Mine gushing water image processing system based on AI identification

Abstract

The present application relates to the technical field of disaster monitoring, in particular to a mine gushing water image processing system based on AI recognition, which comprises a fissure water flow tracking module, a voiceprint disturbance extraction module, an acceleration mapping module, a shape analysis module and a risk output module. In the present application, the boundary pixel coordinates of fissure water flow are extracted from continuous frame images, the inter-frame differences of gray scale gradient amplitude and direction angle are combined, the water flow micro-displacement is accurately captured and the offset trajectory is generated, the liquid surface disturbance wave peak position and the primary and secondary frequency amplitude value difference of voiceprint are synchronously extracted, the detailed voiceprint disturbance sequence is formed, the pixel displacement velocity is calculated to obtain the acceleration and the acceleration is mapped to the time axis, the boundary motion trend and the shape change are synchronously analyzed in the time dimension, the multi-source signals are completed risk frame locking under the double threshold value judgment in the same time period, the water flow dynamics is stereoscopically presented, and the accuracy and reliability of gushing water risk determination in complex mine environment are significantly improved, and the missed detection and misjudgment probability is reduced.

Description

Technical Field

[0001] This invention relates to the field of disaster monitoring technology, and in particular to an AI-based image processing system for mine water inrush. Background Technology

[0002] The field of water control technology in metal mines involves dynamic monitoring and early warning of mines and surrounding aquifers and aquifer structures during mining operations. Core aspects include mine hydrogeological surveys, aquifer permeability analysis, identification of water inrush channels, and monitoring of mine water inrush processes. It also involves identifying and controlling potential water hazards in mines through methods such as on-site sensor acquisition, image recording, and analysis. Traditional AI-based mine water inrush image processing systems utilize camera equipment to acquire images of the mine working face and roadways, combine them with a water inrush image feature set from training samples for feature extraction and pattern matching, and then analyze the presence of signs of water inrush using preset image feature comparison methods. Typically, optical image acquisition and image segmentation, feature parameter statistics, and comparison methods are used to distinguish and identify water inrush-related images from ordinary operation images.

[0003] Current technologies for identifying mine water inrush images rely on optical image acquisition and a comparison model based on sample features. The identification of water flow states is highly dependent on the matching effect of image appearance features. In actual operation, this method is weak in capturing subtle flows or atypical water hazard patterns. When water flow changes are not significant or differ greatly from sample features, risk signs are easily overlooked. Single-dimensional visual monitoring methods are easily interfered with in low-light, dusty, or obstructed environments, resulting in decreased image quality and unstable identification accuracy. Furthermore, this model cannot deeply analyze water flow dynamics over time, lacking the ability to quantitatively analyze movement trends, surface disturbances, and the correlation between multi-dimensional signals. The monitoring of sudden water flow evolution is delayed, potentially missing the optimal early warning opportunity in the case of rapidly developing risks, affecting the timeliness and effectiveness of mine safety protection. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an AI-based image processing system for mine water inrush.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A mine water inrush image processing system based on AI recognition includes:

[0006] The fracture water flow tracking module acquires the pixel coordinates of the boundary of the mine fracture water flow in continuous frame images, calculates the gray-level gradient magnitude and gradient direction angle of the boundary pixels, calculates the difference of gray-level gradient magnitude and gradient direction angle of pixels at the same position between adjacent frames, combines them to form a set of micro-displacement vectors, arranges them in time order to statistically distribute the vector direction changes, and generates a water flow offset image.

[0007] The acoustic disturbance extraction module collects the amplitude values ​​of the main frequency and secondary frequency of the water flow acoustic pattern based on the water flow offset image, calculates the difference in amplitude of the main frequency, extracts the sequence of liquid surface disturbance wave peak positions from the image and calculates the displacement difference between adjacent wave peaks to generate an acoustic disturbance sequence.

[0008] The acceleration mapping module extracts the pixel coordinates of the mine fracture water flow boundary according to the acoustic perturbation sequence, calculates the pixel displacement velocity of adjacent frames and obtains the pixel edge acceleration value, and summarizes the maximum, minimum and average values ​​to form an acceleration triplet to generate an acceleration change map.

[0009] The morphological analysis module calls the acceleration change map and the water flow migration image to perform time axis synchronization matching, extracts time periods with consistent trends, counts the morphological change range at the boundaries of the time periods, and generates a morphological change image.

[0010] As a further aspect of the present invention, the water flow offset image includes a vector direction change distribution, a set of micro-displacement vectors arranged in time sequence, and pixel coordinates of the boundary of mine fracture water flow; the acoustic disturbance sequence includes the main frequency amplitude difference, the peak displacement difference, and the liquid surface disturbance peak position sequence; the acceleration change spectrum includes the edge acceleration change curve mapped by the time axis and the edge acceleration triplet arranged in time sequence; and the morphological change image includes a consistent trend time period and the boundary morphological change range.

[0011] The grayscale gradient magnitude refers to the rate of change of pixel grayscale at spatial location;

[0012] The gradient direction angle represents the direction of pixel grayscale change;

[0013] The dominant frequency and secondary frequency of the water flow acoustic pattern refer to the two frequencies with the highest amplitude in the spectral components of the water flow acoustic wave signal, which are calculated by fast Fourier transform and are measured in Hertz.

[0014] The position of the liquid surface disturbance peak can be determined by analyzing the liquid surface brightness curve frame by frame, and the displacement difference refers to the difference in pixel coordinates between the positions of consecutive peaks.

[0015] The pixel displacement speed is the amount of displacement of a pixel per unit time, which is obtained by dividing the difference in pixel position between adjacent frames by the inter-frame time.

[0016] The pixel edge acceleration value is the time rate of change of the pixel displacement velocity;

[0017] The edge acceleration triplet consists of the maximum acceleration value, the minimum acceleration value, and the average acceleration value;

[0018] The range of boundary morphology changes refers to the spatial positional change interval of the fractured water flow boundary over time, calculated by the difference between the maximum and minimum offsets of the boundary coordinates.

[0019] As a further aspect of the present invention, the fissure water flow tracking module includes:

[0020] The boundary localization submodule obtains the corresponding grayscale value based on the pixel coordinates of the boundary of the mine fracture water flow in the continuous frame image, calculates the grayscale gradient magnitude and gradient direction angle of each boundary pixel, calculates the overall gradient magnitude feature of the boundary pixels in the same frame, and obtains the average boundary gradient magnitude.

[0021] The micro-vector construction submodule calculates the time sequence difference based on the difference in gray-level gradient magnitude and gradient direction angle of the boundary pixels at the same position between adjacent frames, accumulates them by pixel index, obtains the vector direction variability, sorts them according to the variability values ​​and records the time reference, and generates the sequence identifier of the raster mapping.

