Combustible gas non-inductive detection camera system integrated with spectrum sensor

Abstract

The invention discloses a combustible gas non-inductive detection camera system integrated with a spectrum sensor, and relates to the technical field of gas imaging detection, and the combustible gas non-inductive detection camera system comprises the following steps: establishing multi-source observation through a reference measurement frame which shares at least part of an optical link with a main imaging light path, and outputting a fixed reference area statistical abstract, a temperature observation vector and a reference data effective mark; updating the drift state quantity based on the statistical abstract and the temperature observation vector and generating a stability field; performing drift compensation on the infrared image sequence according to the drift state quantity and the stability field, and generating an operation envelope field, a quantization result, a confidence interval and error source decomposition; and finally, executing gating output and writing back scheduling parameters. According to the scheme, continuous connection of reference observation, drift estimation, quantitative judgment and feedback scheduling is achieved, and the output stability, the quantitative credibility and the continuous operation capacity under the fixed online detection condition are improved.

Description

Technical Field

[0001] This invention relates to the field of gas imaging detection technology, specifically to a non-contact detection camera system for combustible gases with integrated spectral sensors. Background Technology

[0002] Leakage sources are concentrated in chemical industrial parks, storage and transportation stations, pipe corridors, and loading and unloading areas, and are mostly located at high altitudes or in explosion-proof areas, making manual inspection difficult. PTZ is often used to achieve multi-position 24 / 7 remote alarm and recording, and can be linked with DCS / SCADA to form event logs, operation ledgers, and data archives. On-site temperature day and night cycles, heat source disturbances, rain, fog, dust, window contamination, condensation, etc. are engineering constraints that affect long-term performance and quantitative accuracy.

[0003] For example, US Patent Document US10684216B2 discloses a multi-band gas quantization and differentiation method for optical gas imaging cameras. This method involves a multi-band optical gas imaging camera acquiring a single multi-band radiation image, where each image pixel includes probe radiation information from different bands. During the calculation process, background radiation is first estimated for at least one pixel. The gas concentration-optical path integral value (OM~m·m along the optical path integral value) is then calculated based on the probe radiation and background radiation. An alarm is issued when the alarm conditions are met. The multi-band configuration may include a reference band located outside the target gas absorption window and a working band covering the absorption window. Background radiation estimation is achieved by establishing a relationship model between the working band probe radiation and the reference band probe radiation, using this model to extrapolate the working band background radiation in the presence / absence of gas. In species differentiation scenarios, a species differentiation threshold can be set to avoid false alarms caused by interfering gases. Alarms are output via display or audio-visual / communication methods.

[0004] The existing technologies mentioned above are only applicable to fixed 24 / 7, low-maintenance deployment conditions. The stability of quantization calculation requires the stability of the camera radiation response and multi-band relationship model. When the system uses an uncooled detector and lacks manual calibration for a long time, the detector sensitivity and bias will experience thermal drift with ambient temperature and slow drift with device aging. Window contamination / condensation changes optical transmittance, and the filter is affected by temperature and stress, resulting in effective spectral band shift.

[0005] The relative response relationship between the reference band and the working band caused by this drift deviates from the modeling assumptions. The changes in system response are mixed into the changes in detected radiation, resulting in systematic deviations or long-period fluctuations in the concentration-optical path integral value. When superimposed with strong non-steady-state background (heat source, solar reflection, steam / dust obstruction, wind-induced plume morphology changes, etc.), model mismatch is more easily amplified, leading to inconsistencies between event verification and report standards.

[0006] Therefore, the current technical problem to be solved is: how to ensure verifiable consistency of multi-spectral / spectral imaging quantization output across temperature ranges and during long-term operation under complex environmental disturbances and low maintenance constraints. Summary of the Invention

[0007] (a) Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides a non-contact combustible gas detection camera system integrating a spectral sensor. It updates drift state variables and generates a stability field based on statistical summaries and temperature observation vectors. Subsequently, it performs drift compensation on the infrared image sequence based on the drift state variables and stability field, generating an operation envelope field, quantization results, confidence intervals, and error source decomposition. Finally, it performs gated output and writes back the scheduling parameters. This scheme achieves seamless integration of reference observation, quantization judgment, and feedback scheduling, improving output stability, quantization reliability, and continuous operation capability under fixed online detection conditions; and solves the technical problems described in the background art.

