Machine vision-based ventilator waveform anomaly monitoring method

Abstract

The present application relates to the medical equipment waveform monitoring technical field, especially ventilator waveform abnormality monitoring method based on machine vision, through the acquisition of ventilator display screen multiple frames of continuous video frames, the video frame is projected along the longitudinal direction to generate a longitudinal pixel distribution curve, and the waveform pixel contour trajectory is extracted based on the wave trough position cutting to obtain an independent waveform channel; In the effective matching extraction area, the feature points are extracted and the inter-frame matching is completed, the waveform transverse rolling pixel displacement amount is calculated, the continuous waveform curve covering multiple complete breathing cycles is spliced to generate, the breathing cycle point is calibrated through the flow rate and pressure waveform cross verification, and the inspiration phase and expiration phase phase interval are divided, the phase constraint and multi-cycle rhythm consistency verification are combined, the waveform abnormality corresponding to invalid trigger and reverse trigger is identified, and the monitoring result is output. The present application is adapted to the rolling display characteristics of the ventilator, improves the identification accuracy of weak feature abnormalities, and can assist in reducing the risk of ventilation-related complications.

Description

Technical Field

[0001] This invention relates to the field of waveform monitoring technology for medical devices, and in particular to a method for monitoring abnormal waveforms in ventilators based on machine vision. Background Technology

[0002] With the rapid development of machine vision technology, the identification of abnormal waveforms on ventilator screens through non-invasive image acquisition has been gradually applied in the medical field. Because ventilator screens generally use a horizontal scrolling refresh display mechanism, the waveform scrolls along the horizontal axis over time. Waveform segments in adjacent frames can be displaced, truncated, or overlapped, resulting in temporal breaks. For common and high-risk waveform abnormalities in clinical practice (such as invalid triggering or reverse triggering), their determination relies on the temporal context and rhythmic characteristics of multiple consecutive respiratory cycles. A single frame image may only show extremely weak curve dips or shifts. Existing technologies often employ end-to-end monitoring methods for single frames, which, due to the loss of temporal context information, result in a very high rate of missed detection for these weak abnormalities. Directly using general video temporal analysis models cannot adapt to the special display logic of ventilator waveform scrolling refresh. Spatiotemporal alignment of waveforms between frames is difficult and easily introduces additional temporal noise, which can actually increase the false detection rate of abnormalities and raise the risk of ventilation-related complications for patients.

[0003] Therefore, it is necessary to propose a machine vision-based method for monitoring abnormal waveforms in ventilators to solve the above problems. Summary of the Invention

[0004] The purpose of this invention is to provide a machine vision-based method for monitoring abnormal ventilator waveforms, in order to solve the problem of increased false detection rate caused by the loss of timing context information and the special display logic of ventilator waveform scrolling refresh.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A machine vision-based method for monitoring abnormal waveforms in ventilators includes the following steps: S1: Acquire video frames corresponding to multiple consecutive screen images on the ventilator display screen; S2: Project the video frame horizontally along the longitudinal direction to generate a longitudinal pixel distribution curve. Based on the trough position in the longitudinal pixel distribution curve, divide the image horizontally to segment independent waveform channels in which physiological parameters change over time, and extract the waveform pixel contour trajectory within the waveform channel. S3: Along the horizontal time axis of the waveform channel, perform waveform integrity verification from the left and right screen boundaries toward the center of the image. Remove the waveform generation section on the right that has not formed a complete cycle shape and the waveform truncated section on the left that has been truncated. Define the remaining middle section as the effective matching extraction area. Extract feature points from the waveform pixel contour trajectory corresponding to the effective matching extraction area. Perform feature point matching in the same waveform channel in adjacent video frames to calculate the pixel displacement of the waveform horizontal scrolling. Then, stitch together the video frames according to the pixel displacement to generate a continuous waveform curve covering multiple complete breathing cycles. S4: On the waveform curve, the cycle point corresponding to the respiratory cycle is determined by cross-verification of the flow-time waveform curve and the pressure-time waveform curve. The time interval of the inspiratory phase and the expiratory phase is obtained by dividing the cycle point. Weak characteristic signals are monitored in the corresponding time interval. Combined with the rhythm consistency verification of multiple consecutive respiratory cycles, waveform abnormalities corresponding to invalid triggering and reverse triggering are identified. S5: Based on the waveform anomalies, the type of anomalies is labeled and the time of occurrence is located to obtain the monitoring results.

[0006] Preferably, the step of segmenting video frames to obtain waveform channels includes: For each line of the video frame, the total number of waveform pixels with grayscale values ​​lower than the background threshold is counted row by row to generate a vertical pixel distribution curve; Local minima are detected on the vertical pixel distribution curve. Vertical coordinates that simultaneously satisfy the local minimum of a 3-pixel neighborhood and whose projected value is less than 10% of the maximum projected value are identified as troughs. Based on the trough position, the image within the video frame is segmented along the horizontal direction, and the segmented image is subjected to continuous curve pixel trajectory detection to obtain waveform channels corresponding to different physiological parameters such as pressure, flow rate, and volume.

[0007] Preferably, the extracted waveform pixel contour trajectory needs to meet the following conditions, including: Extending continuously along the horizontal time axis of the waveform channel, the length of the horizontal breakpoint pixels within a single trajectory does not exceed the preset breakpoint threshold. The vertical pixel coordinates of the trajectory are all within the effective display range of the waveform channel; wherein, the effective display range is the area after the vertical upper and lower boundaries of the waveform channel are offset inward by a preset pixel margin, corresponding to the normal display range of respiratory physiological parameters.

[0008] Preferably, the feature point extraction step includes: On the waveform pixel contour trajectory corresponding to the effective matching extraction area of ​​the waveform channel, local amplitude extreme points and slope inflection points are extracted as candidate feature points. Among the candidate feature points, the candidate feature points with a horizontal pixel spacing greater than a preset spacing threshold and an amplitude difference greater than a preset fluctuation threshold are retained to obtain the feature points.

[0009] Preferably, the step of delineating the effective matching extraction region includes: Along the horizontal time axis of the waveform channel, with a step size of 1 pixel, the waveform contour is traversed from the right boundary to the left. When a complete set of peak-valley extreme pairs is detected for the first time and the amplitude difference is not less than 10% of the effective longitudinal range of the channel, the corresponding horizontal coordinate is recorded as the left boundary of the waveform generation segment, and the right generation segment is removed. Along the horizontal time axis of the waveform channel, with a step size of 1 pixel, traverse the waveform contour from the left boundary to the right. When a complete set of peak-valley extreme pairs is detected for the first time and the amplitude difference is not less than 10% of the effective longitudinal range of the channel, record the corresponding horizontal coordinate as the right boundary of the waveform truncation segment and remove the left truncation segment. The middle waveform segment between the left and right boundaries is defined as the effective matching extraction area.

