Measurement equipment, method and storage medium for the dry-wet state conversion position of a boiler

Abstract

This application relates to the field of boiler monitoring technology, and in particular to a measuring device, method, and storage medium for determining the dry-wet transition position of a boiler. The device includes: an acoustic sensor array for collecting acoustic information from the space surrounding a collection point on the boiler; a collection device connected to the acoustic sensor array, receiving the analog signals output by the array and converting them into electrical signals; and a processing device acquiring the electrical signals from the collection device, extracting acoustic signature features, processing the acoustic signature data based on a prediction model, identifying the working fluid flow regime and dryness measurement values ​​at various locations within the boiler's water-cooled wall tubes under different operating conditions, determining the distribution and dynamic changes of the dry-wet boundary line based on the working fluid flow regime and dryness measurement values, and thus determining the boiler's dry-wet transition position. This solves the problems of large measurement errors caused by relying on indirect temperature calculations for boiler dry-wet transition position measurement, due to the large thermal inertia of the water-cooled wall, uncertain heat transfer coefficient, and large local pressure fluctuations of the working fluid.

Description

Technical Field

[0001] This application relates to the field of boiler monitoring technology, and in particular to a measuring device, method and storage medium for measuring the dry-wet state transition position of a boiler. Background Technology

[0002] In related technologies, the measurement of the dry-wet transition position of a boiler is indirectly determined by monitoring the temperature distribution on the back-fire side of the boiler's water-cooled wall and combining it with thermodynamic parameters. Therefore, the measurement relies on indirect calculation based on temperature. However, the water-cooled wall has high thermal inertia, an uncertain heat transfer coefficient, and large local pressure fluctuations in the working fluid, leading to significant errors in the determination of the transition position. Summary of the Invention

[0003] This application provides a measuring device, method, and storage medium for measuring the dry-wet transition position of a boiler, to solve the problems in related technologies where the measurement of the dry-wet transition position of a boiler relies on indirect calculation based on temperature, resulting in large measurement errors due to the large thermal inertia of the water-cooled wall, the uncertain heat transfer coefficient, and large local pressure fluctuations of the working fluid. A first aspect of this application provides a measuring device for measuring the dry-wet transition position of a boiler, comprising: an acoustic sensor array for collecting acoustic information of the space surrounding a collection point on the boiler; a collection device connected to the acoustic sensor array for receiving analog signals of the acoustic information collected by the acoustic sensor array and converting the analog signals into electrical signals; and a processing device for acquiring the electrical signals from the collection device, extracting acoustic signature features from the electrical signals, processing the acoustic signature data based on a prediction model, identifying the working fluid flow regime and dryness measurement values ​​at different locations within the boiler water-cooled wall tubes under different operating conditions, determining the distribution and change results of the dry-wet boundary line based on the working fluid flow regime and dryness measurement values, and determining the dry-wet transition position of the boiler based on the distribution and change results of the dry-wet boundary line.

[0004] Optionally, the acoustic sensor array includes multiple acoustic sensors, which are installed at the acquisition points. The installation method of the acoustic sensors includes magnetic attraction. The mounting surface corresponding to the acquisition point is at least one of the plane that is attached to the boiler exterior and the semi-circular surface on the backfire side that is attached to the finned tube. The acquisition hole corresponding to the acquisition point is arranged on the magnetic attraction surface.

[0005] Optionally, the acquisition device includes a data transmission line and a signal converter. The signal converter is connected to the acoustic sensor array via the data transmission line to realize the conversion between analog signals and electrical signals.

[0006] The second aspect of this application provides a method for measuring the dry-wet transition position of a boiler, comprising the following steps: acquiring an electrical signal from a data acquisition device and extracting acoustic signature features from the electrical signal; processing the acoustic signature data based on a prediction model to identify the working fluid flow regime and dryness measurement values ​​at different locations within the boiler water-cooled wall tubes under different operating conditions; determining the distribution and change results of the dry-wet boundary line based on the working fluid flow regime and dryness measurement values; and determining the dry-wet transition position of the boiler based on the distribution and change results of the dry-wet boundary line.

[0007] Optionally, before identifying the working fluid flow regime and dryness at different locations within the boiler water-cooled wall tubes under different operating conditions, the method further includes: extracting acoustic fingerprint data from acoustic information; and performing at least one processing step of data cleaning and data enhancement on the acoustic fingerprint data.

[0008] Optionally, the data cleaning method includes at least one of the following: separating acoustic fingerprint data of adjacent pipes through a blind source separation algorithm; filtering acoustic fingerprint data based on the correlation of working fluid state; and performing time-domain and frequency-domain analysis on the acoustic fingerprint data to remove acoustic fingerprint data from non-normal operating sections.

[0009] Optionally, data augmentation methods include at least one of the following: applying time and frequency masks to the cleaned acoustic data; adjusting the signal amplitude of the cleaned acoustic data to simulate changes in sensor spacing; performing variable-rate resampling of the cleaned acoustic data combined with spectrum shifting to simulate changes in working fluid velocity; and adding background noise of different intensities to the cleaned acoustic data to simulate fluctuations in flame interference within the boiler environment.

[0010] Optionally, the prediction model includes at least one of the following: a native clustering model without explicit feature learning, a deep generative clustering model, and a clustering model based on representation learning. The clustering model based on representation learning includes a voiceprint feature compression and embedding layer, a spatiotemporal multi-scale coding layer, a reconstruction loss decoder layer, a probability distribution self-supervised clustering layer, and a dryness soft measurement regression layer. The voiceprint feature compression and embedding layer extracts, compresses, and spatiotemporally fuses voiceprint data. The spatiotemporal multi-scale coding layer extracts and learns representations from the voiceprint data at spatiotemporal multi-scale. The reconstruction loss decoder layer outputs reconstructed voiceprints through a symmetrical decoder structure and calculates the mean square error reconstruction loss. The probability distribution self-supervised clustering layer automatically classifies the high-dimensional features extracted by the coding layer into different working fluid flow states. The dryness soft measurement regression layer converts the flow state probabilities obtained from clustering into dryness measurement values.

[0011] Optionally, before using the trained prediction model for the recognition of actual boiler acoustic fingerprint data, the method further includes: performing weak and strong enhancement on the original acoustic fingerprint data, generating pseudo-labels through model prediction; performing semi-supervised fine-tuning on the prediction model based on the weakly and strongly enhanced acoustic fingerprint data and pseudo-labels; and evaluating the fine-tuned prediction model using unsupervised evaluation metrics.

[0012] Therefore, this application has at least the following beneficial effects: In this embodiment, an acoustic sensor array is used to collect acoustic information from the space surrounding the collection point on the boiler. A collection device is connected to the acoustic sensor array to receive analog signals from the acoustic information and convert them into electrical signals. A processing device obtains these electrical signals from the collection device, extracts acoustic signature features, and processes the acoustic signature data based on a prediction model. This identifies the working fluid flow regime and dryness measurement values ​​at different locations within the boiler's water-cooled wall tubes under different operating conditions. Furthermore, based on the working fluid flow regime and dryness measurement values, the distribution and changes of the dry-wet boundary line are determined. Finally, the dry-wet transition position of the boiler is determined based on these distribution and changes. By directly identifying the working fluid flow regime and dryness measurement values ​​based on acoustic information, the errors introduced by indirect temperature estimation due to the large thermal inertia of the water-cooled wall, the uncertainty of the heat transfer coefficient, and the large local pressure fluctuations of the working fluid are effectively avoided, thus improving the accuracy of dry-wet transition position measurement. Therefore, this solves the problems in related technologies where boiler dry-wet transition position measurement relies on indirect temperature estimation, and is subject to large errors due to the large thermal inertia of the water-cooled wall, the uncertainty of the heat transfer coefficient, and the large local pressure fluctuations of the working fluid.