[0009] (II) Technical Solution

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] A non-contact imaging system for combustible gas detection with integrated spectral sensors, comprising acquiring infrared image sequences in at least two spectral bands and performing gas detection, including:

[0012] Trigger reference sampling to generate a reference measurement frame that shares at least part of the optical link with the main imaging optical path, and extract a statistical summary of the fixed reference area, temperature observation vector, and valid reference data flags;

[0013] Based on the statistical summary, temperature observation vector, and valid reference data flags, update the drift state variables including channel gain drift and channel bias drift and generate a stability field. Freeze the update when the valid reference data flags are invalid.

[0014] Drift compensation is performed on infrared image sequences based on drift state variables and stability fields, and operation envelope fields, quantization results, confidence intervals, and error source decomposition are generated.

[0015] The gating output is decomposed based on the operation envelope field, stability field, confidence interval, and error source, and the updated scheduling parameters are used as the input for the next loop.

[0016] Furthermore, a reference measurement frame is generated in response to events such as reaching a preset reference sampling time, detecting an ambient temperature change rate exceeding a threshold, and triggering events during device operation state switching.

[0017] A fixed reference region is set in the reference measurement frame, and the pixel position of the fixed reference region remains unchanged between consecutive reference samples. When saturation, occlusion or insufficient number of pixels in the fixed reference region occurs, the current reference sample is marked as invalid, and delayed resampling, switching to the backup reference path and outputting the reference unavailable state are performed in sequence.

[0018] Furthermore, when generating the reference measurement frame, the reference measurement frame and the corresponding infrared image are kept to use the same integration time, the same analog gain, and the same non-uniform correction version; and the reference measurement frame index, fixed reference area statistical summary, temperature observation vector, and reference data validity flag are assembled into a unified input record according to a fixed field order.

[0019] Furthermore, based on the reference measurement frame index, the statistical summary of the fixed reference area and the temperature observation vector are aligned, and a drift state variable containing channel gain drift and channel bias drift is constructed. The drift state variable is then decomposed into a slow variable corresponding to the change in environmental thermal equilibrium and a fast variable corresponding to short-term disturbances.

[0020] Then, a stability field is generated based on the updated drift state variables. The stability field includes at least the residual statistics of the fixed reference area and the cross-spectral consistency index.

[0021] Furthermore, when the reference data validity flag is invalid, or the stability field exceeds the preset threshold, the drift state update is frozen, while the channel gain drift and channel offset drift corresponding to the previous valid version remain unchanged; at the same time, the version identifier and statistical summary corresponding to the current drift state are output, and the reason for this freeze is recorded.

[0022] Furthermore, the infrared image sequence is drift compensated based on the drift state quantity to form a multi-spectral radiation characterization; then, the background contrast, distance and field of view configuration, and plume motion proxy quantity are extracted from the multi-spectral radiation characterization to generate the operation envelope field; and when the stability field indicates instability, the instability information is written into the operation envelope field, and the plume region is identified and segmented based on the operation envelope field.

[0023] Furthermore, the concentration path length and velocity field are calculated based on the plume region, and the boundary flux is determined based on the velocity field when the leakage rate is required; then the drift residual statistics, segmentation uncertainty and velocity field uncertainty are written into the error source decomposition, and quantification results and confidence intervals are formed.

[0024] When the confidence level of the plume segmentation is insufficient, the velocity field does not converge, or the operation envelope field exceeds the valid boundary, only qualitative alarms and positioning information are output, and the current dominant error term is output simultaneously.

[0025] Furthermore, gating decisions are performed based on the stability field, operation envelope field, confidence interval, and error source decomposition, and quantization output status, restricted output status, and invalid output status are generated. In the quantization output status, the quantization result, confidence interval, operation envelope field, stability field, and event record index are output; in the restricted output status, qualitative alarms and location information are output; and in the invalid output status, alarms and maintenance prompts are output.

[0026] Furthermore, the current sampled event is written to the event log, which includes at least the device serial number, event timestamp, version number, stability field index, operation envelope field index, gating status identifier, and anomaly source field index. If the event log writing fails, the current sampled event is switched to a restricted output state, and the quantization result is prohibited from entering the report link, while the corresponding location information and anomaly source field are retained.

