Downhole temperature field reconstruction and thermal disaster early warning method and system based on acoustic tomography

Abstract

A kind of well temperature field reconstruction and thermal disaster early warning method and system based on acoustic tomography, method: staggeredly arranging explosion-proof loudspeaker array and microphone array on the two sides of roadway in monitoring area;Utilize frequency division multiplexing technology to emit / receive multiple frequency band acoustic waves synchronously, form entity acoustic path data set, and construct overdetermined equation group;Adopt virtual sound line method, realize virtual path enhancement;Realize high-resolution temperature field reconstruction of roadway cross section temperature field in combination with sound velocity-temperature linear model;Continuous multiple frame temperature field images are input into TimesNet model, and multi-scale features are extracted by adaptive period decomposition and two-dimensional space-time convolution, and fire warning grade and abnormal area coordinates are output.The system: including explosion-proof loudspeaker array, microphone array, edge computing node and host computer;Edge computing node is connected with explosion-proof loudspeaker array and host computer respectively.The present application can realize global temperature perception and concealed fire early warning in target monitoring area.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring technology for mine thermal hazards, specifically relating to a method and system for underground temperature field reconstruction and thermal disaster early warning based on acoustic tomography. Background Technology

[0002] Downhole temperature measurement is a specialized type of temperature measurement. Existing contact-based solutions, such as thermocouples, resistance temperature detectors (RTDs), and expansion sensors, are prone to wire breakage, thermal current interference, and material fatigue failure in mining areas with high dust levels, strong vibrations, and frequent equipment movement. This leads to data distortion and an inability to dynamically cover the entire cross-section. Non-contact technologies that utilize optical principles can avoid the risks of physical contact; however, infrared imaging technology is significantly affected by high-concentration dust obstruction, while laser temperature measurement suffers from positioning errors due to reflections from tunnel walls and has a limited effective detection range, failing to achieve full coverage of large-scale spaces.

[0003] In fire propagation modeling, traditional machine learning models rely on manual feature extraction, which makes it difficult to capture complex nonlinear characteristics and spatiotemporal dependencies of data; ordinary RNNs are prone to gradient vanishing and cannot effectively learn long-term dynamic changes; 3DCNNs have high computational cost and insufficient flexibility, while Transformers have high computational complexity and their spatial feature extraction is not intuitive.

[0004] To address the shortcomings of existing technologies, there is an urgent need to provide a method and system for downhole temperature field reconstruction and thermal disaster early warning based on acoustic tomography. Summary of the Invention

[0005] To address the problems of the existing technologies mentioned above, this invention provides a method and system for underground temperature field reconstruction and thermal disaster early warning based on acoustic tomography. This method is simple to implement and low in cost. It can achieve full-area temperature sensing and early warning of hidden fire sources under complex working conditions such as dust pollution, strong vibration, and large cross-sections without contact or damage to the surrounding rock of the mine roadway. It provides an efficient and accurate technical means for underground temperature field monitoring and has high practicality and adaptability. The system has a simple structure and a high degree of intelligence, and it can autonomously achieve full-area temperature sensing and early warning of hidden fire sources in the target monitoring area.

[0006] To achieve the above objectives, this invention provides a method for downhole temperature field reconstruction and thermal disaster early warning based on acoustic tomography, comprising the following steps:

[0007] Step 1: Acoustic signal acquisition and matrix equation construction;

[0008] S11: Explosion-proof speaker arrays and microphone arrays are arranged alternately on both sides of the roadway in the monitoring area;

[0009] S12: Using frequency division multiplexing technology, the explosion-proof speaker array can simultaneously emit sound waves of different frequency bands; at the same time, the sound wave signals are synchronously collected and received through the microphone array and sent to the edge computing node.

[0010] S13: Edge computing nodes perform bandpass filtering and cross-correlation operations on acoustic signals, extract the flight time of each sound path, and generate high-precision TOF measurement values ​​by filtering the real sound path through time consistency verification, forming a real sound path dataset, and constructing an overdetermined system of equations.

[0011] Step 2: Virtual path enhancement and temperature field reconstruction;

[0012] S21: Based on the TOF data of the microphone array, a virtual sound ray method is used to interpolate and generate a virtual path between real microphones. The length of the virtual path is calculated based on the cosine theorem, and weighted coefficients are used to fuse the virtual and physical TOF data to achieve virtual path enhancement.

[0013] S22: Extend the measurement matrix and TOF vector, solve the extended overdetermined equations by least squares method, and combine the sound velocity-temperature linear model to realize high-resolution temperature field reconstruction of the temperature field of the tunnel cross section;

[0014] Step 3: Fire early warning and anomaly detection;

[0015] Edge computing nodes input multiple consecutive frames of temperature field images into the TimesNet model. A Fast Fourier Transform (FFT) is used to convert the one-dimensional time series to the frequency domain, revealing the dominant periodic pattern. The one-dimensional time series data is folded into a stacked matrix containing complete spatiotemporal information according to different periods. Inception convolution is used to simultaneously capture multi-scale features in both time and space dimensions on the stacked matrix. A parameter-sharing mechanism maintains powerful feature extraction capabilities while achieving model lightweighting. Stable feature transfer and deep evolution are achieved through residual connections and multi-layer stacking. Coordinate mapping is used to convert abnormal pixels in the image into meaningful underground physical coordinates. Finally, the TimesNet model outputs the fire warning level and the coordinates of the abnormal area.

[0016] As a preferred embodiment, in step S12 of step one, frequency division multiplexing technology is used to enable the explosion-proof speaker array to simultaneously emit sound waves of different frequency bands; simultaneously, the sound wave signals are synchronously acquired and received through a microphone array and sent to the edge computing node, as follows:

[0017] S12-1: Frequency band division; The transmitted signal bandwidth is divided into multiple non-overlapping sub-bands in the frequency domain, and a segment is set between adjacent sub-bands as isolation protection to avoid spectrum interference;

[0018] S12-2: Multi-signal synchronous transmission and reception; enabling different explosion-proof loudspeakers to simultaneously transmit sub-signals of different frequency bands, achieving parallel transmission through frequency domain isolation, and eliminating the time interval between channel switching; simultaneously, receiving aliased sound wave signals through a microphone array and sending them to the edge computing node; wherein, the minimum measurement time is obtained according to formula (1). ;

[0019] (1);

[0020] In the formula, Indicates the pulse width of the sound source signal. Indicates the path of sound waves Flight time on the plane.