[0013] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0014] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a schematic diagram of the structure of a measuring device for the dry-wet state transition position of a boiler, according to an embodiment of this application; Figure 2 This is a schematic diagram of an acoustic sensor mounting structure according to an embodiment of this application; Figure 3 This is a schematic diagram of an acoustic sensor array arrangement according to an embodiment of this application; Figure 4 This is a flowchart of a method for measuring the dry-wet state transition position of a boiler, according to an embodiment of this application. Figure 5 This is a schematic diagram illustrating the construction and application process of a method for measuring the dry-wet state transition position of a boiler, provided according to an embodiment of this application. Detailed Implementation

[0015] The embodiments of this application are described in detail below. Examples of the 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.

[0016] In related technologies, fiber optic temperature sensors are arranged on the backfire side of the boiler water-cooled wall to measure the temperature outside the tube. Combined with the heat transfer coefficient determined by the water-cooled wall material and wall thickness, the temperature of the working fluid inside the tube is indirectly calculated. Then, this temperature is compared with the saturation temperature of the working fluid under the current pressure. If the difference between the two is within the set error range, the position is determined to be the dry-wet state transition point.

[0017] However, this method has significant shortcomings: First, the water-cooled wall is composed of thick-walled steel pipes (5mm to 10mm). Steel has a large specific heat capacity and density, resulting in high thermal inertia and poor response performance of the external temperature to changes in the internal operating conditions. Second, the heat transfer coefficient is only set based on the material and wall thickness, which cannot reflect the actual situation of convective heat transfer inside the pipe as the properties of the working fluid and the flow state change, leading to large calculation errors. Third, during the dry-wet transition, the two-phase flow causes drastic local pressure fluctuations, and the use of the constant pressure assumption to determine the saturation temperature further introduces significant errors. Fourth, the temperature measuring device is fixed to the finned tube by welding a U-shaped base, which is an invasive installation. This not only introduces contact thermal resistance and changes the heat transfer boundary conditions, but also leads to the need for boiler shutdown for maintenance, difficult replacement, and high costs. Fifth, the determination of the dry-wet boundary relies heavily on empirical formulas or simple thresholds, without considering the dynamic changes in the working fluid flow state and dryness, making it difficult to meet the accurate monitoring needs under varying operating conditions of the unit.

[0018] The measuring equipment, method, and storage medium for the boiler dry-wet state switching position according to embodiments of this application are described below with reference to the accompanying drawings. To address the issues mentioned in the background art regarding the measurement of boiler dry-wet transition position relying on indirect temperature estimation, the large thermal inertia of the water-cooled wall, the uncertain heat transfer coefficient, and the large local pressure fluctuations of the working fluid leading to significant errors, this application provides a measurement device for the dry-wet transition position of a boiler. In this device, an acoustic sensor array is used to collect acoustic information from the space surrounding the collection point on the boiler. The acquisition device is connected to the acoustic sensor array to receive analog signals of the acoustic information and convert them into electrical signals. The processing device obtains the electrical signals from the acquisition device, extracts acoustic fingerprint features, and processes the acoustic fingerprint data based on a prediction model. It identifies the working fluid flow regime and dryness measurement values ​​at different locations within the boiler water-cooled wall tubes under different operating conditions. Then, based on the working fluid flow regime and dryness measurement values, it determines the distribution and changes of the dry-wet boundary line. Finally, based on these distribution and changes, it determines the dry-wet transition position of the boiler. By directly identifying the working fluid flow regime and dryness measurement values ​​based on acoustic information, it effectively avoids the errors introduced by indirect temperature estimation due to the large thermal inertia of the water-cooled wall, the uncertain heat transfer coefficient, and the large local pressure fluctuations of the working fluid, thus improving the accuracy of dry-wet transition position measurement. This solves the problems in related technologies, such as the reliance on indirect temperature calculations for measuring the dry-wet transition position of boilers, and the large errors caused by the large thermal inertia of water-cooled walls, uncertain heat transfer coefficients, and large local pressure fluctuations of the working fluid.

[0019] Specifically, Figure 1 This is a schematic diagram of the structure of a boiler dry-wet state switching measuring device provided in an embodiment of this application.

[0020] like Figure 1 As shown, the measuring equipment for the dry-wet state transition position of the boiler includes: an acoustic sensor array 100, a data acquisition device 200, and a processing device 300.

[0021] The system includes an acoustic sensor array 100 for collecting acoustic information about the space surrounding the collection point on the boiler; a collection device 200 connected to the acoustic sensor array for receiving analog signals of the acoustic information collected by the acoustic sensor array and converting the analog signals into electrical signals; and a processing device 300 for acquiring electrical signals from the collection device, extracting acoustic features from the electrical signals, processing the acoustic data based on a prediction model, identifying the working fluid flow state and dryness measurement values ​​at different locations within the boiler water-cooled wall tubes under different operating conditions, determining the distribution and change results of the dry-wet boundary line based on the working fluid flow state and dryness measurement values, and determining the dry-wet state transition location of the boiler based on the distribution and change results of the dry-wet boundary line.

[0022] Specifically, the acoustic sensor array 100 includes multiple acoustic sensors, which are piezoelectric acoustic sensors that are resistant to high temperatures and interference, and can collect typical industrial acoustic information from 50Hz to 4000Hz. The acquisition device consists of a series of data transmission lines and signal converters. The processing device is connected to the acquisition device through communication, and identifies the working fluid flow state and dryness measurement values ​​at different locations inside the water-cooled wall tube under different operating conditions based on the collected acoustic data, and determines the distribution and changes of the dry and wet boundary.

[0023] It is understood that the measuring equipment in this embodiment of the application, through the coordinated operation of the acoustic sensor array, the acquisition device and the processing device, can directly and non-invasively acquire the flow state and dryness information of the working fluid inside the boiler water-cooled wall tubes, effectively avoiding errors introduced by heat transfer coefficient and pressure fluctuations, and improving the accuracy of dry-wet state transition position identification.

[0024] In some embodiments, the acoustic sensor array includes multiple acoustic sensors, which are installed at the acquisition points. The installation method of the acoustic sensors includes magnetic attraction. The mounting surface corresponding to the acquisition point is at least one of the plane that is attached to the outer facade of the boiler and the semi-circular surface on the backfire side that is attached to the finned tube. The acquisition hole corresponding to the acquisition point is arranged on the magnetic attraction surface.

[0025] Among them, the acquisition point refers to the location on the boiler where the acoustic sensor is placed to acquire acoustic information; magnetic attraction refers to the installation method of fixing the acoustic sensor to the metal surface by magnetic attraction; the mounting surface refers to the physical surface of the acoustic sensor in contact with the boiler, which is a plane or a semi-circular surface on the back side; the boiler exterior refers to the surface of the boiler's outer shell; finned tubes refer to heat exchange tubes with fins in the water-cooled wall, used to connect adjacent tubes and transfer heat; and the magnetic attraction surface refers to the side of the acoustic sensor with a magnetic attraction structure that is attached to the mounting surface.