[0027] Furthermore, the scheduling parameters are updated based on the dominant error term and drift residual level in the error source decomposition. Specifically, the reference sampling period is shortened when the drift residual increases, the execution order of uncertainty propagation is adjusted when the computational load exceeds the threshold, and the updated scheduling parameters are written back, while keeping the field order of the gating decision chain and event record chain unchanged.

[0028] (III) Beneficial Effects

[0029] This invention provides a non-contact detection camera system for combustible gases with an integrated spectral sensor, which has the following advantages:

[0030] Generate a reference measurement frame that is similar to or has a few optical links to the main imaging optical path, and output a statistical summary of the fixed reference area, temperature observation vector, and valid reference data flag. This allows the reference observation to be in the same thermal environment and optical path as the main imaging link. It can integrate the detector response change and the lens housing flux change into a single input chain, avoiding the phenomenon of missing drift sources due to single temperature compensation.

[0031] The drift state quantity is updated and a stability field is generated based on the statistical summary of the fixed reference area, the temperature observation vector, and the validity flag of the reference data. This allows drift compensation to be established under continuous state updates and validity constraints. When the reference data is abnormal, the update can be frozen and the previous valid state can be maintained, suppressing unreliable references from entering the subsequent processing chain. The drift state quantity and the subsequent quantization chain are stably connected.

[0032] Drift compensation is performed on the infrared image sequence based on the drift state quantity and stability field, and an operational envelope field is generated. This enables the plume region identification, concentration path length and velocity field calculation to be performed on the image representation with boundary conditions, preventing interference from background changes, different fields of view and changes in on-site working conditions on the quantization results, and improving the scene adaptability of fixed online scenes.

[0033] By decomposing the operation envelope field, stability field, confidence interval, and error source into a gated output, the quantization result, qualitative alarm, and invalid indication correspond to the state conditions respectively. This avoids outputting only a few boundary constraint quantization results when the reference is unavailable, the state is unstable, or the envelope is limited. It also ensures that the output chain has the same decision logic as the preceding observation chain, estimation chain, and quantization chain, thereby improving the consistency of the overall output.

[0034] Error source decomposition and drift residual level are written back as scheduling parameters. The scheduling parameters are used as input for the next cycle. Reference sampling, drift state update, image compensation and gating output form a closed-loop processing chain to realize single detection in fixed long-term online applications. In fixed long-term online applications, the processing rhythm of the next round is adjusted according to the results of the previous round. Attached Figure Description

[0035] Figure 1 This is a schematic diagram showing the overall composition and board-level collaborative relationship of the integrated spectral sensor combustible gas non-contact detection camera system of the present invention;

[0036] Figure 2 This is a schematic diagram illustrating the introduction of the common optical path reference field and the locking of the fixed reference region in this invention;

[0037] Figure 3 This is a schematic diagram of the unified input encapsulation and dual-time-scale drift state estimation process of the present invention;

[0038] Figure 4 This is a schematic diagram illustrating the generation of stability fields and the handling of abnormal freeze inheritance in this invention;

[0039] Figure 5 This is a schematic diagram illustrating scene boundary recognition, candidate flow region filtering, and operation envelope field generation of the present invention.

[0040] Figure 6 This is a schematic diagram of the feathering, three-state gating output, and recording feedback scheduling of the present invention. Detailed Implementation

[0041] 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.

[0042] Please see Figures 1-6 This invention provides a non-contact detection camera system for combustible gases with an integrated spectral sensor.

[0043] Step 1: Without interrupting the fixed monitoring screen, establish a common optical path reference observation benchmark for each sampling moment that can simultaneously characterize the changes in detector response and lens housing radiation flux, and organize it into a unified input that can be directly called in Step 2.

[0044] In this embodiment, step one follows a single-chain logic: first, the reference is introduced into the main optical path; then, the reference is locked onto a fixed pixel; and finally, the reference is aligned with and encapsulated with multi-source observations. The common optical path reference field refers to the reference radiation field corresponding to the reference measurement frame sharing the lens assembly, window, part of the housing cavity, and the detector's incident path with the monitored scene before entering the detector. This ensures that changes in the reference measurement frame include both changes in the detector response and changes in the radiant flux returned inward from the lens housing.

