Acoustic intelligent sensor for fire-fighting pipe network leakage detection

Abstract

The invention relates to the technical field of fire-fighting pipe network leakage detection, and particularly discloses an acoustic intelligent sensor for fire-fighting pipe network leakage detection, which uses a signal analysis and denoising technology based on deep learning to carry out time-frequency and spatial domain transformation on a sound signal in a pipeline. And feature extraction is carried out on the transformed pipeline internal sound time-frequency domain image and spatial domain image, so that a denoised pipeline internal sound signal is intelligently generated according to cross-modal ablation interaction representation between the pipeline internal sound time-frequency feature and the pipeline internal sound spatial distribution feature. Therefore, the internal parameters can be automatically adjusted according to the actually input sound signal, and useful leakage characteristics are distinguished and reserved according to the context information, so that the dynamic change of the signal can be better adapted. Moreover, subtle changes and modes of time-frequency and spatial distribution in sound signals can be captured, thereby identifying those small-scale or low-intensity leakage events that may be neglected by conventional methods.

Description

Technical Field

[0001] The present application relates to the technical field of fire protection pipe network leakage detection, and more specifically, to an acoustic intelligent sensor for fire protection pipe network leakage detection. Background Art

[0002] In today's society, fire safety is of vital importance. As the key infrastructure of the fire protection system, the fire protection pipe network bears the heavy responsibility of transporting fire protection water. However, with the passage of time, the aging of the pipe network and the influence of external environmental factors, the leakage problem of the fire protection pipe network has become more and more prominent.

[0003] Traditional fire protection pipe network leakage detection methods mainly rely on manual inspections. Workers need to regularly check for leakage along the pipe network by listening to the sound of leakage and observing whether there are water stains on the ground. This method is not only inefficient and consumes a lot of manpower and material resources, but also the detection accuracy is limited by human experience and sensory acuity, making it difficult to detect tiny leaks and leakage points hidden deep underground or inside buildings.

[0004] Therefore, an acoustic intelligent sensor for fire protection pipe network leakage detection is desired. Summary of the invention

[0005] The present application provides an acoustic intelligent sensor for fire protection pipe network leakage detection, which can automatically adjust its internal parameters according to the actual input sound signal, distinguish and retain useful leakage features according to context information, so as to better adapt to the dynamic changes of the signal. In addition, it can capture the subtle changes and patterns of time-frequency and spatial distribution in the sound signal, so as to identify small-scale or low-intensity leakage events that may be ignored by traditional methods.

[0006] According to one aspect of the present application, an acoustic intelligent sensor for fire protection pipe network leakage detection is provided, comprising: a sound signal acquisition module, for collecting real-time sound signals inside the pipe through acoustic intelligent sensors installed at key positions of the fire protection pipe network; a sound signal uploading module, for uploading the sound signals inside the pipe to a cloud platform; a denoising and optimization module, for performing denoising and optimization on the sound signals inside the pipe on the cloud platform to obtain denoised sound signals inside the pipe; a leakage detection module, for performing fire protection pipe network leakage detection on the denoised sound signals inside the pipe on the cloud platform to obtain a detection result indicating whether there is leakage in the fire protection pipe network; and an early warning module, for sending an early warning signal to a terminal device of a maintenance personnel in response to the detection result indicating that there is leakage in the fire protection pipe network.

[0007] The denoising optimization module includes: an image transformation unit, which is used to perform time-frequency and space-domain transformations on the sound signal inside the pipeline to obtain a time-frequency domain image of the sound inside the pipeline and a space-domain image of the sound inside the pipeline.

[0008] The sound image feature extraction unit is used to extract features from the time-frequency domain image of the sound inside the pipeline and the space-domain image of the sound inside the pipeline to obtain the time-frequency features of the sound inside the pipeline and the space distribution features of the sound inside the pipeline.

[0009] A cross-modal ablation interaction unit is used to perform cross-modal ablation attention depth interaction on the time-frequency characteristics of the sound inside the pipe and the spatial distribution characteristics of the sound inside the pipe to obtain the cross-modal ablation time-frequency coding characteristics guided by the spatial distribution of the sound inside the pipe.

[0010] The denoised image generating unit is used to guide the cross-modal ablation time-frequency coding features based on the spatial distribution of the sound inside the pipeline to obtain the denoised sound signal inside the pipeline.

[0011] The present application provides an acoustic intelligent sensor for fire protection pipe network leakage detection, which uses deep learning-based signal analysis and denoising technology to transform the sound signal inside the pipe in the time-frequency and space domains, and then extracts features from the transformed time-frequency domain image and space-domain image of the sound inside the pipe, thereby intelligently generating a denoised sound signal inside the pipe based on the cross-modal ablation interaction representation between the time-frequency features of the sound inside the pipe and the spatial distribution features of the sound inside the pipe. Unlike traditional filters with fixed parameters, the present application can automatically adjust its internal parameters according to the actual input sound signal, distinguish and retain useful leakage features according to contextual information, and thus better adapt to the dynamic changes of the signal. In addition, it can capture subtle changes and patterns in the time-frequency and spatial distribution of sound signals, thereby identifying small-scale or low-intensity leakage events that may be overlooked by traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings of the embodiments of the present application are briefly introduced below. Obviously, the drawings described below only relate to some embodiments of the present application, and are not intended to limit the present application.

[0013] Figure 1 A schematic block diagram of an acoustic intelligent sensor for fire protection pipe network leakage detection according to an embodiment of the present application.

[0014] Figure 2 This is a data flow diagram of an acoustic intelligent sensor for fire protection pipe network leakage detection according to an embodiment of the present application.

[0015] Figure 3 A schematic block diagram of a denoising optimization module in an acoustic intelligent sensor for fire protection pipe network leakage detection according to an embodiment of the present application.

[0016] Figure 4A schematic block diagram of a sound image feature extraction unit in an acoustic intelligent sensor for fire protection pipe network leakage detection according to an embodiment of the present application.

[0017] Figure 5 A schematic block diagram of a cross-modal ablation interaction unit in an acoustic intelligent sensor for fire protection pipe network leakage detection according to an embodiment of the present application. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by ordinary technicians in this field without creative work also fall within the scope of protection of the present application.

