record the spatial coordinates of each sensor node during deployment, and record the first... The coordinates of each node are The coordinates are stored in the controller; during positioning, the coordinate information is directly called to participate in the calculation.

[0031] Select at least four nodes from the sensors to participate in the localization process; denoted as the set of selected nodes. ;from Multiple sets of node pairs are constructed by pairwise combinations to obtain observations of time difference of arrival. Let the set of node pairs be denoted as . , of which This indicates the selection of the first node from the not less than four nodes. The node and the first A pair of nodes consisting of 1 node; through the 1 node pair; The time delay observations of multiple sets of nodes are jointly solved to achieve three-dimensional positioning.

[0032] Preferably, in step S4, for each node participating in the localization... The enhanced time history signal is captured within the locked time window and denoted as... ; where superscript representing the lock time window, , representing the sampling point sequence number within the window; the enhanced leakage signal output by S2 is taken out and windowed at the node ;

[0033] the frequency domain transform is performed on the to obtain its frequency domain representation , wherein represents the angular frequency; for any pair of nodes , the generalized cross-correlation phase transform algorithm is used to calculate the cross-correlation function of the signals corresponding to the nodes and the nodes ; ;

[0034] the time delay at which the peak of is located is taken as the measured time difference of arrival, denoted as ;

[0035] introducing a three-dimensional temperature field and constructing a sound velocity grid: introducing the real-time temperature field data of the furnace , the temperature field data is provided in real time by the existing temperature monitoring system of the furnace, and is interpolated and gridded in space to form a three-dimensional temperature grid, and then the temperature field is converted into a sound velocity field according to the thermodynamic relationship , to obtain a three-dimensional sound velocity grid model;

[0036] calculating the predicted time difference by ray tracing: the upper and lower bounds of the coordinates are given according to the structure size of the boiler furnace, thereby defining the candidate search range; for any candidate point , the is calculated and the of each node pair is obtained; after obtaining the sound velocity field , the ray tracing method is used to calculate the propagation time in the non-uniform medium; for any candidate space point and any sensor node , the propagation time of the sound wave from to is calculated ; the model predicted time difference of arrival of the node is obtained ; wherein represents the spatial coordinates of the th sensor node participating in positioning.

[0037] Preferably, the nonlinear residual error objective function is jointly constructed in S4, and the particle swarm algorithm is used for iterative solution, and the specific process is as follows:

[0038] Computing the measured time delay difference The deviation between the model predicted time delay difference The square sum of the deviation obtained by calculating all selected node pairs

[0039] Adopting the particle swarm optimization algorithm to iteratively search in the defined candidate search range, so that Take the minimum value to obtain the optimal point The point is the three-dimensional position of the leakage source obtained by inversion.

[0040] The innovation points and beneficial effects of the present application include:

[0041] (1) Self-adaptive modal decomposition and high-fidelity leakage wave field reconstruction: a self-adaptive optimization method of variational modal decomposition based on improved sparrow search algorithm is proposed, which can adaptively decompose the complex mixed sound field in the furnace into several eigenmode functions within 0.1 seconds. By minimizing the envelope entropy objective function, the colored background noise generated by combustion and blowing is accurately stripped, and a high-fidelity leakage feature dataset with a signal-to-noise ratio improvement of more than 10dB is constructed, which breaks through the limitation that small leakage features are difficult to extract in a strong interference environment;

[0042] (2) Dual-flow space-time attention diagnosis network enhanced by physical mechanism: a parallel deep neural network architecture is designed, which includes a 4-layer convolutional neural network branch for processing frequency domain spectrum and a bidirectional long short-term memory network branch for processing time domain sequence. By introducing a cross-attention mechanism, the network can adaptively allocate weights between time-frequency two modes according to feature saliency, for example, focusing on time domain features when capturing transient impact, and focusing on frequency domain features when analyzing wideband energy distribution, and combining with a physical constraint loss function, realizing robust prediction of early leakage, moderate leakage and burst state under small sample working conditions;

[0043] (3) Non-uniform medium sound source positioning based on furnace temperature field correction: a dynamic mapping model from furnace temperature distribution to sound propagation speed is established. Unlike the traditional constant temperature and uniform sound speed assumption, the present application divides the furnace space into a 0.5m resolution sound speed grid according to the real-time monitored and simulated three-dimensional temperature field data of the furnace, and corrects the time delay of sound wave propagation in high-temperature non-uniform medium in real time. The generalized cross-correlation phase transform algorithm is used to calculate the time delay and solve the nonlinear positioning equation set, which reduces the positioning error from the traditional 2m to within 0.5m, realizing accurate visualization inversion of the leakage point. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 The overall method flowchart of the present application. ​​​

[0045] Figure 2 This is a flowchart illustrating the signal processing of VMD based on the improved sparrow search algorithm of this invention.

[0046] Figure 3 This is a diagram of the dual-stream spatiotemporal attention fusion neural network structure of the present invention.

[0047] Figure 4 This is a comparison diagram of the time-domain waveforms of the original signal and the adaptive enhancement signal in an embodiment of the present invention;

[0048] (a) represents the original high-noise frequency signal collected on-site; (b) represents the separated background noise component; and (c) represents the leakage characteristic signal after optimized VMD reconstruction.