[0026] It is understood that the embodiments of this application can flexibly arrange acoustic sensors on the boiler exterior or corresponding acquisition points of finned tubes using a magnetic installation method. This allows the acquisition holes to be located on the magnetic surface to be close to the sound source and to block environmental noise, thereby improving the signal-to-noise ratio of the acoustic signal. Simultaneously, the mounting surface is adapted to the plane of the boiler exterior or the semi-circular surface of the finned tubes on the unfired side, balancing reliability and acoustic coupling effect under high-temperature environments, and facilitating deployment and adjustment.

[0027] Specifically, such as Figure 2As shown, the acoustic sensors in the acoustic sensor array 100 are installed magnetically, allowing them to be easily attached to metal surfaces and adjusted at any time. The mounting surface can be a flat surface that fits against the boiler's exterior or a semi-circular surface on the back-fire side that fits against the finned tubes, depending on the requirements of the working environment. When the temperature on the back-fire side of the boiler's water-cooled wall is low, the acoustic sensors in the acoustic sensor array can be fixed to the finned tubes to obtain clearer acoustic data. When the temperature on the back-fire side is high, the acoustic sensors can be fixed to the exterior surface to avoid harsh working environments. The acoustic acquisition holes are arranged on the magnetic surface. When the acoustic sensors in the acoustic sensor array are attached to the acquisition object, they can effectively block ambient noise and improve the signal-to-noise ratio.

[0028] Furthermore, such as Figure 3 As shown, the acoustic sensor array includes multiple acoustic sensors that can collect acoustic information from the surrounding space. Due to the magnetic mounting method, it is easy to adjust. The sensor array can be distributed along the path, or other distribution schemes that meet the measurement requirements can be used. By analyzing the spatial acoustic information collected by the sensor array, different sound sources can be separated and the spatial distribution of the sound sources can be located.

[0029] In practice, the voiceprint data comes from the target unit to be monitored, such as... Figure 3 As shown, the acoustic sensor array can be distributed along the furnace height, with one acoustic sensor in the array placed every 2m. Depending on their actual effectiveness and operating temperature limitations, the sensors can be installed on the backfire side of the water-cooled wall or the boiler exterior. During actual operation, the unit experiences dry, dry-wet transition, and wet states with loads decreasing from high to low. Acoustic data of the unit's operation is collected using specific sampling parameters, a sampling rate of 16kHz, and a duration of 10s per segment.

[0030] In some embodiments, the acquisition device includes a data transmission line and a signal converter. The signal converter is connected to the acoustic sensor array via the data transmission line to realize the conversion between analog signals and electrical signals.

[0031] Among them, the data transmission line refers to the line used to transmit the analog signal output by the acoustic sensor array to the signal converter, which can be high-temperature resistant optical fiber or shielded twisted pair cable; the signal converter refers to the device used to convert the analog signal transmitted from the data transmission line into an electrical signal; the electrical signal refers to the electrical signal output by the signal converter after converting the analog signal of the acoustic sensor and used to transmit it to the processing device for acoustic information extraction.

[0032] It is understood that the embodiments of this application can achieve reliable transmission and analog-to-digital conversion of acoustic analog signals through the cooperation of data transmission lines and signal converters, ensuring that subsequent processing devices can accurately acquire acoustic information. At the same time, the use of high-temperature resistant or shielded transmission lines adapts to the high-temperature and strong interference environment at the boiler site, improving system stability.

[0033] Specifically, the acquisition device consists of a series of data transmission lines and signal converters. The data transmission lines can be made of high-temperature resistant optical fibers, shielded twisted-pair cables, etc., depending on the installation location of the acoustic sensors. The data transmission lines transmit the analog signals from the sensors to the signal converters, which convert them into electrical signals and then transmit them to the processing device.

[0034] According to the boiler dry-wet transition position measurement equipment proposed in this application embodiment, an acoustic sensor array is used to collect acoustic information of the space surrounding the collection point on the boiler. The acquisition device is connected to the acoustic sensor array to receive analog signals of the acoustic information and convert them into electrical signals. The processing device obtains the electrical signals from the acquisition device, extracts acoustic fingerprint features, and processes the acoustic fingerprint data based on a prediction model. It identifies the working fluid flow regime and dryness measurement values ​​at different locations within the boiler water-cooled wall tubes under different operating conditions. Then, based on the working fluid flow regime and dryness measurement values, it determines the distribution and change results of the dry-wet boundary line. Finally, based on the distribution and change results, it determines the boiler dry-wet transition position. By directly identifying the working fluid flow regime and dryness measurement values ​​based on acoustic information, it effectively avoids the errors introduced by indirect temperature estimation due to the large thermal inertia of the water-cooled wall, the uncertainty of the heat transfer coefficient, and the large local pressure fluctuations of the working fluid, thus improving the accuracy of dry-wet transition position measurement. Therefore, it solves the problems in related technologies where boiler dry-wet transition position measurement relies on indirect temperature estimation, and the large errors caused by the large thermal inertia of the water-cooled wall, the uncertainty of the heat transfer coefficient, and the large local pressure fluctuations of the working fluid.

[0035] Next, referring to the accompanying drawings, a method for measuring the dry-wet state transition position of a boiler according to an embodiment of this application is described.

[0036] Figure 4 This is a flowchart of a method for measuring the dry-wet state transition position of a boiler according to an embodiment of this application.

[0037] like Figure 4 As shown, the method for measuring the dry-wet transition position of the boiler includes the following steps: In step S401, an electrical signal is acquired from the acquisition device, and voiceprint features are extracted from the electrical signal.

[0038] Among them, acoustic signature refers to the sound signal that reflects the flow state and dryness of the working fluid in the boiler water-cooled wall tubes, which is collected by acoustic sensors and converted by the acquisition device. It includes frequency, amplitude and temporal characteristics.

[0039] It is understood that the embodiments of this application can directly extract voiceprint features from electrical signals, avoiding dependence on indirect physical quantities such as temperature, and providing a high-dimensional, real-time raw data foundation for subsequent accurate identification of the dry-wet transition location.

[0040] In step S402, acoustic fingerprint data is processed based on a prediction model to identify the working fluid flow regime and dryness measurement values ​​at different locations inside the boiler water-cooled wall tubes under different operating conditions, and the distribution and change results of the dry and wet boundary line are determined based on the working fluid flow regime and dryness measurement values.

[0041] Among them, the working fluid flow state refers to the physical state of the working fluid in the boiler water-cooled wall tubes, including wet, dry and transitional states; the dryness measurement value is a continuous index reflecting the dryness of the working fluid calculated based on the flow state classification results; the dry and wet boundary line is the boundary position between the dry and wet regions on the boiler water-cooled wall, which can be defined as the height position where the dryness first exceeds the threshold and remains stable.

[0042] It is understood that the embodiments of this application can process acoustic fingerprint data through a predictive model, thereby directly identifying the working fluid flow state and dryness measurement values, effectively avoiding the errors in heat transfer coefficient and pressure fluctuation caused by indirect calculation based on temperature, thus accurately determining the location and dynamic changes of the dry-wet boundary, and improving the reliability of dry-wet state conversion monitoring.

[0043] In some embodiments, before identifying the working fluid flow regime and dryness at different locations within the boiler water-cooled wall tubes under different operating conditions, the method further includes: extracting acoustic fingerprint data from acoustic fingerprint features; and performing at least one processing step on the acoustic fingerprint data: data cleaning and data enhancement.