[0045] When a fixed camera continuously faces the tank area, flange area, or valve group area, the temperature rise of the detector substrate, the effect of external heat sources on the lens assembly, and the heat accumulation on the inner wall of the housing will all be superimposed into the imaging link. If only the temperature is read and the reference radiation field is not sampled in the main imaging optical path, subsequent calculations can only be extrapolated based on the temperature, and it is impossible to distinguish between the two sources: the change in the detector's own response and the additional radiation changes of the lens housing.

[0046] Therefore, a common optical path reference field is first constructed, and then a fixed reference region is locked in the detector coordinate system so that the same batch of pixels have a consistent physical meaning when compared across time periods.

[0047] The acquisition control board does not keep the reference image in front of the lens for extended periods. Instead, it briefly introduces the reference radiation field and generates a reference measurement frame when the preset reference sampling time, the temperature change rate exceeds the trigger threshold, or the gimbal pose changes. Prolonged occlusion weakens continuous scene observation, while short-term introduction leaves a reference trace consistent with the main optical path without altering the main monitoring orientation.

[0048] The preferred method for introducing the common optical path reference field is a flip-type reference film scheme: a flip-drive component is arranged on the outer edge of the lens assembly, and the flip-drive component drives the reference film to enter the edge field of view of the lens around the bias axis; a high emissivity coating is formed on the surface of the reference film, and the high emissivity coating is selected from blackened aluminum oxide layer or high emissivity ceramic layer; the back of the reference film is attached to the metal bracket to ensure stable heat capacity.

[0049] In another embodiment, a built-in reference cavity scheme can be used, in which a bypass reference cavity is set between the lens assembly and the window plate. During sampling, the bypass baffle is switched so that the detector can simultaneously see the reference cavity exit and the edge of the main optical path within a short exposure window. Both schemes share the characteristic that the reference radiation field passes through the same lens assembly and window plate as the main scene before entering the detector; therefore, the additional radiation caused by the lens's own heating is synchronously superimposed on the reference measurement frame.

[0050] To avoid localized high-brightness areas skewing the reference statistics, the edge computing board... The reference response is obtained by using a logarithmic compressed average of the reference measurement frames for each spectral channel:

[0051]

[0052] Among them, reference response quantity : No. The compressed average radiative response of each spectral channel on the current reference measurement frame, taking a non-negative real number; spectral channel number. The value of is located in the interval A positive integer used to identify a specific channel in a multi-band infrared image sequence, representing the total number of spectral channels. A positive integer not less than 2; fixed reference region : A set of pixels pre-locked in the detector coordinate system, used to limit reference statistics to be performed only on the same batch of pixels;

[0053] Reference pixel count Fixed reference area The total number of pixels, a positive integer, is used to achieve comparable averaging across time steps; pixel index. Fixed reference area A single pixel location within the range, used to access the raw pixel response at that location; raw pixel response : No. Each spectral channel in pixel index The digital response value at that location is a non-negative real number, used to carry the instantaneous radiation information at that location;

[0054] Normalized reference quantity The normal values ​​preset on the acquisition control board are greater than 0, which are used to reduce the impact of dimensional differences between different spectral channels on the statistical results.

[0055] As a supplement: the reference measurement frame and its corresponding main scene frame use the same integration time, the same analog gain, the same bad pixel mask, and the same non-uniform correction version; if the system needs to temporarily change the exposure when triggering reference sampling, the reference measurement frame in this round must not be directly entered. Instead of writing the exposure switching flag first, the edge computing board should trigger resampling.

[0056] Taking the installation scenario of a tank area walkway as an example, after the acquisition control board reaches the reference sampling time, it first ends the current main scene exposure, and then drives the flip-type reference film to enter the edge field of view of the lens. The edge computing board receives the reference measurement frame obtained from this exposure. The visible action is that the reference film enters and immediately exits, and the main scene image resumes continuity; what remains in the system internally is the reference response of each spectral channel. This is related to the trigger source identifier. With this setting, the subsequent step two does not need to speculate on whether the lens heating is involved in the drift, because the reference measurement frame has already incorporated the thermal effects of the lens assembly and the window into the same set of observation data.