[0022] The offset imaging submodule, based on the sequence identifier of the raster mapping, uses the vector direction variability corresponding to the sequence identifier as the weight for each boundary pixel index, accumulates the gray-level gradient direction angle difference between adjacent frames and maps it to the raster, and outputs the displacement visualization matrix according to the correspondence between the mapped raster and the boundary pixel coordinates to generate a water flow offset image.

[0023] As a further aspect of the present invention, the voiceprint disturbance extraction module includes:

[0024] The acoustic amplitude acquisition submodule acquires the amplitude values ​​of the main frequency and secondary frequency of the water flow acoustic pattern based on the water flow offset image, calculates the difference in the amplitude of the main frequency, and arranges them according to the sampling time to obtain the difference in the amplitude of the main frequency.

[0025] The peak distance calculation submodule calculates the displacement difference between adjacent peaks and the time between peaks based on the difference in amplitude of the main frequency and the position sequence of the disturbance wave peaks on the liquid surface. It calculates the reference values ​​of amplitude and displacement based on the same segment of data, obtains the disturbance coupling strength value, and indexes and collects them into a disturbance coupling strength sequence.

[0026] The pairing and sequencing submodule, based on the disturbance coupling strength sequence, calls the difference in the main frequency amplitude and the difference in the displacement of adjacent peaks, indexes and pairs the same disturbance coupling strength values ​​and arranges them according to the sampling time to obtain the pairing sequence and obtain the acoustic perturbation sequence.

[0027] As a further aspect of the present invention, the acceleration mapping module includes:

[0028] The boundary coordinate submodule extracts the boundary pixel coordinates of the mine fracture water flow according to the acoustic perturbation sequence, matches the boundary pixels of adjacent frames according to the time index, calculates the boundary pixel displacement velocity and compares the velocity distribution with the boundary gray-scale gradient distribution to obtain the average boundary displacement velocity.

[0029] Based on the boundary pixel correspondence, the pixel acceleration submodule uses the average boundary displacement velocity as a benchmark to calculate the velocity difference of adjacent frame edge pixels and obtain the edge acceleration. After correction by combining the neighborhood velocity change and the acoustic text disturbance noise, the edge acceleration value sequence is obtained.

[0030] The triplet mapping submodule calculates the maximum, minimum, and average acceleration values ​​of frame edge pixels based on the edge acceleration value sequence, generates time series triplets, maps them to the time axis, and obtains an acceleration change spectrum.

[0031] As a further aspect of the present invention, the morphological analysis module includes:

[0032] The time synchronization submodule performs time axis synchronization matching based on the acceleration change map and the fracture water flow migration trajectory image, extracts the timestamp and amplitude pole sequence, calculates the difference between adjacent poles and generates a synchronization error sequence, compares the synchronization error sequence with the time difference benchmark value and filters the time period that meets the conditions, and generates a time alignment deviation interval.

[0033] The trend extraction submodule defines a time window based on the time alignment deviation interval, extracts the acceleration change increment sequence and the offset trajectory displacement increment sequence within the window, calculates the consistency score and filters the time period set, merges adjacent or overlapping time periods, and generates a consistent trend time period set.

[0034] The boundary statistics submodule calls the consistent trend time period set to extract boundary frames, calculates the displacement difference of morphological contour points and counts the range intervals, splices each range interval in time order and maps it to the pixel matrix to generate a morphological change image.

[0035] As a further aspect of the present invention, the system further includes:

[0036] The risk output module determines the risk threshold for sudden water inrush based on the morphological change image and the acoustic perturbation sequence, extracts and marks image frames in which the two result values ​​both exceed the threshold within the same time period, and generates a sudden water inrush risk image set.

[0037] The surge water risk image set includes marked surge water risk image frames;

[0038] The threshold for judging the risk of sudden water inrush is set based on historical statistics of sudden water inrush in the area where the mine is located and national coal mine safety technical standards. The judgment criteria are that both the change rate of acoustic amplitude and the change range of boundary morphology exceed the limit simultaneously.

[0039] As a further aspect of the present invention, the risk output module includes:

[0040] The threshold determination submodule extracts the frame-by-frame morphological change amount and the same sampling rate voiceprint amplitude based on the morphological change image and the voiceprint perturbation sequence, compares the morphological threshold and the voiceprint threshold, determines whether the dual-channel exceeds the threshold at the same moment, and calculates the ratio of the dual-channel exceeding the threshold duration to the total duration to obtain the synchronous exceeding the threshold ratio.

[0041] The common frame screening submodule constructs a comprehensive score for each frame at the alignment time based on the synchronization over-threshold ratio. Combining the difference and synchronization between the morphological change amount and the voiceprint amplitude, it locates candidate common frame segments and obtains the average common frame score.

[0042] The marker set construction submodule calls the common frame score mean and the synchronous over-threshold ratio, filters the dual-channel over-threshold image frames in the same segment, and superimposes the timestamp and marker number on the frames. It then aggregates the frame sequence and time index to obtain the sudden water inrush risk image set.

[0043] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0044] In this invention, the pixel coordinates of the fractured water flow boundary are extracted from continuous frame images. Combined with the inter-frame differences in grayscale gradient amplitude and direction angle, the micro-displacement of the water flow is accurately captured and the offset trajectory is generated. The position of the liquid surface disturbance wave peak and the difference in the amplitude of the main and secondary frequencies of the acoustic pattern are extracted simultaneously to form a detailed acoustic pattern disturbance sequence. The acceleration of the pixel displacement velocity is calculated and mapped to the time axis, so that the boundary motion trend and morphological changes are analyzed synchronously in the time dimension. The risk frame is locked under the dual threshold judgment of multiple source signals in the same time period, realizing a three-dimensional presentation of the dynamics of the water flow. This significantly improves the accuracy and reliability of the risk judgment of sudden water inrush in complex mine environments and reduces the probability of missed detection and false judgment. Attached Figure Description

[0045] Figure 1 This is a system flowchart of the present invention;

[0046] Figure 2 This is a flowchart of the fissure water flow tracking module of the present invention;

[0047] Figure 3 This is a flowchart of the voiceprint disturbance extraction module of the present invention;

[0048] Figure 4 This is a flowchart of the acceleration mapping module of the present invention;

[0049] Figure 5 This is a flowchart of the morphology analysis module of the present invention;

[0050] Figure 6 This is a flowchart of the risk output module of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0052] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0053] Please see Figure 1 A mine water inrush image processing system based on AI recognition includes:

[0054] The fracture water flow tracking module acquires the pixel coordinates of the boundary of the mine fracture water flow in continuous frame images, calculates the gray-level gradient magnitude and gradient direction angle of each boundary pixel, calculates the difference in gray-level gradient magnitude and gradient direction angle of pixels at the same position between adjacent frames, combines them to form a set of micro-displacement vectors, and arranges them in time order to statistically distribute the vector direction change, generating a water flow offset image.