[0010] Preferably, the step of calibrating the periodic point includes: The waveform curves of physiological parameters changing over time within the waveform channel include flow velocity-time waveform curves and pressure-time waveform curves. In the flow-time waveform curve, the first zero-crossing point where the flow rate crosses zero from the baseline in the positive direction is marked as the candidate point for the start of inspiration, the zero-crossing point where the flow rate crosses zero from the positive to the negative direction is marked as the candidate point for the start of expiration, and the point where the flow rate returns to the baseline and remains stable is marked as the candidate point for the end of expiration. Each candidate point marked in the flow-time waveform curve is mapped to the pressure-time waveform curve to verify the physiological linkage characteristics of pressure rise during inhalation, pressure decrease during exhalation, and pressure stabilization at the end of exhalation. The cycle point is obtained through cross-validation of the two waveforms.

[0011] Preferably, the waveform curves showing the changes in physiological parameters over time within the waveform channel also include a capacity-time waveform curve. After cross-validation to generate periodic points, the periodic points are mapped onto the capacity-time waveform curve based on the unified time axis of the waveform channel.

[0012] Preferably, the step of identifying waveform anomalies corresponding to invalid triggers and reverse triggers includes: The weak characteristic signals include pressure waveform indentation signals and flow velocity waveform offset signals; When the amplitude decrease of the pressure waveform during the expiratory phase is within the preset micro-threshold range and is not accompanied by a concave shape of the flow waveform crossing zero in the positive direction, it is marked as a single invalid trigger candidate event. When the flow rate waveform exhibits a second positive fluctuation during the transition phase from the end of the inspiratory phase to the early expiratory phase, and the pressure waveform simultaneously shows a corresponding amplitude disturbance, it is marked as a single reverse trigger candidate event.

[0013] Preferably, the feature point matching and waveform horizontal scrolling pixel displacement calculation steps for adjacent video frames include: Centered on each feature point, the vertical coordinates of continuous waveform points with dimensions of 10 pixels to the left and right are used to form a one-dimensional description vector with dimension 21. Calculate the Euclidean distance between the feature point description vectors of frame k and frame (k+1), and take the pair with the smallest distance as the initial matching pair; For the initial matching pairs, with the constraint that the ventilator waveform only undergoes lateral rigid body translation, mismatched pairs with the absolute value of the longitudinal coordinate deviation of the corresponding feature points in the two frames being greater than a preset deviation threshold are eliminated to obtain a set of valid matching pairs; Calculate the horizontal pixel coordinate difference for all valid matching pairs, and take the median of all differences as the horizontal scroll pixel displacement between adjacent frames.

[0014] Preferred steps for verifying rhythm consistency across multiple consecutive respiratory cycles include: The preset number of consecutive verification cycles is 3, and the frequency threshold is 2. If the same type of candidate event occurs more than or equal to 2 times within 3 consecutive respiratory cycles, and the relative position deviation of each event within the corresponding time phase does not exceed 10% of the total duration of the cycle, it is determined to be a valid waveform abnormality of the corresponding type. If a candidate event occurs in only a single cycle, it is considered occasional noise and no anomaly flag is triggered.

[0015] The technical effects and advantages of the present invention in the above technical solution are as follows: 1. This invention extracts feature points of waveform pixel contours and performs inter-frame matching to calculate the pixel displacement of the waveform during horizontal scrolling. This is used to align and stitch multiple consecutive video frames to generate a continuous waveform curve covering multiple complete respiratory cycles. At the same time, considering the scrolling refresh characteristics of the ventilator, which generates on the right side and moves out on the left, the invention defines an effective matching extraction area to remove invalid waveforms at the edges, thereby improving the accuracy of inter-frame displacement calculation.

[0016] 2. By calibrating the respiratory cycle points to divide the inspiratory and expiratory phases, the detection range of weak feature abnormalities such as invalid triggers and reverse triggers is constrained within the corresponding physiological phases. Combined with dual waveform morphology linkage judgment and continuous multi-cycle rhythm consistency verification, this method solves the technical problems of high false negative rate of existing single-frame end-to-end methods for weak feature abnormalities and high false positive rate caused by the introduction of temporal noise into general time series models. It can reliably identify clinically common invalid trigger and reverse trigger events.

[0017] 3. By accurately identifying and pinpointing invalid and reverse triggers, quantitative monitoring data is provided to medical staff, reducing the risk of ventilation-related complications such as respiratory muscle fatigue and ventilator-associated lung injury. Attached Figure Description

[0018] Figure 1 This is a flowchart of the machine vision-based ventilator waveform anomaly monitoring method of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1 As shown, this embodiment provides a machine vision-based method for monitoring abnormal waveforms in ventilators, including the following steps: To obtain continuous image data containing only valid waveform regions and improve the accuracy of rolling displacement calculation and time-series stitching in subsequent methods, video frames corresponding to multiple consecutive screen images on the ventilator display are acquired in step S1; this is specifically implemented through the following method: In this embodiment, the hardware and parameter configuration for image acquisition specifically employs a visible light industrial area scan camera as the acquisition terminal. This terminal is fixedly mounted on a bracket 0.3–0.8m directly in front of the ventilator's display screen. By adjusting the lens focal length and pitch angle, the camera's field of view completely covers the ventilator's waveform display screen. The camera's operating parameters are set to an effective pixel resolution of 1920×1080 pixels and a frame rate of 30fps. The fixed time interval between two adjacent frames is... ms, To achieve the desired frame rate, the exposure time was set to 1 / 120s to match the refresh rate of most ventilator screens. The screen image output format was uniformly converted to an 8-bit single-channel grayscale image with 256 grayscale levels, which reduced the computational load of subsequent image processing while preserving waveform contour information. For screen effective area correction and cropping, specifically, the effective waveform area of ​​the display screen is extracted from the original image output by the camera. The Canny edge detection operator is used to extract the pixel edges in the image. The Hough line transform is used to filter out the four straight lines that constitute the outer boundary of the quadrilateral of the display screen. Then, the pixel coordinates of the four corner points of the display screen are obtained. Perspective transformation correction is performed based on the coordinates of the four corner points to eliminate the trapezoidal distortion caused by the tilted installation of the camera, resulting in a standard rectangular image. Then, the physical border of the screen, the numerical parameter display area, the alarm indicator area, the mode button area, etc. are removed from the rectangular image by cropping. The display area where the three waveform channels of pressure, flow rate, and capacity are arranged vertically is retained. The horizontal pixel resolution of the effective waveform area after cropping is not less than 1600 pixels, and the vertical pixel resolution is not less than 600 pixels.