[0019] Based on this, this application proposes an acoustic intelligent sensor for fire protection pipe network leakage detection, which can realize automatic, high-precision and real-time leakage monitoring to significantly improve detection efficiency and accuracy, while greatly reducing labor costs and resource consumption, ensuring that even the smallest or most hidden leakage can be discovered and handled in a timely manner. Figure 1 and Figure 2 As shown, the acoustic intelligent sensor for fire protection pipe network leakage detection includes: a sound signal acquisition module 10, which is used to collect the sound signal inside the pipeline in real time through the acoustic intelligent sensor installed at the key position of the fire protection pipe network; a sound signal uploading module 20, which is used to upload the sound signal inside the pipeline to the cloud platform; a denoising optimization module 30, which is used to denoise and optimize the sound signal inside the pipeline on the cloud platform to obtain the denoised sound signal inside the pipeline; a leakage detection module 40, which is used to perform fire protection pipe network leakage detection on the denoised sound signal inside the pipeline on the cloud platform to obtain a detection result indicating whether there is leakage in the fire protection pipe network; an early warning module 50, which is used to send an early warning signal to the terminal device of the maintenance personnel in response to the detection result that the fire protection pipe network has leakage.

[0020] For example, in the sound signal acquisition module 10, the acoustic intelligent sensors installed at the key positions of the fire pipe network collect the sound signals inside the pipe in real time. It should be understood that when the fire pipe network leaks, the water suddenly sprays from the high-pressure environment to the low-pressure environment, which will produce unique turbulent noise, and this noise will propagate along the pipe in the form of sound waves. Therefore, the sensors located at the key positions of the pipe network can monitor these abnormal sound patterns and analyze them, thereby realizing the monitoring of the conditions inside the pipe.

[0021] In one embodiment, to ensure effective signal acquisition, acoustic smart sensors are installed in locations where leakage is most likely to occur or where leakage will have the greatest impact, such as near valves, branch connection points, and areas where problems frequently occur in historical records. These places are considered "critical locations" in the pipe network. The sensor needs to have high sensitivity and wide-band response characteristics so that it can accurately sense weak or high-frequency leakage sound waves, while also being able to distinguish between background noise generated by normal water flow and other non-leakage sources (such as pump operation, water flow pulsation). Acoustic smart sensors have built-in microphones or other forms of acoustic sensors to receive sound waves. When the sound wave reaches the sensor, it causes the sensitive material to vibrate, thereby generating an electrical signal change corresponding to the incident sound pressure.

[0022] Exemplarily, in the sound signal uploading module 20, the sound signal inside the pipeline is uploaded to the cloud platform. It should be understood that the fire protection pipe network usually covers a wide geographical area, including multiple branches and nodes, and each location is equipped with an acoustic intelligent sensor for collecting sound signals. If each sensor needs to independently complete complex signal analysis tasks, powerful computing resources need to be equipped locally, which is not only costly, but also difficult to maintain unified data standards and algorithm updates.

[0023] In this application, by uploading the sound signal to the cloud platform, the above problems can be solved and many benefits can be brought. All sound data from different geographical locations can be managed and stored on a centralized platform, which is convenient for global data analysis and pattern recognition. At the same time, the cloud platform provides powerful computing power, storage space and advanced machine learning models. These resources can be shared by multiple sensors, reducing the cost and technical requirements of a single device. In addition, with the powerful computing power of the cloud platform, the uploaded sound signals can be processed quickly, and the detection results can be fed back to the relevant maintenance personnel in a timely manner to ensure that the leakage incident can be handled quickly. With the accumulation of more data, the cloud platform-based algorithm can improve its accuracy and robustness through continuous training and adjustment to adapt to new leakage characteristics or environmental changes.

[0024] In one embodiment, the original sound signal collected by the acoustic smart sensor is processed initially (such as analog-to-digital conversion) to form a data stream in digital format. Then, using network communication protocols (such as HTTP, MQTT, etc.), these data are sent to a designated server address on the Internet, i.e., the service port of the cloud platform, through a wired network (such as Ethernet) or a wireless network (such as Wi-Fi, 4G / 5G cellular network). In order to ensure the security and integrity of the transmission, encryption technology and verification mechanisms can also be used. After arriving at the cloud platform, the received data will be parsed and stored according to the preset process, and enter the subsequent processing stage.

[0025] Exemplarily, in the denoising optimization module 30, on the cloud platform, the sound signal inside the pipeline is denoised and optimized to obtain the sound signal inside the pipeline after denoising. It should be understood that the sound signal inside the pipeline is extremely complex, mixed with many components such as the normal flow sound of water, the opening and closing sound of valves, the external environmental noise, and the potential leakage sound. It is far from enough to simply collect the signal. It must be processed efficiently and accurately to remove noise interference and extract the real leakage signal. However, the traditional filtering methods, such as low-pass, high-pass, and band-pass filters, have fixed filtering parameters. For example, when processing the sound signal of the fire pipe network, these filters cannot automatically adjust the parameters according to the dynamic changes of the sound signal. For complex and changeable internal pipe sounds, such as normal water flow sound and leakage sound, there may be overlapping parts in frequency, and the fixed filtering frequency range may weaken the useful leakage signal while filtering out the noise. In addition, when processing such signals, the traditional filter may not track the time-varying characteristics of the signal well, which may lead to poor filtering effect. For example, when a high-frequency leakage sound suddenly appears in a pipe and the filter cutoff frequency is set low, the high-frequency component that may be used to determine the leakage cannot be effectively retained.

[0026] Accordingly, in the denoising optimization module, the technical concept of the present application is to use deep learning-based signal analysis and denoising technology to transform the sound signal inside the pipeline in the time-frequency and space domains, and then extract features from the transformed time-frequency domain image and space-domain image of the sound inside the pipeline, so as to intelligently generate the denoised sound signal inside the pipeline based on the cross-modal ablation interaction representation between the time-frequency features of the sound inside the pipeline and the spatial distribution features of the sound inside the pipeline. Unlike the fixed parameters of traditional filters, the present application can automatically adjust its internal parameters according to the actual input sound signal, distinguish and retain useful leakage features according to contextual information, so as to better adapt to the dynamic changes of the signal. In addition, it can capture subtle changes and patterns in the time-frequency and spatial distribution of sound signals, thereby identifying small-scale or low-intensity leakage events that may be overlooked by traditional methods.

[0027] In one embodiment, Figure 3As shown, the denoising optimization module 30 includes: an image transformation unit 31, which is used to perform time-frequency and space domain transformations on the sound signal inside the pipeline respectively to obtain a time-frequency domain image of the sound inside the pipeline and a space domain image of the sound inside the pipeline; a sound image feature extraction unit 32, which is used to extract features from the time-frequency domain image of the sound inside the pipeline and the space domain image of the sound inside the pipeline respectively to obtain time-frequency features of the sound inside the pipeline and spatial distribution features of the sound inside the pipeline; a cross-modal ablation interaction unit 33, which is used to perform cross-modal ablation attention depth interaction on the time-frequency features of the sound inside the pipeline and the spatial distribution features of the sound inside the pipeline to obtain cross-modal ablation time-frequency coding features guided by the spatial distribution of the sound inside the pipeline; and a denoising image generation unit 34, which is used to obtain the denoised sound signal inside the pipeline based on the cross-modal ablation time-frequency coding features guided by the spatial distribution of the sound inside the pipeline.