[0049] Figure 5 This is a heatmap of the confusion matrix for multi-model fault diagnosis in an embodiment of the present invention.

[0050] Figure 6 This is a spatial distribution diagram of the three-dimensional positioning error of the sound source in an embodiment of the present invention. Detailed Implementation

[0051] The present invention proposes a method for detecting leaks in the tube walls of boiler heating surfaces, the overall process of which is as follows: Figure 1 As shown.

[0052] S1. High-fidelity wave field acquisition and signal preprocessing of the high-temperature acoustic array in the furnace: This step first involves hardware layout. The analog voltage signals output by the high-temperature acoustic sensor array and the piezoelectric ceramic sensor are pre-amplified and then transmitted to the high-speed synchronous data acquisition card. Second, the acquisition card performs analog-to-digital conversion on the signals to obtain discrete voltage sequences, and converts them into discrete sound pressure sequences based on sensitivity coefficients. These sequences are then spliced ​​together in channel order to generate the original acoustic observation data. Finally, an infinite impulse response bandpass filter is used to filter out power frequency and low-frequency interference, and a one-dimensional preprocessed acoustic time history signal is formed within a locked time window for subsequent algorithm processing, serving as input data for subsequent processing.

[0053] S2. Signal Enhancement Method Based on Improved Sparrow Search Algorithm: This step takes the preprocessed acoustic time history signal as input. First, a parameter optimization model is established with the goal of minimizing the average envelope entropy. Second, an improved sparrow search algorithm incorporating chaotic mapping and Cauchy mutation is used to iteratively search within the parameter space, outputting the optimal number of modes and the optimal penalty factor required for variational mode decomposition. Third, variational mode decomposition is performed on the same preprocessed time history signal using the optimal parameters to obtain several intrinsic mode components, and the correlation coefficient and kurtosis of each component are calculated to characterize their correlation with the leakage impact component. Finally, effective modes that meet the threshold conditions are selected and linearly superimposed for reconstruction, outputting the enhanced leakage signal.

[0054] S3. Constructing a dual-flow spatio-temporal attention deep neural network to realize feature fusion and boiler leakage type identification: This step takes the enhanced leakage signal as input. First, in the frequency domain branch, the signal is subjected to short-time Fourier transform to generate a two-dimensional time-frequency spectrum, and a convolutional neural network is used to extract a frequency domain feature vector. Second, in the time domain branch, 12 physical statistical characteristic quantities of the signal are calculated to construct a time domain feature sequence matrix, which is input into a bidirectional long short-term memory network to extract a time domain feature vector. Third, a cross-attention mechanism is used to align and weight fuse the frequency domain and time domain features to generate a fused feature vector. Finally, the fused feature is input into a fully connected classification layer to output a probability distribution vector containing four states of normal, early leakage, moderate leakage and burst, which serves as the basis for online monitoring and positioning triggering.

[0055] S4. Precise positioning of sound source combined with three-dimensional temperature field correction and online monitoring: This step first monitors the total probability of leakage in the probability distribution vector in real time. When it exceeds a preset threshold in consecutive time, an alarm is triggered and the current time window data is locked. Second, at least four nodes are selected from the deployed sensor array to participate in positioning, and multiple node pairs are constructed by combining the nodes in pairs. The generalized cross-correlation phase transform algorithm is used to process the enhanced leakage signal in the locked time window to extract the measured time difference of arrival of each node pair for joint solution. Third, the real-time temperature field data of the furnace is introduced and converted into a non-uniform three-dimensional sound speed grid model. The ray tracing method is used to calculate the propagation time of sound waves in the non-uniform medium to obtain the model predicted time difference of arrival of each node pair. Finally, a nonlinear residual error objective function is constructed between the measured time difference of arrival and the model predicted time difference of arrival, and a particle swarm algorithm is used to search for the solution that minimizes the residual error in the candidate space range. The three-dimensional position coordinates of the leakage source are inverted and output.

[0056] The specific implementation of the present application will be further described below in conjunction with specific embodiments.

[0057] S1, high-fidelity wave field acquisition and signal preprocessing of high-temperature-resistant acoustic array in furnace

[0058] First, the high-temperature-resistant array hardware layout and acquisition parameters are defined. In order to balance the coverage of the whole space of the furnace and the geometric accuracy of the sound source positioning, the system adopts a distributed topology layout scheme, with 16 acoustic sensor arrays evenly deployed at different elevations on the front wall, rear wall, left wall and right wall of the boiler. For the high-temperature environment of more than 1000°C inside the furnace, a piezoelectric ceramic sensor equipped with an austenitic stainless steel waveguide with a length of 800mm is selected. The waveguide is fixed to the outside of the furnace wall by a flange and extends through the insulation layer to the inner wall of the furnace, and is specially responsible for picking up the high-frequency jet noise generated by the pipe wall rupture.