[0021] Based on the chronological order of acquisition, each corrected and cropped image is assigned a frame number to obtain a continuous video frame sequence. For the first... The acquisition timestamp corresponding to the frame image is obtained through... The calculation yielded the result, where, For frame number, For the first The acquisition time corresponding to the frame It is a fixed time interval between two adjacent frames.

[0022] To achieve spatial separation of waveforms with different physiological parameters, obtain continuous and accurate single-channel waveform pixel contour trajectories, and avoid cross-channel waveform interference, in step S2, the video frame is horizontally projected along the longitudinal direction to generate a longitudinal pixel distribution curve. Based on the trough positions in the longitudinal pixel distribution curve, the image is horizontally segmented to divide independent waveform channels where physiological parameters change over time, and the waveform pixel contour trajectories within the waveform channels are extracted. This is specifically achieved through the following method: In one embodiment of the present invention, the step of segmenting a video frame to obtain waveform channels includes: S21: Count the total number of waveform pixels with gray values ​​lower than the background threshold line by line along the vertical axis of the video frame to generate a vertical pixel distribution curve; S22: Perform local minima detection on the vertical pixel distribution curve, and determine the vertical coordinates that simultaneously satisfy the local minima of a 3-pixel neighborhood and the projection value is less than 10% of the maximum projection value as troughs; S23: Based on the trough position, the image within the video frame is segmented along the horizontal direction, and the segmented image is subjected to continuous curve pixel trajectory detection to obtain waveform channels corresponding to different physiological parameters such as pressure, flow rate, and volume.

[0023] According to steps S21 to S23 above, for the corrected and cropped single-frame grayscale effective waveform image, horizontal pixel projection statistics are performed row by row along the vertical pixel coordinates. Specifically, the generation process for the vertical pixel distribution curve includes: assuming the horizontal pixel coordinates of the image are... The range of values ​​is , The horizontal total number of pixels in the effective area is set to 1600, and the vertical pixel coordinates are set to... value range , The effective area has a total vertical pixel count, with a value of 600, and the image grayscale value is... Calculate the background grayscale threshold A vertically segmented adaptive background threshold calibration method is adopted. Specifically, the effective waveform image is uniformly divided into three segments along the vertical direction: upper, middle, and lower. Within each segment, 15-pixel edge blank areas are selected on both the left and right sides. The gray-level mean value of each area is calculated as the local background reference value for the corresponding segment. Based on the local background reference values ​​of the three segments, a first-order linear fit is performed to obtain the background gray-level baseline function distributed along the vertical direction. For each vertical row y, a tolerance of 5 gray levels is superimposed on the corresponding baseline value to obtain the waveform pixel judgment threshold for that row. Then, for the th row... Horizontal projection value of the row The total number of waveform pixels in a row whose grayscale value is lower than the corresponding row's threshold is counted using the following formula: ; in, This is an indicator function that takes a value of 1 when a condition is met and a value of 0 otherwise. After traversing all vertical rows, a length of [length missing] is obtained. The vertical pixel distribution curve has peaks corresponding to the vertical center of the waveform curve and troughs corresponding to the blank separation bands between channels.

[0024] Local minima detection is performed on the vertical pixel distribution curve, and the vertical coordinates that simultaneously satisfy the following three conditions are identified. Determined as a trough: (1) This point is a local minimum within a 3-pixel neighborhood, that is and ; (2) The projection value satisfies ,in, This is the maximum projection value of the vertical pixel distribution curve, ensuring that this position is a blank separation area between channels; (3) The longitudinal spacing between adjacent candidate valley points deviates from the preset nominal channel spacing by no more than ±15%; wherein, the preset nominal channel spacing is set to approximately 200 pixels based on the prior design of the mainstream ventilator screen layout, in order to eliminate false valleys caused by local noise and ensure that the valleys correspond to the real channel separation zone.

[0025] Based on the ventilator screen layout, two troughs can be detected, which are denoted from top to bottom as follows: , Using this as the dividing boundary, the effective waveform image is divided into three independent waveform channel sub-images along the horizontal direction: Pressure-time waveform channel: Longitudinal range is ; Flow velocity-time waveform channel: Longitudinal range is ; Capacity-time waveform channel: Vertical range is ; It is important to note that a 1-pixel margin should be retained during segmentation to prevent the waveform edges of adjacent channels from blending into each other.

[0026] For each segmented channel sub-image, candidate curve trajectories are extracted using eight-neighbor connectivity detection. Each trajectory is then validated and ranked for quality. Specifically, for trajectory consistency validation, the horizontal direction is divided into intervals with a step size of 100 pixels. The offset of the longitudinal coordinate of the trajectory within each interval is calculated. If the drift in any interval exceeds 20% of the effective longitudinal range of the channel, it is determined to be an interfering trajectory and is removed. For trajectory quality ranking, the candidate trajectories that pass the validation are comprehensively scored according to the horizontal extension length (weight 60%), the number of breakpoints (weight 20%), and the longitudinal coverage and channel range matching degree (weight 20%). The trajectory with the highest score is selected as the main waveform trajectory of the channel, forming waveform channels corresponding to different physiological parameters such as pressure, flow rate, and volume.

[0027] In one embodiment of the present invention, the extracted waveform pixel contour trajectory needs to meet the following conditions: Extending continuously along the horizontal time axis of the waveform channel, the length of the horizontal breakpoint pixels within a single trajectory does not exceed the preset breakpoint threshold. The vertical pixel coordinates of the trajectory are all within the effective display range of the waveform channel; wherein, the effective display range is the area after the vertical upper and lower boundaries of the waveform channel are offset inward by a preset pixel margin, corresponding to the normal display range of respiratory physiological parameters.