[0057] In use, the reference measurement frame and the main imaging optical path share the optical path, and the source of drift change obtained from subsequent drift state estimation is no longer limited to the electronic channel itself; the logarithmic compressed averaging reduces the pull of local spot points on the overall reference value, and the reference response is more stable; the flip-type reference plate scheme and the built-in reference cavity scheme retain the main monitoring direction, which is convenient for continued use in fixed online monitoring scenarios.

[0058] After obtaining the reference measurement frame, if the reference statistical area drifts with the image content, or the selected pixel position is inconsistent each time, then the reference response with the same name will be affected. This would compromise cross-time comparability. Based on this motivation, the edge computing board defines the fixed reference region directly within the detector coordinate system, rather than within the coordinate system of the monitored scene. The fixed reference region remains fixed along with the detector array and does not change with the position of tanks, valves, or plumes.

[0059] The preferred fixed reference area is located close to the edge of the image and free from mechanical obstructions, occupying 5-18 of the total detector pixels. Its shape can be rectangular, rounded rectangle, or annular, with a clearance of at least 10 pixels from the image border to prevent window edges or lens vignetting from directly entering the statistical area. Saturated pixels, obstructed pixels, and abnormally dark pixels are further removed from the fixed reference area. The edge computing board calculates the reference efficiency as follows:

[0060]

[0061] Among them, reference efficiency : No. The pixel ratio of each spectral channel within a fixed reference region satisfies the effective grayscale window, with values ​​falling within the range... Reference pixel count Pixel Index Fixed reference area and original pixel response All are consistent with the aforementioned definitions to ensure that the formulas are consistent with the preceding statistical objects; indicator functions The value is 1 when the condition inside the parentheses is true and 0 when the condition inside the parentheses is false. It is used to convert whether a pixel falls into the effective grayscale window into an accumulative discriminant.

[0062] Lower limit grayscale threshold A preset threshold greater than or equal to 0 is used to exclude abnormally low-response pixels caused by occlusion, loss of light, or cold spots; the upper limit grayscale threshold. For grayscale values ​​greater than the lower limit threshold A preset threshold is used to exclude abnormally high response pixels caused by saturation, bright reflection, or abnormal hot spots;

[0063] When the reference is effective When the data is below the efficiency threshold, the edge computing board does not update the reference statistics for the current spectral channel, but instead marks the channel as unavailable for reference. When multiple consecutive spectral channels are marked as unavailable for reference, the acquisition control board triggers compensation sampling.

[0064] In a preferred embodiment, the fixed reference region consists of a primary fixed reference region and a secondary fixed reference region. The primary fixed reference region is located on the upper left side of the screen, and the secondary fixed reference region is located on the lower right side of the screen. Both use the same area rules and grayscale window rules. The edge computing board first checks the reference efficiency of the primary fixed reference region. When the primary fixed reference region is affected by flying insects, raindrops, or brief occlusion, the system switches to the secondary fixed reference region. This switching action does not change the terminology of "fixed reference region"; it simply changes its internal index from the primary to the secondary fixed reference region. Therefore, subsequent step two still reads the same data field. This process prevents the system from completely discarding the entire reference sample due to a single instance of partial occlusion.

[0065] In use, the fixed reference area is directly locked within the detector coordinate system, ensuring that cross-time comparisons are established at the same pixel location; reference efficiency. Saturation, occlusion, and light loss are transformed into explicit data conditions, making it easier for the acquisition control board to perform compensated sampling; the switching between the main fixed reference area and the auxiliary fixed reference area still uses the same terminology system, reducing subsequent field interpretation splits.

[0066] Furthermore, the reference measurement frame, the multispectral infrared image sequence of the main scene, and the temperature samples of the lens assembly, detector substrate, and inner wall of the housing do not naturally fall at the same point in time. Without processing under the same time scale, the drift state quantity obtained in step two will mix the thermal states at different times with the reference responses at different times.

[0067] To this end, time alignment rules are first established, and then exception handling and output encapsulation rules are established, so that the input received in step two is a unified input under the same sampling event, rather than discrete fragments from loose sources.