[0021] As a preferred embodiment, in step S13 of step one, the process of constructing the overdetermined system of equations is as follows:

[0022] S13-1: Filtering; extracting sub-signals of each measurement path from the aliased acoustic signal using a bandpass filter;

[0023] S13-2: TOF matrix construction; Organizing the TOF measurements into a matrix of TOF measurements. The structural parameter matrix is ​​constructed by the position of the physical microphone and the length of the sound wave path. ;

[0024] S13-3: Matrix equation construction; Based on the measurement path, establish an overdetermined system of equations according to formula (2);

[0025] (2);

[0026] In the formula, For temperature and speed The matrix to be determined.

[0027] As a preferred option, the virtual path enhancement process in step two is as follows:

[0028] S21-1: Generate a virtual acoustic wave path;

[0029] At a spacing of Two physical microphones Insert evenly between The first virtual microphone, according to the law of cosines, is obtained according to formula (3). The virtual sound wave path length corresponding to each virtual microphone ;

[0030] (3);

[0031] In the formula, For two adjacent physical microphones The spacing; and These are two adjacent physical microphones. The sound wave path; The spacing between adjacent microphones;

[0032] S21-2: Calculation of virtual path TOF vector; calculate the first according to formula (4). TOF of the virtual sound wave path corresponding to each virtual microphone;

[0033] (4);

[0034] In the formula, The weighting coefficients represent the virtual sound wave path. ; and The path length between adjacent entities; and This is the Time-of-Flight (TOF) for adjacent entity paths.

[0035] As a preferred embodiment, in step S22 of step two, the process of achieving high-resolution temperature field reconstruction of the tunnel cross-section temperature field is as follows:

[0036] S22-1: Matrix expansion;

[0037] TOF data from the virtual microphone TOF data compared to physical microphones Merge to form the expanded TOF vector Based on the virtual microphone position and path length, expand the structure parameter matrix. for It contains geometric information of both physical entities and virtual paths;

[0038] S22-2: Least squares solution and temperature field reconstruction;

[0039] The extended overdetermined equations are solved using the least squares method according to formula (5); the temperature-velocity relationship is modeled according to the linear relationship between sound wave propagation speed and temperature in formula (6), so as to realize the high-resolution temperature field reconstruction of the temperature field of the tunnel cross section.

[0040] (5);

[0041] (6).

[0042] As a preferred option, the process of outputting the fire warning level and the coordinates of the abnormal area in step three is as follows:

[0043] S31: Data input; Receives multiple consecutive frames of temperature field images as input data. ,in, For time step, It is a variable dimension;

[0044] S32: Periodic Analysis and Frequency Selection;

[0045] First, analyze the input data along the channel dimension. Perform a Fast Fourier Transform column by column and calculate all channels. The average amplitude is used to obtain the energy spectrum according to formula (7). ;

[0046] (7);

[0047] In the formula, This represents the real-valued FFT operator along the time axis, with an output length of... ;

[0048] Next, according to formula (8), from the frequency index set Select the one with the highest energy There are 10 frequency components, of which 10 are present. ;

[0049] (8);

[0050] In the formula, For the first Frequency index corresponding to high energy;

[0051] Then, select the frequency index according to formula (9). Converted to actual period length ;

[0052] (9);

[0053] In the formula, ;

[0054] S33: Sequence zero padding and matrix folding;

[0055] First, calculate the zero-padding length according to formula (10). ;

[0056] (10);

[0057] Next, add to the right side of the sequence. The zero-padded sequence is obtained according to formula (11). ;

[0058] (11);

[0059] Then, the zero-padded sequence Separate according to channel maintenance, for each channel The data is independently reshaped, and the reshaped two-dimensional matrix is ​​obtained according to formula (12). ,in, ; two-dimensional matrix By folding the matrix along the channel dimension, a stacked matrix containing complete spatiotemporal information is obtained according to formula (13). ;

[0060] (12);

[0061] (13);

[0062] S34: Multi-scale spatiotemporal feature extraction;

[0063] First, for stacked matrices A 1×3 column-direction convolution kernel is used. Slide along the column direction to capture detailed changes and short-term gradients within a period, and obtain the short-term change characteristics within a single period according to formula (14). ;in,

[0064] (14);

[0065] Next, for the stacked matrix ,use row direction convolution kernel Slide along the row direction to capture the long-term trend between cycles, and obtain the long-term trend characteristics between cycles according to formula (15). ;in, ;

[0066] (15);

[0067] Then, for stacked matrices ,use Two-dimensional joint convolution kernel Simultaneously sliding in both row and column directions, the joint spatiotemporal features during the period and within the cycle are captured, and the joint spatiotemporal features are obtained according to formula (16). ;

[0068] (16);

[0069] S35: Multi-branch feature fusion and residual connection;

[0070] First, after the periodic detection module Identify The most prominent cyclical patterns, each corresponding to a set of parameters. ,in, Indicates the period length. Indicates frequency;

[0071] Next, corresponding Create one frequency component Each processing branch is a separate processing branch, and each processing branch deals with the input sequence. Zero-padding and folding are performed sequentially, and then the data is sent to the parameter-sharing Inception module for processing. The two-dimensional feature map is obtained according to formula (17). ;

[0072] (17);

[0073] Furthermore, first calculate the energy value corresponding to each processing branch. Then, the fusion weight of each processing branch is calculated according to formula (18). ; (18);

[0074] Next, the two-dimensional feature map of each processing branch is... Flattened into a one-dimensional vector, using fusion weights The flattened features of each processing branch are weighted, and all weighted features are summed to obtain the final output according to formula (19). ;

[0075] (19);

[0076] Then, output according to formula (20) With input Add, based on the new output ;

[0077] (20);

[0078] S36: Multi-layer TimesBlock stacking and feature evolution;

[0079] New output As new input, repeat steps S31 through S35, stacking... The TimesBlock layer gradually extracts more abstract and higher-level feature representations from the original input;

[0080] S37: Location and positioning processing;

[0081] First, select within the underground tunnels Given feature points with known physical coordinates, record the pixel coordinates of each feature point in the temperature field image. and the corresponding real physical coordinates ,in, ;

[0082] Next, based on formula (21), an affine transformation is used to establish the transformation relationship from pixel coordinates to physical coordinates;

[0083] (twenty one);

[0084] In the formula, Let be the calibration matrix to be solved;

[0085] Furthermore, the pixel coordinates and physical coordinates of the P feature points are substituted into formula (21), and the elements of the calibration matrix are solved by the least squares method.