[0059] The analog voltage signals output from all acoustic wave sensor arrays and piezoelectric ceramic sensors are conditioned by a low-noise preamplifier with a gain of 60dB, and then transmitted to a high-speed synchronous data acquisition card with a sampling frequency of 50,000Hz for multi-channel synchronous sampling. This sampling frequency is set to ensure that the effective high-frequency components up to 20,000Hz in the leakage signal can be completely captured. According to the Nyquist sampling theorem, the sampling frequency must be greater than twice the highest frequency of the signal to avoid aliasing distortion. Given the upper limit of the frequency band of interest in this system (20,000Hz), the theoretical lower limit of sampling is 40,000Hz. The 50,000Hz sampling rate selected in this embodiment provides a redundancy safety margin, thereby physically ensuring the integrity and reconfigurability of the original wavefield information.

[0060] The high-speed synchronous data acquisition card performs analog-to-digital conversion on the conditioned analog voltage signals of each channel to obtain the discrete voltage time sequence for each channel. In this embodiment, the sensor array includes 16 high-temperature resistant acoustic wave sensors and 1 piezoelectric ceramic sensor equipped with a waveguide, for a total of 17 sensors corresponding to 17 acquisition channels. Therefore, the total number of channels is taken as... Record the first The discrete voltage sequence of each acquisition channel is as follows ,in , Indicates the channel number. , To record the number of sampling points within the window. Furthermore, when At that time, the first The channel corresponds to the first A high-temperature resistant acoustic wave sensor; when At that time, the 17th channel corresponds to the piezoelectric ceramic sensor.

[0061] To obtain the sound pressure time history data for each channel, the voltage-sound pressure sensitivity coefficient was based on the sensor's factory calibration. (Unit is) ), to the discrete voltage sequence Converted to discrete sound pressure sequence and satisfy The discrete sound pressure sequences of all channels are concatenated in channel order to generate the original acoustic observation vector. .

[0062] To address the electromagnetic interference and low-frequency combustion noise present in the furnace environment, an infinite impulse response bandpass filter is used to filter the... Preprocessing was performed, with the passband range set to 1000Hz to 20000Hz to filter out 50Hz power frequency interference and low-frequency combustion noise. Obtained by bandpass filtering .in, This refers to the one-dimensional preprocessed acoustic time history signal used for VMD decomposition within the locked time window; subsequent steps all use this signal. For input.

[0063] S2. Signal enhancement method based on improved sparrow search algorithm

[0064] The process is as follows: Figure 2 As shown, taking the acoustic time-history signal as input, a parameter optimization model is first established with the goal of minimizing the average envelope entropy. Secondly, an improved sparrow search algorithm incorporating chaotic mapping and Cauchy mutation is used to iteratively search within the parameter space, outputting the optimal number of modes and the optimal penalty factor for variational mode decomposition. Thirdly, the acoustic time-history signal is decomposed using the optimal parameters to obtain several intrinsic mode components, and the correlation coefficient and kurtosis of each component are calculated. Finally, effective modes that meet the threshold conditions are selected and linearly superimposed to reconstruct the signal, outputting an enhanced leakage signal.

[0065] (1) Establish the VMD parameter adaptive optimization objective:

[0066] The preprocessed acoustic signal vector output by S1 is uniformly denoted as... Variational mode decomposition (VMD) is used to... When performing decomposition, two key parameters need to be determined: the number of modes. With penalty factor Since leakage signals typically exhibit sparse impulsive characteristics, while background noise is closer to a high-entropy random process, this embodiment uses the envelope entropy of each modal component obtained from VMD decomposition as an evaluation index: The first... The envelope entropy of each modal component is denoted as... and with average envelope entropy (for all) The average value is used as a fitness index for the Sparrow Search algorithm, which makes it more likely to obtain parameter combinations that highlight the impact structure and suppress random noise during the parameter search process.

[0067] (2) Improved Sparrow Search Algorithm (SSA) Initialization and Parameter Mapping:

[0068] To obtain the optimal parameter combination globally, an improved sparrow search algorithm is adopted. To improve the initial population diversity, a Logistic chaotic mapping is introduced in the initialization phase to obtain the initial sparrow positions. ,in , ;

[0069] Using the first dimension component For the number of modes Encode it and map it to a preset integer range. For example, using ; Utilizing the second dimension component For the penalty factor Encode the values ​​and map them to a preset real number range. For example, using ;

[0070] Therefore, the sparrow's position The corresponding meaning is: to A set of candidate solutions for performing VMD decomposition Furthermore, to avoid the search getting trapped in local optima, Cauchy mutation is introduced during the individual position update process: that is, in the updated parameter combination... A random perturbation is applied to the parameter space to allow the search to escape local minima. The algorithm iteratively searches within the parameter space to ultimately obtain the fitness index. The minimum optimal parameters are denoted as follows: and .

[0071] (3) Perform VMD decomposition using optimal parameters:

[0072] Use optimal parameters and right Perform VMD decomposition, Decomposition The intrinsic mode components. Each modal component is denoted as... And at the same time, its corresponding center frequency is obtained, denoted as . ,Right now For the first Modal center angular frequency.

[0073] (4) Modal screening and leakage signal reconstruction:

[0074] After obtaining each modal component Subsequently, to select modes more likely to contain leakage impact components, two types of indices were calculated:

[0075] No. Modal components and correlation coefficient The larger the correlation coefficient, the more consistent the main changes of the mode are with the original signal;

[0076] No. Modal components cliff The greater the kurtosis, the more likely the mode is to contain sparse impact components of leakage acoustic emission.