[0028] Exemplarily, the image transformation unit 31 is used to perform time-frequency and space-domain transformations on the sound signal inside the pipeline to obtain a time-frequency domain image of the sound inside the pipeline and a space-domain image of the sound inside the pipeline. It should be understood that considering that the sound signal inside the fire pipe network is a non-stationary signal, its frequency component will change dynamically over time. For example, the frequency of the sound of water flow in a normally operating fire pipe network is relatively stable, but when a leak occurs, due to changes in the state of the water flow, turbulence at the leak port, etc., the frequency of the sound signal will change on the time axis, and the frequency of some noise sounds is also different from the normal frequency. This non-stationary characteristic makes it difficult for traditional signal processing methods that only consider frequency or time single dimensions to effectively remove noise and extract useful information. Based on this, in one embodiment, the image transformation unit is used to: perform S transformation on the sound signal inside the pipeline to obtain the time-frequency domain image of the sound inside the pipeline; perform RP transformation on the sound signal inside the pipeline to obtain the space-domain image of the sound inside the pipeline.

[0029] Specifically, the sound signal inside the pipeline is subjected to an S transform to obtain a time-frequency domain image of the sound inside the pipeline. It should be understood that the S transform is a time-frequency analysis method, which is an improvement on the short-time Fourier transform (STFT) and wavelet transform, and has unique advantages in the analysis of non-stationary signals. By performing an S transform on the sound signal inside the pipeline, detailed information on the sound signal in both time and frequency dimensions can be obtained in the time-frequency domain image. For example, for normal water flow sound, a relatively stable low-frequency area distribution may be presented in the time-frequency domain image; while leakage sound may have abnormal peaks or fluctuations at certain specific time points and frequency areas in the time-frequency domain image, and these abnormal parts are likely to contain noise components. In this way, the frequency distribution of the sound signal at different times can be clearly seen, providing a basis for the subsequent distinction between normal and noise sounds.

[0030] Specifically, the sound signal inside the pipeline is RP transformed to obtain the spatial domain image of the sound inside the pipeline. Considering that in the environment of the fire protection pipe network, the noise sources are diverse and have different spatial distribution characteristics. For example, the external environmental noise (such as the sound of surrounding machine operation, traffic noise, etc.) and the noise generated by the pipeline itself (such as the sound of normal water flow, slight vibration of the pipeline) are different in spatial propagation and distribution. Therefore, in order to clearly display the sound signals at different spatial positions, thereby helping to distinguish the noise from different positions and the sound signals generated by pipeline leakage, in the technical solution of the present application, the sound signal inside the pipeline is RP transformed to convert from a one-dimensional time series to a two-dimensional phase space representation, and obtain the spatial domain image of the sound inside the pipeline. It can be understood that the RP transformation (recursion graph) is a tool for analyzing nonlinear time series data. For sound signals, it constructs a graphical representation based on the recursive characteristics of the signal in the time series. The recursion graph converts the time series of the sound signal into a spatial domain image, and each point in this image reflects the similarity of the sound signal between different time points. Specifically, when processing sound signals inside a pipeline, the recursive graph provides the spatial distribution characteristics of the sound signal in the time dimension, which can be used to distinguish noise from useful signals. For example, noise usually has random characteristics and may appear as irregular point distribution in the recursive graph, while useful pipeline sound signals (such as normal water flow sound or leakage sound) may have a relatively more orderly structure.

[0031] In one embodiment, Figure 4 As shown, the sound image feature extraction unit 32 includes: a pipe internal sound time-frequency feature extraction subunit 321, which is used to pass the pipe internal sound time-frequency domain image through a sound time-frequency feature extractor based on an MCNN model to obtain a pipe internal sound time-frequency feature map as the pipe internal sound time-frequency feature; a pipe internal sound space distribution feature extraction subunit 322, which is used to pass the pipe internal sound space domain image through a sound space distribution feature extractor based on a CNN-LSTM model to obtain a pipe internal sound space distribution feature vector as the pipe internal sound space distribution feature.

[0032] In the pipe internal sound time-frequency feature extraction subunit 321, it is considered that the pipe internal sound time-frequency domain image contains complex time-frequency information, in which there are time-frequency features related to the inside of the pipe, but there are also some noise information interference and influence, and these time-frequency information will change complexly over time. Therefore, in the technical solution of the present application, the pipe internal sound time-frequency domain image is extracted by a sound time-frequency feature extractor based on the MCNN model to obtain the pipe internal sound time-frequency feature map. In other words, the MCNN model usually has the advantage of multi-channel processing. For the pipe internal sound time-frequency domain image, there may be multiple different types of time-frequency features that need to be extracted simultaneously. For example, on the one hand, it may be necessary to extract the features of low-frequency stable signals to determine the normal water flow state, and on the other hand, it is necessary to pay attention to the features of high-frequency burst signals to detect possible leakage. The multi-channel structure of the MCNN model can adapt well to this demand, and each channel can focus on extracting a specific type of time-frequency feature, so as to more comprehensively mine the useful information in the pipe internal sound time-frequency domain image and better understand the dynamic time changes of the signal, while reducing the interference and influence of background noise.

[0033] In the pipeline internal sound spatial distribution feature extraction subunit 322, considering that the pipeline internal sound spatial domain image contains complex spatial information, there are spatial associations and change rules between its pixels. Moreover, in the actual monitoring process, these spatial domain images may change dynamically over time and have a certain sequence. For example, with the change of water flow state or the development of leakage in the fire protection pipe network, the propagation and distribution of sound in the internal space of the pipeline will be different at different times. Therefore, in order to capture these spatial feature information and understand the correlation between these spatial features, the present application captures the pipeline internal sound spatial domain image through a sound spatial distribution feature extractor based on the CNN-LSTM model to obtain the pipeline internal sound spatial distribution feature vector. It should be understood that CNN has excellent performance in processing the spatial features of image data. It can automatically extract local spatial features in the image through convolutional layers and pooling layers. For the internal sound spatial domain image of the pipeline, CNN can effectively extract the distribution characteristics of sound intensity in the pipeline space, the possible location information of the sound source, and the spatial change characteristics during the sound propagation process. For example, the filter of the convolution layer can capture the position and range of the sound intensity peak in the spatial domain image. These features are very important for determining the location of the leak point and other information. The LSTM network is good at processing time series data and can learn long-term dependencies in the data. In the pipeline sound monitoring scenario, considering the temporal correlation between spatial domain image sequences, LSTM can learn the change rules between spatial domain images at different times. For example, when a pipeline leaks, the propagation range and intensity distribution of the leaking sound in space will gradually change over time. LSTM can capture the characteristics of this dynamic change, so as to better understand the evolution process of the spatial distribution of sound and provide support for the accurate extraction of spatial distribution features. Therefore, the CNN-LSTM model can simultaneously consider the spatial distribution of sound and its dynamic characteristics over time, providing data support for subsequent image denoising and leakage detection.