[0086] Then, output the set of pixel coordinates of the abnormal region. Then, substituting the calibration matrix, the corresponding set of downhole physical coordinates is calculated. This enables precise location of abnormal areas within underground roadways;

[0087] S38: Output of fire warning level and coordinates of abnormal areas;

[0088] After deep extraction and evolution of temporal features through multiple layers of TimesBlock, the TimesNet model determines the fire risk level by periodic analysis of the temporal changes in the temperature field. The fire risk levels are none, low, medium, high, and extremely high, and the model provides the key features corresponding to each risk level, forming fire warning level data. At the same time, the TimesNet model combines the spatial perception capabilities of periodicity and convolution to provide the coordinate set and thermal distribution matrix of abnormally high temperature areas in the mine, predict the fire spread trend vector, and form fire spatial coordinate location data. Finally, the TimesNet model outputs fire warning level data and fire spatial location data.

[0089] This invention proposes a method for mine temperature field reconstruction and thermal disaster early warning based on acoustic tomography. First, it utilizes frequency division multiplexing (FDM) technology to simultaneously transmit / receive multi-band acoustic waves, overcoming the limitations of traditional time delay estimation algorithms. By allowing multiple signals of different frequency bands to be transmitted in parallel, it eliminates the time-consuming problem of sequential transmission and simultaneously achieves instantaneous sampling of multi-channel signals, ensuring time consistency and improving dynamic response capabilities. It also significantly improves data acquisition efficiency. Alternating acoustic signals are acquired via an explosion-proof microphone array, and then filtered and cross-correlated by edge computing nodes. This process separates the clean signals of each sound path from the aliased sound waves, effectively eliminating multipath interference caused by roadway wall reflections. A further filtering process ensures that only sound path paths conforming to physical laws are retained, effectively improving data quality. Next, virtual paths are generated by interpolation between real microphones, effectively increasing the coverage density of sound path paths and compensating for measurement blind spots caused by the limited number of physical microphones. By combining the cosine theorem to calculate the virtual path length and employing weighted coefficients to fuse virtual and physical Time-of-Flight (TOF) data, the overall accuracy and spatial resolution of acoustic time-of-flight measurements can be significantly improved, resolving the matrix rank deficiency problem caused by insufficient sensor quantity. Simultaneously, the least squares method effectively suppresses measurement noise and errors, enhancing the stability of parameter inversion. Combined with a sound velocity-temperature linear model, high-resolution reconstruction of the temperature field in the tunnel cross-section can be achieved, providing more accurate data support for mine environmental monitoring. Finally, multiple consecutive frames of temperature fields are input into a TimesNet model. The TimesNet model extracts multi-scale features through adaptive periodic decomposition and two-dimensional spatiotemporal convolution, combining the optimization of both temporal and spatial resolution to map abnormal regions from pixel space to physical space. Ultimately, the fire warning level and abnormal region coordinates are output, providing support for mine safety decision-making and realizing a closed loop from monitoring to decision support.

[0090] This method is simple to implement and has low implementation costs. It can achieve full-area temperature sensing and early warning of hidden fire sources under complex working conditions such as dust, strong vibration and large cross-section without contacting or damaging the surrounding rock of the roadway. It can provide efficient and accurate technical means for temperature field monitoring in mines and has high practicality and adaptability.

[0091] The present invention also provides a downhole temperature field reconstruction and thermal disaster early warning system based on acoustic tomography, which is used to realize a method for downhole temperature field reconstruction and thermal disaster early warning based on acoustic tomography, including an explosion-proof speaker array, a microphone array, an edge computing node and a host computer;

[0092] The explosion-proof loudspeakers and microphone arrays are arranged alternately on both sides of the alleyway in the monitoring area;

[0093] The explosion-proof speaker array is used to simultaneously emit sound waves of different frequency bands;

[0094] The microphone array is used to collect sound wave signals and send them to the edge computing node;

[0095] The edge computing nodes are connected to the explosion-proof speaker array and the microphone array respectively. They are used to drive the explosion-proof speaker array to emit sound waves of different frequency bands simultaneously using frequency division multiplexing technology. They are used to realize high-resolution temperature field reconstruction of the temperature field of the tunnel cross section based on the received sound wave signals. At the same time, they are used to input multiple consecutive frames of temperature field images as input data to the host computer.

[0096] The host computer is connected to the edge computing node and has a built-in TimesNet model. The TimesNet model is used to extract multi-scale features from the input data through adaptive periodic decomposition and two-dimensional spatiotemporal convolution. It achieves stable feature transfer and deep evolution through residual connection and multi-layer stacking. At the same time, combined with coordinate mapping, it converts abnormal pixels in the image into meaningful underground physical coordinates, and finally outputs the fire warning level and abnormal area coordinates.

[0097] As a preferred embodiment, the edge computing node is a PLC controller.

[0098] As a preferred option, the host computer is an industrial computer.

[0099] In this invention, explosion-proof loudspeakers and microphone arrays are arranged in a staggered pattern on both sides of the tunnel, effectively covering the monitoring area. The explosion-proof loudspeaker and microphone arrays work together using frequency division multiplexing technology to synchronously transmit / receive multi-band sound waves, achieving instantaneous sampling of multi-channel signals, ensuring time consistency, and significantly improving data acquisition efficiency. Edge computing nodes integrate signal and deep inference capabilities, reducing reliance on monitoring center processing equipment, lowering data transmission volume, and enabling high-resolution temperature field reconstruction at the near end. The host computer has a built-in TimesNet model, which can analyze temperature field data in real time and autonomously and promptly detect potential fire threats, providing intuitive and operable decision support for emergency response and significantly improving fire location accuracy.