[0077] Select those that simultaneously satisfy and The modes are taken as the effective leakage mode set, denoted as S. If S is empty (i.e., no mode simultaneously meets the threshold condition), a fallback strategy is adopted: a comprehensive score is calculated for each mode (the comprehensive score is calculated by taking the comprehensive score of each mode as the threshold condition). and (After normalization and weighting), select the top scorers. The modalities constitute a set S to ensure the stability and executable nature of the process. The correlation coefficient threshold is denoted as... ,set up ; kurtosis threshold is denoted as ,set up .

[0078] Finally, the enhanced leakage signal is obtained by linearly superimposing and reconstructing the modal components in set S, and is uniformly denoted as . Specifically, the linear superposition reconstruction process involves adding the modes in set S point by point along the same time axis. Through the joint screening and reconstruction of correlation and impact mentioned above, steady-state combustion noise and low-frequency airflow interference can be effectively suppressed, while retaining the abrupt acoustic signature features induced by leakage, thereby providing input data with a higher signal-to-noise ratio for subsequent deep learning models.

[0079] S3. Construct a dual-stream spatiotemporal attention fusion neural network to achieve feature fusion and boiler leak type identification.

[0080] Two-stream spatiotemporal attention fusion neural network structure such as Figure 3 As shown, taking the enhanced leakage signal as input, firstly, a short-time Fourier transform is performed on the enhanced leakage signal in the frequency domain branch to generate a two-dimensional time-frequency spectrum, and the frequency domain feature vector is extracted through a convolutional neural network; secondly, 12 physical statistical features of the enhanced leakage signal are calculated in the time domain branch and a time domain feature sequence matrix is ​​constructed, which is then input into a bidirectional long short-time memory network to extract the time domain feature vector; thirdly, a cross-attention mechanism is used to align and weight the frequency domain and time domain features to generate a fused feature vector; finally, the fused features are input into a fully connected classification layer, and the output is a probability distribution vector containing four states: normal, early leakage, moderate leakage, and pipe burst.

[0081] (1) Constructing the frequency domain feature extraction branch: for the output of S2 A short-time Fourier transform with a window length of 256 and an overlap rate of 50% is performed to generate a two-dimensional time-frequency spectrum reflecting the distribution of signal energy over time and frequency, which is defined as the input tensor. .Will The input is fed into a feature extraction network containing four cascaded convolutional blocks. Each convolutional block contains a convolutional layer, a batch normalization layer, and a ReLU activation layer. The convolutional operation extracts local texture features through a learnable filter bank, mathematically expressed as:

[0082] ;

[0083] in This is the weight matrix. For bias terms, Representing the The feature map tensor output by the convolution operation. This represents the input feature map of the previous layer. This represents the linear rectified activation function. After feature extraction in the fourth convolutional block, a feature map containing high-level semantic information is obtained. .

[0084] To reduce dimensionality while preserving key features, Perform a global average pooling operation to calculate the average response value of each channel in the spatial dimension, generating a frequency domain feature vector. This vector highly condenses the broadband energy distribution pattern of the leakage signal in the frequency domain.

[0085] (2) Constructing the temporal feature extraction branch: for the output of S2 A full-dimensional time-domain statistical analysis was performed, calculating a total of 12 physical statistical characteristics of the signal, including mean, root mean square, variance, standard deviation, peak value, peak-to-peak value, rectified average, skewness, kurtosis, waveform factor, impulse factor, and margin factor. These characteristics quantified the dynamic patterns of the leakage signal from four dimensions: central tendency, dispersion, distribution pattern, and impulse characteristics.

[0086] The 12-dimensional feature data obtained from the above calculations are stacked in chronological order to construct a time-domain feature sequence matrix with a time step of 50. The sequence is then input into a bidirectional long short-term memory (BiLSTM) network. This network utilizes a gating mechanism to address the long sequence dependency problem, for each time step... The forward layer calculates the hidden state. Backward layer calculates hidden states The hidden states from the two directions are concatenated to obtain the combined state vector at that moment. Extract the output of the last time step as the temporal feature vector. This vector physically fully characterizes the dynamic impact law and statistical distribution characteristics of the leakage signal as it evolves over time.

[0087] (3) In order to make full use of the complementarity between time domain and frequency domain information, this embodiment designs a cross-attention feature fusion module: firstly, the time domain feature vector is fused with the frequency domain information. A fully connected layer is used to map to 512 dimensions to achieve the same effect as the frequency domain feature vector. Dimensional alignment yields aligned temporal features. ; then with As a query vector, As the key vector (Key) and the value vector (Value), the cross-attention weight matrix is calculated , which is used to quantify the attention degree of the frequency domain features to each component of the time domain features in fusion, The calculation formula is:

[0088] ;

[0089] Among them is a scaling factor (corresponding to the dimension of the key vector) to stabilize the attention distribution; further multiply the weight matrix with the value vector and introduce the residual connection to obtain the fusion feature vector , specifically:

[0090] ;

[0091] Among them, the residual term is used to retain the frequency domain discriminative information and improve the fusion stability.

[0092] Finally, the fusion feature vector is input into the fully connected layer, and the probability distribution vector corresponding to the normal, early leakage, moderate leakage and burst pipe four states is output .

