A deep learning-based online determination method for vent glue blocking

Abstract

The application discloses a kind of online determination method of air vent glue block based on deep learning, belong to cigarette air vent online detection technical field, for solving the existing air vent glue block online determination depends on single branch absolute response or fixed reference, it is difficult to distinguish the problem of working condition change and glue block. The method will be measured air vent branch and reference orifice branch access same negative pressure interface and form bridge type detection unit, in the establishment, maintenance and release phase of short time negative pressure pulse, synchronous acquisition branch pressure change and calculate branch differential pressure, determine bridge anchoring baseline and generate bridge imbalance counterbalance trajectory;Bridge anchoring baseline and bridge imbalance counterbalance trajectory are input into deep learning model, with bridge anchoring baseline as anchoring reference, according to unblocked sample, bridge imbalance counterbalance trajectory is regressed, and bridge state deviation is obtained;Bridge state deviation is compared with preset threshold value, determine air vent for glue block state or unobstructed state.The application is favorable to improve the stability and consistency of online determination.

Description

Technical Field

[0001] This invention relates to the field of online detection technology for cigarette ventilation holes, and in particular to an online method for determining whether adhesive is obstructing ventilation holes based on deep learning. Background Technology

[0002] In cigarette production, the ventilation holes in the cigarette are used to regulate ventilation rate and draw resistance, and their unobstructed state directly affects the consistency of product processes and the quality of the finished product. During molding, assembly, and subsequent processing, adhesive may spread to the ventilation hole area, causing local obstruction or even blockage, thereby changing the actual permeability of the ventilation hole area. Due to the fast production cycle, continuous changes in cigarette posture, and the presence of negative pressure fluctuations, air path disturbances, and differences in clamping and bonding in the testing environment, the adhesive obstruction status of the ventilation holes cannot be reliably judged solely by visual inspection or a single instantaneous aerodynamic volume. Therefore, how to reliably determine the true permeability status of the ventilation hole area under online conditions has always been one of the key issues in cigarette quality inspection.

[0003] In existing technologies, online detection of cigarette ventilation holes typically employs pneumatic detection methods or a combination of pneumatic and visual inspection methods. Pneumatic detection generally involves connecting the cigarette to be tested to a negative pressure detection channel. Under suction, it collects pressure, flow, or differential pressure values ​​related to the ventilation hole area and determines whether the ventilation hole is abnormal based on pre-set standard ranges. Some solutions also introduce standard orifices, standard air resistance components, or reference branches for device calibration, benchmark correction, or operational condition compensation. Other solutions extract peak values, steady-state values, slope values, and recovery values ​​at different stages of negative pressure establishment, maintenance, or release, and then provide detection results through rule-based judgment, empirical models, or classification models. Visual inspection methods often identify the surface condition of the ventilation hole area through visual imaging and combine position matching, image segmentation, or defect classification to determine whether there is glue coverage, misalignment, or contamination. These solutions have been applied to some extent in online cigarette detection scenarios.

[0004] However, in the specific scenario of adhesive blockage of ventilation holes, the existing solutions mentioned above often rely on the absolute response of a single branch, a small number of feature points, or a fixed benchmark, making it difficult to fully reflect the continuous changes in the permeability of the ventilation hole area relative to the reference channel during the same suction process. When there are fluctuations in the shared negative pressure source, changes in the air path state, or differences in clamping and bonding, the original pressure value or a small number of feature quantities are easily affected by both operating conditions and the actual blockage state of the ventilation hole. This makes it difficult to distinguish between response offsets from two different sources, namely "changes in operating conditions" and "adhesive blockage," in online judgment, thereby affecting the stable closed loop of the ventilation hole state judgment link.

[0005] Therefore, an online method for determining whether ventilator adhesive blockage can overcome the shortcomings of the prior art is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0006] This invention provides a deep learning-based online determination method for adhesive obstruction in ventilation holes. This method addresses the technical problem in existing technologies where online determination of adhesive obstruction in ventilation holes relies on the absolute response of a single branch, a small number of feature points, or a fixed benchmark. This makes it difficult to fully reflect the continuous changes in the permeability of the ventilation hole area relative to the reference channel during the same suction process. Furthermore, when there are fluctuations in the shared negative pressure source, changes in the air path state, or differences in clamping and bonding, it is difficult to distinguish between response offsets from two different sources: "changes in operating conditions" and "adhesive obstruction." This results in insufficient stability of the online determination link for the ventilation hole status.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A deep learning-based online method for determining adhesive blockage in ventilation holes includes:

[0009] Step S1: Receive the cigarette to be tested, connect the ventilation hole area of ​​the cigarette to be tested to the ventilation hole branch, and connect the reference orifice to the reference orifice branch, so that the ventilation hole branch to be tested and the reference orifice branch form a bridge detection unit.

[0010] Step S2: Apply a negative pressure pulse to the bridge detection unit under the constraint of a shared negative pressure source, and collect the pressure changes of the ventilation hole branch under test and the reference orifice branch during the establishment, maintenance and release phases of the negative pressure pulse to obtain the branch pressure changes.

[0011] Step S3: Calculate the branch differential pressure based on the branch pressure change, determine the bridge anchorage baseline based on the branch differential pressure, and generate the bridge imbalance and rebalancing trajectory based on the continuous change of the branch differential pressure.

[0012] Step S4: Input the bridge anchoring baseline and the bridge imbalance recovery trajectory into the deep learning model. The deep learning model uses the bridge anchoring baseline as the anchoring reference and performs regression calculation on the bridge imbalance recovery trajectory based on the regression relationship established in advance based on the unoccluded samples to obtain the unoccluded bridge state corresponding to the bridge anchoring baseline. Based on the difference between the unoccluded bridge state and the bridge imbalance recovery trajectory, the bridge state deviation is obtained.

[0013] Step S5: Compare the bridge state deviation with the preset occlusion determination threshold to obtain the state determination result;

[0014] Step S6: Determine whether the ventilation hole is blocked by glue or unobstructed based on the status judgment result.

[0015] Further, step S1 includes:

[0016] Step S11: Send the cigarette to be tested into the positioning clamping part, and limit the cigarette to be tested axially and orient it circumferentially so that the coverage area of ​​the ventilation hole area corresponds to the air intake window of the ventilation hole branch to be tested, and form a local detection cavity surrounding the ventilation hole area.

[0017] Step S12: Press the sealing boundary of the local detection cavity onto the outer surface of the cigarette to be tested, and close the non-detection air intake path outside the ventilation hole area, so that the gas received by the ventilation hole branch to be tested only enters through the ventilation hole area.

[0018] Step S13: Install the reference orifice at the air inlet end of the reference orifice branch, and seal and fix the connection position between the reference orifice and the reference orifice branch so that the reference orifice becomes the only air inlet boundary of the reference orifice branch. Then connect the ventilation hole branch to be tested and the reference orifice branch to the same negative pressure interface.

[0019] Step S14: Based on the synchronous pressure relationship between the ventilation hole branch to be tested and the reference hole branch under the same negative pressure interface, a bridge detection unit is formed, so that the reference hole branch participates in the formation of branch pressure change.

[0020] Further, step S2 includes:

[0021] Step S21: Connect the shared negative pressure source to the bridge detection unit through the same negative pressure control path, and apply unified negative pressure start and stop control to the ventilation hole branch to be tested and the reference orifice branch at the same time, so that the ventilation hole branch to be tested and the reference orifice branch are subjected to the same suction action in the same detection cycle.

[0022] Step S22: Use negative pressure pulses to drive the bridge detection unit, so that the negative pressure pulses sequentially form a setup phase, a maintenance phase and a release phase, and keep the ventilation hole branch under test and the reference orifice branch under synchronous pressure and synchronous pressure release during the setup phase, maintenance phase and release phase.

[0023] Step S23: During the establishment phase, continuous pressure data are collected from the branch of the ventilation hole to be tested and the branch of the reference orifice at the same sampling starting point to obtain the pressure change during the establishment phase that characterizes the negative pressure propagation process.

[0024] Step S24: During the maintenance phase, the sampling process of the establishment phase continues to continuously collect pressure data on the ventilation hole branch to be tested and the reference orifice branch to obtain the pressure change during the maintenance phase, which characterizes the continuous suction process.