[0100] The system has a simple structure and a high degree of intelligence. It can autonomously achieve full-area temperature perception of the target monitoring area and early warning of hidden fire sources. Attached Figure Description

[0101] Figure 1 This is a flowchart of the method in this invention;

[0102] Schematic diagram of a downhole temperature monitoring and early warning system based on acoustic tomography;

[0103] Figure 2This is a schematic diagram of the synchronous measurement of temperature and velocity based on the acoustic FDM measurement method of the present invention;

[0104] Figure 3 This is a schematic diagram of the virtual voice in the original text;

[0105] Figure 4 This is a structural diagram of the TimesNet model in this invention;

[0106] Figure 5 This is a schematic diagram of the system in this invention;

[0107] Figure 6 This is a schematic diagram of the system layout in the tunnel according to the present invention. Detailed Implementation

[0108] The invention will now be further described with reference to the accompanying drawings.

[0109] like Figures 1 to 6 As shown, this invention provides a method for downhole temperature field reconstruction and thermal disaster early warning based on acoustic tomography, comprising the following steps:

[0110] Step 1: Acoustic signal acquisition and matrix equation construction;

[0111] S11: Explosion-proof speaker arrays and microphone arrays are arranged alternately on both sides of the roadway in the monitoring area;

[0112] S12: Using frequency division multiplexing (FDM) technology, the explosion-proof speaker array can simultaneously emit sound waves of different frequency bands; at the same time, the sound wave signals are synchronously collected and received through the microphone array and sent to the edge computing node.

[0113] S13: Edge computing nodes perform bandpass filtering and cross-correlation operations on acoustic signals to extract the Time of Flight (TOF) of each acoustic path. They also filter the real acoustic paths through time consistency verification to generate high-precision TOF measurements, effectively removing multipath interference and clock synchronization errors, forming a physical acoustic path dataset, and constructing an overdetermined set of equations.

[0114] Step 2: Virtual path enhancement and temperature field reconstruction;

[0115] S21: Based on the TOF data of the microphone array, a virtual sound ray method is used to interpolate and generate a virtual path between real microphones. The length of the virtual path is calculated based on the cosine theorem, and weighted coefficients are used to fuse the virtual and physical TOF data to achieve virtual path enhancement.

[0116] S22: Extend the measurement matrix and TOF vector, solve the extended overdetermined equations by least squares method, and combine the sound velocity-temperature linear model to realize high-resolution temperature field reconstruction of the temperature field of the tunnel cross section;

[0117] This technology improves the spatial resolution of traditional acoustic thermometry to the sub-meter level while suppressing the interference of non-uniform media on the speed of sound.

[0118] Step 3: Fire early warning and anomaly detection;

[0119] Edge computing nodes input multiple consecutive frames of temperature field images into the TimesNet model. They then transform the one-dimensional time series to the frequency domain using Fast Fourier Transform to obtain the dominant periodic patterns. The one-dimensional time series data is folded into a stacked matrix containing complete spatiotemporal information according to different periods. Inception convolution is used to simultaneously capture multi-scale features in both time and space dimensions on the stacked matrix. A parameter sharing mechanism maintains powerful feature extraction capabilities while achieving model lightweighting. Stable feature transfer and deep evolution are achieved through residual connections and multi-layer stacking. Coordinate mapping is used to convert abnormal pixels in the image into meaningful underground physical coordinates. Finally, the TimesNet model outputs the fire warning level and the coordinates of the abnormal area.

[0120] As a preferred embodiment, in step S12 of step one, frequency division multiplexing (FDM) technology is used to enable the explosion-proof speaker array to simultaneously emit sound waves of different frequency bands; simultaneously, the sound wave signals are synchronously acquired and received through a microphone array and sent to the edge computing node, as follows:

[0121] S12-1: Frequency band division; The transmit signal bandwidth (20 Hz~20 kHz) is divided into multiple non-overlapping sub-bands in the frequency domain, and a segment is set between adjacent sub-bands as isolation protection to avoid spectrum interference;

[0122] In this way, within each transmission cycle, the transmitted signal consists of several linearly swept sub-signals with non-overlapping frequency domains. The number of sub-signals, their frequency band range, and time delay can be flexibly configured according to the tunnel length and resolution requirements. Due to the superposition of multiple frequency signals, the amplitude of the aliased waveform may increase, but the frequency domains of the sub-signals in each measurement path are independent, and they can be effectively separated by subsequent bandpass filtering.

[0123] S12-2: Synchronous transmission and reception of multiple signals;

[0124] Since there is no overlap between two adjacent frequency bands, this allows different explosion-proof loudspeakers to simultaneously emit sub-signals of different frequency bands without causing frequency interference in the spectrum. Figure 3 In the middle, the speaker and Simultaneously transmitting sound source signals of different frequency ranges, via microphone and Simultaneous sampling is performed. Parallel transmission is achieved through frequency domain isolation, eliminating the time interval between channel switching; simultaneously, aliased acoustic signals are received via a microphone array and sent to the edge computing node; the minimum measurement cycle time is determined by the longest TOF acoustic path. Figure 3 In For the longest sound path, the minimum measurement time for temperature and velocity measurements is... It can be obtained according to formula (1);

[0125] (1);

[0126] In the formula, Indicates the pulse width of the sound source signal. Indicates the path of sound waves Flight time on the plane.

[0127] As a preferred embodiment, in step S13 of step one, the process of constructing the overdetermined system of equations is as follows:

[0128] S13-1: Filtering; extracting sub-signals of each measurement path from the aliased acoustic signal using a bandpass filter;

[0129] Because the sub-band frequency domains are independent, filtering can effectively separate superimposed signals and suppress the influence of aliasing amplitude. For example, a microphone. and The corresponding frequency band signals are received respectively, and after filtering, four independent acoustic path data are obtained; Figure 3 A speaker was shown. , With microphone , Relative position and longest sound path To verify the robustness of this method under low signal-to-noise ratio conditions, 10dB Gaussian white noise was added to the signal in the simulation. In practical applications, there is no need to actively introduce noise.

[0130] S13-2: TOF matrix construction; Organizing the TOF measurements into a matrix of TOF measurements. The structural parameter matrix is ​​constructed by the position of the physical microphone and the length of the sound wave path. ;

[0131] S13-3: Matrix Equation Construction;

[0132] The FDM-based measurement method involves multiple measurement paths, and the parameters to be solved are usually gas velocity and temperature. Taking four measurement paths as an example, an overdetermined equation set is established based on the measurement paths and according to formula (2);

[0133] (2);

[0134] In the formula, For temperature and speed The matrix to be determined.