[0093] S4, sound source accurate positioning and online monitoring method combined with three-dimensional temperature field correction

[0094] (1) Real-time leakage state discrimination and alarm triggering:

[0095] During online monitoring, the controller reads the four-class probability vector output by step S3 at a fixed time step , wherein represents the probability of the normal state, , , respectively represent the probabilities of early leakage, moderate leakage and burst pipe state. In order to avoid inconsistent judgments caused by jumps between early / moderate / burst pipe three categories, the probabilities of the three leakage related categories are combined into the total leakage probability , wherein . When is greater than the preset safety threshold 0.85 (i.e. always satisfies >0.85 in the 3-second continuous discrimination window), the system determines that a leakage event is occurring, immediately triggers a level one sound and light alarm, and automatically locks the acoustic data in the time window, and starts the positioning program.

[0096] (2) Sensor node and position information calling:

[0097] The positioning phase selects a set of sensor nodes for positioning from the deployed acoustic / piezoelectric sensor array, denoted as wherein denotes the sensor node participating in positioning. To ensure the reproducibility of positioning, the spatial coordinates of each sensor node are recorded when the system is deployed, denoted as the coordinates of the node are recorded and stored in the controller; the coordinate information is directly called to participate in the calculation during positioning.

[0098] In addition, to improve the stability of three-dimensional positioning, the embodiment preferably selects no less than 4 nodes from the deployed sensors to participate in positioning, denoted as On this basis, a plurality of node pairs are constructed from for obtaining time difference of arrival observations, denoted as wherein any denotes a node pair consisting of the node and the node selected from the no less than 4 nodes; by jointly solving the time delay observations of the plurality of node pairs in , three-dimensional positioning is achieved, rather than relying on the positioning of any two nodes alone.

[0099] (3) Time difference of arrival extraction based on GCC-PHAT:

[0100] For each node participating in positioning , its enhanced time history signal is intercepted within a lock time window, denoted as .Wherein the superscript denotes the lock time window, , denotes the sampling point number in the window, is the number of sampling points in the time window; the enhanced leakage signal output by step S2 is taken out at the node and windowed to obtain.

[0101] The is subjected to frequency domain transformation to obtain its frequency domain representation , wherein denotes the angular frequency. For any pair of nodes , the generalized cross-correlation phase transform algorithm (GCC-PHAT) is used to calculate the cross-correlation function of the signals corresponding to the node and the node :Specifically, the algorithm constructs the cross-power spectrum and amplitude normalization, and then transform back to time delay domain to get , so as to suppress the interference of reverberation and amplitude fluctuation on time delay estimation. Wherein represents the complex conjugate.

[0102] Finally, the time delay where the peak of is located is taken as the measured time difference of arrival, denoted as (through searching to make maximum). ).

[0103] (4) Introduce three-dimensional temperature field and construct sound speed grid:

[0104] Considering that the temperature distribution inside the furnace is uneven (for example, the combustion core area can reach about 1500°C, and the water wall area is about 400°C), if the constant temperature sound speed is directly used, it will cause significant positioning error, therefore, the real-time temperature field data of the furnace is introduced to correct the sound speed. The temperature field data is provided by the existing temperature monitoring system of the furnace in real time, and is interpolated / grided in space to form a three-dimensional temperature grid with a resolution of 0.5m. Then, according to the thermodynamic relationship, the temperature field is converted to the sound speed field:

[0105] ;

[0106] , wherein is the specific heat ratio of the gas, is the gas constant, is the real-time temperature field data, and is the absolute temperature (K). The local sound speed is calculated for each grid point, so as to obtain a three-dimensional sound speed grid model for subsequent propagation time modeling.

[0107] (5) Calculate the predicted time difference by ray tracing method:

[0108] According to the structure and size of the boiler furnace, the upper and lower bounds of are given, so as to define the candidate search range;

[0109] For any candidate point , calculate , and get the of each node pair; the process includes: after obtaining the sound speed field , the propagation time in non-uniform medium is calculated by ray tracing method: for any candidate space point and any sensor node , the propagation time of sound wave from to is calculated . According to this, the node The model predicted time difference of arrival can be obtained . Wherein represents the spatial coordinates of the first participating in positioning sensor nodes, the coordinates are measured and stored in the sensor deployment stage.

[0110] (6) Construct a positioning objective function, and solve the positioning:

[0111] Based on And Calculate the nonlinear residual error objective function To invert the leakage source position; Specifically, calculate the deviation between the measured time delay difference And the model predicted time delay difference Between , and calculate the sum of squares of the deviation obtained for all selected node pairs ; ;

[0112] The particle swarm optimization algorithm is used to search iteratively within the defined candidate search range, so that Take the minimum value, get the optimal point , which is the three-dimensional position of the inversion of the leakage source.

[0113] In summary, after the multi-channel acoustic signals are collected and preprocessed by S1 and enhanced by S2, the deep learning model of S3 is used to output Realize the online judgment of boiler heating surface pipe wall leakage; When the total probability of leakage Trigger threshold condition, the system automatically locks the abnormal period and completes the three-dimensional positioning combined with the temperature field corrected propagation time model, so as to obtain the specific abnormal position in the furnace , Provide basis for maintenance decision and rapid disposal.