[0034] Exemplarily, the cross-modal ablation interaction unit 33 is used to perform cross-modal ablation attention depth interaction on the time-frequency characteristics of the sound inside the pipeline and the spatial distribution characteristics of the sound inside the pipeline to obtain the time-frequency coding characteristics of the cross-modal ablation guided by the spatial distribution of the sound inside the pipeline. It should be understood that, considering that the time-frequency characteristics of the sound inside the pipeline reflect the characteristics of the sound signal in the time and frequency dimensions, and the spatial distribution characteristics of the sound inside the pipeline reflect the relevant characteristics of the sound in the space inside the pipeline, these two characteristics describe the sound inside the pipeline from different angles, and there is a mutual fine-grained relationship between the two. Therefore, in order to be able to fully and accurately grasp the sound conditions inside the pipeline, so as to dig out the potential correlation between them and make the characterization of the pipeline sound more comprehensive and detailed, in the technical solution of the present application, the time-frequency characteristics of the sound inside the pipeline and the spatial distribution characteristics of the sound inside the pipeline are deeply interacted by cross-modal ablation attention to obtain the time-frequency coding characteristics of the cross-modal ablation guided by the spatial distribution of the sound inside the pipeline. In this way, the attention to different parts of the time-frequency characteristics and spatial distribution characteristics can be adaptively adjusted according to the actual situation of the current pipeline sound. Let the model focus on the more discriminative time-frequency features and the associated spatial features in the current context, so as to more accurately capture the mutual connection between the two modalities, achieve more effective feature fusion, and extract features with more specific representation capabilities.

[0035] In one embodiment, Figure 5 As shown, the cross-modal ablation interaction unit 33 includes: a pipe internal sound time-frequency decomposition subunit 331, which is used to perform feature fine-grained decomposition on the pipe internal sound time-frequency feature map to obtain a set of pipe internal local sound time-frequency feature matrices; a pipe internal sound feature weight calculation subunit 332, which is used to calculate the multidimensional fine-grained ablation weight factor of each pipe internal local sound time-frequency feature matrix in the set of pipe internal local sound time-frequency feature matrices based on the pipe internal sound spatial distribution feature vector to obtain a set of multidimensional fine-grained ablation weight factors of pipe internal sound; a pipe internal sound cross-modal encoding subunit 333, which is used to perform fine-grained ablation modulation and feature aggregation on the set of pipe internal local sound time-frequency feature matrices based on the set of multidimensional fine-grained ablation weight factors of pipe internal sound to obtain a pipe internal sound spatial distribution guided cross-modal ablation time-frequency coding feature map as the pipe internal sound spatial distribution guided cross-modal ablation time-frequency coding feature.

[0036] In one embodiment, the pipe internal sound time-frequency decomposition subunit 331 is used to: perform feature fine-grained decomposition of the pipe internal sound time-frequency feature map along the channel dimension of the pipe internal sound time-frequency feature map to obtain a set of local sound time-frequency feature matrices inside the pipe. Specifically, the process can be expressed as follows: ;in, is the time-frequency characteristic diagram of the sound inside the pipeline, For feature decomposition operations, and is the first, second, and third in the set of local sound time-frequency feature matrices inside the pipeline. and The local sound time-frequency feature matrix inside the pipe.

[0037] Specifically, in the pipe internal sound time-frequency decomposition subunit 331, the feature map is finely decomposed along the channel dimension of the pipe internal sound time-frequency feature map in order to extract more detailed and specific local sound time-frequency features from complex multi-channel signals. This process can be understood as dividing the originally integrated sound signals into multiple smaller and easier to analyze parts according to different frequencies or time windows, thereby forming a collection of a series of local sound time-frequency feature matrices. In this way, each local sound time-frequency feature matrix can capture the acoustic characteristics within a specific frequency range in a specific time period, such as the normal flow sound of water flow, the turbulent noise generated by possible leakage, etc. This fine-grained decomposition helps to more accurately identify those weak but critical sound features, especially when faced with low-intensity or small-scale leakage events, the sound patterns that may not be effectively detected by traditional methods become easier to be discovered and analyzed after this step. In addition, since the sound environment of the fire pipe network is complex and changeable, the sound signals at different locations and times will be different, so the fine-grained decomposition of the time-frequency feature map can also help the model better adapt to these changes.

[0038] In one embodiment, the pipe internal sound feature weight calculation subunit 332 is used to: use each pipe internal local sound time-frequency feature matrix in the set of the pipe internal local sound time-frequency feature matrix as a key matrix, perform multi-dimensional encoding of the pipe internal sound spatial distribution feature vector and the key matrix to simulate the converter ablation factor to obtain a set of multi-dimensional fine-grained ablation factors of the pipe internal sound. Specifically, the process can be expressed by the formula: ;in, is the spatial distribution feature vector of the sound inside the pipe, is matrix multiplication, and They are respectively the embedding matrix for querying the spatial distribution features of sound inside the pipeline and the bias vector for querying the spatial distribution features of sound inside the pipeline. is the query vector of the spatial distribution characteristics of the sound inside the pipe, and are respectively the embedding matrix of the eigenvalue of the spatial distribution of sound inside the pipeline and the bias vector of the eigenvalue of the spatial distribution of sound inside the pipeline, is the eigenvalue vector of the spatial distribution of sound inside the pipe, and They are respectively the embedding matrix of the local sound time-frequency feature key inside the pipeline and the bias vector of the local sound time-frequency feature key inside the pipeline, yes The corresponding time-frequency characteristic key matrix of the local sound inside the pipeline, yes The transposed matrix of yes The scale is the matrix width multiplied by the matrix height, yes function, It is the first in the set of multi-dimensional fine-grained interaction vectors of local sound inside the pipeline. The multi-dimensional fine-grained interaction vector of the local sound inside the pipe, To extract the maximum value, and They are The mean and variance of is a hyperparameter, For calculation The fine-grained ablation factor, is the first in the set of multi-dimensional fine-grained ablation factors for the sound inside the pipeline. A multi-dimensional fine-grained ablation factor of the sound inside the pipe.