[0135] As a preferred option, the virtual path enhancement process in step two is as follows:

[0136] S21-1: Generate virtual sound wave paths; the virtual sound wave path method is used when the number of sound wave paths is small. New sound wave paths can be generated using the known positions and parameters of the microphones. For virtual microphones at positions other than the first and last virtual microphones, the length of their corresponding virtual sound wave paths can be calculated based on the lengths of the sound wave paths corresponding to the two adjacent microphones.

[0137] At a spacing of Two physical microphones Insert evenly between The first virtual microphone, according to the law of cosines, is obtained according to formula (3). The virtual sound wave path length corresponding to each virtual microphone ;

[0138] (3);

[0139] In the formula, For two adjacent physical microphones The spacing; and These are two adjacent physical microphones. The sound wave path; The spacing between adjacent microphones;

[0140] Based on the above formula, the solution can be obtained. Figure 4 The virtual sound wave path length corresponding to all virtual microphones in the program.

[0141] S21-2: Calculation of the Time-of-Flight (TOF) vector for the virtual path; after determining the coordinates of the virtual microphone and its corresponding acoustic path length, the TOF of adjacent virtual acoustic paths can be used. For example... Figure 4 As shown, the first step is calculated according to formula (4). TOF of the virtual sound wave path corresponding to each virtual microphone;

[0142] (4);

[0143] In the formula, The weighting coefficients represent the virtual sound wave path. ; and The path length between adjacent entities; and This is the Time-of-Flight (TOF) for adjacent entity paths.

[0144] As a preferred embodiment, in step S22 of step two, the process of achieving high-resolution temperature field reconstruction of the tunnel cross-section temperature field is as follows:

[0145] S22-1: Matrix expansion;

[0146] TOF data from the virtual microphone TOF data compared to physical microphones Merge to form the expanded TOF vector Based on the virtual microphone position and path length, expand the structure parameter matrix. for It contains geometric information of both physical entities and virtual paths;

[0147] S22-2: Least squares solution and temperature field reconstruction;

[0148] Due to the matrix Since it is a non-square matrix, it is calculated... A positive definite square matrix can be obtained, whose rank is full column rank, therefore the corresponding solution is unique. The extended overdetermined system of equations is solved using the least squares method according to formula (5); in the underground environment of a mine, the medium temperature... ( (and the speed of sound) There is a known single-valued functional relationship between them, which forms the physical basis for retrieving the temperature field using acoustic tomography. Based on the linear relationship between acoustic propagation speed and temperature in formula (6), a temperature-velocity relationship model is performed to realize high-resolution temperature field reconstruction of the tunnel cross-section temperature field;

[0149] (5);

[0150] (6).

[0151] correspond The reference speed of sound at that time; This refers to the temperature coefficient. For potential gas and humidity deviations occurring underground, the formula can be compensated using conventional linear correction methods. The compensated model still maintains a linear form and can be directly used for temperature field inversion, meeting accuracy requirements without additional on-site calibration. The above process demonstrates that the FDM-based measurement method can solve for the unknowns of medium velocity and temperature based on acoustic signals along all acoustic paths, thereby achieving effective measurement of gas temperature and velocity.

[0152] As a preferred option, the process of outputting the fire warning level and the coordinates of the abnormal area in step three is as follows:

[0153] exist Figure 5 The complete model process and structure can be seen from this. Figure 5 It showcases a streamlined model structure, including periodic mining, 2D folding, parametrically efficient convolution, and Softmax fusion. The details of each part are described below.

[0154] S31: Data input; Receives multiple consecutive frames of temperature field images as input data. ,in, is the time step, representing the number of samples taken from multiple consecutive temperature field images, i.e., the length of the temperature field sequence in the input model. For example, if one frame of the temperature field image is acquired every 5 seconds, and the input model sequence contains 10 frames, then... The corresponding total time span is Second; The spatial resolution of the temperature field image is: 1) **Variable dimension**, containing two types of information: temperature value and temperature gradient, which together constitute the variable dimension of the model input. The temperature value dimension is the real-time temperature value of each spatial pixel, used to capture the magnitude of the temperature. The temperature gradient dimension is the temperature gradient of each spatial pixel in the horizontal and vertical directions (calculated from the temperatures of adjacent pixels), used to capture the trend of temperature changes. If the spatial resolution of the temperature field image is... (M is the number of vertical pixels, N is the number of horizontal pixels), then after unfolding a single frame temperature field image, the variable dimensions are... (“ (This corresponds to two types of information: temperature value and temperature gradient).

[0155] S32: Periodic Analysis and Frequency Selection;

[0156] First, analyze the input data along the channel dimension. Perform a Fast Fourier Transform (FFT) column by column and compute all channels. The average amplitude is used to obtain the energy spectrum according to formula (7). After taking the mold Each element represents the average energy of that frequency component across all variables;

[0157] (7);

[0158] In the formula, This represents the real-valued FFT operator along the time axis, with an output length of... ;

[0159] Next, according to formula (8), from the frequency index set Select the one with the highest energy There are 10 frequency components, of which 10 are present. ;

[0160] (8);

[0161] In the formula, For the first The frequency index corresponding to high energy. The relationship with physical frequency is ;

[0162] Then, select the frequency index according to formula (9). Convert to actual period length Period length The unit is time-step, which facilitates subsequent geometric folding. Rounding up during calculations ensures... This provides an integer multiple condition for zero-complement reshape;

[0163] (9);

[0164] In the formula, ;

[0165] S33: Sequence zero padding and matrix folding;

[0166] First, calculate the zero-padding length according to formula (10). ;

[0167] (10);

[0168] Next, to ensure that the length of the 1D sequence is exactly equal to the total number of units in the selected periodic matrix, an array is added to the right side of the sequence. The zero-padded sequence is obtained according to formula (11). ;

[0169] (11);

[0170] Zero padding does not introduce frequency domain energy, and the subsequent FFT energy weights automatically suppress its contribution, ensuring that periodic folding has no spectral leakage.