[0044] Among them, voiceprint data refers to the original sound signal data extracted from voiceprint features that reflects the changes in the working fluid's flow state and dryness; data cleaning refers to voiceprint preprocessing to remove interference, abnormal segments, and noise; and data enhancement refers to voiceprint augmentation processing that simulates changes in working conditions.

[0045] It is understood that the embodiments of this application can extract voiceprint data from voiceprint features and perform at least one data cleaning and data enhancement process to improve the signal-to-noise ratio and generalization ability of voiceprint data, so that the model can more accurately identify the fluid flow state and dryness measurement values ​​under different working conditions.

[0046] In some embodiments, data cleaning methods include at least one of the following: separating acoustic fingerprint data of adjacent pipes using a blind source separation algorithm; filtering acoustic fingerprint data based on the correlation of working fluid state; and performing time-domain and frequency-domain analysis on the acoustic fingerprint data to remove acoustic fingerprint data from non-normal operating sections.

[0047] Among them, blind source separation algorithm refers to the algorithm that can separate the characteristic acoustic fingerprint of a single pipe from the mixed signal without knowing the propagation path of the sound source in advance, such as Independent Component Analysis (ICA) or beamforming technology; working fluid state correlation refers to the degree of correlation between acoustic fingerprint data and working fluid flow state and dryness; time domain analysis refers to analyzing acoustic fingerprint data in the time dimension to identify abnormal working sections; frequency domain analysis refers to analyzing acoustic fingerprint data in the frequency dimension to identify abnormal working sections.

[0048] It is understood that the embodiments of this application can filter acoustic data of adjacent pipes through blind source separation algorithm, filter acoustic data according to the correlation of working fluid state, and remove acoustic data of non-normal working section through time domain and frequency domain analysis, effectively eliminating interference and invalid signals, retaining acoustic data that are strongly correlated with working fluid state, and improving the accuracy of flow state and dryness identification.

[0049] In some embodiments, data augmentation methods include at least one of the following: applying time and frequency masks to the cleaned acoustic data; adjusting the signal amplitude of the cleaned acoustic data to simulate changes in sensor spacing; performing variable-rate resampling of the cleaned acoustic data combined with spectrum shifting to simulate changes in working fluid velocity; and adding background noise of different intensities to the cleaned acoustic data to simulate fluctuations in flame interference within the boiler environment.

[0050] Among them, time masking refers to the processing method of partially masking the acoustic data in the time dimension; frequency masking refers to the processing method of partially masking the acoustic data in the frequency dimension; sensor arrangement spacing refers to the installation position interval of acoustic sensors on the boiler; variable rate sampling refers to changing the duration of the signal by changing the sampling frequency; spectrum shifting refers to translating the spectrum of the entire signal along the frequency axis; working fluid velocity refers to the flow velocity of the working fluid in the water-cooled wall tubes; flame interference refers to the acoustic environment fluctuations caused by the flame inside the boiler.

[0051] It is understood that the embodiments of this application can enhance the model's adaptability to the real operating environment by applying time masks, frequency masks, signal amplitude adjustments, variable rate sampling, spectrum shifting, and adding background noise to the cleaned acoustic data to simulate actual working conditions such as changes in sensor spacing, changes in working fluid flow rate, and flame interference.

[0052] In some embodiments, the prediction model includes at least one of a native clustering model without explicit feature learning, a deep generative clustering model, and a clustering model based on representation learning. The clustering model based on representation learning includes a voiceprint feature compression and embedding layer, a spatiotemporal multi-scale coding layer, a reconstruction loss decoder layer, a probability distribution self-supervised clustering layer, and a dryness soft measurement regression layer. The voiceprint feature compression and embedding layer extracts, compresses, and spatiotemporally fuses voiceprint data. The spatiotemporal multi-scale coding layer extracts and learns representations from the voiceprint data at spatiotemporal multi-scale. The reconstruction loss decoder layer outputs reconstructed voiceprints through a symmetrical decoder structure and calculates the mean square error reconstruction loss. The probability distribution self-supervised clustering layer automatically classifies the high-dimensional features extracted by the coding layer into different working fluid flow states. The dryness soft measurement regression layer converts the flow state probabilities obtained from clustering into dryness measurement values.

[0053] Among them, the native clustering model without explicit feature learning refers to a clustering model that completes cluster partitioning directly based on the inherent metric rules of the original data without introducing an independent deep feature extractor; the deep generative clustering model refers to a clustering model that combines a deep generative model with a clustering task, completing feature extraction and cluster assignment simultaneously by modeling the generation distribution of the data; and the clustering model based on representation learning refers to a clustering model that first maps high-dimensional original data to a low-dimensional, high-discrimination feature space through a deep representation learning model, and then applies traditional clustering algorithms on this feature space.

[0054] It is understood that the prediction model in this application embodiment includes at least one of the following: a native clustering model without explicit feature learning, a deep generative clustering model, and a clustering model based on representation learning. The clustering model based on representation learning includes a voiceprint feature compression and embedding layer, a spatiotemporal multi-scale coding layer, a reconstruction loss decoder layer, a probability distribution self-supervised clustering layer, and a dryness soft measurement regression layer. The voiceprint feature compression and embedding layer performs dimensionality reduction and spatiotemporal fusion on the original voiceprint data, preserving the flow-related temporal characteristics. The spatiotemporal multi-scale coding layer extracts high-frequency bubble bursts through a physically constrained attention mechanism. The system identifies multi-scale features such as cracks and low-frequency pipe vibrations, and models the physical continuity of adjacent working fluid sections. The reconstruction loss decoder layer uses deconvolution to gradually restore the feature sequence to the original acoustic signature dimension and calculates the mean square error reconstruction loss. The probability distribution self-supervised clustering layer automatically clusters high-dimensional features into three categories—wet, transition, and dry—based on the Student's t-distribution, without requiring manual labeling. The dryness soft measurement regression layer uses the soft probability weighting of the clustering output to calculate the dryness value and applies time continuity constraints to ensure physical rationality. This enables non-invasive, highly robust, and real-time soft measurement of the wet-dry transition position of the water-cooled wall.

[0055] The clustering model based on representation learning adopts probabilistic distribution self-supervised clustering, outputs the soft probabilities of each flow state, and obtains the dryness measurement value through probability weighted summation. This model can capture long-term time-series dependencies and is suitable for analyzing the dynamic process of flow state transition. Its self-attention mechanism can model the spatial correlation of multi-sensor acoustic signatures, effectively ensuring the rationality of spatiotemporal continuity when judging the working fluid flow state and dryness in a single tube. For example, the working fluid flow state in the tube usually shows an evolution law of wet state → transition state → dry state along the furnace height direction, and the dryness also increases monotonically.

[0056] Specifically, the embodiments of this application include an embedding layer, an encoding layer, a decoder layer, a clustering layer, and a quantization layer. The embedding layer corresponds to the voiceprint feature compression and embedding layer, which is a feature processing module that denoises, normalizes, compresses features, fuses spatial coding and temporal location coding, and reduces dimensionality of the voiceprint signal. The encoding layer corresponds to the spatiotemporal multi-scale encoding layer, which is a core encoding module that fuses deep and shallow layer information and strengthens the physical correlation of the working fluid to efficiently extract flow-related semantic features. The decoder layer corresponds to the reconstruction loss decoder layer, which is a module that restores the encoded low-dimensional features to the original voiceprint signal and calculates the reconstruction loss, providing a constraint module for self-supervised learning optimization objectives. The clustering layer corresponds to the probability distribution self-supervised clustering layer, which is a clustering module that realizes unsupervised aggregation of flow-like features and outputs flow-like probabilities. The quantization layer corresponds to the dryness soft measurement regression layer, which is a dryness measurement module that realizes smooth quantization of dryness and avoids unreasonable physical jumps. At the same time, unsupervised clustering is an automatic classification of data by automatically mining the inherent structure and similarity of data in the case of unlabeled data and dividing data samples with similar features into the same cluster.