[0028] In this embodiment of the invention, for continuity verification, all pixels on the trajectory are traversed along the horizontal time axis of the waveform channel, and the horizontal pixel length of consecutive missing trajectory points, i.e., the breakpoint length, is counted. The preset breakpoint threshold is 5 pixels, and the breakpoints are processed in stages according to their size. Specifically: (1) If the length of the breakpoint is ≤3 pixels, it is determined to be a minor breakpoint caused by isolated pixel noise. The trajectory is directly completed by the linear difference between the adjacent pixels on both sides of the breakpoint and is not included in the statistics of valid breakpoints. (2) If 3 pixels < breakpoint length ≤ 5 pixels, it is determined to be a slight breakpoint caused by screen scan line interference. The trajectory segment is completed by fitting a quadratic polynomial with 5 trajectory points before and after the breakpoint. The completed segment is marked as the repair segment. (3) If the breakpoint length is greater than 5 pixels, it is determined that the trajectory continuity is broken. The candidate trajectory is directly removed and the trajectory is extracted again.

[0029] For valid physiological range verification, the effective display range of the preset channel is determined by statistically analyzing the vertical pixel coordinates of all pixels in the trajectory. ,in, This represents the vertical coordinate after the lower boundary of the channel is offset upwards by 3 pixels. The vertical coordinate is the vertical coordinate after shifting downwards by 3 pixels from the upper boundary of the channel. In this embodiment, the preset pixel margin is 3 pixels, corresponding to the normal display range of respiratory physiological parameters, to perform a two-level judgment, including: (1) Boundary compliance judgment: if all vertical pixel coordinates of the trajectory are within the boundary... Within the interval, the trajectory is directly determined as valid through interval verification; if there are a few out-of-bounds pixels, and the continuous length of the out-of-bounds points is ≤2 pixels and the vertical deviation is ≤3 pixels, it is determined as edge noise interference, and the out-of-bounds points are cropped to the interval boundary and retained; if the continuous length of the out-of-bounds points is >2 pixels or the vertical deviation is >3 pixels, it is determined as a cross-channel interference trajectory and is removed. (2) Determine the reasonableness of the amplitude by calculating the range of the longitudinal coordinates of the trajectory (the difference between the peak value and the valley value). If the range is within 15% to 90% of the effective longitudinal range of the channel, it is determined to be reasonable.

[0030] It should be understood that in this embodiment, the effective longitudinal range of the channel refers to the complete pixel span between the upper and lower boundaries of a single waveform channel. The effective display range is the area obtained by offsetting a preset pixel margin inward based on the effective longitudinal range of the channel, which is used to eliminate noise interference at the channel edge and corresponds to the normal display range of respiratory physiological parameters.

[0031] After the above two conditions are verified, the waveform pixel contour trajectory is obtained.

[0032] It should be understood that, considering the continuous and smooth characteristics of the physiological waveform of the ventilator, this embodiment further verifies the morphology of the physiological waveform, including local slope constraints, length proportion constraints, and no-crossing constraints. Specifically, the local slope constraint is the first-order slope of the trajectory calculated in a 5-pixel sliding window along the horizontal direction. The absolute value of the slope within any window does not exceed the preset upper limit of 0.3, so as to exclude non-physiological morphological trajectories with right-angle turns. The length proportion constraint is that the total horizontal extension length of the trajectory is not less than 70% of the horizontal length of the channel, so as to exclude short-segment pseudo-trajectories existing in local areas. The no-crossing constraint is that the candidate trajectory does not have any horizontal intersection points with other candidate trajectories in the same channel, so as to exclude the interference of multiple trajectory entanglement.

[0033] Considering that the ventilator screen uses a horizontal scrolling refresh mechanism with right-side generation, left-side removal, and pure translation in the middle, if feature point matching is performed on the entire waveform area, incomplete drawing of the edge areas and truncated waveforms will lead to insufficient accuracy in scrolling displacement calculation, resulting in timing splicing misalignment and long waveform timing breaks. Therefore, in step S3, along the horizontal time axis of the waveform channel, waveform integrity verification is performed from the left and right screen boundaries towards the center of the image. Waveform generation segments that have not formed a complete cycle on the right and waveform truncated segments that have been truncated on the left are removed. The remaining middle segment is designated as the effective matching extraction area. Feature points are extracted from the waveform pixel contour trajectory corresponding to the effective matching extraction area. Feature point matching is performed on the same waveform channel in adjacent video frames to calculate the pixel displacement of the horizontal scrolling of the waveform. Video frames are then spliced ​​based on the pixel displacement to generate a continuous waveform curve covering multiple complete respiratory cycles. This is specifically implemented through the following method: In one embodiment of the present invention, the step of delineating the effective matching extraction region includes: S31: Along the horizontal time axis of the waveform channel, with a step size of 1 pixel, traverse the waveform contour from the right boundary to the left. When a complete set of peak-valley extreme pairs is detected for the first time and the amplitude difference is not less than 10% of the effective longitudinal range of the channel, record the corresponding horizontal coordinate as the left boundary of the waveform generation section and remove the right generation section. S32: Along the horizontal time axis of the waveform channel, with a step size of 1 pixel, traverse the waveform contour from the left boundary to the right. When a complete set of peak-valley extreme pairs is detected for the first time and the amplitude difference is not less than 10% of the effective longitudinal range of the channel, record the corresponding horizontal coordinate as the right boundary of the waveform truncation segment and remove the left truncation segment. S33: Define the middle waveform segment between the left and right boundaries as the effective matching extraction area.

[0034] As shown in steps S31 to S33 above, along the horizontal time axis of the waveform channel, with a step size of 1 pixel, waveform integrity verification is performed from the left and right screen boundaries towards the center of the image. Specifically, let the vertical pixel coordinates of the waveform outline be... , The horizontal pixel coordinates are used. Since the three waveform channels are arranged vertically and fill the entire effective area horizontally, the total horizontal width of a single waveform channel is equal to the total number of horizontal pixels in the effective area in the above embodiment. Pixel; from the right boundary ( Traverse the waveform profile to the left, counting the number of local extrema along the traversal path. When a complete peak-trough extremum pair is detected for the first time, and the longitudinal amplitude difference of the extremum pair is not less than 10% of the effective longitudinal range of the channel, determine the lateral coordinate corresponding to the extremum pair. The left boundary of the waveform generation segment, and the right screen boundary to... The sections between these points are waveform generation sections that do not form a complete cycle and are discarded. The effective longitudinal range of a channel is the pixel difference between its upper and lower boundaries. The range for pressure and flow velocity channels is approximately 200 pixels, corresponding to a 10% amplitude difference threshold of 20 pixels. For the left-side waveform truncation section, the range is calculated from the left boundary (…). Traverse the waveform profile to the right, using the same judgment rules as the right side. When a complete peak-trough extreme value pair is detected for the first time, record the corresponding horizontal coordinate. The right boundary of the waveform truncation segment, and the left screen boundary to... The segments between these points are the waveform truncation segments that have already undergone boundary truncation, and these segments are then removed; The middle waveform segment is defined as the effective matching extraction area.