[0055] Gray-level gradient magnitude refers to the rate of change of pixel gray level in spatial location;

[0056] The gradient direction angle represents the direction of pixel grayscale change;

[0057] The acoustic perturbation extraction module acquires the amplitude of the main frequency and secondary frequency of the water flow acoustic pattern based on the water flow offset image, calculates the amplitude difference of the main frequency, extracts the position sequence of the liquid surface disturbance wave peaks from the image and calculates the displacement difference between adjacent wave peaks, and pairs the main frequency amplitude difference with the wave peak displacement difference to form an acoustic perturbation sequence.

[0058] The dominant frequency and secondary frequency of the water flow acoustic pattern refer to the two frequencies with the highest amplitude in the spectral components of the water flow acoustic wave signal. They are calculated by Fast Fourier Transform (FFT) and the unit is Hertz (Hz).

[0059] The position of the liquid surface disturbance peak can be determined by analyzing the liquid surface brightness curve frame by frame. The displacement difference refers to the difference in pixel coordinates between the positions of consecutive peaks.

[0060] The acceleration mapping module extracts the pixel coordinates of the boundary of the mine fracture water flow based on the acoustic perturbation sequence, calculates the pixel displacement velocity in adjacent frames, and obtains the pixel edge acceleration value. The maximum, minimum and average values ​​are summarized to form an edge acceleration triplet. The triplets are arranged in chronological order and mapped to the time axis to generate an acceleration change map.

[0061] Pixel displacement speed is the amount of displacement of a pixel per unit time (pixels / second), which is obtained by dividing the difference in pixel position between adjacent frames by the inter-frame time.

[0062] Pixel edge acceleration is the rate of change of pixel displacement velocity over time, expressed in pixels per second².

[0063] The edge acceleration triplet consists of the maximum acceleration value, the minimum acceleration value, and the average acceleration value;

[0064] The morphological analysis module calls the acceleration change map and the fracture water flow migration trajectory image to perform time axis synchronous matching, extracts time periods with consistent trends, counts the boundary morphological change range of the time period, and generates morphological change images.

[0065] The range of boundary morphology change refers to the spatial positional change of the fractured flow boundary over a time series, which can be calculated by the difference between the maximum and minimum offsets of the boundary coordinates.

[0066] The risk output module determines the threshold for judging the risk of sudden water inrush based on morphological change images and acoustic perturbation sequences, extracts and marks image frames in which both result values ​​exceed the threshold within the same time period, and generates a sudden water inrush risk image set.

[0067] The threshold for judging the risk of sudden water inrush is set based on historical statistics of sudden water inrush in the area where the mine is located and national coal mine safety technical standards. The judgment criteria are the simultaneous exceeding of the limits of the change rate of acoustic amplitude and the change range of boundary morphology.

[0068] The water flow migration images include vector direction change distribution, time-sequential micro-displacement vector set, and mine fracture water flow boundary pixel coordinates. The acoustic disturbance sequence includes the main frequency amplitude difference, wave crest displacement difference, and liquid surface disturbance wave crest position sequence. The acceleration change map includes edge acceleration change curves mapped by the time axis and time-sequential edge acceleration triples. The morphological change images include time periods with consistent trends and boundary morphological change ranges. The sudden water inrush risk image set includes marked sudden water inrush risk image frames.

[0069] Please see Figure 2 The fissure water flow tracking module includes:

[0070] The boundary localization submodule obtains the corresponding grayscale value based on the pixel coordinates of the boundary of the mine fracture water flow in the continuous frame image, calculates the grayscale gradient magnitude and gradient direction angle of each boundary pixel, calculates the overall gradient magnitude feature of the boundary pixels in the same frame, and obtains the average boundary gradient magnitude.

[0071] Based on the pixel coordinates of the mine fracture water flow boundary in continuous frame images, the grayscale values ​​of the images are read pixel by pixel, and the coordinates in the grayscale matrix are used to... Corresponding grayscale As input data, the grayscale difference in the horizontal and vertical directions is calculated for each pixel. The horizontal gradient is calculated by subtracting the grayscale value of the left neighbor from the grayscale value of the right neighbor and dividing by two. The vertical gradient is calculated by subtracting the grayscale value of the upper neighbor from the grayscale value of the lower neighbor and dividing by two. For example, if pixel (50, 80) has a grayscale value of 120, its right neighbor (51, 80) has a grayscale value of 130, and its left neighbor (49, 80) has a grayscale value of 110, then the horizontal gradient is... Assuming the gray level of the lower neighbor (50, 81) is 125 and the gray level of the upper neighbor (50, 79) is 115 in the vertical direction, then the vertical gradient is: The overall gradient magnitude is Repeat this process to calculate the gradient magnitude of all boundary pixels. For example, the gradient magnitudes of the other four boundary pixels are 14.32, 12.50, 9.85, and 13.10. Store these values ​​in the set vector [11.18, 14.32, 12.50, 9.85, 13.10]. Sum the sets to get 60.95, and then divide by the total number of pixels, 5, to get the overall average gradient magnitude of 12.19. When the gray value is in the range of 0 to 255, the acceptable range of gradient magnitude is usually 0 to 50. Therefore, this calculation result is within the normal value range, and finally the average gradient magnitude of the boundary pixels is obtained.