[0171] Then, the zero-padded sequence Separate according to channel maintenance, for each channel The data is independently reshaped, and the reshaped two-dimensional matrix is ​​obtained according to formula (12). ,in, This process improves processing efficiency through parallel computing, and the reshaped two-dimensional matrix is ​​stored contiguously in memory; the two-dimensional matrix By folding the matrix along the channel dimension, a stacked matrix containing complete spatiotemporal information is obtained according to formula (13). ;

[0172] (12);

[0173] (13);

[0174] Among them, row index This refers to inter-period changes, specifically column indexes. This refers to intra-period changes;

[0175] S34: Multi-scale spatiotemporal feature extraction;

[0176] First, adjacent column indexes and Corresponding to the original time series Such time points refer to adjacent time steps within the same period. Based on this, for stacked matrices... A 1×3 column-direction convolution kernel is used. Slide along the column direction to capture detailed changes and short-term gradients within a period, and obtain the short-term change characteristics within a single period according to formula (14). ;in,

[0177] (14);

[0178] Next, the adjacent row indexes and Corresponding to the time points in the original time series and This represents time points with different periods but the same phase. Based on this, for stacked matrices... ,use row direction convolution kernel Slide along the row direction to capture the long-term trend between cycles, and obtain the long-term trend characteristics between cycles according to formula (15). ;in, ;

[0179] (15);

[0180] Then, for stacked matrices ,use Two-dimensional joint convolution kernel Simultaneously sliding in both row and column directions, it captures joint spatiotemporal features within the period and the cycle (capturing) The joint receptive field is used to obtain the joint spatiotemporal features according to formula (16). ; When the convolution kernel slides once on the feature matrix, it involves The elements at each position correspond to 9 time points in the original time series. Calculate the output. At that time, you need to input the middle The surrounding area (extending one position to the top, bottom, left, and right) These elements are combined. From a row perspective, this represents the intervals between the points in time involved. From a column perspective, it captures short-term changes within the same period.

[0181] (16);

[0182] S35: Multi-branch feature fusion and residual connection;

[0183] First, after the periodic detection module Identify The most prominent cyclical patterns, each corresponding to a set of parameters. ,in, Indicates the period length. Indicates frequency;

[0184] Next, corresponding Create one frequency component Each processing branch is a separate processing branch, and each processing branch deals with the input sequence. Zero-padding and folding are performed sequentially, and then the data is sent to the parameter-sharing Inception module for processing. This utilizes a parameter-efficient Inception structure. Each independent processing branch shares the same set of Inception convolutional kernel parameters, significantly reducing the number of parameters compared to traditional multi-path Inception networks, thus achieving lightweight deployment. Inception blocks share weights, ensuring that the model size does not increase with the number of processing branches. Growth; only the input shape is different, realizing "same core, multiple cycles reuse".

[0185] Therefore, the number of output channels for each processing branch is Two-dimensional feature map Specifically, the two-dimensional feature map is obtained according to formula (17). ;

[0186] (17);

[0187] Furthermore, first calculate the energy value corresponding to each processing branch. Then, the fusion weight of each processing branch is calculated according to formula (18). That is, the energy corresponding to the processing branch cycle. Take the exponent, and then divide by the sum of the exponents of the energies corresponding to all k branch cycles; (18);

[0188] Next, the two-dimensional feature map of each processing branch is... Flattened into a one-dimensional vector, using fusion weights The flattened features of each processing branch are weighted, and all weighted features are summed to obtain the final output according to formula (19). Its shape is ;

[0189] (19);

[0190] By using this weighted summation, significant periods will receive greater weight due to their higher energy levels, thus becoming more prominent in the final result and preventing gradient vanishing; while noisy periods, due to their lower energy levels and smaller weights, will have their influence suppressed.

[0191] Then, output according to formula (20) With input Add, based on the new output Its shape is ;

[0192] (20);

[0193] The output Z has two main roles: first, it can be used as input to the next layer of TimesBlock; second, it can be used to form a deep network by stacking layers like L.

[0194] S36: Multi-layer TimesBlock stacking and feature evolution;

[0195] New output As new input, repeat steps S31 through S35, stacking... The TimesBlock layer gradually extracts more abstract and higher-level feature representations from the original input;

[0196] Each TimeBlock layer further processes and transforms the input features, thereby achieving deep evolution of the features.

[0197] S37: Location and positioning processing;

[0198] First, select within the underground tunnels For each feature point with known physical coordinates (such as a tunnel corner, equipment installation point, etc.), record the pixel coordinates of each feature point in the temperature field image. and the corresponding real physical coordinates Unit: meters, based on tunnel design drawings or on-site measurements, where, ;

[0199] Next, based on formula (21), an affine transformation is used to establish the transformation relationship from pixel coordinates to physical coordinates;

[0200] (twenty one);

[0201] In the formula, Let be the calibration matrix to be solved;

[0202] Furthermore, the pixel coordinates and physical coordinates of the P feature points are substituted into formula (21), and the elements of the calibration matrix are solved by the least squares method.

[0203] Then, output the set of pixel coordinates of the abnormal region. (Q represents the number of abnormal pixels) After substituting into the calibration matrix, the corresponding set of downhole physical coordinates is calculated. This enables precise location of abnormal areas within underground roadways;

[0204] S38: Output of fire warning level and coordinates of abnormal areas;

[0205] After deep extraction and evolution of temporal features through multiple layers of TimesBlock, the TimesNet model determines the fire risk level by periodic analysis of the temporal changes in the temperature field. The fire risk levels are none, low, medium, high, and extremely high, and the model provides key features corresponding to each risk level (such as maximum temperature and heating rate) to form fire warning level data. At the same time, the TimesNet model combines the spatial perception capabilities of periodicity and convolution to provide the coordinate set and thermal distribution matrix of abnormally high temperature areas in the mine, predict the fire spread trend vector, and form fire spatial coordinate location data. Finally, the TimesNet model outputs fire warning level data and fire spatial location data to support mine safety decisions.

[0206] Fire warning level data: By analyzing the periodic changes in the temperature field over time, the fire risk level is determined, and the key characteristics corresponding to each risk level are given; the fire risk levels are none, low, medium, high, and extremely high.