[0039] The set of multi-dimensional fine-grained ablation factors of the internal sound of the pipeline is normalized and masked to obtain a set of multi-dimensional fine-grained ablation weight factors of the internal sound of the pipeline. Specifically, the process can be expressed by a formula as follows: ;in, For calculation The fine-grained ablation weight factor, represents the exponential function value with the natural constant e as the base, is the set of multi-dimensional fine-grained ablation factors of the normalized internal sound of the pipeline. Normalized multi-dimensional fine-grained ablation factor of the internal sound of the pipeline, is a mask operation, is the preset threshold, is the first in the set of multi-dimensional fine-grained ablation weight factors for the sound inside the pipeline. Multi-dimensional fine-grained ablation weight factors for the sound inside the pipeline.

[0040] Specifically, in the subunit 332 for calculating the weight of the sound feature inside the pipeline, each local sound time-frequency feature matrix in the set of the local sound time-frequency feature matrix inside the pipeline is used as a key matrix, and the spatial distribution feature vector of the sound inside the pipeline and the local sound time-frequency feature matrix inside the pipeline are multi-dimensionally encoded with the transformer-like ablation factor to obtain a set of fine-grained ablation factors. The use of the transformer-like structure to calculate the fine-grained ablation weight factor actually draws on the idea of ​​the self-attention mechanism, and has been adaptively improved to be suitable for multimodal data processing. In this process, a set of multi-dimensional fine-grained ablation weight factors of the sound inside the pipeline will be calculated, which reflects the importance of the removal of a specific local sound time-frequency feature on the sound spatial distribution feature. In other words, it measures the impact on the final feature cross-modal interaction when a local sound feature is ignored, thereby helping to determine which features are critical.

[0041] Next, the obtained set of fine-grained ablation factors needs to be normalized and masked to generate the final set of fine-grained ablation weight factors. The normalization step ensures that all ablation factors are compared under a unified standard to avoid misjudgment due to scale differences; at the same time, it also helps to adjust the weight distribution so that those truly important features can receive higher attention in subsequent processing. Masking is to further highlight key features and suppress irrelevant or redundant information. For example, in practical applications, sound signals in certain frequency ranges may be more important for detecting leaks, while other parts are relatively unimportant. Through masking operations, it can ensure that the model focuses more on these important features and improves detection accuracy.

[0042] In one embodiment, the pipe internal sound cross-modal encoding subunit 333 is used to: based on the set of multi-dimensional fine-grained ablation weight factors of the pipe internal sound, perform fine-grained ablation modulation on the set of local sound time-frequency feature matrices in the pipe internal to obtain a set of local feature matrices of cross-domain guided ablation pipe internal sound; perform feature aggregation on the set of local feature matrices of cross-domain guided ablation pipe internal sound to obtain the pipe internal sound spatial distribution guided cross-modal ablation time-frequency encoding feature map. Specifically, the process can be expressed by the formula: ;in, and is the first, second, and third in the set of multi-dimensional fine-grained ablation weight factors for the sound inside the pipeline. and Multi-dimensional fine-grained ablation weight factors of the internal sound of the pipeline, It is the time-frequency coding feature map of the spatial distribution of sound inside the pipeline guiding cross-modal ablation.

[0043] Specifically, in the pipe internal sound cross-modal encoding subunit 333, the previously calculated fine-grained ablation weight factor set is used as a guide to apply corresponding weight adjustments to each local sound time-frequency feature matrix. This means that if a local sound feature is determined to be very important for detecting leakage (i.e., the corresponding ablation weight factor is high), its importance in subsequent processing will be amplified; on the contrary, those features that are considered less relevant or redundant will be appropriately suppressed. This fine-grained modulation ensures that the model can focus more on key information that is truly helpful in identifying leakage while avoiding interference from irrelevant noise. Next, when all modulated cross-domain guided ablation pipe internal sound local feature matrices are aggregated, a comprehensive cross-modal feature representation is actually constructed, that is, the pipe internal sound spatial distribution guided cross-modal ablation time-frequency encoding feature map. This feature map combines rich information from the time-frequency domain and the spatial domain, and contains carefully selected and optimized sound features that not only retain the uniqueness of the original data, but also enhance the synergy between different modalities through cross-modal interaction.

[0044] Exemplarily, the denoised image generation unit 34 is used to obtain the denoised internal sound signal of the pipeline based on the time-frequency coding features of the spatial distribution of sound inside the pipeline. In one embodiment, the denoised image generation unit is used to: expand the time-frequency coding feature map of the spatial distribution of sound inside the pipeline guiding cross-modal ablation into a time-frequency coding feature vector of the spatial distribution of sound inside the pipeline guiding cross-modal ablation, add generation prompt information to the tail of the time-frequency coding feature vector of the spatial distribution of sound inside the pipeline guiding cross-modal ablation, and input it into the sound signal denoising generator based on AIGC to obtain the denoised internal sound signal of the pipeline, wherein the generation prompt information is: based on the input corpus, the internal sound signal of the pipeline is denoised and optimized. That is, the spatial distribution of sound inside the pipeline obtained by cross-modal ablation interaction using the time-frequency feature map of the sound inside the pipeline and the spatial distribution feature vector of the sound inside the pipeline guides the generation and processing of cross-modal ablation time-frequency coding features, thereby utilizing AIGC (artificial intelligence generated content) technology to demonstrate powerful capabilities in processing and generating complex content, deeply mining and analyzing the rich sound information contained in the feature vector, and intelligently generating the denoised internal pipeline sound signal based on the denoising optimization goal required in the prompt information.

[0045] Specifically, by adding specific generation hint information to the tail of the time-frequency encoding feature vector of the cross-modal ablation guided by the spatial distribution of sound inside the pipeline, the AIGC-based generator can be clearly informed of the task to be performed next, that is, to optimize the denoising of the sound signal. This approach helps the generator understand the specific needs of the input data, so as to adjust its internal parameters more specifically and improve the quality of denoising. The types of noise in different application scenarios may be different. For example, the background noise in some environments may be a constant hum, while in other environments it may be intermittent impact noise. Providing specific generation hint information allows the model to flexibly adjust the processing strategy according to the actual scenario, rather than relying on fixed algorithms or preset conditions. At the same time, the generation hint information helps maintain the contextual consistency from the original sound signal to the final output, which means that even in complex and changing environments, the generator can maintain sensitivity to the characteristics of the original signal and avoid the loss of useful information due to excessive denoising. By describing the input corpus (i.e., the original sound signal), the generator can better understand the characteristics of the signal, thereby producing higher quality denoising results.