[0072] The micro-vector construction submodule calculates the temporal differences based on the difference in grayscale gradient magnitude and gradient direction angle between boundary pixels at the same position in adjacent frames, and then accumulates these differences by pixel index using the following specific formula:

[0073] ;

[0074] The vector direction variability is obtained by calculation, based on Numerical sorting and time index recording are used to generate sequence identifiers for raster mapping;

[0075] in, Represents the variability of vector direction. Representing pixels In time The grayscale gradient magnitude (unit: Gray / pixel). Representing pixels The grayscale gradient magnitude at the next time step (unit: Gray / pixel). Represents the reference gradient magnitude (unit: Gray / pixel). Representing pixels The gradient direction angle difference (in rad) between adjacent frames. Represents the difference in reference angle (unit: rad). Representing pixels The boundary curvature and angle difference are coupled with weights (dimensionless). Representing pixels Time stability measure (dimensionless). Representing pixels Gray variance of the neighborhood (unit: Gray) ), Representative reference grayscale variance (unit: Gray) ), Represents the index of all boundary pixels Summation;

[0076] Based on the boundary pixel position, extract the grayscale gradient magnitude at the same coordinates in adjacent frames. and Calculate the absolute value of the difference and divide it by the reference gradient magnitude. Take the square root and multiply by the boundary curvature and angle difference coupling weight. The first part is obtained by summing all pixels, and then dividing by the sum of all weights. The second part is the time-stability metric. Difference in gradient direction angle with adjacent frames Divide by the reference angle difference Multiply and sum the results, then take the absolute value and divide by the grayscale variance. Divide by reference gray variance The square root of the sum, plus the two parts, gives the vector direction variability. Assuming there are 3 boundary pixels, The values ​​are 12.0, 15.0, and 14.5. The values ​​are 13.2, 14.0, and 15.5. , The values ​​are 1.2, 1.0, and 1.4. The values ​​are 0.9, 1.0, and 0.8. The values ​​are 0.15, 0.18, and 0.12 (rad). , For 25.0, 30.0, 20.0 (Gray) ), As shown in Table 1:

[0077]

[0078] Part 1 Calculation: The difference in pixel 1 is... The square root is 0.2449, multiplied by a weight of 1.2, resulting in 0.2939, which is the difference between two pixels. Taking the absolute value of 0.05, the square root of 0.2236, multiplied by 1.0, yields 0.2236, the difference between 3 pixels. The square root 0.2236 multiplied by 1.4 equals 0.3130. The sum of the three is 0.8305. Dividing by the weighted sum 3.6 gives 0.2307. The second part of the calculation: the sum of the angles is... The absolute value remains 2.055, and the variance is... The square root is 1.2247. The second part of the result is 2.055 / 1.2247 = 1.6782. Finally... This value is at a medium level within the preset baseline range [0, 5], and can be used as the weighting basis for generating the sequence identifier of the raster map;

[0079] The formula's operational logic involves decomposing the changes in boundary pixels across different time frames into two weighted components and merging them. The first part calculates the difference in grayscale gradient magnitude between the current and next time steps for each pixel, then divides it by a reference gradient magnitude for normalization to eliminate dimensional differences. The absolute value is then taken to ensure the difference is unaffected by positive or negative directions. Taking the square root weakens the impact of extreme differences on the overall result, scaling the change magnitude proportionally to the square root of the magnitude. This value is then multiplied by the boundary curvature and angle difference coupling weights of the pixel, reflecting its importance in the overall boundary structure. Finally, the weighted results of all pixels are summed and divided by the total weights to obtain normalization. The first part calculates the average change; the second part calculates the product of the time stability metric of each pixel and the gradient direction angle difference, and normalizes it with the reference angle difference to make the angle difference compare under a unified dimension. The absolute value of the sum of the values ​​of all pixels is taken to ensure that the directional differences do not cancel each other out during synthesis. Then it is divided by the square root of the ratio of gray variance to reference variance. The square root operation is also to suppress the excessive influence of extreme variance values ​​on the whole. Finally, the results of the first and second parts are added together to unify the two different dimensions of gradient magnitude change and direction change into a whole index, namely vector direction variability, which is used to reflect the dynamic change of the boundary in space and direction.

[0080] Vector directional variability is a comprehensive quantitative indicator of the changes in the boundary of mine fracture water flow between adjacent time frames. It not only reflects the degree of change of the gray-level gradient amplitude of each pixel on the boundary over time, but also combines the magnitude of the gradient direction change. By simultaneously considering the spatial differences in gradient strength and directional shift, this indicator can express the dynamic stability or degree of disturbance of the boundary morphology in a continuous image sequence on a unified numerical scale. The larger the value, the more significant the temporal changes in the gray-level or directional features of the boundary. The smaller the value, the more stable the boundary morphology and flow direction remain over time. Therefore, it can serve as an important basis for analyzing and comparing the water flow shift in different time periods.

[0081] The offset imaging submodule is based on the raster mapping sequence identifier. For each boundary pixel index, the vector direction variation corresponding to the sequence identifier is used as the weight. The gray-level gradient direction angle difference between adjacent frames is accumulated and mapped to the raster. According to the correspondence between the mapped raster and the boundary pixel coordinates, the displacement visualization matrix is ​​output to generate the water flow offset image.

[0082] The raster-mapped sequence identifier method maps the variation weight of each pixel in the sequence identifier array to the original boundary pixel coordinates. It iterates through each boundary pixel, reads its weight, and multiplies it by the gradient direction angle difference between adjacent frames to obtain the weighted angle difference, which is then stored in the corresponding position of the mapping matrix. For example, the weight of pixel (50, 80) is 1.9089, and the angle difference is 0.16 rad, so the weighted angle difference is... The values ​​are recorded at position [50, 80] in the matrix. This process is repeated to form a complete weighted angle difference matrix. All values ​​in the matrix are arranged according to the boundary pixel positions to generate a two-dimensional raster map. Finally, the map is converted into a two-dimensional visual matrix and rendered as a grayscale or pseudo-color displacement image file. At this time, each value in the matrix represents the displacement change amplitude of the pixel between frames. For example, the color level corresponding to the value 0.3054 at position [50, 80] in the matrix directly reflects the intensity of its displacement change, thus obtaining the water flow displacement image.

[0083] Please see Figure 3 The voiceprint perturbation extraction module includes:

[0084] The acoustic amplitude acquisition submodule acquires the amplitude values ​​of the main frequency and secondary frequency of the water flow acoustic pattern based on the water flow offset image, calculates the difference in the amplitude of the main frequency, and arranges them according to the sampling time to obtain the difference in the amplitude of the main frequency.

[0085] To acquire the dominant and secondary frequency amplitudes of water flow acoustic signatures based on water flow migration images, the acquisition device needs to be fixed above the water flow cross-section, maintaining a stable angle and lighting conditions. Images of surface ripples in the water flow are continuously acquired at a frequency of 30 frames per second. Spectral decomposition is performed on each frame to obtain the dominant and secondary frequency amplitude data. The acquisition frequency band is set from 0Hz to 2000Hz to ensure coverage of the entire water flow acoustic signature characteristic range. Amplitude data is recorded in decibels (dB), for example, at the sampling time... The detected main frequency amplitude is 82.5dB and the secondary frequency amplitude is 74.3dB. Therefore, the difference in main frequency amplitude is... During the sampling process, it is necessary to determine whether the peak values ​​of the spectrum meet the selection criteria. The criterion is that the peak amplitude is greater than the average background noise amplitude plus a 3dB threshold. If the average background noise measured in the experiment is 65dB, then the threshold is 68dB. Only peak values ​​with amplitudes higher than this threshold are retained for recording, and the amplitude difference array is continuously collected. Meanwhile, the sampling time is recorded to ensure that the amplitude difference corresponds one-to-one with the time, and the values ​​are arranged in chronological order to finally obtain the main frequency amplitude difference.