[0114] Experimental analysis:

[0115] In order to comprehensively verify the effectiveness and reliability of the boiler heating surface tube wall leakage detection method based on adaptive modal decomposition and spatio-temporal attention fusion network proposed in the present application, a set of high-fidelity boiler leakage acoustic simulation experiment platform is constructed in the present embodiment. The platform is built based on the 1:10 scale of the furnace geometry of a 600 MW supercritical coal-fired unit, and the hardware system includes 16 high-temperature sound wave sensors and 1 piezoelectric ceramic sensor, a multi-channel high-speed data acquisition card, and a heating device simulating the high-temperature environment of the furnace. The experimental objects cover the key heating surface areas such as water wall, superheater and reheater, and simulate the micro leakage of different aperture from 0.5mm to 5.0mm and the complex working conditions of background noise intensity from 80dB to 110dB. A total of 10000 test data are collected, and the three core indicators of signal enhancement ability in strong noise background, intelligent recognition accuracy of different leakage degrees and accuracy of three-dimensional space sound source positioning are evaluated.

[0116] 1. Weak leakage signal enhancement experiment analysis under strong noise background

[0117] The present experiment aims to quantitatively evaluate the ability of the improved sparrow search algorithm optimized variational modal decomposition technology to extract weak leakage features in low signal-to-noise ratio environment. The test set selects extreme working condition samples with background noise intensity up to 105dB and leakage sound intensity of only 90dB, at this time the signal-to-noise ratio is -15dB, and the leakage feature is completely submerged by combustion and airflow noise. We compare the adaptive signal processing method proposed in the present application with the traditional band-pass filtering method and the fixed parameter variational modal decomposition method, and calculate the kurtosis value and signal-to-noise ratio gain of the processed signal.

[0118] As shown in the original signal and adaptive enhanced signal time domain waveform comparison chart shown in Figure 4 , the gray curve in the figure represents the original strong noise signal collected on site, and the red solid line represents the reconstructed signal after processing by the method of the present application. The experimental waveform clearly shows the denoising effect of different methods. In the original signal, the impact component generated by the leakage is completely covered by the high-amplitude random noise and cannot be identified by the naked eye or simple threshold. After processing by the method of the present application, the low-frequency combustion noise and high-frequency electromagnetic interference in the background are effectively stripped, and the periodic impact feature caused by the leakage is significantly preserved and enhanced. The data statistics show that the output signal-to-noise ratio of the present application is improved by 18.5dB compared with the original signal, and the kurtosis value is improved from 2.4 to 8.9, proving the excellent extraction ability of the algorithm for weak fault features in strong interference environment.

[0119] 2. Multi-stage leakage state intelligent diagnosis precision comparison experiment analysis

[0120] The experiment focuses on the classification performance of the dual-flow space-time attention fusion network in distinguishing between normal operation, early micro leakage, moderate leakage and severe pipe burst. The experiment selects 2000 test samples, including early pinhole leakage samples (hole diameter less than 1mm) that are difficult to distinguish. Under the same test conditions, the dual-flow fusion network proposed in the application, the single convolutional neural network and the long short-term memory network are respectively run, and the performance of different models on the confusion matrix is recorded.

[0121] As Figure 5 shown in the confusion matrix heat map of the dual-flow fusion network fault diagnosis of the application, the horizontal axis represents the predicted class, the vertical axis represents the true class, and the value on the diagonal line represents the classification accuracy. The single modal network model performs poorly in processing early micro leakage, with an accuracy rate of less than 75%, and is easily mistaken for early leakage as background noise fluctuation. In contrast, the experimental results show that the dual-flow fusion network proposed in the application exhibits high diagnostic accuracy, with an overall recognition accuracy of 93.5%. Especially in the most challenging early micro leakage class, thanks to the deep mining of cross-attention mechanism for time-frequency domain complementary features, the model can sensitively capture the weak high-frequency jet soundprint, and the recognition accuracy remains above 87.6%. This proves that the application can effectively overcome the problem of insufficient single feature expression ability, and significantly reduce the false negative rate.

[0122] 3. Sound source positioning accuracy experiment analysis based on temperature field correction

[0123] The experiment aims to verify the actual effect of introducing the furnace temperature field correction model on improving the sound source positioning accuracy. The experiment sets 5 simulated leakage points with known coordinates, which are located in the center of the front wall, the corner area and the burner of the furnace. A non-uniform temperature field is established inside the furnace, with a temperature distribution range from 400 degrees Celsius to 1200 degrees Celsius. The experiment compares two positioning strategies: the traditional positioning method based on constant sound speed (340m / s) and the positioning method based on real-time temperature field correction proposed in the application.

[0124] As Figure 6The three-dimensional positioning error space distribution of the sound source is shown, the blue scattered points in the figure represent the positioning results of the traditional method, the red pentagram represents the positioning results of the method of the application, and the black cross represents the true leak source position. Experimental data show that the positioning results of the traditional method show significant systematic deviation due to the neglect of the change of the propagation speed of sound waves in the high-temperature non-uniform medium, the average positioning error is as high as 2.3m, and the error is further expanded to 3.5m in the burner area with large temperature gradient. On the contrary, the method of the application accurately compensates the propagation time delay by dividing the furnace into a high-resolution sound velocity grid, and the positioning results are closely gathered around the real sound source, and the average positioning error is reduced to 0.42m. The results confirm the effectiveness of the nonlinear positioning equation set constructed in step S4, and realize sub-m level accurate positioning sufficient to guide the maintenance personnel to quickly lock the fault point.