[0057] Specifically, due to the long-term temporal characteristics and strong noise environment of boiler acoustic fingerprint data, the clustering model based on representation learning can adopt a CNN-Transformer hybrid architecture with an encoder-decoder symmetric structure, including: an acoustic fingerprint feature compression and embedding layer, a spatiotemporal multi-scale coding layer, a reconstruction loss decoder layer, a probability distribution self-supervised clustering layer, and a dryness soft measurement regression layer. This can be further elaborated as follows: (1) Voiceprint feature compression and embedding layer. To achieve feature dimensionality reduction and spatiotemporal fusion, the voiceprint feature compression and embedding layer includes: 1. Data preprocessing. Raw input N A single acoustic signature sensor is used to acquire a one-dimensional time series vector. The sequence length is set to a preset length, for example, 160,000, the sampling frequency is 16 kHz, the sampling duration is 10 seconds, and each sequence segment contains 160,000 sampling points. The data is first subjected to wavelet denoising to suppress environmental noise other than pipeline vibration, and then amplitude normalization is performed to map the signal amplitude to a preset interval through linear mapping. The preset interval can be [-1, 1], thereby preserving the original temporal characteristics of the signal.

[0058] 2. Feature Extraction. A 4-layer 1D-CNN residual network is used to perform "convolutional kernel scanning - nonlinear activation - normalization - residual fusion" steps on each sensor, transforming the one-dimensional time-series signals from all sensors into dimensional features. The feature vectors are compressed to a preset time step. Finally, the compressed feature sequence of the sensor is output. .

[0059] Optionally, for scenarios with sufficient computing resources, a self-supervised model fine-tuned by industrial voiceprints (such as Wav2Vec2.0) can be used instead of 1D-CNN as the feature extractor. This embodiment uses a 4-layer 1D-CNN residual network as the main implementation method to extract temporal and spectral features from the original waveform, and obtains a fixed-length feature vector after pooling or normalization. Furthermore, after data normalization, the effectiveness of the feature dimensions is verified based on principal component analysis to ensure that the 256-dimensional features retain more than 95% of the variance information of the original signal.

[0060] 3. Spatiotemporal fusion. N The coordinate values ​​of each sensor along the furnace height are normalized to a preset interval, which can be [0,1]. This preset interval is then mapped to a 256-dimensional spatial encoding vector through a fully connected layer. And it is replicated to all time steps via a broadcast mechanism. The sequence of 500 time steps is identified using sine / cosine position coding, resulting in a 256-dimensional time coding vector. It is replicated to all sensors via a broadcast mechanism. After aligning the spatiotemporal coding dimensions, the time-frequency features of the voiceprint and the spatiotemporal location coding are fused to obtain the input of the Transformer model:

[0061] in, The feature sequence for the feature extraction step; This is a random regularization operation; Encode for time; Encoding for space; A learnable scaling factor; This is a learnable scaling factor.

[0062] Understandable , The initial values ​​for all parameters are 0.1. The relative weights of temporal dependency and spatial prior are controlled separately. By setting smaller initial values, the model prioritizes learning essential voiceprint features in the early stages of training, while the scaling factor adaptively increases in later training, gradually introducing the physical constraints of temporal causality and furnace height distribution. Furthermore, a Dropout function (random deactivation) is introduced to randomly mask the overall information of the location encoding with a 10% probability, preventing the model from over-relying on spatiotemporal location information and improving the model's robustness. Simultaneously, element-wise addition instead of concatenation allows each feature point to carry spatiotemporal background information without changing the vector length.

[0063] (2) Spatiotemporal multi-scale coding layer. To achieve the fusion of deep and shallow layer information and the enhancement of the physical correlation of the working medium, the spatiotemporal multi-scale coding layer includes: 1. Attention Mechanism Modification. To modify the physically constrained Transformer, a spatiotemporal multi-scale encoding layer is stacked with 6 Transformer Encoder Blocks. Each Encoder Block consists of two main sub-layers: a multi-head self-attention mechanism and a feedforward neural network, supplemented by residual connections and normalization layers. The multi-head self-attention mechanism has a preset number of attention heads. h It can have 8 dimensions. D =256 is split into subspaces, with a preset number of 8 subspaces, each with a dimension of 32, to capture different features of high-frequency and low-frequency components in the voiceprint in parallel; the feedforward neural network adopts a two-layer fully connected structure (256→1024→256), and uses the Gaussian Error Linear Unit (GELU) activation function to enhance the nonlinear fitting ability; the normalization layer adopts layer normalization, and normalization is performed before and after each sublayer to ensure the training stability of the deep network; the residual connection layer directly adds the input of the sublayer to the output, thereby preventing the gradient vanishing problem during the stacking of 6 layers.

[0064] To strengthen physical correlation, a physical constraint mask is introduced on top of the standard Transformer multi-head self-attention mechanism sublayer. Among these, a temporal mask... It is a lower triangular matrix. This is used to mask future information that exceeds the current time window, thereby ensuring that the model focuses on the current flow characteristics and avoids interference across time windows; spatial mask. Based on two sensors i and j Actual spacing at the furnace height (Unit: meters), apply Gaussian attenuation: ,in σ =5m is the attenuation coefficient, when the distance between the two sensors is...d When the distance exceeds 10 m, the weights approach 0, indicating a significant decrease in the attention weights between the two sensors. This allows the model to prioritize learning the physical continuity between adjacent working fluid sections, such as the smooth evolution from a wet state to a transitional state, avoiding meaningless feature associations spanning tens of meters. Furthermore, global average pooling or global virtual nodes are introduced to assist in capturing the overall low-frequency background noise within the furnace, ensuring that the spatial mask, while severing long-distance feature associations, does not lose the macroscopic semantics of the overall operating state.

[0065] Based on the aforementioned physical constraint mask, the formula for calculating the attention score is modified as follows:

[0066] Among them, attention matrix yes Q and K Similarity scoring table; It is a normalized exponential function; This is the attention scaling factor; For timing mask; For space mask; Q The input vector is processed by the weight function The resulting vector after linear projection; K The input vector is processed by the weight function The resulting vector after linear projection; V The input vector is processed by the weight function The resulting vector after linear projection; This indicates element-wise multiplication; This is the normalization function.