[0086] The peak distance calculation submodule calculates the displacement difference and inter-peak time between adjacent peaks based on the difference in the amplitude of the main frequency and the position sequence of the liquid surface disturbance wave peaks. The specific calculation formula for the reference quantities of amplitude and displacement based on the same segment of data is as follows:

[0087] ;

[0088] The perturbation coupling strength value is obtained through calculation, and the indexes are collected into a perturbation coupling strength sequence;

[0089] in, Represents the perturbation coupling strength of segment i. Represents the difference in main frequency amplitude. The root mean square of the non-peak amplitude of the dominant frequency adjacent band in segment i is represented by... The difference in wave crest displacement between adjacent segments i represents the segment i. Represents the amplitude of the secondary frequency. The time interval between peaks in segment i represents the time interval between peaks. The number of jitter samples in segment i represents the number of jitter samples. The instantaneous displacement residual of sample j within segment i. Within segment i The median reference amplitude, Within segment i The median reference displacement, Within segment i The median reference time, This indicates summing over j. Represents absolute value. Represents the square root;

[0090] To calculate the displacement difference and inter-peak time between adjacent peaks based on the difference in the amplitude of the main frequency and the position sequence of the liquid surface disturbance peaks, it is necessary to extract the local maxima of the brightness curve in the image sequence as the peak positions and record the horizontal pixel position difference between adjacent peaks, combined with pixel scale. Convert to millimeters, for example, in segments In the middle, the position of the first peak is Pixel, the position of the second peak is Pixels The displacement difference is Peak time Calculated by dividing the frame rate by the frame difference; for example, if the interval is 13 frames... The root mean square amplitude of the non-peak adjacent band of the main frequency Calculated based on non-peak amplitude samples, for example, sample but Second frequency amplitude The amplitude of the second frequency peak detected in the section is 4.5 dB, and the reference amplitude is taken. Reference displacement Baseline time Take the section respectively , , The median value, for example , , Combined with the number of jitter samples and displacement residual The calculation results are shown in Table 2:

[0091]

[0092] Based on the data in Table 2, substitute each parameter into the formula:

[0093] ;

[0094] in The calculation process is as follows:

[0095] ;

[0096] ;

[0097] ;

[0098] Get section The perturbation coupling strength value is 1.664, and it is added to the perturbation coupling strength sequence by index;

[0099] The calculation logic of this formula is to integrate different disturbance characteristic dimensions into a whole index. First, the normalized result of the difference between the main frequency amplitude and the root mean square difference of the amplitude of the adjacent non-peak is multiplied with the normalized result of the displacement difference between adjacent peaks. This represents the degree of coupling between amplitude disturbance and displacement disturbance. The multiplication reflects that both must change simultaneously to amplify the disturbance intensity. Second, the secondary frequency amplitude and the inter-peak time are normalized and multiplied and squared. This aims to balance the common influence of frequency components and fluctuation period on the disturbance. The introduction of the square root can reduce the impact of a sudden increase in a single parameter on the result. Finally, the displacement residual is normalized, averaged, and subtracted. This represents the removal of the influence of random and non-periodic components from the total disturbance intensity, making the calculated value more accurately reflect the core characteristics of periodic disturbance.

[0100] The disturbance coupling strength value represents the comprehensive correlation between the change in the amplitude of the dominant frequency and the change in the peak displacement of the liquid surface disturbance within a specific water flow section. It is calculated by processing various disturbance characteristics such as amplitude difference, displacement difference, secondary frequency amplitude, inter-peak time and displacement residual with unified dimensions. The larger the value, the stronger the synchronicity between the dominant frequency energy fluctuation and the change in the geometric shape of the liquid surface in time and space within the section, while the smaller the value, the weaker the coordination between the two. This index can quantify the overall coupling level between the acoustic signature of the water flow and the disturbance of the liquid surface morphology, and provide a unified numerical basis for subsequent dynamic analysis and disturbance sequence construction based on acoustic signature.

[0101] The pairing and sequencing submodule is based on the perturbation coupling strength sequence. It calls the difference in the amplitude of the main frequency and the difference in the displacement of adjacent peaks, indexes and pairs the same perturbation coupling strength values ​​and arranges them according to the sampling time to obtain the pairing sequence and obtain the acoustic perturbation sequence.

[0102] Call the difference in the main frequency amplitude and the difference in the displacement of adjacent peaks, and use the same index... The amplitude difference is paired with the displacement difference, for example, in the segment. Pairing as Section Pairing as Arranged in chronological order as follows And bind the corresponding time series. This ensures that the amplitude and displacement correspondence of each pair of data are preserved in subsequent processing, ultimately yielding the acoustic perturbation sequence.

[0103] Please see Figure 4 The acceleration mapping module includes:

[0104] The boundary coordinate submodule extracts the boundary pixel coordinates of the mine fracture water flow based on the acoustic perturbation sequence, matches the boundary pixels of adjacent frames according to the time index, calculates the boundary pixel displacement velocity and compares the velocity distribution with the boundary gray-scale gradient distribution, and obtains the average boundary displacement velocity after removing velocity outliers.

[0105] When extracting the boundary pixel coordinates of mine fracture water flow based on acoustic perturbation sequences, acoustic wave acquisition equipment is first deployed at actual monitoring points in the mine to record continuous acoustic perturbation signals at a sampling interval of 0.01 seconds. This data is then converted into a two-dimensional pixel matrix. Boundary extraction is performed on the pixel matrices of adjacent frames, identifying the edge points of the water flow fractures through grayscale gradient detection. Next, the boundary pixels are matched one-to-one according to their timestamps to ensure that pixels at the same physical location can be matched at different times. After obtaining the correspondence, the displacement distance of the boundary pixels in adjacent frames is calculated. This displacement distance is determined by combining the pixel coordinate difference with the pixel's spatial ratio. For example, if the horizontal difference between two pixels is 3 pixels and the vertical difference is 4 pixels, then the spatial displacement is 5 pixel units. If each pixel represents 1 millimeter, then the displacement is 5 millimeters. The displacement is then divided by the time interval between two frames to obtain the boundary pixel velocity value. To ensure the stability of the result, the velocity of all pixels is compared with the grayscale gradient of the frame. The gradient threshold is set to 50, and pixels with a grayscale gradient below the threshold and drastic velocity changes are removed. For example, if a pixel has a gradient of 45 and a velocity greater than 200 millimeters per second, it is determined to be an outlier and deleted. Finally, the remaining valid velocity samples are weighted and averaged to obtain the average boundary displacement velocity.