[0125] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Those skilled in the art can make various modifications and changes to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0126] Although the specific embodiments of the present application have been described above, they are not intended to limit the scope of protection of the present application. Those skilled in the art should understand that various modifications or changes made on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A method for detecting leakage in the tube wall of a boiler heating surface, characterized in that, Includes the following processes: S1, based on the omnidirectional deployment of acoustic wave sensor array and piezoelectric ceramic sensor array, collects analog voltage signals, which are then preprocessed to form a one-dimensional acoustic time history signal; S2 takes the acoustic time history signal as input, constructs a parameter optimization model that minimizes the average envelope entropy, and uses a sparrow search algorithm that introduces chaotic mapping and Cauchy mutation to search for the optimal parameters of variational mode decomposition. The optimal parameters are then used to perform variational mode decomposition on the acoustic time history signal to obtain the intrinsic mode components, perform mode screening and leakage signal reconstruction, and output the enhanced leakage signal. The parameter optimization model for minimizing the average envelope entropy is constructed using a sparrow search algorithm incorporating chaotic mapping and Cauchy mutation to search for the optimal parameters of variational mode decomposition. The specific process is as follows: Variational Mode Decomposition (VMD) is used to analyze acoustic time-history signals. The decomposition process identifies two key parameters: the number of modes. With penalty factor ; The envelope entropy of each modal component obtained from VMD decomposition is used as an evaluation index: the first... The envelope entropy of each modal component is denoted as... and with average envelope entropy As a fitness index for the sparrow search algorithm; In the initialization phase, a Logistic chaotic mapping is introduced to obtain the initial sparrow positions. ; Using the first dimension component For the number of modes Encode it and map it to a preset integer range. ; Using the second dimension component For the penalty factor Encode the values ​​and map them to a preset real number range. ; Sparrow position The corresponding meaning is, to A set of candidate solutions for performing VMD decomposition ; Introducing Cauchy variation during individual location updates, i.e., in the updated parameter combinations... Apply random perturbations to enable the search to escape local minima; iterative search ultimately yields a fitness index that satisfies the fitness metric. The minimum optimal parameters are denoted as follows: and ; S3 performs dual-branch processing on the enhanced leakage signal, fusing the feature vectors extracted by the frequency domain feature extraction branch and the time domain feature extraction branch through the cross-attention feature fusion module, and then outputting the probability distribution vector of the leakage state through the fully connected layer; S4. When the total probability of leakage exceeds the threshold, an alarm is triggered and the data is locked. Sensor node pairs are selected to participate in the localization process, and the signal arrival time difference is extracted using the generalized cross-correlation phase transformation algorithm. A non-uniform three-dimensional sound velocity grid model is constructed by combining the furnace temperature field. The predicted time difference is calculated by the ray tracing method. A nonlinear residual objective function is jointly constructed and solved iteratively using the particle swarm algorithm to invert the three-dimensional location coordinates of the leakage source. Sensor node pairs are selected to participate in the localization process, specifically as follows: Record the spatial coordinates of each sensor node during deployment, and record the first... The coordinates of each node are The coordinates are stored in the controller; during positioning, the coordinate information is directly called to participate in the calculation. Select at least four nodes from the sensors to participate in the localization process; denoted as the set of selected nodes. ;from Multiple sets of node pairs are constructed by pairwise combinations to obtain observations of time difference of arrival. Let the set of node pairs be denoted as . , of which This indicates the selection of the first node from the not less than four nodes. The node and the first A pair of nodes consisting of 1 node; through the 1 node pair; The time delay observations of multiple sets of nodes are jointly solved to achieve three-dimensional positioning; For each node participating in the positioning The enhanced time history signal is captured within the locked time window and denoted as... ; where superscript This refers to the locking time window. , Indicates the sampling point number within the window; the Enhanced leakage signal output by S2 At the node Extract and window the result; The Perform a frequency domain transformation to obtain its frequency domain representation. ,in Represents angular frequency; for any pair of nodes The generalized cross-correlation phase transform algorithm is used to calculate the node. With nodes Cross-correlation function of corresponding signals ; Will The time delay at which the peak occurs is taken as the measured arrival time difference, denoted as . ; Introducing a three-dimensional temperature field and constructing a sound velocity grid: Incorporating real-time temperature field data from the furnace. The sound velocity is corrected. The temperature field data is provided in real time by the existing temperature monitoring system in the furnace, and is spatially interpolated and rasterized to form a three-dimensional temperature grid. Then, the temperature field is converted into a sound velocity field according to thermodynamic relationships. A three-dimensional sound velocity mesh model is obtained; The predicted time difference was calculated using the ray tracing method: coordinates were given based on the structural dimensions of the boiler furnace. The upper and lower bounds are used to define the candidate search range; for any candidate point... ,calculate And obtain each node pair ; after obtaining the sound speed field Then, the propagation time in a non-uniform medium is calculated using the ray tracing method for any candidate spatial point. With any sensor node Calculate sound waves from spread to transmission time ; for nodes Obtain the predicted time difference of arrival from the model. ;in Indicates the first The spatial coordinates of the sensor nodes involved in the positioning.