[0068] The edge computing board writes a timestamp to each multi-spectral image frame, each temperature sample, and each reference measurement frame using a unified clock source, and uses the timestamp of the reference measurement frame as the central time stamp for the current observation encapsulation. Around this central time stamp, the system checks whether the main frame of each spectral channel and the temperature observation vector are within the allowed alignment window. The alignment window is preset by the acquisition control board. When the channel sampling frequency is higher than the reference sampling frequency, it selects one of the nearest main frames before and after the reference measurement frame for pairing; when the temperature sampling frequency is lower than the reference sampling frequency, the edge computing board only accepts the most recent temperature observation vector falling within the allowed temperature alignment window, without filling in across windows. This is because the reference measurement frame itself is triggered for drift observations; placing it in the center helps ensure that the reference, image, and temperature point to the same physical state.

[0069] The temperature observation vector preferably consists of three items: lens barrel temperature, detector substrate temperature, and housing inner wall temperature, arranged in a fixed order. When the system is configured with window edge temperature sampling points, the edge computing board adds this temperature as an extension item to the end of the temperature observation vector.

[0070] Taking the installation scenario on the pipe rack support as an example, the afternoon sunlight causes the temperature of the mirror tube to rise first, and the temperature of the inner wall of the casing continues to change with a lag after the air is stopped at night. If the time alignment only depends on a single temperature point, the source of drift will be compressed into a single heat source. However, by using a fixed sequence of temperature observation vectors and aligning them with the reference measurement frame, step two can observe the hierarchical changes in the thermal state from a unified entry point.

[0071] When in use, the reference measurement frame is used as the central time stamp, which can constrain multi-spectral infrared image sequences and temperature observation vectors to the same sampling event; the temperature observation vectors are entered into a unified input in a fixed order, which makes it easier for the subsequent drift state quantities to separate the effects of different heat sources; the alignment window refuses to fill across windows, avoiding the mixing of thermal states at different times in step two.

[0072] After timing alignment is completed, the system faces another type of problem: although the reference measurement frames are acquired on time, they may contain anomalies such as saturation, occlusion, localized dirt coverage, incomplete placement of the flipped reference chip, delayed temperature sampling, and board-level bus retransmission failure. If these anomaly observations are directly fed into step two, the drift state variables will be incorrectly manipulated. Therefore, the edge computing board performs unified reference anomaly handling before output. The judgment order is fixed as follows: first check the reference efficiency of the fixed reference area. Next, it checks whether each spectral channel has a full-frame saturation band, then checks whether the position feedback of the flip-drive device matches the timestamp of the reference measurement frame, and finally checks whether the temperature observation vector has arrived completely. If any of these checks fails, the edge computing board writes the reference data validity flag for this sampling event as invalid and records the anomaly source field.

[0073] After the reference data validity flag is generated, the acquisition control board performs compensatory sampling based on the source of the anomaly. If the source of the anomaly is short-term occlusion or insufficient positioning of the flipped reference film, reference sampling is retried in the next exposure cycle. If the source of the anomaly is mechanical jamming of the reference film, the built-in reference cavity is switched as a backup reference path. If the source of the anomaly is a missing term in the temperature observation vector, the edge computing board maintains the current multi-spectral infrared image sequence cache, waits for the next temperature sampling to complete it, and then reassembles the unified input according to the principle of nearest to the original center time scale. To ensure that the interface is fixed when calling step two, the edge computing board encapsulates the output into a structured record. The structured record contains at least the reference measurement frame index and the reference response of each spectral channel. Reference efficiency of each spectral channel The system outputs the temperature observation vector, trigger source identifier, and reference data validity flag. If the reference data validity flag is invalid, the structured record is still output, but the invalid state remains unchanged and is not masked or replaced in this step.

[0074] For example, the structured record uses a combination of a fixed-length header field and a variable-length channel data segment. The fixed-length header field sequentially contains the device serial number, reference measurement frame index, center time stamp, trigger source identifier, and reference data validity flag; the variable-length channel data segment sequentially contains the reference response values ​​for each spectral channel. and the reference efficiency of each spectral channel Finally, the temperature observation vector and verification field are placed. The fixed-length header result is written locally through a local circular cache and pushed to the drift state estimation corresponding to step two. When reading data in step two, there is no need to reverse-guess why there was no update last time; it only determines whether to freeze the drift state quantity based on the valid flag of the reference data.