[0106] The pixel acceleration submodule uses the boundary pixel correspondence as a reference, calls the average boundary displacement velocity as a benchmark, calculates the velocity difference of edge pixels in adjacent frames and obtains the edge acceleration, and obtains the edge acceleration value sequence after correction by combining the neighborhood velocity change and acoustic text disturbance noise.

[0107] Based on the correspondence of boundary pixels, the average boundary displacement velocity is used as the velocity benchmark. First, the velocity value of each edge pixel in adjacent frames is extracted, and the velocity difference is calculated. This velocity difference reflects the degree of local motion change of the pixel. For example, if the velocity of a pixel in the previous frame is 120 mm / s and in the current frame is 150 mm / s, the velocity difference is 30 mm / s. Then, the acceleration of the velocity difference is calculated over the inter-frame time interval. To eliminate the influence of neighborhood disturbances, the velocity change of other pixels within a 3×3 neighborhood of each pixel is extracted, and these changes are weighted according to distance, such as the weight of the nearest neighbor pixel. The weights are 0.6 for the second nearest neighbor, 0.3 for the second nearest neighbor, and 0.1 for the farthest neighbor. The weighted sum of the neighborhood velocity changes is added to the velocity difference of the pixel itself to obtain the corrected acceleration. Then, the influence of the acoustic disturbance noise is subtracted from the corrected acceleration. This noise is calculated from the energy of the collected background acoustic signal. Assuming that the background energy is 2.5 dB, the equivalent velocity influence is 5 mm / s², so it is directly subtracted from the corrected acceleration. Finally, the acceleration values ​​are normalized according to the grayscale gradient so that the edge acceleration under different brightness conditions can be compared. After all the processing is completed, an edge acceleration value sequence is formed.

[0108] The triplet mapping submodule calculates the maximum, minimum and average acceleration values ​​of frame edge pixels based on the edge acceleration value sequence, generates time series triplets and maps them to the time axis to obtain an acceleration change map.

[0109] Based on the edge acceleration value sequence, the acceleration values ​​of all edge pixels are first scanned in each frame to find the maximum, minimum, and average values. These three data points are then combined into a triplet. For example, if the maximum acceleration of a frame is 260 mm / s², the minimum is 20 mm / s², and the average is 145 mm / s², then the triplet for that frame is (260, 20, 145). Subsequently, combined with the frame's time index, the triplets of all frames are arranged in chronological order to form a time series set. This set is then mapped onto a time axis, with the horizontal axis representing time and the vertical axis representing the acceleration value change curve, to visually reflect the trend of edge acceleration over time, ultimately resulting in an acceleration change map.

[0110] Please see Figure 5 The morphological analysis module includes:

[0111] The time synchronization submodule performs time axis synchronization matching based on the acceleration change spectrum and the fracture water flow migration trajectory image, extracts the timestamp and amplitude pole sequence, calculates the difference between adjacent poles and generates a synchronization error sequence, compares the synchronization error sequence with the time difference benchmark value and filters the time period that meets the conditions, and generates the time alignment deviation interval.

[0112] Based on time-axis synchronization matching between acceleration variation maps and fracture flow migration trajectory images, the timestamps of each monitoring point in the acceleration variation map are first extracted and matched with the corresponding time information in the fracture flow migration trajectory image. For example, in a field monitoring, the sampling frequency of the acceleration variation map is 10Hz, and the acquisition frequency of the fracture flow migration trajectory image is 5Hz. Therefore, interpolation is used to align the two at the same time reference. Subsequently, amplitude extreme point sequences are extracted from the two matched sequences. For example, in the acceleration variation... Local maximum points were detected in the data, corresponding to times of 12.4s, 15.6s, and 18.9s, while the corresponding extreme times in the fracture water flow migration trajectory data were 12.5s, 15.7s, and 19.0s. The interval difference was obtained by subtracting the two sets of extreme times pairwise and accumulated to form a synchronization error sequence. This synchronization error sequence was compared with a time difference benchmark. The time difference benchmark was determined by statistically analyzing samples from different time periods in a single experiment, and taking the median error plus the standard deviation as the benchmark value. The statistical results of this experiment are shown in Table 3 below.

[0113]

[0114] As shown in Table 3, the reference value ranges from 0.17 to 0.19 s. In this embodiment, 0.20 s is selected as the final reference value. If several consecutive time differences in the candidate time period are less than or equal to the reference value, the time period is selected as the synchronization segment. The minimum error and the maximum error in the synchronization segment are recorded as feature values, and finally the time alignment deviation interval is generated.

[0115] The trend extraction submodule limits the time window based on the time alignment deviation interval, extracts the acceleration change increment sequence and the offset trajectory displacement increment sequence within the window, calculates the consistency score and filters the time period set, merges adjacent or overlapping time periods, and generates a consistent trend time period set.

[0116] The time window is defined based on the time alignment deviation interval. The acceleration change increment sequence and the displacement trajectory increment sequence are extracted from this window. For example, in an interval, the acceleration increment is [0.12, 0.18, -0.05, 0.20] m / s², and the displacement increment is [0.010, 0.015, -0.004, 0.018] m. The consistency of the two sets of data is calculated point by point according to the time correspondence to obtain the difference value at each moment, and the consistency score is calculated accordingly. The score threshold is determined by the 75th percentile value of the score distribution of the entire window. For example, in a batch of sample scores, the range is 0.35~0.92, and the 75th percentile value is 0.78. Then, the threshold is set to 0.78. Time periods below the threshold are filtered out, and time periods above or equal to the threshold are retained. Then, adjacent segments in the time period set with an interval of no more than 1 second are merged to enhance the continuity. The merged set forms a consistent trend time period set.

[0117] The boundary statistics submodule calls the consistent trend time period set to extract boundary frames, calculates the displacement difference of shape contour points and counts the range intervals, splices each range interval in time order and maps it to the pixel matrix to generate a shape change image.

[0118] The boundary frames of each segment are extracted by calling the consistent trend time period set. The first and last frames of each segment are matched with points. For example, the first frame contour points of a segment are [(12, 45), (15, 48), (20, 50)], and the last frame contour points are [(13, 46), (16, 50), (22, 53)]. The displacement difference of each corresponding point is calculated to obtain the displacement range of the segment, and it is divided into three levels: 0.5mm, 5mm and above. The ranges of all time periods are spliced ​​in chronological order to form a continuous time series. The time series is mapped into a two-dimensional pixel matrix and different ranges are represented by gray values. Finally, a morphological change image is generated.