[0207] Fire spatial coordinate positioning data: Combining the spatial perception capabilities of periodicity and convolution, the system outputs the coordinate set and thermal distribution matrix of abnormally high-temperature areas within the mine, predicts the fire spread trend vector, and provides support for mine safety decisions. After these data are transmitted to the monitoring system in real time through edge nodes, they are presented intuitively in the form of "early warning level visualization panel + fire heat map (overlaid with mine map) + key parameter list", allowing safety personnel to quickly grasp the fire situation and assist in formulating response strategies.

[0208] like Figure 5 and Figure 6 As shown, this invention proposes a method for mine temperature field reconstruction and thermal disaster early warning based on acoustic tomography. First, frequency division multiplexing (FDM) technology is used to simultaneously transmit / receive multi-band acoustic waves, overcoming the limitations of traditional time delay estimation algorithms. By allowing multiple signals of different frequency bands to be transmitted in parallel, the time-consuming problem of sequential transmission is eliminated. Simultaneously, instantaneous sampling of multi-channel signals is achieved, ensuring time consistency and improving dynamic response capabilities. Furthermore, data acquisition efficiency is significantly improved. Altered acoustic signals are acquired via an explosion-proof microphone array, and then filtered and cross-correlated by edge computing nodes. This separates the clean signals of each sound path from the aliased acoustic waves, effectively eliminating multipath interference caused by roadway wall reflections. A further filtering process ensures that only sound path paths conforming to physical laws are retained, effectively improving data quality. Next, virtual paths are generated by interpolation between real microphones, effectively increasing the coverage density of sound path paths and compensating for measurement blind spots caused by the limited number of physical microphones. By combining the cosine theorem to calculate the virtual path length and employing weighted coefficients to fuse virtual and physical Time-of-Flight (TOF) data, the overall accuracy and spatial resolution of acoustic time-of-flight measurements can be significantly improved, resolving the matrix rank deficiency problem caused by insufficient sensor quantity. Simultaneously, the least squares method effectively suppresses measurement noise and errors, enhancing the stability of parameter inversion. Combined with a sound velocity-temperature linear model, high-resolution reconstruction of the temperature field in the tunnel cross-section can be achieved, providing more accurate data support for mine environmental monitoring. Finally, multiple consecutive frames of temperature fields are input into a TimesNet model. The TimesNet model extracts multi-scale features through adaptive periodic decomposition and two-dimensional spatiotemporal convolution, combining the optimization of both temporal and spatial resolution to map abnormal regions from pixel space to physical space. Ultimately, the fire warning level and abnormal region coordinates are output, providing support for mine safety decision-making and realizing a closed loop from monitoring to decision support.

[0209] This method is simple to implement and has low implementation costs. It can achieve full-area temperature sensing and early warning of hidden fire sources under complex working conditions such as dust, strong vibration and large cross-section without contacting or damaging the surrounding rock of the roadway. It can provide efficient and accurate technical means for temperature field monitoring in mines and has high practicality and adaptability.

[0210] The present invention also provides a downhole temperature field reconstruction and thermal disaster early warning system based on acoustic tomography, which is used to realize a method for downhole temperature field reconstruction and thermal disaster early warning based on acoustic tomography, including an explosion-proof speaker array, a microphone array, an edge computing node and a host computer;

[0211] The explosion-proof loudspeakers and microphone arrays are arranged alternately on both sides of the alleyway in the monitoring area;

[0212] The explosion-proof speaker array is used to simultaneously emit sound waves of different frequency bands;

[0213] The microphone array is used to collect sound wave signals and send them to the edge computing node;

[0214] The edge computing nodes are connected to the explosion-proof speaker array and the microphone array respectively. They are used to drive the explosion-proof speaker array to emit sound waves of different frequency bands simultaneously using frequency division multiplexing technology. They are used to realize high-resolution temperature field reconstruction of the temperature field of the tunnel cross section based on the received sound wave signals. At the same time, they are used to input multiple consecutive frames of temperature field images as input data to the host computer.

[0215] The host computer is connected to the edge computing node and has a built-in TimesNet model. The TimesNet model is used to extract multi-scale features from the input data through adaptive periodic decomposition and two-dimensional spatiotemporal convolution. It achieves stable feature transfer and deep evolution through residual connection and multi-layer stacking. At the same time, combined with coordinate mapping, it converts abnormal pixels in the image into meaningful underground physical coordinates, and finally outputs the fire warning level and abnormal area coordinates.

[0216] As a preferred embodiment, it also includes an explosion-proof power supply, which is used to supply power to various electrical devices;

[0217] As a preferred embodiment, the edge computing node is a PLC controller.

[0218] As a preferred option, the host computer is an industrial computer.

[0219] In this invention, explosion-proof loudspeakers and microphone arrays are arranged in a staggered pattern on both sides of the tunnel, effectively covering the monitoring area. The explosion-proof loudspeaker and microphone arrays work together using frequency division multiplexing technology to synchronously transmit / receive multi-band sound waves, achieving instantaneous sampling of multi-channel signals, ensuring time consistency, and significantly improving data acquisition efficiency. Edge computing nodes integrate signal and deep inference capabilities, reducing reliance on monitoring center processing equipment, lowering data transmission volume, and enabling high-resolution temperature field reconstruction at the near end. The host computer has a built-in TimesNet model, which can analyze temperature field data in real time and autonomously and promptly detect potential fire threats, providing intuitive and operable decision support for emergency response and significantly improving fire location accuracy.

[0220] The system has a simple structure and a high degree of intelligence. It can autonomously achieve full-area temperature perception of the target monitoring area and early warning of hidden fire sources.