[0046] In one example, adding generation prompt information to the tail of the time-frequency encoding feature vector of the spatial distribution of sound inside the pipeline and inputting it into the sound signal denoising generator based on AIGC to obtain the denoised sound signal inside the pipeline, including: first constructing a concise generation prompt information according to the task requirements, such as "based on the input corpus, denoise and optimize the sound signal inside the pipeline". This information should clearly express the work expected to be completed by the generator, as well as any necessary parameter settings or restrictions. Then convert the constructed generation prompt information into a form suitable for machine processing (such as an embedded vector) and attach it to the end of the time-frequency encoding feature vector of the spatial distribution of sound inside the pipeline. This ensures that the generator can obtain the data itself and its corresponding instructions at the same time when receiving the input. Next, the spliced ​​feature vector is input as a whole to the sound signal denoising generator based on AIGC. In this process, the generator will use the powerful capabilities of the deep learning model to conduct in-depth analysis of the input data and perform corresponding operations according to the requirements of the prompt information, that is, intelligent denoising of the sound signal. After the above steps, the generator will output an optimized and denoised internal pipe sound signal, which not only retains the original important features but also significantly reduces the impact of background noise, thereby improving the accuracy and reliability of subsequent leakage detection.

[0047] Preferably, in one example, generating prompt information is added at the tail of the time-frequency coding feature vector of the spatial distribution of sound inside the pipeline guiding cross-modal ablation, and inputting it into the sound signal denoising generator based on AIGC to obtain the denoised internal pipeline sound signal, including: determining the number of isolated eigenvalues ​​of the time-frequency coding feature vector of the spatial distribution of sound inside the pipeline guiding cross-modal ablation whose sum of distances to adjacent eigenvalues ​​is greater than a predetermined threshold value as a hyperparameter. and its corresponding spatial distribution of sound inside the pipe to guide cross-modal ablation time-frequency coding smoothing factor , that is: ;in, represents a predetermined threshold as a hyperparameter, They represent the first order of the time-frequency encoding feature vector of the cross-modal ablation guided by the spatial distribution of sound inside the pipeline. , , The eigenvalues ​​at the positions, Represents a counting function.

[0048] Calculate the square root of the sum of squares of all eigenvalues ​​of the time-frequency coding feature vector of the internal sound spatial distribution of the pipeline guiding cross-modal ablation and divide it by the length of the time-frequency coding feature vector of the internal sound spatial distribution of the pipeline guiding cross-modal ablation to obtain the principal component domain value of the time-frequency coding feature vector of the internal sound spatial distribution of the pipeline guiding cross-modal ablation ,in, The length of the time-frequency encoding feature vector that represents the spatial distribution of sound inside the pipeline and guides cross-modal ablation.

[0049] The number of isolated eigenvalues ​​of cross-modal ablation time-frequency coding guided by the spatial distribution of sound inside the pipeline The spatial distribution of sound inside the pipeline is calculated as a power function of the time-frequency encoding feature vector guided by cross-modal ablation and the spatial distribution of sound inside the pipeline guides cross-modal ablation time-frequency coding smoothing factor Perform point multiplication to obtain the spatial distribution of sound inside the pipe to guide cross-modal ablation time-frequency coding orthogonal basis eigenvector ,in, Represents dot product.

[0050] Calculate the time-frequency coding feature vector of the sound spatial distribution inside the pipeline guiding cross-modal ablation and the smooth factor of the time-frequency coding of the sound spatial distribution inside the pipeline guiding cross-modal ablation The spatial distribution of sound inside the pipeline guides cross-modal ablation of the principal component domain value of time-frequency coding The dot product of the inverse of obtains the spatial distribution of sound inside the pipeline to guide cross-modal ablation time-frequency coding projection channel vector .

[0051] Calculate the spatial distribution of sound inside the pipeline to guide cross-modal ablation time-frequency coding orthogonal basis eigenvectors Guided cross-modal ablation with spatial distribution of sound inside the pipeline using time-frequency coded projection channel vectors The weighted sum of the time-frequency encoding feature vectors is used to obtain the optimized spatial distribution of sound inside the pipeline to guide cross-modal ablation.

[0052] Generate prompt information by adding it to the tail of the optimized internal sound spatial distribution of the pipeline to guide the cross-modal ablation time-frequency coding feature vector, and input it into the AIGC-based sound signal denoising generator to obtain the denoised internal sound signal of the pipeline.

[0053] Here, when the time-frequency feature map of the sound inside the pipe and the spatial distribution feature vector of the sound inside the pipe respectively represent the image semantic features of the time-frequency domain image of the sound inside the pipe and the local image semantic short-range-long-range bidirectional correlation features of the internal sound spatial domain image, when performing cross-domain attention joint encoding based on fine-grained ablation, the cross-domain ablation attention misalignment under the differentiated image semantic feature expression mode of the source image semantics will also cause the joint encoding richness of the internal sound spatial distribution guided cross-modal ablation time-frequency coding feature vector obtained after the internal sound spatial distribution of the sound inside the pipe is expanded, resulting in redundancy of the probability distribution information in the mapping domain, affecting the accuracy of the generated results.

[0054] Therefore, based on the vector field approximation theorem, the discrete singularities and smoothness factors in the time-frequency coding feature vector guided by the sound spatial distribution inside the pipeline are used as the differentiable dimensions of the multidimensional topological structure. By fixing the normal directions of the orthogonal basis vectors in the local coordinate system of the principal component aggregation domain of the time-frequency coding feature vector guided by the sound spatial distribution inside the pipeline, an orthogonal projection channel from the feature space to the probability measure manifold is established. Thus, the principal component aggregation domain of the time-frequency coding feature vector guided by the sound spatial distribution inside the pipeline is effectively aligned, and the random noise interference caused by the multimodal coupling strength is eliminated, thereby avoiding the feature distortion problem caused by the failure of information redundancy elimination in the probability mapping process, and improving the accuracy of the generation result obtained by the sound signal denoising generator based on AIGC.

[0055] Exemplarily, in the leakage detection module 40, on the cloud platform, the sound signal inside the pipeline after denoising is subjected to fire pipe network leakage detection to obtain a detection result for indicating whether there is leakage in the fire pipe network. It should be understood that the sound signal inside the pipeline after denoising has significantly reduced the influence of background noise, making the leakage characteristics contained therein more obvious. Then, by using the powerful computing power and data processing advantages of cloud computing, the sound signal inside the pipeline after denoising is subjected to fire pipe network leakage detection to obtain a detection result for indicating whether there is leakage in the fire pipe network.