[0119] Please see Figure 6 The risk output module includes:

[0120] The threshold determination submodule extracts the frame-by-frame morphological change amount and the same sampling rate voiceprint amplitude based on the morphological change image and the voiceprint perturbation sequence. It compares the morphological threshold and the voiceprint threshold to determine whether the dual channels exceed the threshold at the same moment. It calculates the ratio of the dual-channel threshold exceedance duration to the total duration to obtain the synchronous threshold exceedance ratio.

[0121] Based on morphological change images and acoustic perturbation sequences, frame-by-frame image data is first acquired from monitoring points during the synchronous acquisition phase and converted into an array of morphological change quantities. For example, one frame is acquired every second for 10 consecutive seconds. The morphological change quantity for each frame is obtained by summing the pixel grayscale differences between adjacent frames and taking the absolute value. Simultaneously, an acoustic perturbation sequence is acquired. The acoustic pressure signal from the underwater sensor is sampled, and the amplitude sequence is extracted as acoustic perturbation amplitude data. Subsequently, the morphological change quantity and acoustic perturbation amplitude at each time step are compared with preset morphological thresholds and acoustic perturbation thresholds, respectively. The thresholds are set with reference to the abrupt change critical data collected in the water flow surge experiment. For example, the morphological threshold is 2. The threshold for sound pressure level is 5 Pa. This value is derived from gradually increasing the flow rate in an experimental water tank and recording the changes in flow rate and sound pressure level. When both the change in flow rate and the sound pressure level reach or exceed this value, the water flow exhibits a sudden surge characteristic. The judgment criteria are that the change in flow rate ≥ 20 and the sound pressure level ≥ 5, which is considered a dual-channel over-threshold condition. The number of dual-channel over-threshold moments is counted across all frames of data, and the ratio of this count to the total number of sampled frames is used as the synchronous over-threshold ratio. For example, in 100 frames of data, if frames 2, 3, and 5 meet both conditions, the count is 3, and the synchronous over-threshold ratio is calculated as 3 / 100 = 0.03. The data example is as follows:

[0122]

[0123] The common frame screening submodule constructs a comprehensive score for each frame at the alignment time based on the proportion of synchronization exceeding the threshold. Combining the difference and synchronization between morphological changes and voiceprint amplitude, it locates candidate common frame segments and obtains the average score of common frames.

[0124] Based on the proportion of synchronous over-threshold, the morphological change amount and voiceprint amplitude are aligned by time index. A comprehensive score is calculated for each frame. The comprehensive score is composed of the absolute difference between the morphological change and the voiceprint amplitude, the dual-channel product, and the neighborhood smoothing value weighted. For example, in the 2nd frame, the morphological change amount is 22 and the voiceprint amplitude is 6. The difference between these values ​​and the thresholds 20 and 5 is 2 and 1, respectively, and the product is 132. The neighborhood smoothing value is the average of the morphological changes in the 1st to 3rd frames, which is (15+22+21) / 3=19.33, and the average of the voiceprint amplitude, which is (4+6+7) / 3=5.67. Substituting these values ​​into the weighted calculation, the comprehensive score is 85.4. The average score of all frames is obtained by traversing all frames and taking the average score. For example, the average value of 100 frames is 82.6. Frames with values ​​higher than this average value are marked as candidate common frame segments. In the example data, the 2nd, 3rd, and 5th frames fall into the segment, and the common frame score average is obtained.

[0125] The marker set construction submodule calls the common frame score mean and synchronous over-threshold ratio, filters the dual-channel over-threshold image frames in the same segment, and superimposes the timestamp and marker number on the frames. It then aggregates the frame sequence and time index to obtain the sudden water inrush risk image set.

[0126] The average score of common frames and the proportion of synchronous over-threshold are used to filter dual-channel over-threshold image frames within the candidate common frame segment. Each frame is marked with a timestamp and a tag number. For example, the second frame is marked as "T02-1", the third frame as "T03-2", and the fifth frame as "T05-3". These image frames are aggregated in chronological order to form a continuous sequence. At the same time, their time index intervals, such as [2,3] and [5,5], are recorded. These marked frame sequences are merged to form a sudden water inrush risk image set. In the example data above, the image set consists of three frames corresponding to time indices 2, 3, and 5, and the final sudden water inrush risk image set is obtained.

[0127] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A mine water inrush image processing system based on AI recognition, characterized in that, The system includes: The fracture water flow tracking module acquires the pixel coordinates of the boundary of the mine fracture water flow in continuous frame images, calculates the gray-level gradient magnitude and gradient direction angle of the boundary pixels, calculates the difference of gray-level gradient magnitude and gradient direction angle of pixels at the same position between adjacent frames, combines them to form a set of micro-displacement vectors, arranges them in time order to statistically distribute the vector direction changes, and generates a water flow offset image. The acoustic disturbance extraction module collects the amplitude values ​​of the main frequency and secondary frequency of the water flow acoustic pattern based on the water flow offset image, calculates the difference in amplitude of the main frequency, extracts the sequence of liquid surface disturbance wave peak positions from the image and calculates the displacement difference between adjacent wave peaks to generate an acoustic disturbance sequence. The acceleration mapping module extracts the pixel coordinates of the mine fracture water flow boundary according to the acoustic perturbation sequence, calculates the pixel displacement velocity of adjacent frames and obtains the pixel edge acceleration value, and summarizes the maximum, minimum and average values ​​to form an acceleration triplet to generate an acceleration change map. The morphological analysis module calls the acceleration change map and the water flow migration image to perform time axis synchronization matching, extracts time periods with consistent trends, counts the morphological change range at the boundaries of the time periods, and generates a morphological change image.