[0035] In one embodiment of the present invention, the feature point extraction step includes: S34: On the waveform pixel contour trajectory corresponding to the effective matching extraction area of ​​the waveform channel, extract local amplitude extreme points and slope inflection points as candidate feature points; S35: Among the candidate feature points, retain the candidate feature points whose horizontal pixel spacing is greater than a preset spacing threshold and whose amplitude difference is greater than a preset fluctuation threshold, and obtain the feature points.

[0036] As shown in steps S34 and S35 above, for candidate feature points, on the waveform pixel contour trajectory corresponding to the effective matching extraction area, candidate feature points are extracted, including local amplitude extrema points and slope inflection points. Specifically, for amplitude extrema points, a neighborhood window of 3 pixels in the horizontal coordinate is used. If the vertical coordinate of a point is the maximum value in the neighborhood, it is marked as a peak point; if it is the minimum value in the neighborhood, it is marked as a trough point. For slope inflection points, the first-order slope of the waveform contour is calculated point by point. ,in, These are the horizontal pixel coordinates of points on the waveform contour. For the horizontal coordinate The function representing the horizontal coordinate. The vertical pixel coordinates corresponding to the waveform (excluding product operations). , They are respectively The vertical coordinates of adjacent pixels on the right and left sides, and the denominator 2 corresponding to the horizontal distance between two adjacent pixels, are calculated to obtain... for The first-order slope of the waveform reflects the ratio of the longitudinal change to the transverse change. When the sign of the slope changes between two adjacent points, and the absolute value of the slope change is greater than a preset slope threshold... At that point, the slope inflection point is marked, corresponding to the location where the waveform trend undergoes a significant reversal. Specifically, a slope changing from positive to negative is marked as an inflection point from rising to falling, and a slope changing from negative to positive is marked as a node from falling to rising. All candidate feature points are arranged in ascending order according to their horizontal coordinates, forming a candidate feature point sequence. Specifically, for the ascending order arrangement according to the horizontal coordinates, a preset spacing threshold is used. 20 pixels, fluctuation threshold For each 8-pixel interval, adjacent candidate feature points are traversed. If the horizontal pixel distance between two points is less than or equal to 20 pixels, only the point with the more significant amplitude is retained. If the vertical amplitude difference between two adjacent points is less than or equal to 8 pixels, it is determined to be a noise feature point and deleted. Valid feature points for inter-frame matching are obtained by filtering feature points.

[0037] For the validity verification of candidate feature points, specifically, for peak points, the slopes of the three consecutive pixels to the left are all positive and the slopes of the three consecutive pixels to the right are all negative; for trough points, the slopes of the three consecutive pixels to the left are all negative and the slopes of the three consecutive pixels to the right are all positive; for slope inflection points, the slopes of the three pixels before and after the inflection point change monotonically; points that do not meet the above rules are judged as false feature points and are removed. For extreme points, the difference between the longitudinal amplitude of the extreme point and the baselines on both sides is not less than 5% of the effective longitudinal range of the channel, and the difference between the longitudinal amplitudes on both sides of the inflection point is not less than 3% of the effective longitudinal range of the channel. Candidate feature points that do not meet these requirements are judged as noise points and are removed.

[0038] In one embodiment of the present invention, the feature point matching and waveform horizontal scrolling pixel displacement calculation steps of adjacent video frames include: S36: Taking each feature point as the center, take the vertical coordinates of the continuous waveform points 10 pixels to the left and right to form a one-dimensional description vector with dimension 21. S37: Calculate the Euclidean distance between the feature point description vectors of frame k and frame (k+1), and take the pair with the smallest distance as the initial matching pair; S38: For the initial matching pairs, with the constraint that the ventilator waveform only undergoes lateral rigid body translation, mismatched pairs with the absolute value of the longitudinal coordinate deviation of the corresponding feature points in the two frames being greater than a preset deviation threshold are eliminated to obtain a set of valid matching pairs. S39: Calculate the difference in horizontal pixel coordinates for all valid matching pairs, and take the median of all differences as the horizontal scroll pixel displacement between adjacent frames.

[0039] As shown in steps S36 to S39 above, the pixel displacement of the waveform horizontal scrolling is calculated based on effective feature points. In this embodiment, with each feature point as the center, a vertical coordinate sequence of 21 consecutive waveform points, 10 pixels to the left and right, is taken to form a one-dimensional description vector with dimension 21, and the type label of the feature point is attached (e.g., peak, trough, rising-to-falling inflection point, falling-to-rising inflection point). Same-type priority matching is performed on the feature points of the k-th frame and the k+1-th frame, that is, the Euclidean distance between the pairwise description vectors is calculated between feature points with the same type label. The maximum allowed Euclidean distance for matching feature points of the same type is preset to be 4 pixels. The pair with the smallest distance that meets the threshold distance is taken as the initial matching pair to obtain the initial matching pair set. If the number of feature points of a certain type is insufficient, cross-type matching is added as an alternative, but the matching distance threshold is tightened to 70% of the original threshold (i.e., 2.8 pixels). Assuming the ventilator waveform only undergoes lateral rigid body translation, for any initial set of matched pairs, if the absolute value of the longitudinal coordinate deviation between corresponding feature points in two frames is greater than a preset deviation threshold (2 pixels), it is determined to be a mismatch and removed. The preliminary median value of the lateral displacement of all initial matched pairs is calculated. If the displacement of a single matched pair deviates from the preliminary median value by more than ±25%, it is determined to be an outlier mismatch and removed, thus obtaining the set of valid matched pairs. The lateral pixel coordinate difference of all valid matched pairs is calculated. Let the lateral coordinate of the i-th valid matched pair in the k-th frame be... The horizontal coordinate of the (k+1)th frame is Then the displacement of a single pair is The median of all single-pair displacements is taken as the horizontal scroll pixel displacement of the waveform between the current adjacent frames. Based on the inter-frame scroll pixel displacement The waveform segments of multiple consecutive frames are aligned, spliced, and deduplicated to reconstruct a waveform curve covering multiple complete respiratory cycles. This is specifically achieved through the following method: For the first... Frame waveform, based on scroll pixel displacement Shift the whole thing to the left Pixel, making it similar to the first Timing alignment is achieved by overlapping regions of two frame waveforms, with the width of the overlapping region being [missing information]. For each pixel, the vertical coordinate of the same horizontal position within the overlapping area is averaged over two frames; the first pixel is retained. The newly added non-overlapping waveform segments on the right side of the frame are spliced ​​to the end of the existing long waveform to complete a single round of splicing. The above splicing process is repeated until the waveform curve is reconstructed.