[0025] Step S25: During the release phase, the sampling process of the maintenance phase is continued to continuously collect pressure data on the branch of the ventilation hole to be tested and the branch of the reference orifice to obtain the pressure change during the release phase, which characterizes the negative pressure release and fall process.

[0026] Step S26: Connect the pressure changes during the establishment, maintenance, and release phases of the ventilation hole branch to be tested in the sampling order to obtain the continuous pressure change of the ventilation hole branch to be tested. Connect the pressure changes during the establishment, maintenance, and release phases of the reference orifice branch in the sampling order to obtain the continuous pressure change of the reference orifice branch. The continuous pressure changes of the ventilation hole branch to be tested and the continuous pressure changes of the reference orifice branch are determined as the branch pressure changes.

[0027] Further, step S3 includes:

[0028] Step S31: Based on the pressure values ​​of the ventilation hole branch to be tested and the reference orifice branch at the corresponding sampling positions in the branch pressure change, calculate the difference in the branch differential pressure according to the unified sampling order.

[0029] Step S32: Extract the common reference position before the negative pressure pulse is applied and after the negative pressure pulse is released, and determine the bridge anchorage baseline based on the common reference position;

[0030] Step S33: Using the bridge anchorage baseline as a reference, continuously characterize the deviation direction, deviation magnitude, and duration of the branch differential pressure relative to the bridge anchorage baseline to obtain the branch differential pressure deviation characterization results.

[0031] Step S34: Based on the branch differential pressure deviation characterization results, extract the bridge imbalance initiation process in the establishment phase, extract the bridge imbalance persistence process in the maintenance phase, and extract the bridge rebalancing regression process in the release phase. The bridge rebalancing regression process is the continuous change process of branch differential pressure from deviation from the bridge anchoring baseline to entering the preset rebalancing judgment threshold range.

[0032] Step S35: Connect the bridge imbalance initiation process, bridge imbalance duration process, and bridge rebalancing regression process in the original sampling order to obtain the bridge imbalance rebalancing trajectory.

[0033] Furthermore, the deep learning model is a bridge-anchored complex equilibrium score-based diffusion model, and step S4 includes:

[0034] Step S41: Divide the bridge imbalance recovery trajectory into three trajectory inputs according to the establishment phase, maintenance phase and release phase, and input them into the three trajectory encoding layers respectively to obtain the establishment phase features, maintenance phase features and release phase features;

[0035] Step S42: Input the bridge anchoring baseline into the bridge anchoring coding layer to obtain the anchoring features, and input the diffusion step conditions into the diffusion step coding layer to obtain the diffusion step features;

[0036] Step S43: Input the establishment phase features, maintenance phase features, release phase features, anchoring features and diffusion step features into the bridge state fusion layer, and splice them in a fixed order to obtain unified bridge state fusion features;

[0037] Step S44: Establish the regression space of the unoccluded bridge state trajectory based on the unoccluded samples, and perform regression calculation on the unified bridge state fusion features by the rebalancing regression layer to obtain the unoccluded bridge state trajectory corresponding to the bridge anchorage baseline.

[0038] Step S45: Combine the difference expression between the unobstructed bridge trajectory and the bridge imbalance recovery trajectory with the anchoring feature input to the bridge deviation output layer to map the degree of deviation of the bridge imbalance recovery trajectory relative to the unobstructed bridge trajectory, and obtain the bridge deviation amount.

[0039] Further, step S5 includes:

[0040] Step S51: Compare the bridge state deviation with the preset occlusion judgment threshold once to determine whether the bridge state deviation is greater than or equal to the preset occlusion judgment threshold.

[0041] Step S52: When the bridge state deviation is greater than or equal to the preset occlusion determination threshold, the state determination result is determined as the occlusion determination result.

[0042] Step S53: When the bridge state deviation is less than the preset occlusion judgment threshold, the state judgment result is determined to be the smooth judgment result.

[0043] Furthermore, based on the distinction between occlusion determination results and unobstructedness determination results in the state determination results, step S6 includes:

[0044] Step S61: When the state determination result is the occlusion determination result, the ventilation hole is determined to be in the glue occlusion state;

[0045] Step S62: When the status determination result is unobstructed, the ventilation hole is determined to be unobstructed.

[0046] Further, in step S32, the common reference position includes the stable sampling interval before the negative pressure pulse is applied and the stable sampling interval after the negative pressure pulse is released. The continuous sampling interval in which the branch differential pressure change between adjacent sampling positions before the negative pressure pulse is applied is less than the preset baseline fluctuation threshold is determined as the stable sampling interval before the negative pressure pulse is applied. The continuous sampling interval in which the branch differential pressure change between adjacent sampling positions after the negative pressure pulse is released is less than the preset baseline fluctuation threshold is determined as the stable sampling interval after the negative pressure pulse is released. The bridge anchorage baseline is determined based on the stable sampling interval before the negative pressure pulse is applied and the stable sampling interval after the negative pressure pulse is released.

[0047] Furthermore, in step S34, during the release phase, when the branch differential pressure enters the preset rebalancing judgment threshold range, and the number of sampling positions within the preset rebalancing judgment threshold range in the continuous sampling positions reaches the preset sampling number, it is determined that the bridge rebalancing is completed, and the sampling position that first satisfies the bridge rebalancing completion is determined as the rebalancing endpoint of the bridge imbalance rebalancing trajectory.

[0048] Further, in step S44, the bridge anchoring baseline is matched with multiple pre-set bridge anchoring baseline value ranges, the unobstructed sample corresponding to the matching result is called, and the unobstructed bridge state trajectory corresponding to the bridge anchoring baseline is established based on the called unobstructed sample.

[0049] Compared with the prior art, the beneficial effects of the present invention are:

[0050] First, this invention proposes a bridge state regression judgment method based on a deep learning model. In the online detection scenario of glue obstruction of cigarette vents, the original branch pressure, a few momentary feature values, or a fixed template are no longer directly used as the judgment basis. Instead, a bridge detection relationship is first established between the branch of the vent to be tested and the reference orifice branch. Then, the bridge anchoring baseline is determined based on the branch differential pressure of the current detection cycle, and the continuous differential pressure changes in the establishment, maintenance, and release phases are organized into a bridge imbalance recovery trajectory. On this basis, the deep learning model does not directly classify the original signal as anomalies, but uses the bridge anchoring baseline as a conditional reference to regress the unobstructed bridge state and output the bridge state deviation. This transforms the model's processing object from the original response, which is susceptible to negative pressure fluctuations, changes in air path state, and differences in clamping and adhesion, into an intermediate quantity that is closer to the actual transparency state of the vent, namely the bridge state change of the branch to be tested relative to the reference branch. This helps to distinguish the response deviation caused by working condition disturbances and glue obstruction under realistic online constraints. In a preferred embodiment, the deep learning model employs a bridge-anchored balanced score-based diffusion model, which regresses the trajectory of the unoccluded bridge state by decreasing the diffusion step. This allows the regression process to better adapt to the characteristics of the unoccluded bridge state itself changing with the bridge anchoring baseline, rather than simply pulling the regression results toward a single fixed template.

[0051] Secondly, this invention proposes a pneumatic pulse detection method for bridge imbalance and rebalancing. Negative pressure pulses are simultaneously applied to the branch of the ventilation hole under test and the branch of the reference hole through the same negative pressure interface. Data from the entire process of negative pressure establishment, maintenance, and release is continuously collected. It does not extract single-point values ​​to replace the dynamic process, nor does it separately determine thresholds for each of the three stages. Instead, it retains the sequential evolution relationship consistent with the actual suction process. Furthermore, this invention uniformly represents the deviation direction, deviation magnitude, and duration of the branch differential pressure relative to the bridge anchoring baseline as the branch differential pressure deviation characterization result. Based on this characterization result, the bridge imbalance initiation process, bridge imbalance duration process, and bridge rebalancing regression process corresponding to the establishment, maintenance, and release stages are extracted separately. This ensures that the process extraction of the three stages is based on the same characterization benchmark, making the generation logic of the bridge imbalance and rebalancing trajectory coherent and traceable step by step, avoiding the inconsistency in characterization standards caused by independently processing branch differential pressure at different stages. The bridge anchorage baseline is generated jointly by the stable interval before negative pressure and the stable interval after release within the current detection cycle, without relying on fixed constants or device calibration parameters. The establishment of the unobstructed bridge trajectory is also constrained by the range of bridge anchorage baseline values ​​and the anchorage distance, and is not simply pulled towards a single fixed template. Therefore, under the condition that only online acquisition of dual-branch pressure sequences is available, this invention establishes a dynamic characterization mechanism for microscopic bridge state changes, enabling the bridge state deviation, a key intermediate quantity, to be constructed and stably estimated around the current operating conditions.