[0075] In use, the reference anomaly handling process unifies mechanical non-position, image obstruction, and temperature missing data into a single valid reference data flag, clearly defining the entry conditions for the subsequent step two. Compensation sampling executes three paths along the anomaly source: resampling, backup path switching, and buffer reassembly, ensuring consistency between on-site actions and data flow. Structured recording includes the reference measurement frame index and reference response quantity. Reference efficiency The temperature observation vector and the reference data validity flag are placed in the same record to ensure that the input object name and field order received in step two are completely fixed.

[0076] Step 2: Based on the unified input formed in Step 1, the drift effects in the multi-spectral infrared image sequence are decomposed into continuously updatable drift state variables, and the drift state variables are further transformed into stability fields that can be directly called in Step 3.

[0077] In this embodiment, step two does not involve using the reference response quantity. Instead of directly writing it back to the image as a compensation parameter, it is first fed into the drift state estimation chain along with the temperature observation vector, and then the drift state estimation chain outputs the channel gain drift and channel bias drift with physical orientation.

[0078] When a fixed camera is pointed at a flange, pipeline, or tank area for an extended period, the drift is not a slow offset from a single source, but rather a superposition of changes in detector response, additional flux changes in the lens assembly, heat accumulation inside the housing, and transient heat exchange following attitude changes. If we consider the reference response quantity from step one... If directly mapped to static correction coefficients, the multi-spectral radiation characterization obtained in the subsequent step three will mix the slow-changing thermal equilibrium and the fast-changing thermal shock in the same compensation amount.

[0079] The reference response quantity given in step one Reference efficiency The temperature observation vector still belongs to the observation layer and is not directly equal to the parameters required by the image compensation layer. If the observation layer data is directly written into the compensation chain, the system can only operate under the mechanism of an immediate jump in compensation amount when the observation changes, which can easily lead to short-term thermal disturbances, local cooling, or tube heat transfer being mistakenly written as long-term drift.

[0080] To address this, a drift state variable is introduced to transform the discrete changes of the observation layer into the continuous evolution of the state layer, and then the state layer outputs usable drift compensation parameters to step three.

[0081] The drift solver first establishes a state vector according to the spectral channels. Each spectral channel corresponds to a set of channel gain drift and channel bias drift, and then further decomposes it into slow-varying components and fast-varying components.

[0082] The slow-varying component corresponds to the overall thermal equilibrium of the housing and the flux migration after the lens assembly is slowly heated, while the fast-varying component corresponds to the transient heat exchange near the reference sampling trigger, the short-term thermal shock after the gimbal is rotated, or the sharp shift caused by the passage of an external heat source. A dual timescale is used instead of a single timescale because fixed systems exhibit two types of shift phenomena with significantly different durations when switching between day and night, when ventilation stops, when supplementary lighting equipment is started or stopped, or when inspection vehicles pass by.

[0083] In order to truly couple the fields output from step one into the state variables, the drift solver plate performs a certain step on the first step. Each spectral channel is constructed with a reference deviation:

[0084]

[0085] Among them, reference deviation : No. The normalized deviation of the current reference response of each spectral channel from the reference response of that channel is represented by a real number; the reference deviation is... Used to characterize the current reference response. Relative to the reference response The degree of normalization deviation; the temperature convergence when the equipment is under high heat. Increase, thereby reducing the impact of reference changes on the same amplitude. The direct amplification effect;

[0086] Reference response quantity The meaning, source, and function are consistent with step one, and are used to characterize the current reference measurement frame in the [missing information]. Compressed average radiative response across each spectral channel;

[0087] Reference response : No. The reference response values ​​for each spectral channel, retained after initial calibration following equipment installation, are real numbers. After installation, under normal operating conditions of the reference field components, with integration time locked and analog gain locked, multiple sets of reference measurement frames are continuously acquired. The reference response value for each spectral channel... Take the median or truncated mean as the benchmark reference response. Then, based on the uniform field-of-view response during the initial installation phase, the fixed offset of each channel is calculated to form a baseline compensation term. . and All versions have version numbers and will only be updated during equipment maintenance, lens replacement, detector replacement, or recalibration.