[0067] Furthermore, the data operation logic of the attention mechanism is as follows: the input is a global spatiotemporal sequence that integrates spatial location and temporal encoding. Equivalent to each spatiotemporal location 256-dimensional global spatiotemporal sequence Through a learnable shared weight matrix , , (All are 256×256 dimensional) Generate query vectors Key vector Value vector Finally, a three-dimensional tensor is obtained. These represent the voiceprint features at the current time step / location, the feature labels for all time / spatial locations, and the real physical information carried by each location, respectively. A scaling dot product operation is performed to calculate the inner product of each pair of spatiotemporal points, thereby representing the feature similarity. Divided by attention scaling factor To prevent the softmax gradient from vanishing due to excessively large dot product values, causality constraints are applied, and a temporal mask is added. Strictly shield future information; perform normalization processing using a normalized exponential function. ,in This represents an element in one row of a matrix. This indicates that all elements in the row are traversed. This process converts the logits into a probability distribution and outputs a strictly lower triangular matrix. After processing by the Softmax function, it is combined with the space mask. Element-wise multiplication achieves a soft constraint where spatial influence gradually diminishes with distance; finally, a function is used... Normalization is performed again, where This indicates the nth element in the attention matrix after mask modulation. i OK j The column elements are used to ensure that the final attention weights follow a valid probability distribution; finally, information extraction is performed by multiplying the attention weight matrix by a reshaping matrix. The output is , representing the characteristics after the fusion of spatiotemporal correlations.

[0068] 2. Feature fusion. Calculate attention score. Then, a feedforward neural network sublayer is added, supplemented with residual connections and normalization layers, to form a complete Encoder Block. Each layer outputs feature representations. A feature fusion module is added after the outputs of layers 3 and 6 of the 6-layer Encoder Block to integrate shallow temporal detail features. With deep semantic features spliced ​​along the channel dimension Then, a 1×1 convolutional layer (512 input channels, 256 output channels, 1×1 kernel size) is used to reduce the dimension back to 256, generating fused features. This preserves acoustic signature information at different scales (such as detailed features of high-frequency bubble bursts and global features of low-frequency pipe vibrations). The final feature vector after processing by the above spatiotemporal multi-scale coding layer... z The data will be fed into the reconstruction loss decoding and probability distribution self-supervised clustering layer, which are used to calculate the mean square error reconstruction loss to constrain feature quality and for flow regime identification and dryness prediction, respectively.

[0069] (3) Reconstruction loss decoder layer: To calculate the reconstruction loss, a symmetric decoder structure is adopted. The decoder converts the feature vectors into a single layer by stacking 6 TransformerDecoder Blocks and 4 1D-Transposed CNNs. zGradually restore the original voiceprint dimensions (160,000 sampling points) and output the reconstructed voiceprint. X' And calculate the mean square error reconstruction loss. .

[0070] (4) Probability Distribution Self-Supervised Clustering Layer. To achieve unsupervised aggregation and output the flow probability, the probability distribution self-supervised clustering layer includes: 1. Soft Assignment Probability Calculation. Based on the Deep Embedded Clustering (DEC) concept, unsupervised flow state recognition is achieved through probabilistic soft assignment. Specifically, this includes calculating sample probabilities using the Student's t-distribution. i Features Belongs to the j Posterior probability of flow-like states:

[0071] in The first output of the coding layer i Each sample feature vector; For the first j Learnable prototype centers for flow-like states (wet / transitional / dry); For degrees of freedom; It is the square of the Euclidean distance.

[0072] It should be noted that degrees of freedom We can take 1, at which point the student t-distribution degenerates into a Cauchy distribution, which has the characteristics of no mean, no variance, and thick tails, and is robust to abnormal soundprint samples in a boiler noise environment. The square of the Euclidean distance indicates that the closer the distance, the larger the numerator and the higher the probability of belonging to the numerator.

[0073] 2. Joint Loss Function. To enable the model to be trained without human annotation, a combination of two losses is used: reconstruction loss. The reconstruction loss is calculated by the decoder layer described above. If the encoded features The ability to reconstruct the original voiceprint indicates Without losing crucial physical details, the encoder can retain the most important information from the original voiceprint. KL divergence clustering loss. Calculate the current probability distribution With a pre-defined high-confidence distribution Differences between Target distribution P Based on the current probability distribution Q Generated after sharpening: ,in The squaring operation amplifies high-confidence samples and compresses low-confidence samples, forcing the model to be more confident in purely wet / dry samples and maintaining reasonable ambiguity for transitional samples, consistent with the asymptotic physical characteristics of working fluid phase transitions. By minimizing the KL divergence clustering loss, the model continuously "compresses" the feature space, making the acoustic features of similar flow states closer to their prototype centers in space. The characteristics of different flow states are far apart, ultimately achieving automatic differentiation between dry and wet states and accurate identification of transition boundaries.

[0074] (5) Dryness Soft Measurement Regression Layer. To achieve smooth quantification of dryness and meet thermodynamic physical constraints, the dryness soft measurement regression layer includes: 1. Calculation of the expected probability baseline. Ensuring smoothness of dryness variation. Based on soft-assignment probabilities from the clustering layer output. Calculate samples using the expectation operator i Initial dryness:

[0075] in, For the first j Dryness reference value for flow-like states; For the sample Belongs to class j The soft assignment probability; For the sample The soft allocation probability belongs to the wet state; For the sample The soft assignment probability that belongs to the transition state; For the sample The soft assignment probability belonging to the dry state; For the sample The initial dryness value.

[0076] Specifically, the dryness reference values ​​for wet, transitional, and dry states are set sequentially as follows: This weighted summation operation ensures that the dryness output changes continuously within the [0,1] interval, avoiding the step jump caused by hard classification, and conforming to the gradual physical characteristics of the phase change of the working fluid in the boiler water-cooled wall.

[0077] 2. Neural Network Regression Fine-tuning. To compensate for clustering discretization errors—that is, setting only three baselines cannot finely characterize the continuous evolution between [0,1]—a lightweight regression fine-tuning head is introduced. A two-layer fully connected network (256→64→1) is used, with the input being the feature vector. The output is a fine-tuning amount. The final dryness form is:

[0078] in, This is a truncation function; For the sample The initial dryness value.

[0079] It should be noted that the truncation function can ensure that the dryness value is strictly within the physically valid range [0,1], preventing the neural network from overflowing and causing unreasonable outputs of negative dryness or dryness exceeding 1.

[0080] 3. Time Continuity Constraint. Due to the thermal inertia of the working fluid phase transition, it is impossible to abruptly change from a purely wet state to a purely dry state within 10 seconds. To prevent drastic jumps in the dryness value within adjacent time windows, constraints are imposed on adjacent time windows. t and t Apply a first-order difference penalty to the dryness output of +1:

[0081] Understandably, this loss function uses the absolute change in working fluid dryness in adjacent time windows as a penalty measure, forcing the output curve to be smooth, which is consistent with the inertial delay characteristics caused by the heat capacity of the water-cooled wall metal and the flow of the working fluid.

[0082] 4. Total Loss Function. The model training employs a three-objective joint optimization, and the total loss function is... for: .

[0083] in, The reconstruction loss is obtained from the above decoder layer; The KL divergence clustering loss is obtained from the above clustering layers; The loss weights for KL divergence clustering; For the aforementioned time continuity constraints; The loss weights are for time continuity constraints.

[0084] It should be noted that the model will update the neural network parameters based on the total loss function during training. (Lower weights for KL divergence clustering) The loss weight can be 1.0, which represents the time continuity constraint. It can be 0.1.

[0085] Specifically, based on the above hybrid architecture, the training process includes two stages: warm-up and alternation. First, the Transformer parameters are frozen, and the CNN feature extractor is pre-trained using pure wet and pure dry historical data and 10% transitional data generated by computational fluid dynamics (CFD) simulation until the validation set reconstruction loss converges. This solves the problem that the deep attention mechanism of the Transformer is difficult to learn effectively when the feature space is not initially formed, and also supplements the scarce transitional samples with CFD simulation data, avoiding the feature extraction bias towards wet / dry states due to class imbalance in the pre-training stage, thus ignoring critical transition features.