[0052] Third, this invention proposes an online determination method for ventilator glue obstruction based on bridge state deviation. It uses the one-dimensional bridge state deviation output by the model as the sole input for determination, and completes the semantic mapping between the unobstructed state and the glue obstruction state through a single comparison with a preset obstruction determination threshold. This eliminates the need for point-by-point repeated discrimination of the original pressure, branch differential pressure, or complete bridge imbalance recovery trajectory. Compared to existing technologies that rely on multi-feature splicing, staged experience thresholds, or direct anomaly identification, this invention forms a complete link from bridge acquisition, bridge state construction, unobstructed bridge state regression to single-value determination. This ensures consistency between the online determination criteria, data sources, intermediate processing quantities, and output results, facilitating the convergence of ventilator glue obstruction determination from a scattered signal comparison process into a closed-loop bridge state regression determination process. This makes it more suitable for continuous delivery and consistent execution in online production scenarios. Attached Figure Description

[0053] Figure 1 This is a flowchart of an online method for determining adhesive blockage in ventilation holes based on deep learning, as proposed in this invention.

[0054] Figure 2 The flowchart shows the formation of a bridge-type detection unit in the online determination method for adhesive blockage of ventilation holes based on deep learning proposed in this invention.

[0055] Figure 3 This is a flowchart of the branch pressure change acquisition process for an online determination method for ventilator adhesive blockage based on deep learning proposed in this invention.

[0056] Figure 4 This is a flowchart of the bridge anchorage baseline and bridge imbalance recovery trajectory generation process for an online determination method for ventilation hole glue obstruction based on deep learning proposed in this invention.

[0057] Figure 5 This is a flowchart of the bridge anchorage balance Score-based Diffusion model processing for an online determination method of adhesive obstruction of ventilation holes based on deep learning proposed in this invention.

[0058] Figure 6 This is a flowchart illustrating the state determination result generation process of an online determination method for adhesive obstruction of ventilation holes based on deep learning proposed in this invention.

[0059] Figure 7 This is a flowchart illustrating the final online determination result of a deep learning-based online determination method for determining adhesive blockage in ventilation holes, as proposed in this invention.

[0060] Figure 8 This is a distribution diagram of the bridge imbalance and rebalancing trajectory in the reduced-dimensional bridge state space under different ventilation hole states of the online determination method for ventilation hole glue blockage based on deep learning proposed in this invention.

[0061] Figure 9 This is a scatter plot showing the joint distribution of bridge anchorage baseline and bridge state deviation in a deep learning-based online determination method for ventilator adhesive obstruction proposed in this invention.

[0062] Figure 10 The image shows the mean and standard deviation of the unobstructed bridge trajectory under different bridge anchorage baseline conditions for the online determination method of ventilation hole glue obstruction based on deep learning proposed in this invention. Detailed Implementation

[0063] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that the following embodiments are merely preferred embodiments of the present invention, and the scope of protection of the present invention is not limited thereto.

[0064] refer to Figures 1 to 10 This embodiment provides an online method for determining adhesive obstruction in ventilation holes based on deep learning, including steps S1 to S6, specifically:

[0065] Step S1: Receive the cigarette to be tested, connect the ventilation hole area of ​​the cigarette to be tested to the ventilation hole branch, and connect the reference orifice to the reference orifice branch, so that the ventilation hole branch to be tested and the reference orifice branch form a bridge detection unit.

[0066] Step S2: Apply a negative pressure pulse to the bridge detection unit under the constraint of a shared negative pressure source, and collect the pressure changes of the ventilation hole branch under test and the reference orifice branch during the establishment, maintenance and release phases of the negative pressure pulse to obtain the branch pressure changes.

[0067] Step S3: Calculate the branch differential pressure based on the branch pressure change, determine the bridge anchorage baseline based on the branch differential pressure, and generate the bridge imbalance and rebalancing trajectory based on the continuous change of the branch differential pressure.

[0068] Step S4: Input the bridge anchoring baseline and the bridge imbalance recovery trajectory into the deep learning model. The deep learning model uses the bridge anchoring baseline as the anchoring reference and performs regression calculation on the bridge imbalance recovery trajectory based on the regression relationship established in advance based on the unoccluded samples to obtain the unoccluded bridge state corresponding to the bridge anchoring baseline. Based on the difference between the unoccluded bridge state and the bridge imbalance recovery trajectory, the bridge state deviation is obtained.

[0069] Step S5: Compare the bridge state deviation with the preset occlusion determination threshold to obtain the state determination result;

[0070] Step S6: Determine whether the ventilation hole is blocked by glue or unobstructed based on the status judgment result.

[0071] In this embodiment, to meet the actual needs of fast online production cycle and short detection cycle, the negative pressure pulse is specifically a short-time negative pressure pulse. That is, the total duration of the establishment, maintenance and release phases of the short-time negative pressure pulse is limited to the same detection cycle, so that the bridge detection unit can complete a complete negative pressure establishment, maintenance and release process without significantly affecting the production cycle.

[0072] In this embodiment, step S1 includes steps S11 to S14, as referenced. Figure 2 The details are as follows:

[0073] Step S11: Send the cigarette to be tested into the positioning clamping part, and limit the cigarette to be tested axially and orient it circumferentially so that the coverage area of ​​the ventilation hole area corresponds to the air intake window of the ventilation hole branch to be tested, and form a local detection cavity surrounding the ventilation hole area.

[0074] The specific implementation method of step S11 is as follows: Let For the branch of the ventilation hole to be tested, As the reference orifice branch, This is the gas intake window. For local detection cavity, As the reference orifice, For the same negative pressure interface, This is the axial stop position. For circumferential orientation, the positioning sampling sequence of the cigarette to be tested is denoted as:

[0075] ;

[0076] in, To determine the number of sampling points, For the first Each location sampling time is measured in seconds. , The axial position of the cigarette to be tested in the positioning clamp is given in millimeters. The circumferential position of the cigarette to be tested in the positioning clamp is expressed in degrees. Gas intake window The axial coverage area is denoted as Gas intake window The circumferential coverage area is denoted as , and From the gas intake window The structural dimensions are determined;

[0077] When the cigarette to be tested is inserted into the positioning clamp, press... Read in ascending order When the condition is met and The delivery is stopped when the sampling point is located, so that the cigarette to be tested is axially... Limit, in the circumferential direction Oriented and surrounded by positioning clamps around the air intake window. To form a closed boundary, obtain the same as Local detection chamber directly connected to the air intake end ;

[0078] Through the above-mentioned limiting and orientation process, the coverage area of ​​the ventilation hole area and the relative position of the air intake window W are kept consistent in each test, providing a stable air path structure foundation for subsequent pressure acquisition.

[0079] Step S12: Press the sealing boundary of the local detection cavity onto the outer surface of the cigarette to be tested, and close the non-detection air intake path outside the ventilation hole area, so that the gas received by the ventilation hole branch to be tested only enters through the ventilation hole area.

[0080] The specific implementation method of step S12 is as follows: the local detection cavity The sealing boundary is denoted as The set of air intake paths for the branch of the ventilation hole to be tested is denoted as... ,Will After being pressed onto the outer surface of the cigarette to be tested, the air intake path located outside the ventilation hole area is from... The cavity was sealed off, with only the area through the ventilation holes allowing access to the local testing chamber. The intake path makes The received gas enters only through the ventilation hole area, therefore, A unique correspondence is established between the air intake boundary and the ventilation hole area, and a local detection cavity is established. No more air intake outside the ventilation hole area will be accepted; air will not enter through non-detection air intake paths on the outer surface of the cigarette being tested. This avoids air leakage outside the ventilation hole area from interfering with subsequent pressure acquisition.