[0056] In one example, the de-noised internal pipe sound signal is used to detect fire pipe network leakage using a feature matching and classification method based on machine learning. Specifically, the feature matching and classification method based on machine learning relies on a pre-trained model for feature matching and classification. This method starts with building a classifier that can identify abnormal patterns within a specific frequency range. These patterns are usually associated with leakage, such as the unique turbulent noise generated when water is suddenly ejected from a high-pressure environment to a low-pressure environment. It has unique spectral characteristics. By learning from a large number of annotated sample data sets (including normal operating conditions and different types of leakage events), the system learns to distinguish between normal water flow sounds and other sound features that may indicate leakage. For each de-noised internal pipe sound signal, key attributes such as the energy distribution, periodic changes, harmonic components, and mutation points in the time series of the sound wave are calculated. These feature values ​​together constitute a multi-dimensional description of the current pipe network status. Then, the extracted features are compared and analyzed with the preset standard threshold or known leakage cases in the historical database. If the detected features meet the characteristics of a certain type of leakage event, the corresponding alarm information is generated; otherwise, the pipe network is considered to be in normal working condition.

[0057] In another example, the time series prediction and anomaly detection method based on a deep neural network is used to detect the leakage of the fire pipe network on the sound signal inside the pipe after denoising. Specifically, the time series prediction and anomaly detection method based on a deep neural network uses a network architecture that is good at capturing long-term series dependencies, such as a long short-term memory network LSTM or a gated recurrent unit GRU, which is very suitable for processing non-stationary signals. The method first converts the denoised sound signal into a time series format and inputs it into a fully trained deep neural network. The network has been optimized on a large amount of historical data and can learn the time series pattern of normal water flow sound. The network attempts to predict the sound signal at the next moment and attempts to reconstruct the entire time series. Under normal circumstances, the difference between the predicted value and the actual value should be small; once a leak occurs, the prediction error will increase significantly due to the change in the signal pattern. By setting a reasonable threshold, when the prediction error exceeds this threshold, it is considered that an abnormal situation has occurred, which may be due to a leak.

[0058] In another example, the denoised sound signal inside the pipe is analyzed using classical signal processing techniques, and leak detection is performed using predefined thresholds and rules. First, for each denoised sound signal inside the pipe, a fast Fourier transform (FFT) or short-time Fourier transform (STFT) is used to convert it from the time domain to the frequency domain. This step can reveal the distribution of different frequency components, especially those high-frequency components that are abnormally prominent compared to normal water flow sounds. Because leaks usually produce specific high-frequency noise, these noises appear as increased energy in certain frequency ranges on the spectrum graph. Next, calculate the total energy or peak energy in a specific frequency range. For example, select a frequency interval where a known leak event often occurs, such as between 400Hz and 1kHz, and then solve for the average energy or peak energy in this interval. This process can help quantify the intensity of the potential leak. In addition, the energy distribution of the entire spectrum can be calculated to find out whether there are significant energy peaks, which may indicate special noises generated by leaks. In addition to frequency domain analysis, feature extraction of time domain signals is also required. Calculate the root mean square (RMS) value, zero crossing rate (ZCR), short-time energy and other features of the sound wave. These features can reflect the dynamic changes of the sound signal and help distinguish normal water flow fluctuations from abnormal leakage sounds. For example, the zero crossing rate reflects the speed at which the signal crosses the zero point. The turbulent noise generated by the leakage often has a high zero crossing rate; while the short-time energy can capture the sudden change of energy in a short time, which is particularly important for identifying sudden leakage. According to historical data and experimental results, a series of thresholds are pre-set to determine whether leakage has occurred. These thresholds include frequency domain energy thresholds, time domain feature thresholds, and combined thresholds. When the energy in a certain frequency interval exceeds the set threshold, it is considered that there may be a leakage; if the calculated RMS value, ZCR or short-time energy exceeds the normal range, an alarm is triggered; an alarm is only issued when all selected features meet the corresponding threshold conditions, which can reduce the false alarm rate. Once any one or more features are detected to exceed the set threshold, the system immediately generates an early warning message. In order to improve the reliability and accuracy of the system, the effectiveness of the alarm can be verified through multiple continuous sampling to avoid false alarms caused by instantaneous interference.

[0059] Exemplarily, in the early warning module 50, in response to the detection result that the fire pipe network has leakage, an early warning signal is sent to the terminal device of the maintenance personnel. It should be understood that by sending an early warning signal to the terminal device of the maintenance personnel in response to the detection result that the fire pipe network has leakage, the time interval from the discovery of the problem to the action can be greatly shortened, avoiding possible negligence or delays in manual inspections, and ensuring that every potential problem can be discovered and handled in a timely manner. Accurate early warning information can help the maintenance team reasonably arrange work priorities and concentrate on solving the most urgent problems without having to conduct a comprehensive inspection of the entire pipe network, thereby improving work efficiency and resource utilization. At the same time, real-time early warning can promote information sharing between different departments, strengthen collaboration, and form a more efficient emergency response mechanism.

[0060] In one example, an integrated communication interface is used to establish a stable communication connection between the leakage detection module on the cloud platform and the terminal device (such as a smartphone, tablet or other mobile device) used by maintenance personnel. This is achieved through standard Internet protocols (such as HTTP / HTTPS, MQTT, etc.) to ensure the security and reliability of data transmission. In a preferred example, different alarm levels and trigger conditions can be preset in the system according to actual needs. For example, for minor leaks, only logs can be recorded; for serious cases, high-priority alarms are issued immediately. In addition, personalized early warning strategies can be set for specific areas or time periods. Alternatively, combined with a geographic information system (GIS), the early warning signal can not only inform that a leakage has occurred, but also accurately point out its location, so that maintenance personnel can arrive at the site faster and improve response speed. In addition to simple alarms, the system can also integrate more automated functions, such as automatically generating work orders, assigning tasks to the nearest maintenance team, providing detailed fault descriptions and recommended solutions, etc., to simplify the operation process and improve the overall service level.

[0061] In summary, the acoustic intelligent sensor for fire protection pipe network leakage detection according to the embodiment of the present application is explained. Unlike the fixed parameters of traditional filters, the present application can automatically adjust its internal parameters according to the actual input sound signal, distinguish and retain useful leakage features according to context information, so as to better adapt to the dynamic changes of the signal. In addition, it can capture the subtle changes and patterns of time-frequency and spatial distribution in the sound signal, so as to identify small-scale or low-intensity leakage events that may be ignored by traditional methods.

[0062] The basic principle of the present invention is described above in conjunction with specific embodiments. However, it should be pointed out that the advantages, strengths, effects, etc. mentioned in the present invention are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. must be possessed by each embodiment of the present invention. In addition, the specific details of the above embodiments are only for the purpose of illustration and facilitation of understanding, rather than limitation, and the above details do not limit the present invention to being implemented by adopting the above specific details.