2. The mine water inrush image processing system based on AI recognition according to claim 1, characterized in that, The water flow migration image includes vector direction change distribution, a set of micro-displacement vectors arranged in time sequence, and pixel coordinates of the boundary of mine fracture water flow. The acoustic disturbance sequence includes the main frequency amplitude difference, peak displacement difference, and liquid surface disturbance peak position sequence. The acceleration change spectrum includes the edge acceleration change curve mapped by the time axis and the edge acceleration triplet arranged in time sequence. The morphological change image includes a consistent trend time period and the boundary morphological change range. The grayscale gradient magnitude refers to the rate of change of pixel grayscale at spatial location; The gradient direction angle represents the direction of pixel grayscale change; The dominant frequency and secondary frequency of the water flow acoustic pattern refer to the two frequencies with the highest amplitude in the spectral components of the water flow acoustic wave signal, which are calculated by fast Fourier transform and are measured in Hertz. The position of the liquid surface disturbance peak can be determined by analyzing the liquid surface brightness curve frame by frame, and the displacement difference refers to the difference in pixel coordinates between the positions of consecutive peaks. The pixel displacement speed is the amount of displacement of a pixel per unit time, which is obtained by dividing the difference in pixel position between adjacent frames by the inter-frame time. The pixel edge acceleration value is the time rate of change of the pixel displacement velocity; The edge acceleration triplet consists of the maximum acceleration value, the minimum acceleration value, and the average acceleration value; The range of boundary morphology changes refers to the spatial positional change interval of the fractured water flow boundary over time, calculated by the difference between the maximum and minimum offsets of the boundary coordinates.

3. The mine water inrush image processing system based on AI recognition according to claim 1, characterized in that, The fissure water flow tracking module includes: The boundary localization submodule obtains the corresponding grayscale value based on the pixel coordinates of the boundary of the mine fracture water flow in the continuous frame image, calculates the grayscale gradient magnitude and gradient direction angle of each boundary pixel, calculates the overall gradient magnitude feature of the boundary pixels in the same frame, and obtains the average boundary gradient magnitude. The micro-vector construction submodule calculates the time sequence difference based on the difference in gray-level gradient magnitude and gradient direction angle of the boundary pixels at the same position between adjacent frames, accumulates them by pixel index, obtains the vector direction variability, sorts them according to the variability value and records the time index, and generates the sequence identifier of the raster mapping. The offset imaging submodule, based on the sequence identifier of the raster mapping, uses the vector direction variability corresponding to the sequence identifier as the weight for each boundary pixel index, accumulates the gray-level gradient direction angle difference between adjacent frames and maps it to the raster, and outputs the displacement visualization matrix according to the correspondence between the mapped raster and the boundary pixel coordinates to generate a water flow offset image.

4. The mine water inrush image processing system based on AI recognition according to claim 1, characterized in that, The voiceprint disturbance extraction module includes: The acoustic amplitude acquisition submodule acquires the amplitude values ​​of the main frequency and secondary frequency of the water flow acoustic pattern based on the water flow offset image, calculates the difference in the amplitude of the main frequency, and arranges them according to the sampling time to obtain the difference in the amplitude of the main frequency. The peak distance calculation submodule calculates the displacement difference between adjacent peaks and the time between peaks based on the difference in amplitude of the main frequency and the position sequence of the disturbance wave peaks on the liquid surface. It calculates the reference values ​​of amplitude and displacement based on the same segment of data, obtains the disturbance coupling strength value, and indexes and collects them into a disturbance coupling strength sequence. The pairing and sequencing submodule, based on the disturbance coupling strength sequence, calls the difference in the main frequency amplitude and the difference in the displacement of adjacent peaks, indexes and pairs the same disturbance coupling strength values ​​and arranges them according to the sampling time to obtain the pairing sequence and obtain the acoustic perturbation sequence.

5. The mine water inrush image processing system based on AI recognition according to claim 1, characterized in that, The acceleration mapping module includes: The boundary coordinate submodule extracts the boundary pixel coordinates of the mine fracture water flow according to the acoustic perturbation sequence, matches the boundary pixels of adjacent frames according to the time index, calculates the boundary pixel displacement velocity and compares the velocity distribution with the boundary gray-scale gradient distribution to obtain the average boundary displacement velocity. Based on the boundary pixel correspondence, the pixel acceleration submodule uses the average boundary displacement velocity as a benchmark to calculate the velocity difference of adjacent frame edge pixels and obtain the edge acceleration. After correction by combining the neighborhood velocity change and the acoustic text disturbance noise, the edge acceleration value sequence is obtained. The triplet mapping submodule calculates the maximum, minimum, and average acceleration values ​​of frame edge pixels based on the edge acceleration value sequence, generates time series triplets, maps them to the time axis, and obtains an acceleration change spectrum.

6. The mine water inrush image processing system based on AI recognition according to claim 1, characterized in that, The morphological analysis module includes: The time synchronization submodule performs time axis synchronization matching based on the acceleration change map and the fracture water flow migration trajectory image, extracts the timestamp and amplitude pole sequence, calculates the difference between adjacent poles and generates a synchronization error sequence, compares the synchronization error sequence with the time difference benchmark value and filters the time period that meets the conditions, and generates a time alignment deviation interval. The trend extraction submodule defines a time window based on the time alignment deviation interval, extracts the acceleration change increment sequence and the offset trajectory displacement increment sequence within the window, calculates the consistency score and filters the time period set, merges adjacent or overlapping time periods, and generates a consistent trend time period set. The boundary statistics submodule calls the consistent trend time period set to extract boundary frames, calculates the displacement difference of morphological contour points and counts the range intervals, splices each range interval in time order and maps it to the pixel matrix to generate a morphological change image.

7. The mine water inrush image processing system based on AI recognition according to claim 1, characterized in that, The system also includes: The risk output module determines the risk threshold for sudden water inrush based on the morphological change image and the acoustic perturbation sequence, extracts and marks image frames in which the two result values ​​both exceed the threshold within the same time period, and generates a sudden water inrush risk image set. The surge water risk image set includes marked surge water risk image frames; The threshold for judging the risk of sudden water inrush is set based on historical statistics of sudden water inrush in the area where the mine is located and national coal mine safety technical standards. The judgment criteria are that both the change rate of acoustic amplitude and the change range of boundary morphology exceed the limit simultaneously.

8. The mine water inrush image processing system based on AI recognition according to claim 7, characterized in that, The risk output module includes: The threshold determination submodule extracts the frame-by-frame morphological change amount and the same sampling rate voiceprint amplitude based on the morphological change image and the voiceprint perturbation sequence, compares the morphological threshold and the voiceprint threshold, determines whether the dual-channel exceeds the threshold at the same moment, and calculates the ratio of the dual-channel exceeding the threshold duration to the total duration to obtain the synchronous exceeding the threshold ratio. The common frame screening submodule constructs a comprehensive score for each frame at the alignment time based on the synchronization over-threshold ratio. Combining the difference and synchronization between the morphological change amount and the voiceprint amplitude, it locates candidate common frame segments and obtains the average common frame score. The marker set construction submodule calls the common frame score mean and the synchronous over-threshold ratio, filters the dual-channel over-threshold image frames in the same segment, and superimposes the timestamp and marker number on the frames. It then aggregates the frame sequence and time index to obtain the sudden water inrush risk image set.

Images