[0081] Step S13: Install the reference orifice at the air inlet end of the reference orifice branch, and seal and fix the connection position between the reference orifice and the reference orifice branch so that the reference orifice becomes the only air inlet boundary of the reference orifice branch. Then connect the ventilation hole branch to be tested and the reference orifice branch to the same negative pressure interface.

[0082] The specific implementation method of step S13 is as follows: Let As the reference orifice, For the same negative pressure interface, Installed at The intake end, and for and The connection points are sealed and fixed to ensure that... become The unique intake boundary, the branch connection relationship is denoted as ,in, express Connect to , express Connect to ,according to Will and Merging connections to the same negative pressure interface back, and Under the same suction source, synchronous pressure conditions are established. The reference orifice branch is not used as a device calibration branch, but as a comparison branch that participates in the pressure response synchronously with the ventilation orifice branch under test.

[0083] Step S14: Based on the synchronous pressure relationship between the ventilation hole branch to be tested and the reference hole branch under the same negative pressure interface, a bridge detection unit is formed, so that the reference hole branch participates in the formation of branch pressure change.

[0084] The specific implementation of step S14 is as follows: After the bridge detection unit is formed, a unified sampling time sequence is constructed using a shared sampling clock:

[0085] ;

[0086] in, , This is the start time of negative pressure sampling. This is the sampling time interval, in seconds. The number of sampling points within the same short-duration negative pressure pulse detection cycle. ;

[0087] On this unified sampling time sequence, Output the pressure sequence of the ventilation hole branch to be tested:

[0088] ;

[0089] in, for The pressure sampling value of the ventilation branch under test at any given time, in kPa. Output reference orifice branch pressure sequence:

[0090] ;

[0091] in, for The reference orifice branch pressure sampling value at a given time, in kilopascals.

[0092] because and Shared negative pressure interface Two pressure sequences are generated synchronously at the same sampling time, allowing the reference orifice branch to directly participate in the formation of pressure changes in subsequent branches. Together, they constitute the source of raw data output by the bridge detection unit to subsequent steps.

[0093] In this embodiment, step S2 includes steps S21 to S26, as referenced. Figure 3 The details are as follows:

[0094] Step S21: Connect the shared negative pressure source to the bridge detection unit through the same negative pressure control path, and apply unified negative pressure start and stop control to the ventilation hole branch to be tested and the reference orifice branch at the same time, so that the ventilation hole branch to be tested and the reference orifice branch are subjected to the same suction action in the same detection cycle.

[0095] The specific implementation method of step S21 is as follows:

[0096] set up For negative pressure control records shared by a negative pressure source on the same negative pressure control path,

[0097] in, , The number of sampling points is controlled by negative pressure. For the first Each negative pressure control sampling time is measured in seconds. For the first Each negative voltage control voltage sample value is in volts;

[0098] Connect the shared negative pressure source to the same negative pressure control path. Then by Simultaneously connected to and ,make and Receives the same suction action within the same detection cycle, because... and Shared And the same negative pressure control path, The negative pressure start-stop changes recorded in the data are synchronously applied to the same detection cycle. and This establishes the synchronous pressure conditions for the bridge-type detection unit;

[0099] Step S22: Use negative pressure pulses to drive the bridge detection unit, so that the negative pressure pulses sequentially form a setup phase, a maintenance phase and a release phase, and keep the ventilation hole branch under test and the reference orifice branch under synchronous pressure and synchronous pressure release during the setup phase, maintenance phase and release phase.

[0100] The specific implementation method of step S22 is as follows: Let To unify the sampling time sequence,

[0101] in, , ;

[0102] This is the sampling start time for this detection cycle. This is the sampling time interval, in seconds. This represents the total number of sampling points for this detection cycle. ;

[0103] Record negative pressure control With uniform sampling time sequence The timing is matched, and the stage boundaries of the short-time negative pressure pulse establishment stage, maintenance stage, and release stage are determined based on the sampled values ​​of the negative pressure control voltage on the same negative pressure control path. The specific calculation formula is as follows:

[0104] ;

[0105] ;

[0106] ;

[0107] ;

[0108] ;

[0109] ;

[0110] ;

[0111] ;

[0112] ;

[0113] In the formula, To be related to the sampling time The corresponding negative pressure control sampling point index, To standardize sampling time The negative voltage control voltage sample value at the location is in volts. The time corresponds to the allowed window, in seconds. Establish a phase start index for short-duration negative pressure pulses. This is the starting index for the short-term negative pressure pulse maintenance phase. This is the starting index for the short-term negative pressure pulse release phase. To create a stage index window, To maintain the stage index window, To release the phase index window, To establish the phase initiation threshold, the unit is volts. To maintain a stable threshold during the phase, the unit is volts. The threshold for the pullback during the release phase, expressed in volts. To maintain the phase conduction determination threshold, the unit is volts. This represents the absolute value operation. This represents the index of the sampling point with the smallest absolute value of the time difference. This means that both conditions are met simultaneously;

[0114] This calculation formula directly binds the stage boundaries of short-duration negative pressure pulses to the same negative pressure control path, without inferring the stage boundaries from the pressure sampling values ​​of the tested ventilation hole branch and the reference orifice branch. By setting the conduction duration and sampling time interval of the same negative pressure control path, , and In one embodiment, there are 32 sampling positions; thus, the short-term negative pressure pulse sequentially forms a setup phase, a maintenance phase, and a release phase, and keeps the two branches in a state of synchronous pressure and synchronous pressure release during the three phases.

[0115] Step S23: During the establishment phase, continuous pressure data are collected from the branch of the ventilation hole to be tested and the branch of the reference orifice at the same sampling starting point to obtain the pressure change during the establishment phase that characterizes the negative pressure propagation process.

[0116] The specific implementation method of step S23 is as follows: in the unified sampling time sequence Above, to and Continuous pressure data was collected using the same sampling starting point and the same sampling interval, resulting in:

[0117] and ;

[0118] in, , for The pressure sampling value of the ventilation branch under test at any given time, in kPa. for The reference orifice branch pressure sampling value at a given time, in kilopascals;

[0119] Will and The pressure change during the setup phase is identified as a continuous process of pressure propagation into the two branches at the start of the same suction cycle. During data collection, the index order is kept consistent with the time order, and single-point pressure values ​​during the setup phase are not used to replace the pressure change during the setup phase.

[0120] Step S24: During the maintenance phase, the sampling process of the establishment phase continues to continuously collect pressure data on the ventilation hole branch to be tested and the reference orifice branch to obtain the pressure change during the maintenance phase, which characterizes the continuous suction process.

[0121] The specific implementation of step S24 is as follows: After the establishment phase is completed, the sampling starting point is not changed, the sampling sequence number is not reset, and the unified sampling time sequence is directly continued. right and Continue collecting, and The pressure change during the maintenance phase is determined to reflect the continuous passage process of the measured ventilation hole branch and the reference orifice branch under continuous suction. The single-point pressure value of the maintenance phase is not taken to replace the pressure change during the maintenance phase.

[0122] Step S25: During the release phase, the sampling process of the maintenance phase is continued to continuously collect pressure data on the branch of the ventilation hole to be tested and the branch of the reference orifice to obtain the pressure change during the release phase, which characterizes the negative pressure release and fall process.

[0123] The specific implementation of step S25 is as follows: after the maintenance phase ends, the unified sampling time sequence continues. right and Collection, will and The pressure change during the release phase is identified as a continuous rebalancing process of the pressure drop in the two branches after the negative pressure is released. The pressure changes during the maintenance phase and the pressure changes during the release phase retain the original sampling order and the original adjacency relationship. The single-point pressure value of the release phase is not taken to replace the pressure change during the release phase.

[0124] Step S26: Connect the pressure changes during the establishment, maintenance, and release phases of the ventilation hole branch to be tested in the sampling order to obtain the continuous pressure change of the ventilation hole branch to be tested. Connect the pressure changes during the establishment, maintenance, and release phases of the reference orifice branch in the sampling order to obtain the continuous pressure change of the reference orifice branch. The continuous pressure changes of the ventilation hole branch to be tested and the continuous pressure changes of the reference orifice branch are determined as the branch pressure changes.