[0086] Then, alternating optimization was performed, first fixing the cluster centers. Update CNN and Transformer parameters to optimize reconstruction loss and KL divergence loss. Then, fixing the parameters of the CNN and Transformer, and applying the momentum formula... Update the cluster center locations, where Indicates the first Cluster centers at the next iteration; The instantaneous center is represented by the soft-assignment probability of the current batch of samples. The characteristics of the weighted average are calculated; It is the momentum coefficient, which is the weight for retaining historical memory; The new information weights are the weights of the current observations; this momentum update mechanism is used to smooth the iterative trajectory of cluster centers, effectively avoiding center point oscillations or collapse to local optima caused by hard updates. The above-mentioned alternating optimization strategy is adopted until the following convergence conditions are met: cluster allocation entropy change < 0.1% and the validation set silhouette coefficient no longer increases, or the preset maximum number of iterations is reached. Wherein, cluster allocation entropy... Entropy change The relative change in entropy between two adjacent iterations; a clustering assignment entropy change of <0.1% indicates stable soft assignment; silhouette coefficient. ,in For the sample i The average distance to other points in the same cluster. For the sample i The average distance to the nearest heterogeneous cluster. The closer to 1, the better the clustering effect; set the maximum number of iterations, and force stop if the first two conditions are not met to prevent infinite loop.

[0087] In some embodiments, before using the trained prediction model for the identification of actual boiler acoustic signature data, the method further includes: semi-supervised fine-tuning of the model based on pseudo-labels, and then evaluating the fine-tuned prediction model using unsupervised evaluation metrics.

[0088] Specifically, semi-supervised fine-tuning includes: performing weak and strong enhancement processing on the same original acoustic signature data; inputting weakly enhanced samples into a pre-trained model, selecting samples with prediction confidence greater than a threshold, and using their predicted categories as pseudo-labels; forcing the model to match the predictions of the strongly enhanced version with the pseudo-labels, and updating network parameters through cross-entropy consistency loss to achieve semi-supervised fine-tuning. Here, the original acoustic signature data refers to the original data set collected and pre-processed during the boiler's dry, wet, and dry-wet transition processes; weak enhancement refers to adding slight interference to the original acoustic signature data, and strong enhancement refers to adding strong interference; pseudo-labels refer to the prediction results generated by the model for weakly enhanced acoustic signature samples with confidence higher than a threshold; and cross-entropy consistency loss is... ,in The flow regime is categorized as follows: 1 = wet, 2 = transitional, 3 = dry. It is a pseudo-label. To enhance the prediction results, the network parameters are adjusted using gradient descent. Towards To ensure the stability of prediction results for strong and weak enhanced versions of the model, we need to converge and force a unified model. Specifically, unsupervised evaluation metrics include: clustering quality metrics, which assess whether the model clusters voiceprint data with similar operating conditions into one class by calculating the silhouette coefficient (as above) and the Davies-Bouldin index; among which the Davies-Bouldin index... , The number of fluid clusters, For the first i The average distance from all points within a cluster to its centroid. For the first i Cluster and the first j The distance between the centroids of each cluster; the closer the silhouette coefficient is to 1 or the smaller the Davies-Bouldin exponent, the better the clustering effect. Consistency index: Multiple sets of small noises generated by different random seeds are added to the original speaker data, and multiple prediction results are output. The variance or coefficient of variation (CV) of the multiple prediction results is calculated. ,in The standard deviation of the prediction results is denoted as . The variance or coefficient of variation is the arithmetic mean of multiple predictions. The robustness of the model's predictions is evaluated based on the variance or coefficient of variation. The smaller the variance or the smaller the coefficient of variation, the stronger the robustness of the model.

[0089] It is understood that the embodiments of this application can construct weak and strong enhancement versions by acquiring boiler acoustic data, perform semi-supervised fine-tuning by combining pseudo-labels, and verify the model performance with unsupervised evaluation indicators, thereby improving the accuracy and robustness of the prediction model without the need for manual annotation.

[0090] In step S403, the dry-wet state transition position of the boiler is determined based on the distribution and change results of the dry-wet boundary line.

[0091] The dry-wet transition position refers to the water-cooled wall height position when the dryness first exceeds the threshold and remains stable.

[0092] It is understood that the embodiments of this application can determine the dry-wet state transition position by the distribution and change results of the dry-wet boundary line, which can reflect the dynamic process of the boiler's dry-wet state transition under changing operating conditions in real time and accurately, and provide a basis for load regulation and safe operation.

[0093] Specifically, such as Figure 5 As shown, the boiler dry-wet state transition position measurement process includes: In step 501, acoustic fingerprint data is collected when the boiler is operating stably in dry and wet conditions and during the transition between dry and wet conditions.

[0094] In step 502, the collected voiceprint data is preprocessed.

[0095] Specifically, preprocessing of voiceprint data includes: (1) Data cleaning: Time and frequency domain analysis of acoustic data is performed. Non-normal operating sections, such as the soot blower operating section, are removed by short-time energy threshold and spectrum anomaly detection. Acoustic patterns of adjacent pipes are separated by blind source separation algorithms, such as independent component analysis (ICA) or beamforming technology, so that the model training objects are concentrated on the acoustic patterns of a single pipe. Acoustic patterns with weak correlation to the working medium state, such as environmental noise, mechanical vibration, and flue gas entrainment, are filtered by filtering algorithms (bandpass filter and adaptive noise cancellation algorithm) to retain data with strong correlation to the working medium state and improve the data signal-to-noise ratio.

[0096] (2) Data augmentation: The cleaned acoustic data is resampled at a variable rate and then shifted to simulate the Doppler frequency shift effect caused by the change in working fluid velocity; background noise of different intensities is added to simulate the fluctuations such as flame interference in the boiler internal environment; the signal amplitude is adjusted to simulate the sound attenuation caused by the change in the spacing between sensors; time mask and frequency mask are used to enhance the robustness of the model to local information loss.

[0097] In step 503, based on the preprocessed acoustic fingerprint data, a prediction model for the working fluid flow pattern and dryness measurement of the water-cooled wall is constructed and pre-trained.

[0098] In step 504, the prediction model is evaluated and fine-tuned.

[0099] In step 505, the acoustic fingerprint data of the boiler in real time is collected and preprocessed.

[0100] Specifically, multiple acoustic sensor arrays identical to those used when collecting training data can be deployed to collect acoustic data of the boiler during actual operation with the same parameters. Based on the attenuation characteristics of sound propagation, the phase difference of signals received by different sensors, and the similarity of acoustic data change patterns of adjacent pipes, acoustic data of different pipes can be separated, and the acoustic data of different pipes can be processed in the same way as the preprocessing described above.

[0101] In step 506, the preprocessed acoustic signature data is analyzed using a prediction model, outputting the working fluid flow regime and dryness at different locations within the water-cooled wall, determining the location of the dry-wet transition, and visualizing the results.

[0102] Specifically, the preprocessed acoustic signature data is analyzed using the trained prediction model to output the working fluid flow regime and dryness measurement values ​​at different locations within the boiler water-cooled wall tubes. When the boiler is operating under low-load peak shaving conditions, the dry-wet boundary can be defined as "the dryness first exceeds the threshold". For boilers operating under high load conditions, the threshold can be appropriately relaxed to a height position that remains stable for three or more consecutive time windows. By combining boiler load changes, the boundary curves under different loads are plotted, the load-location correspondence is established, and the furnace height-dryness distribution cloud map and dynamic curve of the conversion position are displayed in real time through a visualization interface.