[0040] In this embodiment, the reconstructed waveform curve needs to meet the requirement that the total horizontal pixel length is not less than 4000 pixels and covers no less than 6 complete breathing cycles, corresponding to a time span of approximately 15 to 30 seconds.

[0041] To improve the accuracy of identifying weak feature anomalies and reduce the risk of missed and false detections, in step S4, the cycle points corresponding to the respiratory cycle are calibrated by cross-validation of the flow-time waveform curve and the pressure-time waveform curve on the waveform curve. The time intervals of the inspiratory and expiratory phases are obtained by dividing the time intervals. Weak feature signals are monitored within the corresponding time intervals, and combined with the rhythm consistency check of multiple consecutive respiratory cycles, waveform anomalies corresponding to invalid triggers and reverse triggers are identified. This is specifically implemented through the following method: In one embodiment of the present invention, the step of calibrating the periodic point includes: S41: In the flow-time waveform curve, the first zero-crossing point where the flow rate crosses the zero value from the baseline in the positive direction is marked as the candidate point for the start of inspiration, the zero-crossing point where the flow rate crosses the zero value from the positive to the negative direction is marked as the candidate point for the start of expiration, and the point where the flow rate returns to the baseline and remains stable is marked as the candidate point for the end of expiration. S42: Map each candidate point marked in the flow-time waveform curve to the pressure-time waveform curve, respectively verifying the physiological linkage characteristics of pressure rise during the inspiratory phase, pressure decrease during the expiratory phase, and pressure stabilization at the end of the expiratory phase. The cycle point is obtained through cross-validation of the two waveforms.

[0042] As shown in steps S41 and S42 above, based on the reconstructed flow-time waveform curve obtained in the above embodiment, the determination of the longitudinal pixel position of the zero flow baseline is achieved by selecting the end-expiratory steady segment of three consecutive complete respiratory cycles, taking the average longitudinal coordinate of the middle 10 pixels of each steady segment as the single-cycle baseline sample, and taking the average of the longitudinal pixel coordinates of all samples as the flow baseline when the longitudinal coordinate deviation between samples is ≤2 pixels. The preset baseline tolerance band is Pixel, i.e., the vertical coordinate is located at When the flow rate is within the range, it is determined that the flow velocity is at the zero flow velocity baseline level; Extract the inhalation candidate point from the flow velocity waveform curve (i.e., when the flow velocity crosses the upper boundary of the tolerance band from within the baseline tolerance band upwards, and the average upward slope of the three consecutive pixels after the crossing is ≥0.5 pixels / pixel, mark the first crossing point as the inhalation candidate point). (This corresponds to the moment when the flow rate changes from zero to positive and the inhalation begins.) For example, after passing through, three consecutive horizontal pixels are: x=100, y=20; x=101, y=21; x=102, y=23. The horizontal distance traveled a total of 102-100=2 pixels, and the vertical rise was a total of 23-20=3 pixels. The average slope is 3÷2=1.5 pixels / pixel, which is greater than 0.5. Therefore, the condition is met, and it is determined to be the start of a real inhalation. The extracted candidate points also include exhalation initiation candidate points (i.e., when the flow rate crosses the lower boundary of the tolerance band from the positive peak downwards, enters the negative direction from the positive direction, and the average descent slope of the three consecutive pixels after the crossing is ≤-0.5 pixels / pixel, the crossing point is marked as an exhalation initiation candidate point). The candidate points correspond to the time from the end of inspiratory flow to the beginning of expiration. Candidate points also include end-expiratory candidate points (i.e., when the flow rate returns from negative to within the baseline tolerance band, and the longitudinal coordinate range of 15 consecutive pixels is ≤2 pixels, and remains within the tolerance band without significant fluctuations, the starting point of this stable segment is marked as an end-expiratory candidate point). (This corresponds to the moment between the end of this exhalation and the beginning of the next inhalation). It is important to understand that a negative slope indicates that the waveform changes downward along the time axis, corresponding to the flow rate falling back during the exhalation phase; a positive slope indicates that the waveform rises upward, corresponding to the inhalation initiation process.

[0043] Each candidate point marked in the flow-time waveform is mapped to the pressure-time waveform. Single-point physiological linkage verification and phase-time ratio verification are then performed sequentially. Valid cycle points are obtained through dual-waveform cross-verification. The single-point physiological linkage verification includes verification of the inspiratory start candidate point, the expiratory start candidate point, and the expiratory end candidate point, specifically including: For the verification of the inhalation start candidate point, the pressure waveform should enter a continuous rising phase after the time point of the candidate point. Therefore, the cumulative increase in the vertical coordinate of 10 consecutive pixels should not be less than 5% of the effective vertical range of the channel. In this embodiment, the vertical range of the pressure channel is 200 pixels, corresponding to the cumulative increase threshold. for: ; For the verification of candidate points of exhalation initiation, the first 3 pixels of the corresponding time point should be in the neighborhood of the pressure peak. After this point, the pressure waveform enters a continuous decline phase. The cumulative decline of 10 consecutive pixels should not be less than 10 pixels, which is 5% of the effective longitudinal range of the channel. For end-expiratory candidate point verification, the pressure waveform at the corresponding time point falls back to the end-expiratory baseline level, and the vertical coordinate is within ±5 pixels of the pressure baseline and remains stable. For phase timing ratio verification, the ratio of the inspiratory phase duration to the expiratory phase duration within the same cycle is calculated. If the ratio falls within the physiological rhythm range of 30:70 to 40:60, the timing of the candidate points is deemed valid. Among them, 30:70 to 40:60 is the typical I:E ratio reference range in adult mechanical ventilation scenarios. Physiological priors of ventilator ventilation are used to filter out false respiratory cycle candidate points generated by waveform noise. Candidate points that pass all the above checks are taken as valid respiratory cycle points. Based on the valid cycle points, time phase intervals are divided. Specifically, the inspiratory phase interval is... The expiratory phase interval is .