[0125] The specific implementation method of step S26 is as follows: , and Connect them in ascending order of index to form a continuous pressure change of the branch of the ventilation hole to be tested. , ;Will , and Connect them in ascending order of index to form a continuous pressure variation of the reference orifice branch. , Maintain during this connection process. , and The preceding and following boundaries remain unchanged, ensuring that the sequential relationship between the establishment phase, maintenance phase, and release phase in the complete sequence remains unchanged. and Together, they constitute the branch pressure change, which is represented by two one-dimensional pressure sequences. The dimension of each one-dimensional pressure sequence is... In one implementation, the dimension is ninety-six. The pressure changes during the establishment, maintenance, and release stages correspond to thirty-two continuous subsequences in two one-dimensional pressure sequences, respectively. Branch pressure changes are not included. , and Instead of performing threshold determinations separately, the data is used as direct input for branch differential pressure calculation. The branch differential pressure is then used to determine the bridge anchorage baseline and generate the bridge imbalance recovery trajectory.

[0126] In this embodiment, step S3 includes steps S31 to S35, as referenced. Figure 4 The details are as follows:

[0127] Step S31: Based on the pressure values ​​of the ventilation hole branch to be tested and the reference orifice branch at the corresponding sampling positions in the branch pressure change, calculate the difference in the branch differential pressure according to the unified sampling order.

[0128] The specific implementation method of step S31 is as follows: Let ( ( ) represents the continuous pressure change in the branch of the ventilation opening to be measured. ( The continuous pressure variation in the reference orifice branch is represented by [value]. For each sampling location... ,Will and By subtracting corresponding values ​​at the same location according to a unified sampling starting point and a unified sampling order, the branch differential pressure sequence is obtained. ,in, For the first The differential pressure sampling values ​​of the branches at each sampling location, in kPa, represent the relative response of the two branches of the bridge detection unit under the same short-duration negative pressure pulse, and replace the single-point pressure value in the bridge state characterization link. ;

[0129] Step S32: Extract the common reference position before the negative pressure pulse is applied and after the negative pressure pulse is released, and determine the bridge anchorage baseline based on the common reference position;

[0130] The specific implementation method of step S32 is as follows: Let This is a sequence of branch differential pressure changes between adjacent sampling locations;

[0131] in, For the first The sampling location and the first The absolute value of the branch differential pressure change between each sampling location, in kPa, is given. To set a preset baseline fluctuation threshold, in kilopascals, let... The preset stable sampling number is expressed in units of sampling locations;

[0132] The candidate sampling positions before the application of the short-duration negative pressure pulse are limited to continuous sampling positions near the initial boundary of the setup phase, and from... Selecting from those that meet the requirements And the number of consecutive sampling locations is not less than The continuous sampling interval was determined as the stable sampling interval before the application of the short-time negative pressure pulse. The candidate sampling positions after the release of the short-term negative pressure pulse are limited to continuous sampling positions near the termination boundary of the release phase. Continuous sampling intervals that meet the conditions are selected in the same manner and determined as the stable sampling interval after the release of the short-term negative pressure pulse. When multiple consecutive sampling intervals satisfy the conditions, the consecutive sampling interval closest to the boundary of the short-time negative pressure pulse application is determined as... The continuous sampling interval closest to the termination boundary of the short-time negative pressure pulse release is defined as... . and Together, they form a common reference position before the application of the short-term negative pressure pulse and after the release of the short-term negative pressure pulse.

[0133] set up To establish the bridge anchorage baseline, the unit is kPa, and the stable sampling interval is determined. and stable sampling interval The differential pressure samples from the branches within the bridge are used together to calculate the bridge anchorage baseline, and the stability is weighted according to the change in differential pressure at each sampling location, using the following calculation formula:

[0134] ;

[0135] In the formula, To sum up each sample location within a specified stable sampling interval, each sample location within the stable sampling interval must satisfy... ,therefore The value is positive and serves as a stable weight for the corresponding sampling location. The bridge anchorage baseline is directly generated by the branch differential pressure of the current detection cycle, without using a fixed constant or device calibration parameters.

[0136] Step S33: Using the bridge anchorage baseline as a reference, continuously characterize the deviation direction, deviation magnitude, and duration of the branch differential pressure relative to the bridge anchorage baseline to obtain the branch differential pressure deviation characterization results.

[0137] The specific implementation method of step S33 is as follows: After the bridge anchorage baseline is determined, for each sampling position Will and By subtracting, we obtain the bridge state deviation sequence:

[0138] ;

[0139] in, For the first The bridge state deviation values ​​of each sampling location relative to the bridge anchorage baseline, in kPa. The symbolic representation deviates from the direction. The numerical value represents the magnitude of the deviation, and the number of sampling positions that continuously maintain the same deviation direction represents the duration. This refers to the branch differential pressure deviation characterization result obtained after continuously characterizing the deviation direction, deviation magnitude, and duration of the branch differential pressure relative to the bridge anchorage baseline. This characterization result serves as the common basis for extracting the three stages of the process in step S34.

[0140] Step S34: Based on the branch differential pressure deviation characterization results, extract the bridge imbalance initiation process in the establishment phase, extract the bridge imbalance persistence process in the maintenance phase, and extract the bridge rebalancing regression process in the release phase. The bridge rebalancing regression process is the continuous change process of branch differential pressure from deviation from the bridge anchoring baseline to entering the preset rebalancing judgment threshold range.

[0141] The specific implementation method of step S34 is as follows: based on the branch differential pressure deviation characterization results A stage index window will be established. China meets the first time The sampling location was determined as the starting point of bridge imbalance. ,Will This was determined to be the initiation process of bridge imbalance. It was determined to be a continuous process of bridge imbalance.

[0142] Index window during the release phase The internally determined bridge equilibrium regression process is as follows: Let... To preset the rebalancing threshold, in kilopascals, let... The preset number of samples is expressed in units of sampling locations. To rebalance the endpoint index, the release phase index window is... within Perform continuous counting, when Continuously located in the interval The number of sampling locations within the area reached a new high for the first time. When this condition is first met, the sampling location will be determined as... This confirms that the bridge rebalancing is complete. Thus, the continuous change in branch differential pressure from deviation from the bridge anchorage baseline to entering the preset rebalancing threshold range is limited to the sampling location. To the sampling location The continuous change process between these points is called the bridge equilibrium regression process.

[0143] Step S35: Connect the bridge imbalance initiation process, bridge imbalance duration process, and bridge rebalancing regression process in the original sampling order to obtain the bridge imbalance rebalancing trajectory.

[0144] The specific implementation of step S35 is as follows: connect the bridge imbalance initiation process, the bridge imbalance duration process, and the bridge rebalancing regression process in the original sampling order to generate the bridge imbalance rebalancing trajectory. and keep , and The segment boundaries remain unchanged. Bridge anchorage baseline. The anchor input that forms the basis of the subsequent deep learning model has a dimension of 1; the bridge imbalance and rebalancing trajectory. The main input to the subsequent deep learning model has dimension Q, which is ninety-six dimensions in one implementation. These dimensions correspond sequentially to the thirty-two-dimensional trajectory inputs for the setup phase, the maintenance phase, and the release phase. The deep learning model receives the bridge state change of the ventilation hole branch under test relative to the reference orifice branch, and does not directly compare the branch differential pressure as the final judgment value.

[0145] refer to Figure 8 , Figure 8 The figure shows the distribution of the bridge imbalance recovery trajectory in the bridge state space after dimensionality reduction under different ventilation hole conditions. The horizontal axis represents the first principal component PC1, the vertical axis represents the second principal component PC2, hollow circles represent samples in the unobstructed state, and solid triangles represent samples in the glue-blocked state. Solid and dashed lines represent the 95% confidence ellipses of the two types of samples, respectively. As can be seen from the figure, the unobstructed state samples are mainly concentrated in the narrow left region, while the glue-blocked state samples are mainly distributed in the right region. The two types show a clear spatial separation, indicating that glue blockage does not only cause a single point of pressure anomaly, but also leads to a systematic shift in the entire bridge imbalance recovery trajectory. This verifies the rationality of using the complete bridge imbalance recovery trajectory rather than a single point of pressure value for judgment, which is beneficial for improving the ability to distinguish the blocking state and the noise resistance stability.