[0103] It should be noted that the explanation of the above-mentioned embodiment of the measuring equipment for the boiler dry-wet state transition position also applies to the measuring method of the boiler dry-wet state transition position in this embodiment, and will not be repeated here.

[0104] According to the boiler dry-wet transition position measurement method proposed in this application, an acoustic sensor array is used to collect acoustic information of the space surrounding the collection point on the boiler. A collection device is connected to the acoustic sensor array to receive analog signals of the acoustic information and convert them into electrical signals. A processing device obtains the electrical signals from the collection device, extracts acoustic fingerprint features, and processes the acoustic fingerprint data based on a prediction model. This identifies the working fluid flow regime and dryness measurement values ​​at different locations within the boiler water-cooled wall tubes under different operating conditions. Then, based on the working fluid flow regime and dryness measurement values, the distribution and changes of the dry-wet boundary line are determined. Finally, the boiler dry-wet transition position is determined based on these distribution and changes. This method directly identifies the working fluid flow regime and dryness measurement values ​​based on acoustic information, effectively avoiding errors introduced by indirect temperature estimation due to the large thermal inertia of the water-cooled wall, uncertain heat transfer coefficient, and large local pressure fluctuations of the working fluid. This improves the accuracy of dry-wet transition position measurement. Therefore, it solves the problems in related technologies where boiler dry-wet transition position measurement relies on indirect temperature estimation, and where the large thermal inertia of the water-cooled wall, uncertain heat transfer coefficient, and large local pressure fluctuations of the working fluid lead to large errors.

[0105] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be found in any one or more of these embodiments or examples. N The various embodiments or examples are combined in a suitable manner. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described herein, as well as the features of those different embodiments or examples, without contradiction.

[0106] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0107] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0108] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any of the following techniques known in the art, or a combination thereof: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (FPGAs), field-programmable gate arrays (FPGAs), etc.

[0109] Those skilled in the art will understand that all or part of the steps of the methods implementing the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0110] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A measuring apparatus of a wet-dry state conversion position of a boiler, characterized by, include: An acoustic sensor array is used to collect acoustic information about the space surrounding the collection point on the boiler. A data acquisition device is connected to the acoustic sensor array and is used to receive analog signals of acoustic information acquired by the acoustic sensor array and convert the analog signals into electrical signals. The processing device is used to acquire electrical signals from the acquisition device, extract acoustic features from the electrical signals, process the acoustic data based on a prediction model, identify the working fluid flow regime and dryness measurement values ​​at different locations inside the boiler water-cooled wall tubes under different operating conditions, determine the distribution and dynamic changes of the dry-wet boundary line based on the working fluid flow regime and dryness measurement values, and determine the dry-wet state transition position of the boiler based on the distribution and change results of the dry-wet boundary line.

2. The boiler wetness transition position measuring apparatus according to claim 1, characterized by, The acoustic sensor array includes multiple acoustic sensors, which are installed at the acquisition point. The installation method of the acoustic sensors includes magnetic attraction. The mounting surface corresponding to the acquisition point is at least one of the plane that is attached to the boiler exterior and the semi-circular surface on the backfire side that is attached to the finned tube. The acquisition hole corresponding to the acquisition point is arranged on the magnetic attraction surface.

3. The apparatus for measuring the transition position of a boiler from dry to wet state according to claim 1, characterized in that, The acquisition device includes a data transmission line and a signal converter. The signal converter is connected to the acoustic sensor array through the data transmission line to realize the conversion between the analog signal and the electrical signal.

4. A method of measuring the wet-dry transition position of a boiler, characterized by, The method is applied to the processing device of the boiler dry-wet state switching position measuring equipment according to any one of claims 1-3, wherein the method includes the following steps: The electrical signal is acquired from the acquisition device, and the voiceprint feature is extracted from the electrical signal; Based on the prediction model, the acoustic data is processed to identify the working fluid flow regime and dryness measurement values ​​at different locations inside the boiler water-cooled wall tubes under different operating conditions. The distribution and change results of the dry and wet boundary line are determined based on the working fluid flow regime and dryness measurement values. The location of the boiler's dry-wet transition is determined based on the distribution and changes of the dry-wet boundary line.

5. The method of claim 4, wherein Before identifying the working fluid flow regime and dryness at different locations within the boiler water-cooled wall tubes under different operating conditions, the following steps are also included: Extract voiceprint data from the voiceprint features; The voiceprint data is subjected to at least one of the following processes: data cleaning and data enhancement.

6. The method of claim 5, wherein The data cleaning methods include at least one of the following: The acoustic signature data of adjacent pipes are separated using a blind source separation algorithm; The acoustic signature data is filtered based on the correlation of the working medium state. The voiceprint data is analyzed in both the time and frequency domains to remove voiceprint data from non-normal operating segments.

7. The method for measuring the boiler dry-wet state transition position according to claim 5, characterized in that, The data augmentation methods include at least one of the following: Time and frequency masks are used on the cleaned voiceprint data; The signal amplitude of the cleaned voiceprint data is adjusted to simulate changes in the sensor spacing. Variable-rate resampling of cleaned voiceprint data combined with spectrum shifting is used to simulate changes in the working fluid velocity. Background noise of varying intensities was added to the cleaned voiceprint data to simulate fluctuations in the flame interference environment inside the boiler.

8. The method of claim 5, wherein the dryness state of the boiler is determined by the following equation: ###0001### where, P is the pressure of the boiler, T is the temperature of the boiler, and K is a constant. The prediction model includes at least one of the following: a native clustering model without explicit feature learning, a deep generative clustering model, and a clustering model based on representation learning. The clustering model based on representation learning includes a voiceprint feature compression and embedding layer, a spatiotemporal multi-scale coding layer, a reconstruction loss decoder layer, a probability distribution self-supervised clustering layer, and a dryness soft measurement regression layer. Specifically, the voiceprint feature compression and embedding layer extracts, compresses, and spatiotemporally fuses voiceprint data; the spatiotemporal multi-scale coding layer extracts and learns representations from the voiceprint data at spatiotemporal scales; the reconstruction loss decoder layer outputs reconstructed voiceprints through a symmetrical decoder structure and calculates the mean square error reconstruction loss; the probability distribution self-supervised clustering layer automatically classifies the high-dimensional features extracted by the coding layer into different working fluid flow states; and the dryness soft measurement regression layer converts the flow state probabilities obtained from clustering into dryness measurement values.

9. The method of claim 5, wherein the dryness state of the boiler is determined by the following equation: ###0001### where, P is the pressure of the boiler, T is the temperature of the boiler, and K is a constant. Before using the trained prediction model for recognizing actual boiler acoustic signature data, the following steps are also included: The original voiceprint data is weakly and strongly enhanced, and pseudo-labels are generated through model prediction; the prediction model is then semi-supervised and fine-tuned based on the weakly and strongly enhanced voiceprint data and the pseudo-labels. The fine-tuned prediction model was evaluated using unsupervised evaluation metrics.

10. A computer readable storage medium having stored thereon a computer program or instructions, characterized in that, When the computer program or instructions are executed, they implement the method for measuring the dry-wet transition position of the boiler as described in any one of claims 4-9.

Images