[0044] In one embodiment of the present invention, the waveform curve of physiological parameters changing with time in the waveform channel also includes a capacity-time waveform curve. After cross-validation to generate periodic points, the periodic points are mapped onto the capacity-time waveform curve based on the unified time axis of the waveform channel.

[0045] In this embodiment, since the pressure, flow rate and capacity waveform channels share a unified horizontal time axis, global time alignment has been completed after time-series splicing and reconstruction. Therefore, it is not necessary to separately calibrate the period points of the capacity-time waveform curve. The time coordinates of the effective period points obtained by the dual waveform verification are directly mapped onto the capacity-time waveform curve to complete the temporal interval division of the capacity waveform.

[0046] In one embodiment of the present invention, the step of identifying waveform anomalies corresponding to invalid triggering and reverse triggering includes: S43: The weak characteristic signals include pressure waveform indentation signals and flow velocity waveform offset signals; S44: When the amplitude decrease of the pressure waveform during the expiratory phase is within the preset micro-threshold range and is not accompanied by a concave shape of the flow waveform crossing zero in the positive direction, it is marked as a single invalid trigger candidate event. S45: When the flow rate waveform exhibits a second positive fluctuation during the transition phase from the end of the inspiratory phase to the early expiratory phase, and the pressure waveform simultaneously exhibits a corresponding amplitude disturbance, it is marked as a single reverse trigger candidate event.

[0047] In one embodiment of the present invention, the rhythm consistency verification step for multiple consecutive respiratory cycles includes: S46: The preset number of consecutive verification cycles is 3, and the frequency threshold is 2 times; S47: Statistically analyze the same type of candidate events within three consecutive respiratory cycles. If the same type of candidate event occurs more than or equal to 2 times, and the relative position deviation of each event within the corresponding time phase does not exceed 10% of the total duration of the cycle, it is determined to be a valid waveform abnormality of the corresponding type. S48: If a candidate event occurs in only a single cycle, it is determined to be occasional noise and no anomaly flag is triggered.

[0048] Based on steps S43-S45 and S46-S48 above, for the determination of waveform anomalies in invalid triggers, a local indentation pattern is searched in the pressure waveform of the expiratory phase. When the following conditions are met, it is marked as a single invalid trigger candidate event, including: (1) The horizontal width of the recess is in the range of 5 to 30 pixels to eliminate pixel noise that is too narrow and baseline drift that is too wide; (2) The flow rate waveform at the corresponding time position did not cross the baseline in the positive direction, that is, the indentation did not trigger the ventilator to deliver air; For waveform anomaly determination in reverse triggering, a single reverse triggering candidate event is marked when the following conditions are met during the transition phase from the end of the inspiratory phase to the early expiratory phase: (1) The flow velocity waveform shows a second positive fluctuation, that is, the flow velocity should have been continuously decreasing but then a brief upward bulge occurs, and the amplitude difference between the peak value and the valley values ​​on both sides is not less than 6 pixels; (2) Within the same time window, the pressure waveform synchronously exhibits corresponding amplitude disturbances, and the amplitude difference of the disturbances is not less than 4 pixels; (3) The time deviation between the peak value of the flow velocity waveform and the extreme value of the pressure disturbance does not exceed 3 pixels.

[0049] In this embodiment, the preset number of consecutive verification cycles is 3, and the frequency threshold is 2 times. The same type of candidate events within 3 consecutive respiratory cycles are statistically analyzed. If the same type of candidate event appears ≥ 2 times, and the relative position deviation of each event in the corresponding time phase does not exceed 10% of the total duration of the cycle, it is determined to be a valid waveform abnormality of the corresponding type. If a candidate event appears in only a single cycle, the abnormality marker is not triggered.

[0050] It is important to understand that invalid triggering occurs when the patient spontaneously exerts an inspiratory effort during the expiratory phase, but the intensity of the inspiratory effort is insufficient and fails to reach the ventilator's preset trigger threshold; reverse triggering occurs when the ventilator delivers air and triggers the patient's spontaneous breathing, which is the opposite of the normal logic where the patient's inhalation triggers the ventilator's air delivery.

[0051] In one embodiment of the present invention, in step S5, the monitoring result is obtained by labeling the type of waveform anomaly and locating the time of occurrence. This is specifically achieved through the following method: For waveform anomalies deemed valid, invalid triggers or reverse triggers are labeled according to the pathological type of the anomaly. At the same time, the horizontal pixel coordinates of the starting point of the anomaly event on the reconstructed long-time waveform are extracted. Based on the acquisition time of the first frame video, combined with the temporal offset of multi-frame splicing, the actual occurrence time and duration of the anomaly event are calculated. Information such as anomaly type, occurrence time, anomaly amplitude, and respiratory cycle are integrated into the monitoring results, and the corresponding position of the anomaly segment on the original waveform curve is marked to generate an anomaly monitoring record.

[0052] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A machine vision-based method for monitoring abnormal waveforms in ventilators, characterized in that, Includes the following steps: S1: Acquire video frames corresponding to multiple consecutive screen images on the ventilator display screen; S2: Project the video frame horizontally along the longitudinal direction to generate a longitudinal pixel distribution curve. Based on the trough position in the longitudinal pixel distribution curve, divide the image horizontally to segment independent waveform channels in which physiological parameters change over time, and extract the waveform pixel contour trajectory within the waveform channel. S3: Along the horizontal time axis of the waveform channel, perform waveform integrity verification from the left and right screen boundaries toward the center of the image. Remove the waveform generation section on the right that has not formed a complete cycle shape and the waveform truncated section on the left that has been truncated. Define the remaining middle section as the effective matching extraction area. Extract feature points from the waveform pixel contour trajectory corresponding to the effective matching extraction area. Perform feature point matching in the same waveform channel in adjacent video frames to calculate the pixel displacement of the waveform horizontal scrolling. Then, stitch together the video frames according to the pixel displacement to generate a continuous waveform curve covering multiple complete breathing cycles. S4: On the waveform curve, the cycle point corresponding to the respiratory cycle is determined by cross-verification of the flow-time waveform curve and the pressure-time waveform curve. The time interval of the inspiratory phase and the expiratory phase is obtained by dividing the cycle point. Weak characteristic signals are monitored in the corresponding time interval. Combined with the rhythm consistency verification of multiple consecutive respiratory cycles, waveform abnormalities corresponding to invalid triggering and reverse triggering are identified. S5: Based on the waveform anomalies, the type of anomalies is labeled and the time of occurrence is located to obtain the monitoring results.