[0146] In this embodiment, step S4 is generally implemented as follows: the deep learning model uses the bridge anchoring baseline as an anchoring reference, and performs regression calculation on the bridge imbalance recovery trajectory based on the regression relationship established in advance based on unobstructed samples to obtain the unobstructed bridge state corresponding to the bridge anchoring baseline; then, based on the difference between the unobstructed bridge state and the bridge imbalance recovery trajectory, the bridge state deviation is obtained. The above regression relationship enables the deep learning model to process the bridge state change of the tested ventilation hole branch relative to the reference orifice branch, without directly identifying anomalies in the original branch pressure change.

[0147] In a preferred embodiment, the deep learning model is a bridge-anchored complex equilibrium score-based diffusion model, and step S4 includes steps S41 to S45, referencing... Figure 5 The details are as follows:

[0148] Step S41: Divide the bridge imbalance recovery trajectory into three trajectory inputs according to the establishment phase, maintenance phase and release phase, and input them into the three trajectory encoding layers respectively to obtain the establishment phase features, maintenance phase features and release phase features;

[0149] The specific implementation method of step S41 is as follows: Let ( (This is the trajectory of the bridge's unbalanced and rebalanced state.) , , These are the establishment phase index window, the maintenance phase index window, and the release phase index window, which together correspond to 96 sampling positions, with 32 sampling positions for the establishment phase, 32 for the maintenance phase, and 32 for the release phase. The bridge imbalance and rebalancing trajectory is then... , and The input is divided into three 32-dimensional trajectory inputs: setup phase, maintenance phase, and release phase. The three trajectory inputs are fed into three trajectory encoding layers. Each of the three branches adopts a two-layer fully connected structure, with the first layer having 64 neurons and the second layer having 32 neurons, to obtain setup phase features, maintenance phase features, and release phase features, respectively.

[0150] Step S42: Input the bridge anchoring baseline into the bridge anchoring coding layer to obtain the anchoring features, and input the diffusion step conditions into the diffusion step coding layer to obtain the diffusion step features;

[0151] The specific implementation method of step S42 is as follows: Let To determine the baseline for bridge anchorage, For the diffusion step condition, , This represents the total number of diffusion steps. The bridge is anchored to the baseline. The input is the bridge anchoring coding layer, which employs a two-layer fully connected structure. The first layer has sixteen neurons, and the second layer has thirty-two neurons, resulting in thirty-two-dimensional anchoring features. The diffusion step condition is then applied. The input is a diffusion step encoding layer, which adopts a two-layer fully connected structure. The first layer has sixteen neurons, and the second layer has sixteen neurons, resulting in a sixteen-dimensional diffusion step feature.

[0152] Step S43: Input the establishment phase features, maintenance phase features, release phase features, anchoring features and diffusion step features into the bridge state fusion layer, and splice them in a fixed order to obtain unified bridge state fusion features;

[0153] The specific implementation of step S43 is as follows: the establishment phase features, maintenance phase features, release phase features, anchoring features and diffusion step features are spliced ​​together in a fixed order to form a 144-dimensional unified bridge state fusion feature, which is then input into the rebalancing regression layer. The rebalancing regression layer adopts a three-layer fully connected structure with 128, 128 and 64 neurons respectively.

[0154] Step S44: Establish the regression space of the unoccluded bridge state trajectory based on the unoccluded samples, and perform regression calculation on the unified bridge state fusion features by the rebalancing regression layer to obtain the unoccluded bridge state trajectory corresponding to the bridge anchorage baseline.

[0155] The specific implementation method of step S44 is as follows: Let the first... The range of values ​​for the bridge anchorage baseline is as follows: , ,in, The number of values ​​for the bridge anchorage baseline. For the first The lower boundary of the range of values ​​for the bridge anchorage baseline. For the first The upper boundary of the range of values ​​for the bridge anchorage baseline, in units of kPa, is given. For the pre-stored first The unobstructed bridge-state trajectory in the first The sampled values ​​at each sampling location, For the first The bridge anchorage baseline corresponding to the unobstructed bridge trajectory , This represents the number of unoccluded samples. In this embodiment, the unoccluded bridge state trajectory is the specific manifestation of the unoccluded bridge state in this embodiment, that is, the point-by-point sequence corresponding to the unoccluded bridge state within the current detection period.

[0156] anchoring the bridge to the baseline By comparing each of the multiple bridge anchorage baseline ranges, the unique one that satisfies the condition is determined. The range of bridge anchorage baseline values ​​is then used to determine all values ​​that meet the criteria. The unobstructed bridge trajectory is determined by the constraint of both the range of bridge anchorage baseline values ​​and the numerical distance of the bridge anchorage baseline. The unobstructed bridge state corresponding to the current bridge anchorage baseline is denoted as... , ( ), and is constructed using the following calculation formula for each sampling position:

[0157] ;

[0158] In the formula, The attenuation factor for the bridge anchorage baseline distance, expressed in kilopascals. This is an indicator function; its value is [value] when the condition is true. When the condition is not met, the value is taken as... , It is an exponential function.

[0159] Through this calculation formula, the unoccluded bridge state trajectory is directly established by the invoked unoccluded samples without using a fixed template. This process is to establish the regression space of the unoccluded bridge state trajectory based on the unoccluded samples.

[0160] Within the current detection cycle, prediction is performed using a decreasing diffusion step approach: Let... The diffusion step condition is Time The one-step regression trajectory value at each sampling location, in kPa, is first... Set as Press again The following steps are performed sequentially: three-segment trajectory splitting, three-segment trajectory encoding, bridge anchoring encoding, diffusion step encoding, bridge state fusion, and rebalancing regression (i.e., repeating steps S41 to S43, and using...). (using the regression reference to perform the calculation of the balanced regression layer), resulting in... When decreasing to At that time, , ( ),in, For the first The unoccluded bridging trajectory values ​​at each sampling location are expressed in kPa. The 64-dimensional complex equilibrium regression features are input into the unoccluded bridging output layer, which employs 96 output neurons, corresponding to the output... This refers to the unobstructed bridge trajectory corresponding to the bridge anchorage baseline.

[0161] refer to Figure 10 , Figure 10 The figure shows the mean and standard deviation distributions of the unobstructed bridge trajectory under different bridge anchorage baseline conditions. The horizontal axis represents the sampling location, and the vertical axis represents the bridge anchorage baseline bins. The dashed lines divide the sampling process into the setup, maintenance, and release phases. As can be seen from the figure, the mean distribution of the unobstructed bridge trajectory differs significantly under different bridge anchorage baseline conditions, indicating that the normal bridge state is not a fixed, unchanging template. Simultaneously, the standard deviation exhibits different levels at different stages, indicating that the stability and fluctuation characteristics of each stage are not the same. This verifies the necessity of using the bridge anchorage baseline as the anchorage input, segmenting and encoding the three-stage trajectory, and regressing it to the corresponding unobstructed bridge trajectory. This allows the model to establish a more accurate normal reference for the current working condition, improving the adaptability and reliability of online judgment.

[0162] Step S45: Combine the difference expression between the unobstructed bridge trajectory and the bridge imbalance recovery trajectory with the anchoring feature input to the bridge deviation output layer to map the degree of deviation of the bridge imbalance recovery trajectory relative to the unobstructed bridge trajectory, and obtain the bridge deviation amount.

[0163] The specific implementation method of step S45 is as follows: and Differential expression sequences were obtained by subtracting each point from the uniform sampling location. , ( ),in, For the first Bridge-state difference values ​​at each sampling location, in kPa. The ninety-six-dimensional differential expression sequence... The 32-dimensional anchoring features are input into the bridging state deviation output layer, which uses 32 neurons and one output neuron to obtain the one-dimensional bridging state deviation. The input of the bridge anchorage rebalancing score-based diffusion model consists of a 96-dimensional bridge imbalance rebalancing trajectory, a 1-dimensional bridge anchorage baseline, and a 1-dimensional diffusion step condition. The output consists of a 96-dimensional unobstructed bridge state trajectory and a 1-dimensional bridge state deviation. This allows the model to perform regression calculations around the bridge state changes of the tested ventilation hole branch relative to the reference orifice branch, without directly identifying anomalies in the original branch pressure changes.