Claims

1. A method for downhole temperature field reconstruction and thermal disaster early warning based on acoustic tomography, characterized in that, Includes the following steps: Step 1: Acoustic signal acquisition and matrix equation construction; S11: Explosion-proof speaker arrays and microphone arrays are arranged alternately on both sides of the roadway in the monitoring area; S12: Using frequency division multiplexing technology, the explosion-proof speaker array can simultaneously emit sound waves of different frequency bands; at the same time, the sound wave signals are synchronously collected and received through the microphone array and sent to the edge computing node. S12-1: Frequency band division; The transmitted signal bandwidth is divided into multiple non-overlapping sub-bands in the frequency domain, and a segment is set between adjacent sub-bands as isolation protection to avoid spectrum interference; S12-2: Multi-signal synchronous transmission and reception; enabling different explosion-proof loudspeakers to simultaneously transmit sub-signals of different frequency bands, achieving parallel transmission through frequency domain isolation, and eliminating the time interval between channel switching; simultaneously, receiving aliased sound wave signals through a microphone array and sending them to the edge computing node; wherein, the minimum measurement time is obtained according to formula (1). ; （1）； In the formula, Indicates the pulse width of the sound source signal. Indicates the path of sound waves Flight time on; S13: Edge computing nodes perform bandpass filtering and cross-correlation operations on acoustic signals, extract the time of flight of each sound path, and filter the TOF measurement values ​​of the real sound path by verifying time consistency, forming a real sound path dataset and constructing an overdetermined set of equations. Step 2: Virtual path enhancement and temperature field reconstruction; S21: Based on the TOF data of the microphone array, a virtual sound ray method is used to interpolate and generate a virtual path between real microphones. The length of the virtual path is calculated based on the cosine theorem, and weighted coefficients are used to fuse the virtual and physical TOF data to achieve virtual path enhancement. S22: Extend the measurement matrix and TOF vector, solve the extended overdetermined equations by least squares method, and combine the sound velocity-temperature linear model to realize high-resolution temperature field reconstruction of the temperature field of the tunnel cross section; Step 3: Fire early warning and anomaly detection; Edge computing nodes input multiple consecutive frames of temperature field images into the TimesNet model. A Fast Fourier Transform (FFT) is used to convert the one-dimensional time series to the frequency domain, revealing the dominant periodic pattern. The one-dimensional time series data is folded into a stacked matrix containing complete spatiotemporal information according to different periods. Inception convolution is used to simultaneously capture multi-scale features in both time and space dimensions on the stacked matrix. A parameter-sharing mechanism maintains powerful feature extraction capabilities while achieving model lightweighting. Stable feature transfer and deep evolution are achieved through residual connections and multi-layer stacking. Coordinate mapping is used to convert abnormal pixels in the image into meaningful underground physical coordinates. Finally, the TimesNet model outputs the fire warning level and the coordinates of the abnormal area.

2. The method for downhole temperature field reconstruction and thermal disaster early warning based on acoustic tomography according to claim 1, characterized in that, In step S13 of step one, the process of constructing the overdetermined system of equations is as follows: S13-1: Filtering; extracting sub-signals of each measurement path from the aliased acoustic signal using a bandpass filter; S13-2: TOF matrix construction; Organizing the TOF measurements into a matrix of TOF measurements. The structural parameter matrix is ​​constructed by the position of the physical microphone and the length of the sound wave path. ; S13-3: Matrix equation construction; Based on the measurement path, establish an overdetermined system of equations according to formula (2); （2）； In the formula, For temperature and speed The matrix to be determined.

3. The method for downhole temperature field reconstruction and thermal disaster early warning based on acoustic tomography according to claim 2, characterized in that, In step S21 of step two, the process of implementing virtual path enhancement is as follows: S21-1: Generate a virtual acoustic wave path; At a spacing of Two physical microphones Insert evenly between The first virtual microphone, according to the law of cosines, is obtained according to formula (3). The virtual sound wave path length corresponding to each virtual microphone ; （3）； In the formula, For two adjacent physical microphones The spacing; and These are two adjacent physical microphones. The sound wave path; The spacing between adjacent microphones; S21-2: Calculation of the virtual path TOF vector; Calculate the first according to formula (4) TOF of the virtual sound wave path corresponding to each virtual microphone; （4）； In the formula, The weighting coefficients represent the virtual sound wave path. ; and The path length between adjacent entities; and For the Time-of-Flight (TOF) of adjacent entity paths.

4. The method for downhole temperature field reconstruction and thermal disaster early warning based on acoustic tomography according to claim 3, characterized in that, In step S22 of step two, the process of achieving high-resolution temperature field reconstruction of the tunnel cross-section temperature field is as follows: S22-1: Matrix expansion; TOF data from the virtual microphone TOF data compared to physical microphones Merge to form the expanded TOF vector Based on the virtual microphone position and path length, expand the structure parameter matrix. for It contains geometric information of both physical entities and virtual paths; S22-2: Least squares solution and temperature field reconstruction; The extended overdetermined equations are solved using the least squares method according to formula (5); the temperature-velocity relationship is modeled according to the linear relationship between sound wave propagation speed and temperature in formula (6), so as to realize the high-resolution temperature field reconstruction of the temperature field of the tunnel cross section. （5）； （6）。 5. A downhole temperature field reconstruction and thermal disaster early warning system based on acoustic tomography, used to implement the downhole temperature field reconstruction and thermal disaster early warning method based on acoustic tomography as described in any one of claims 1 to 4, characterized in that, This includes explosion-proof speaker arrays, microphone arrays, edge computing nodes, and host computers; The explosion-proof loudspeakers and microphone arrays are arranged alternately on both sides of the alleyway in the monitoring area; The explosion-proof speaker array is used to simultaneously emit sound waves of different frequency bands; The microphone array is used to collect sound wave signals and send them to the edge computing node; The edge computing nodes are connected to the explosion-proof speaker array and the microphone array respectively. They are used to drive the explosion-proof speaker array to emit sound waves of different frequency bands simultaneously using frequency division multiplexing technology. They are used to realize high-resolution temperature field reconstruction of the temperature field of the tunnel cross section based on the received sound wave signals. At the same time, they are used to input multiple consecutive frames of temperature field images as input data to the host computer. The host computer is connected to the edge computing node and has a built-in TimesNet model. The TimesNet model is used to extract multi-scale features from the input data through adaptive periodic decomposition and two-dimensional spatiotemporal convolution. It achieves stable feature transfer and deep evolution through residual connection and multi-layer stacking. At the same time, combined with coordinate mapping, it converts abnormal pixels in the image into meaningful underground physical coordinates, and finally outputs the fire warning level and abnormal area coordinates.

6. The downhole temperature field reconstruction and thermal disaster early warning system based on acoustic tomography according to claim 5, characterized in that, The edge computing node is a PLC controller.

7. The downhole temperature field reconstruction and thermal disaster early warning system based on acoustic tomography according to claim 6, characterized in that, The host computer is an industrial computer.

Images