2. The method for detecting leakage in the boiler heating surface tube wall as described in claim 1, characterized in that, S1 specifically includes: deploying an acoustic sensor array and a piezoelectric ceramic sensor using a distributed topology layout scheme to collect analog voltage signals containing high-frequency jet noise generated by pipe wall rupture, which are then conditioned by a preamplifier and transmitted to a high-speed synchronous data acquisition card; secondly, the acquisition card performs analog-to-digital conversion on the signal to obtain a discrete voltage sequence, and converts it into a discrete sound pressure sequence based on the sensitivity coefficient, and then splices it in channel order to generate the original acoustic observation data; finally, an infinite impulse response bandpass filter is used to filter out power frequency and low-frequency interference, and a one-dimensional preprocessed acoustic time history signal is formed within a locked time window for subsequent algorithm processing.

3. The method for detecting leakage in the boiler heating surface tube wall as described in claim 1, characterized in that: In step S2, variational mode decomposition is performed on the acoustic time history signal using optimal parameters to obtain intrinsic mode components, followed by mode screening and leakage signal reconstruction. The specific process is as follows: Use optimal parameters and right Perform VMD decomposition, Decomposition The intrinsic mode components; the first Each modal component is denoted as... And at the same time, its corresponding center frequency is obtained, denoted as . ,Right now For the first Modal center angular frequency; Then calculate the first one respectively. Modal components and correlation coefficient , and the Modal components cliff Select those that simultaneously satisfy and The modes are taken as the effective leakage mode set, denoted as S. If S is empty, a fallback strategy is adopted: calculate the comprehensive score for each mode and select the mode with the highest score. The modalities constitute a set S; where the correlation coefficient threshold is denoted as . ,set up ; kurtosis threshold is denoted as ,set up ; Finally, the enhanced leakage signal is obtained by linear superposition and reconstruction of the modal components in set S. .

4. The method for detecting leakage in the boiler heating surface tube wall as described in claim 1, characterized in that: The frequency domain feature extraction branch is specifically as follows: Enhance the leakage signal output of S2 Perform a short-time Fourier transform to generate a two-dimensional time-frequency spectrum reflecting the distribution of signal energy over time and frequency, which is defined as the input tensor. ;Will The input is fed into a feature extraction network containing four cascaded convolutional blocks. Each convolutional block contains a convolutional layer, a batch normalization layer, and a ReLU activation layer. The convolutional operation extracts local texture features through a learnable filter bank, and its mathematical expression is as follows: ； in This is the weight matrix. For bias terms, Representing the The feature map tensor output by the convolution operation. This represents the input feature map of the previous layer. This represents the linear rectified activation function; after feature extraction in the fourth convolutional block, a feature map containing high-level semantic information is obtained. ;right Perform a global average pooling operation to calculate the average response value of each channel in the spatial dimension, generating a frequency domain feature vector. .

5. The method for detecting leakage in the boiler heating surface tube wall as described in claim 1, characterized in that: The temporal feature extraction branch is specifically as follows: Enhance the leakage signal output of S2 Perform full-dimensional time-domain statistical analysis, and calculate a total of 12 physical statistical features of the signal in sequence, including mean, root mean square value, variance, standard deviation, peak value, peak-to-peak value, rectified average value, skewness, kurtosis, waveform factor, impulse factor and margin factor. The calculated 12-dimensional feature data are stacked in chronological order to construct a time-domain feature sequence matrix with a time step of 50. The sequence matrix is ​​input into a bidirectional long short-term memory network (BiLSTM) for each time step. The forward layer calculates the hidden state. Backward layer calculates hidden states The hidden states in the two directions are concatenated to obtain the combined state vector at the current time step. Extract the output of the last time step as the temporal feature vector. .

6. The method for detecting leakage in the boiler heating surface tube wall as described in claim 2, characterized in that: S3 outputs a four-class probability vector. ,in This represents the probability of the normal state. , , These represent the probabilities of early leakage, moderate leakage, and pipe rupture, respectively; the probabilities of the three leakage-related categories are combined to form the total leakage probability. ,in ;when If the acoustic data exceeds the preset safety threshold of 0.85 for three consecutive seconds, a leakage event is determined to have occurred. An audible and visual alarm is immediately triggered, and the acoustic data within that time window is automatically locked, initiating the location procedure.

7. The method for detecting leakage in the boiler heating surface tube wall as described in claim 1, characterized in that: In S4, a nonlinear residual objective function is jointly constructed and solved iteratively using a particle swarm optimization algorithm. The specific process is as follows: Calculated time delay difference Time delay difference with model prediction Deviation between ( ), and for all selected node pairs The calculated deviations are summed by squares to obtain... ; The particle swarm optimization algorithm is used to iteratively search within the defined candidate search range, so that... Find the minimum value to obtain the optimal point. This point is the three-dimensional location of the leakage source obtained through inversion.

Images