[0164] In this embodiment, a preset occlusion determination threshold is used. Before online determination, it is pre-determined as follows: Let... For the first Bridge state deviation corresponding to a known unobstructed state sample. ,set up For the first Bridge state deviation corresponding to a known occlusion state sample ,set up Given the number of samples in the known unobstructed state. Given the number of known occlusion state samples, each known unobstructed state sample is input into the bridge-anchored complex equilibrium Score-based Diffusion model, and the corresponding values ​​are recorded. Each known occlusion state sample is input into the bridge-anchored complex equilibrium Score-based Diffusion model, and the corresponding values ​​are recorded. Take all again Take the maximum value in the range. The minimum value in the sample is obtained, and when the maximum value of the bridge state deviation in the known unobstructed state is less than the minimum value of the bridge state deviation in the known occluded state, the arithmetic mean of these two values ​​is calculated. ,Finish Once confirmed, Write to the parameter storage area and retain it during online detection. constant;

[0165] refer to Figure 9 , Figure 9 The diagram illustrates the joint distribution relationship between the bridge anchorage baseline and the bridge state deviation. The horizontal axis represents the bridge anchorage baseline, the vertical axis represents the bridge state deviation, hollow circles represent samples in the unobstructed state, solid triangles represent samples in the glue-occluded state, solid and dashed lines represent the isodense distribution contours of the two types of samples, and dotted lines represent the preset occlusion judgment threshold. As shown in the figure, the samples in the unobstructed state are mainly clustered below the threshold line and are relatively concentrated, while the samples in the glue-covered state are mainly distributed near and above the threshold line, forming a density region that is significantly different from that of the unobstructed samples. This indicates that the method of first using the bridge anchoring baseline to complete the working condition anchoring and then using the bridge state deviation as a unified judgment quantity can still maintain a good state differentiation effect under different baseline conditions, thereby reducing the impact of device fluctuations, clamping differences and gas path disturbances on the final judgment result.

[0166] Step S5 includes steps S51 to S53, see reference. Figure 6 The details are as follows:

[0167] Step S51: Compare the bridge state deviation with the preset occlusion judgment threshold once to determine whether the bridge state deviation is greater than or equal to the preset occlusion judgment threshold.

[0168] The specific implementation method of step S51 is as follows: Let The bridge state deviation output by the score-based Diffusion model for bridge anchorage. For a one-dimensional real number, To and A one-dimensional real number residing in the same output scalar space. Within the current detection period, The data is sent to the comparison unit and simultaneously read from the parameter storage area. The comparison unit performs only one numerical comparison, does not set comparison thresholds for the setup, maintenance, and release phases, does not compare the unobstructed bridge trajectory point by point, and does not compare the bridge imbalance and rebalancing trajectory point by point. The comparison rule is fixed as a judgment. Is it greater than or equal to? This allows the comparison result to directly correspond to the degree of deviation of the current bridge imbalance recovery trajectory from the unobstructed bridge trajectory. The state determination step only reads... and Two values ​​are used: continuous pressure change of the branch of the ventilation hole to be tested, continuous pressure change of the branch of the reference hole, differential pressure of the branch, and point-by-point sampling value in the bridge imbalance recovery trajectory. Therefore, the judgment object is limited to the bridge state deviation output by the bridge anchorage recovery Score-based Diffusion model, instead of directly comparing the branch pressure change and the branch differential pressure with thresholds.

[0169] Step S52: When the bridge state deviation is greater than or equal to the preset occlusion determination threshold, the state determination result is determined as the occlusion determination result.

[0170] The specific implementation method of step S52 is as follows: Let The result of the state determination. It is represented using one-dimensional discrete numerical representation. When At that time, Assigned value and will The corresponding definition is the occlusion determination result.

[0171] Step S53: When the bridge state deviation is less than the preset occlusion judgment threshold, the state judgment result is determined to be the smooth judgment result.

[0172] The specific implementation method of step S53 is as follows: when At that time, Assigned value and will The corresponding definition is the smoothness determination result. Therefore, the state determination result... Only by one-dimensional bridge state deviation Compared with a one-dimensional preset occlusion detection threshold In a single comparison generation process, the state determination stage only adds a one-dimensional discrete result after the model output. This allows the judgment object to remain as the bridge state change of the ventilation hole branch to be tested relative to the reference orifice branch.

[0173] In this embodiment, based on the distinction between occlusion determination results and unobstructedness determination results in the state determination results, step S6 includes steps S61 to S62, as referred to... Figure 7 The details are as follows:

[0174] Step S61: When the state determination result is the occlusion determination result, the ventilation hole is determined to be in the glue occlusion state.

[0175] The specific implementation method of step S61 is as follows: Let The final online determination result for the ventilation hole of the cigarette to be tested. For a one-dimensional discrete result, let... For the current semantic judgment, To and One-to-one correspondence of one-dimensional semantic tags. When When =1, the current semantic judgment is... The semantics for determining obstruction are defined, and the ventilation hole of the cigarette under test is determined to be obstructed by glue. The value is assigned to 1, and 1 is defined as the final online judgment result corresponding to the glue obscuring state. This determination process no longer introduces a second comparison threshold or manual correction rules, but directly uses the existing method. =1 corresponds to the occlusion determination semantics.

[0176] Step S62: When the status determination result is unobstructed, the ventilation hole is determined to be unobstructed.

[0177] The specific implementation method of step S62 is as follows: when At that time, the current judgment semantics The semantics for determining patency are defined, and the ventilation hole of the cigarette being tested is determined to be in a patency state. The value is assigned to 0, and 0 is defined as the final online judgment result corresponding to the unobstructed state. This determination process also no longer introduces a second comparison threshold, and no longer performs repeated discrimination on the unobstructed bridge trajectory, but directly uses the existing method. The corresponding fluency judgment semantics. Because... correspond , correspond Current semantic judgment Deviation from bridge state The comparison conclusions maintain a fixed correspondence, ensuring that the ventilation hole status of the cigarette under test is always determined based on the regression judgment results of the bridge anchorage complex balance Score-based Diffusion model.

[0178] Through the above semantic mapping and result assignment, the state of glue obstruction or unobstructed flow is directly determined as the final online judgment result of the ventilation hole of the cigarette under test. ,in, Based solely on the state determination result Mapped to obtain, Bridge state deviation only Compared with the preset occlusion detection threshold The single comparison generation ensures that the final online determination result of the ventilation hole of the cigarette under test remains consistent with the regression determination link of the bridge anchorage complex balance Score-based Diffusion model.

[0179] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the embodiments of the present invention have been described in detail, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of protection of the claims of the present invention.

Claims

1. A deep learning-based online method for determining adhesive obstruction in ventilation holes, characterized in that, include: Step S1: Receive the cigarette to be tested, connect the ventilation hole area of ​​the cigarette to be tested to the ventilation hole branch, and connect the reference orifice to the reference orifice branch, so that the ventilation hole branch to be tested and the reference orifice branch form a bridge detection unit. Step S2: Apply a negative pressure pulse to the bridge-type detection unit under the constraint of a shared negative pressure source, and collect the pressure changes of the ventilation hole branch to be tested and the reference orifice branch during the establishment, maintenance and release phases of the negative pressure pulse to obtain the branch pressure change. Step S3: Calculate the branch differential pressure based on the branch pressure change, determine the bridge anchorage baseline based on the branch differential pressure, and generate the bridge imbalance recovery trajectory based on the continuous change of the branch differential pressure. Step S4: Input the bridge anchoring baseline and the bridge imbalance recovery trajectory into the deep learning model. The deep learning model uses the bridge anchoring baseline as an anchoring reference and performs regression calculation on the bridge imbalance recovery trajectory based on the regression relationship established in advance based on unobstructed samples to obtain the unobstructed bridge state corresponding to the bridge anchoring baseline. Based on the difference between the unobstructed bridge state and the bridge imbalance recovery trajectory, the bridge state deviation is obtained. Step S5: Compare the bridge state deviation with a preset occlusion determination threshold to obtain the state determination result; Step S6: Determine whether the ventilation hole is blocked by glue or unobstructed based on the state determination result.