[0063] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, please refer to the relevant description of other embodiments. In the several embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the module division is only a logical function division, and there may be other division methods in actual implementation. The modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0064] In addition, each functional module in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of hardware plus software functional modules.

[0065] Finally, it should be noted that the above description has been given for the purpose of illustration and description. In addition, the above embodiments are only used to illustrate the technical solution of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. An acoustic intelligent sensor for fire protection pipe network leakage detection, characterized in that: include: The sound signal acquisition module is used to collect the sound signals inside the pipeline in real time through the acoustic intelligent sensors installed at the key positions of the fire protection pipe network; A sound signal uploading module, used for uploading the sound signal inside the pipeline to the cloud platform; A denoising optimization module is used to perform denoising optimization on the sound signal inside the pipeline on the cloud platform to obtain a denoised sound signal inside the pipeline; a leakage detection module is used to perform fire pipe network leakage detection on the denoised sound signal inside the pipeline on the cloud platform to obtain a detection result indicating whether there is leakage in the fire pipe network; An early warning module is used to send an early warning signal to the terminal device of the maintenance personnel in response to the detection result that there is a leakage in the fire protection pipe network; wherein the denoising optimization module includes: an image transformation unit, which is used to perform time-frequency and space-domain transformations on the sound signal inside the pipeline to obtain a time-frequency domain image of the sound inside the pipeline and a space-domain image of the sound inside the pipeline; a sound image feature extraction unit, which is used to extract features from the time-frequency domain image of the sound inside the pipeline and the space-domain image of the sound inside the pipeline to obtain the time-frequency features of the sound inside the pipeline and the space-domain features of the sound inside the pipeline; a cross-modal ablation interaction unit, which is used to perform cross-modal ablation attention depth interaction on the time-frequency features of the sound inside the pipeline and the space-distribution features of the sound inside the pipeline to obtain the cross-modal ablation time-frequency coding features guided by the space-distribution of the sound inside the pipeline; a denoising image generation unit, which is used to obtain the denoised sound signal inside the pipeline based on the cross-modal ablation time-frequency coding features guided by the space-distribution of the sound inside the pipeline.

2. The acoustic intelligent sensor for fire protection pipe network leakage detection according to claim 1 is characterized in that: The image transformation unit is used to: perform S transformation on the sound signal inside the pipeline to obtain the time-frequency domain image of the sound inside the pipeline; perform RP transformation on the sound signal inside the pipeline to obtain the space domain image of the sound inside the pipeline.

3. The acoustic intelligent sensor for fire protection pipe network leakage detection according to claim 2 is characterized in that: The sound image feature extraction unit includes: a pipe internal sound time-frequency feature extraction subunit, which is used to pass the pipe internal sound time-frequency domain image through a sound time-frequency feature extractor based on an MCNN model to obtain a pipe internal sound time-frequency feature map as the pipe internal sound time-frequency feature; and a pipe internal sound space distribution feature extraction subunit, which is used to pass the pipe internal sound space domain image through a sound space distribution feature extractor based on a CNN-LSTM model to obtain a pipe internal sound space distribution feature vector as the pipe internal sound space distribution feature.

4. The acoustic intelligent sensor for fire protection pipe network leakage detection according to claim 3 is characterized in that: The cross-modal ablation interaction unit includes: a time-frequency decomposition subunit for sound inside the pipe, which is used to perform fine-grained feature decomposition on the time-frequency feature map of sound inside the pipe to obtain a set of time-frequency feature matrices of local sound inside the pipe; a feature weight calculation subunit for sound inside the pipe, which is used to calculate the multidimensional fine-grained ablation weight factor of each local sound time-frequency feature matrix inside the pipe in the set of local sound time-frequency feature matrices inside the pipe based on the spatial distribution feature vector of sound inside the pipe to obtain a set of multidimensional fine-grained ablation weight factors of sound inside the pipe; a cross-modal encoding subunit for sound inside the pipe, which is used to perform fine-grained ablation modulation and feature aggregation on the set of local sound time-frequency feature matrices inside the pipe based on the set of multidimensional fine-grained ablation weight factors of sound inside the pipe to obtain a time-frequency coding feature map of cross-modal ablation guided by spatial distribution of sound inside the pipe as the time-frequency coding feature of cross-modal ablation guided by spatial distribution of sound inside the pipe.

5. The acoustic intelligent sensor for fire protection pipe network leakage detection according to claim 4 is characterized in that: The pipe internal sound time-frequency decomposition subunit is used to perform feature fine-grained decomposition of the pipe internal sound time-frequency feature map along the channel dimension of the pipe internal sound time-frequency feature map to obtain a set of local pipe internal sound time-frequency feature matrices.

6. The acoustic intelligent sensor for fire protection pipe network leakage detection according to claim 5 is characterized in that: The pipe internal sound feature weight calculation subunit is used to: use each pipe internal local sound time-frequency feature matrix in the set of pipe internal local sound time-frequency feature matrices as a key matrix, perform multi-dimensional encoding of the pipe internal sound spatial distribution feature vector and the key matrix to obtain a set of multi-dimensional fine-grained ablation factors of the pipe internal sound; and perform normalization and masking on the set of multi-dimensional fine-grained ablation factors of the pipe internal sound to obtain a set of multi-dimensional fine-grained ablation weight factors of the pipe internal sound.

7. The acoustic intelligent sensor for fire protection pipe network leakage detection according to claim 6 is characterized in that: The cross-modal encoding subunit for sound inside the pipeline is used to: based on the set of multi-dimensional fine-grained ablation weight factors for sound inside the pipeline, perform fine-grained ablation modulation on the set of time-frequency feature matrices of local sound inside the pipeline to obtain a set of local feature matrices for cross-domain guided ablation of sound inside the pipeline; perform feature aggregation on the set of local feature matrices for cross-domain guided ablation of sound inside the pipeline to obtain a time-frequency encoding feature map for guiding cross-modal ablation of sound spatial distribution inside the pipeline.

8. The acoustic intelligent sensor for fire protection pipe network leakage detection according to claim 7 is characterized in that: The denoised image generating unit is used to: expand the time-frequency coding feature map of the spatial distribution of sound inside the pipeline guiding cross-modal ablation into a time-frequency coding feature vector of the spatial distribution of sound inside the pipeline guiding cross-modal ablation, add generation prompt information to the end of the time-frequency coding feature vector of the spatial distribution of sound inside the pipeline guiding cross-modal ablation, and input it into the sound signal denoising generator based on AIGC to obtain the denoised internal pipeline sound signal.

9. The acoustic intelligent sensor for fire protection pipe network leakage detection according to claim 8, characterized in that: The generating prompt information is: based on the input corpus, denoising and optimizing the sound signal inside the pipeline.

Images