2. The machine vision-based ventilator waveform anomaly monitoring method according to claim 1, characterized in that, The steps for segmenting video frames to obtain waveform channels include: For each line of the video frame, the total number of waveform pixels with grayscale values ​​lower than the background threshold is counted row by row to generate a vertical pixel distribution curve; Local minima are detected on the vertical pixel distribution curve. Vertical coordinates that simultaneously satisfy the local minimum of a 3-pixel neighborhood and whose projected value is less than 10% of the maximum projected value are identified as troughs. Based on the trough position, the image within the video frame is segmented along the horizontal direction, and the segmented image is subjected to continuous curve pixel trajectory detection to obtain waveform channels corresponding to different physiological parameters such as pressure, flow rate, and volume.

3. The machine vision-based ventilator waveform anomaly monitoring method according to claim 1, characterized in that, The extracted waveform pixel contour trajectory needs to meet the following conditions, including: Extending continuously along the horizontal time axis of the waveform channel, the length of the horizontal breakpoint pixels within a single trajectory does not exceed the preset breakpoint threshold. The vertical pixel coordinates of the trajectory are all within the effective display range of the waveform channel; wherein, the effective display range is the area after the vertical upper and lower boundaries of the waveform channel are offset inward by a preset pixel margin, corresponding to the normal display range of respiratory physiological parameters.

4. The machine vision-based ventilator waveform anomaly monitoring method according to claim 1, characterized in that, The feature point extraction step includes: On the waveform pixel contour trajectory corresponding to the effective matching extraction area of ​​the waveform channel, local amplitude extreme points and slope inflection points are extracted as candidate feature points. Among the candidate feature points, the candidate feature points with a horizontal pixel spacing greater than a preset spacing threshold and an amplitude difference greater than a preset fluctuation threshold are retained to obtain the feature points.

5. The machine vision-based ventilator waveform anomaly monitoring method according to claim 1, characterized in that, The step of defining the effective matching extraction region includes: Along the horizontal time axis of the waveform channel, with a step size of 1 pixel, the waveform contour is traversed from the right boundary to the left. When a complete set of peak-valley extreme pairs is detected for the first time and the amplitude difference is not less than 10% of the effective longitudinal range of the channel, the corresponding horizontal coordinate is recorded as the left boundary of the waveform generation segment, and the right generation segment is removed. Along the horizontal time axis of the waveform channel, with a step size of 1 pixel, traverse the waveform contour from the left boundary to the right. When a complete set of peak-valley extreme pairs is detected for the first time and the amplitude difference is not less than 10% of the effective longitudinal range of the channel, record the corresponding horizontal coordinate as the right boundary of the waveform truncation segment and remove the left truncation segment. The middle waveform segment between the left and right boundaries is defined as the effective matching extraction area.

6. The machine vision-based ventilator waveform anomaly monitoring method according to claim 1, characterized in that, The steps for calibrating the periodic point include: In the flow-time waveform curve, the first zero-crossing point where the flow rate crosses zero from the baseline in the positive direction is marked as the candidate point for the start of inspiration, the zero-crossing point where the flow rate crosses zero from the positive to the negative direction is marked as the candidate point for the start of expiration, and the point where the flow rate returns to the baseline and remains stable is marked as the candidate point for the end of expiration. Each candidate point marked in the flow-time waveform curve is mapped to the pressure-time waveform curve to verify the physiological linkage characteristics of pressure rise during inhalation, pressure decrease during exhalation, and pressure stabilization at the end of exhalation. The cycle point is obtained through cross-validation of the two waveforms.

7. The machine vision-based ventilator waveform anomaly monitoring method according to claim 1, characterized in that, The waveform curves showing the changes of physiological parameters over time within the waveform channel also include a capacity-time waveform curve. After cross-validation to generate periodic points, the periodic points are mapped onto the capacity-time waveform curve based on the unified time axis of the waveform channel.

8. The machine vision-based method for monitoring abnormal waveforms in ventilators according to claim 1, characterized in that, The step of identifying waveform anomalies corresponding to invalid triggers and reverse triggers includes: The weak characteristic signals include pressure waveform indentation signals and flow velocity waveform offset signals; When the amplitude decrease of the pressure waveform during the expiratory phase is within the preset micro-threshold range and is not accompanied by a concave shape of the flow waveform crossing zero in the positive direction, it is marked as a single invalid trigger candidate event. When the flow rate waveform exhibits a second positive fluctuation during the transition phase from the end of the inspiratory phase to the early expiratory phase, and the pressure waveform simultaneously shows a corresponding amplitude disturbance, it is marked as a single reverse trigger candidate event.

9. The machine vision-based method for monitoring abnormal waveforms in ventilators according to claim 1, characterized in that, The steps for feature point matching between adjacent video frames and calculation of waveform horizontal scroll pixel displacement include: With each feature point as the center, the vertical coordinates of the continuous waveform points 10 pixels to the left and right are used to form a one-dimensional description vector with dimension 21. Calculate the Euclidean distance between the feature point description vectors of frame k and frame (k+1), and take the pair with the smallest distance as the initial matching pair; For the initial matching pairs, with the constraint that the ventilator waveform only undergoes lateral rigid body translation, mismatched pairs with the absolute value of the longitudinal coordinate deviation of the corresponding feature points in the two frames being greater than a preset deviation threshold are eliminated to obtain a set of valid matching pairs; Calculate the horizontal pixel coordinate difference for all valid matching pairs, and take the median of all differences as the horizontal scroll pixel displacement between adjacent frames.

10. The machine vision-based method for monitoring abnormal waveforms in ventilators according to claim 1, characterized in that, The steps for verifying rhythm consistency across multiple consecutive respiratory cycles include: The preset number of consecutive verification cycles is 3, and the frequency threshold is 2. If the same type of candidate event occurs more than or equal to 2 times within 3 consecutive respiratory cycles, and the relative position deviation of each event within the corresponding time phase does not exceed 10% of the total duration of the cycle, it is determined to be a valid waveform abnormality of the corresponding type. If a candidate event occurs in only a single cycle, it is considered occasional noise and no anomaly flag is triggered.

Images