2. The online determination method for adhesive obstruction of ventilation holes based on deep learning according to claim 1, characterized in that, Step S1 includes: Step S11: Send the cigarette to be tested into the positioning clamping part, and perform axial limiting and circumferential orientation on the cigarette to be tested, so that the coverage area of ​​the ventilation hole area corresponds to the air intake window of the ventilation hole branch to be tested, and form a local detection cavity surrounding the ventilation hole area. Step S12: Press the sealing boundary of the local detection cavity onto the outer surface of the cigarette to be tested, and close the non-detection air intake path outside the ventilation hole area, so that the gas received by the ventilation hole branch to be tested only enters through the ventilation hole area; Step S13: Install the reference orifice on the air inlet end of the reference orifice branch, and seal and fix the connection position between the reference orifice and the reference orifice branch so that the reference orifice becomes the only air inlet boundary of the reference orifice branch. Then connect the ventilation hole branch to be tested and the reference orifice branch to the same negative pressure interface. Step S14: Based on the synchronous pressure relationship between the ventilation hole branch to be tested and the reference orifice branch under the same negative pressure interface, the bridge detection unit is formed, so that the reference orifice branch participates in the formation of the branch pressure change.

3. The online determination method for adhesive obstruction of ventilation holes based on deep learning according to claim 1, characterized in that, Step S2 includes: Step S21: Connect the shared negative pressure source to the bridge detection unit through the same negative pressure control path, and apply unified negative pressure start-stop control to the ventilation hole branch to be tested and the reference orifice branch at the same time, so that the ventilation hole branch to be tested and the reference orifice branch are subjected to the same suction action in the same detection cycle. Step S22: Drive the bridge detection unit with the negative pressure pulse so that the negative pressure pulse sequentially forms the establishment phase, the maintenance phase and the release phase, and keep the ventilation hole branch under test and the reference orifice branch in a state of synchronous pressure and synchronous pressure release during the establishment phase, the maintenance phase and the release phase; Step S23: In the establishment phase, continuous pressure data are collected from the branch of the ventilation hole to be tested and the branch of the reference orifice at the same sampling starting point to obtain the pressure change in the establishment phase that characterizes the negative pressure propagation process. Step S24: In the maintenance phase, the sampling process of the establishment phase is continued to continuously collect pressure data on the branch of the ventilation hole to be tested and the branch of the reference orifice to obtain the pressure change in the maintenance phase that characterizes the continuous suction process. Step S25: In the release phase, the sampling process of the maintenance phase is continued to continuously collect pressure data on the branch of the ventilation hole to be tested and the branch of the reference orifice to obtain the pressure change in the release phase that characterizes the negative pressure release and fall process. Step S26: Connect the pressure changes during the establishment, maintenance, and release phases of the ventilation hole branch to be tested in the sampling order to obtain the continuous pressure change of the ventilation hole branch to be tested. Connect the pressure changes during the establishment, maintenance, and release phases of the reference orifice branch in the sampling order to obtain the continuous pressure change of the reference orifice branch. Determine the continuous pressure change of the ventilation hole branch to be tested and the continuous pressure change of the reference orifice branch as the branch pressure change.

4. The online determination method for adhesive obstruction of ventilation holes based on deep learning according to claim 1, characterized in that, Step S3 includes: Step S31: Based on the pressure values ​​of the ventilation hole branch to be tested and the reference orifice branch at the corresponding sampling positions in the branch pressure change, calculate the difference in the sampling order to obtain the branch differential pressure. Step S32: Extract the common reference position before the negative pressure pulse is applied and after the negative pressure pulse is released, and determine the bridge anchorage baseline based on the common reference position; Step S33: Using the bridge anchorage baseline as a reference, continuously characterize the deviation direction, deviation magnitude, and duration of the branch differential pressure relative to the bridge anchorage baseline to obtain the branch differential pressure deviation characterization result; Step S34: Based on the branch differential pressure deviation characterization results, extract the bridge imbalance initiation process in the establishment stage, extract the bridge imbalance persistence process in the maintenance stage, and extract the bridge rebalancing regression process in the release stage. The bridge rebalancing regression process is the continuous change process of the branch differential pressure from deviating from the bridge anchoring baseline to entering the preset rebalancing judgment threshold range. Step S35: Connect the bridge imbalance initiation process, the bridge imbalance duration process, and the bridge rebalancing regression process in the original sampling order to obtain the bridge imbalance rebalancing trajectory.

5. The online determination method for adhesive obstruction of ventilation holes based on deep learning according to claim 1, characterized in that, The deep learning model is a bridge-anchored complex equilibrium score-based Diffusion model, and step S4 includes: Step S41: Divide the bridge imbalance recovery trajectory into three trajectory inputs according to the establishment phase, the maintenance phase and the release phase, and input them into the three trajectory encoding layers respectively to obtain the establishment phase features, maintenance phase features and release phase features; Step S42: Input the bridge anchoring baseline into the bridge anchoring coding layer to obtain anchoring features, and input the diffusion step conditions into the diffusion step coding layer to obtain diffusion step features; Step S43: Input the establishment phase features, the maintenance phase features, the release phase features, the anchoring features, and the diffusion step features into the bridge state fusion layer, and splice them in a fixed order to obtain unified bridge state fusion features; Step S44: Establish a regression space for the unoccluded bridge state trajectory based on the unoccluded samples, and perform regression calculation on the unified bridge state fusion features by the rebalancing regression layer to obtain the unoccluded bridge state trajectory corresponding to the bridge anchoring baseline; Step S45: Combine the difference expression between the unobstructed bridge trajectory and the bridge imbalance recovery trajectory with the anchoring feature input bridge deviation output layer to map the degree of deviation of the bridge imbalance recovery trajectory relative to the unobstructed bridge trajectory, and obtain the bridge deviation amount.

6. The online determination method for adhesive obstruction of ventilation holes based on deep learning according to claim 1, characterized in that, Step S5 includes: Step S51: Compare the bridge state deviation with the preset occlusion determination threshold once to determine whether the bridge state deviation is greater than or equal to the preset occlusion determination threshold. Step S52: When the bridge state deviation is greater than or equal to the preset occlusion determination threshold, the state determination result is determined to be an occlusion determination result; Step S53: When the bridge state deviation is less than the preset occlusion determination threshold, the state determination result is determined to be a smooth state determination result.

7. The online determination method for adhesive obstruction of ventilation holes based on deep learning according to claim 6, characterized in that, Step S6 includes: Step S61: When the state determination result is the obstruction determination result, the ventilation hole is determined to be the glue obstruction state; Step S62: When the state determination result is a smooth state determination result, the ventilation hole is determined to be in the smooth state.

8. The online determination method for adhesive obstruction of ventilation holes based on deep learning according to claim 4, characterized in that, In step S32, the common reference position includes a stable sampling interval before the negative pressure pulse is applied and a stable sampling interval after the negative pressure pulse is released. The continuous sampling interval in which the branch differential pressure change between adjacent sampling positions before the negative pressure pulse is applied is less than the preset baseline fluctuation threshold is determined as the stable sampling interval before the negative pressure pulse is applied. The continuous sampling interval in which the branch differential pressure change between adjacent sampling positions after the negative pressure pulse is released is less than the preset baseline fluctuation threshold is determined as the stable sampling interval after the negative pressure pulse is released. The bridge anchorage baseline is determined based on the stable sampling interval before the negative pressure pulse is applied and the stable sampling interval after the negative pressure pulse is released.

9. The online determination method for adhesive obstruction of ventilation holes based on deep learning according to claim 4, characterized in that, In step S34, during the release phase, when the branch differential pressure enters the preset rebalancing determination threshold range, and the number of sampling positions within the preset rebalancing determination threshold range in the continuous sampling positions reaches a preset sampling number, it is determined that the bridge rebalancing is completed, and the sampling position that first satisfies the bridge rebalancing completion is determined as the rebalancing endpoint of the bridge imbalance rebalancing trajectory.

10. The online determination method for adhesive obstruction of ventilation holes based on deep learning according to claim 5, characterized in that, In step S44, the bridge anchoring baseline is matched with a plurality of pre-set bridge anchoring baseline value ranges, the unobstructed sample corresponding to the matching result is called, and the unobstructed bridge state trajectory corresponding to the bridge anchoring baseline is established based on the called unobstructed sample.

Images