[0088] Temperature coupling coefficient : The coupling weight of the temperature observation vector to the reference deviation normalization process, with a value of a real number greater than or equal to 0; temperature convergence. The thermal state convergence result, extracted from the temperature observation vector, is a non-negative real number used to reflect the current thermal state under the combined influence of the lens barrel temperature, detector substrate temperature, and housing inner wall temperature; in this embodiment, the temperature convergence amount... It is constructed using a weighted summation method in a fixed order;

[0089]

[0090] Among them, the lens barrel temperature Temperature data is collected by a temperature sensor located on the outer wall of the lens barrel or at the lens mount; detector substrate temperature. Temperature data is collected by a temperature sensor near the detector package; temperature of the inner wall of the housing. Temperature data is collected by a temperature sensor near the heat source inside the casing; window ring temperature. The temperature is collected by a temperature sensor near the window retaining ring. If the equipment does not have this measuring point, the window ring temperature is set to... And the corresponding weight Each weight satisfies And it is fixed into the parameter table during the equipment installation and commissioning stage.

[0091] In one specific embodiment, the device is installed on a high-level support in the refining unit area. After sunrise, the outer surface of the lens barrel is exposed to sunlight first, followed by a rise in the temperature of the detector substrate, while the temperature of the inner wall of the casing changes with a lag. After receiving the output from step one, the edge computing board first assembles the lens barrel temperature, detector substrate temperature, and inner wall temperature of the casing into a temperature observation vector in a fixed order, and then writes the temperature convergence quantity. The drift solver then uses the first... Reference response of each spectral channel Compared with the reference response Formation of reference deviation This binding relationship ensures that subsequent dual-timescale updates no longer simply follow grayscale changes, but rather follow reference deviations with thermal state indices.

[0092] In use, the drift state quantity on both time scales separates the slow thermal migration from transient thermal shock; reference deviation quantity Directly inherit the reference response value from step one And temperature observation vector, ensuring that step two and step one are continuous on the data object; temperature convergence amount After normalization, the drift solver plate has the ability to distinguish reference changes of the same amplitude under different thermal states.

[0093] After obtaining the reference deviation Afterwards, the drift solver does not directly provide a single correction value, but updates the slow-changing component and the fast-changing component separately, and then combines the two into a total drift state variable.

[0094] Among them, the slow-varying component needs to absorb the diurnal thermal balance migration, so its update speed should be relatively slow; the fast-varying component needs to track attitude changes or short-term thermal shocks, so its update speed should be relatively fast. If the same recursive coefficient is used for these two types of effects, the system will trail due to slow response when the pipe gallery is windward, and will write transient noise into the long-term baseline due to excessively fast response when the temperature continues to rise in the afternoon.

[0095] Therefore, the update coefficients for the slowly varying components are set to a smaller range, and the update coefficients for the rapidly varying components are set to a larger range. Preferably, the update coefficients for the slowly varying components are set to 0.01 to 0.12, and the update coefficients for the rapidly varying components are set to 0.15 to 0.55.

[0096] The total drift state quantity is expressed in the following form:

[0097]

[0098] Among them, the total drift state quantity : No. Each spectral channel is currently used for the overall drift result of the subsequent compensation chain, and its value is a real number; the slowly varying component... : No. Slowly varying drift state variables for each spectral channel, taking real values; rapidly varying components : No. The rapidly changing drift state variables of each spectral channel are real numbers; the slowly changing weights... Slowly varying components In total drift state quantity The combined weights in the formula are real numbers greater than or equal to 0.

[0099] Quickly change weight : Fast Variable Components In total drift state quantity The combined weights in the formula are real numbers greater than or equal to 0; the channel coupling coefficient... : No. Each spectral channel's deviation from the reference The direct absorption weight, which takes a real value, is used to immediately inject the current reference deviation into the total drift state quantity. Reference deviation The definition and function remain consistent with those described above; it is used to measure the reference response quantity. Relative to the reference response The deviation is introduced into this formula;

[0100] In terms of software implementation, the drift solver plate calculates the slowly varying components of each spectral channel. and fast variable components It is stored in a circular state cache, and a version number and timestamp are provided for each update. If the reference validity flag output in step one is invalid, or the reference validity of a certain spectral channel is invalid... If the value is below the threshold, the freeze rule will be applied first: the slow-changing components of the previous version. and fast variable components The design remains unchanged, only recording the reason for the current non-update. The engineering purpose of this design is to ensure that the freeze occurs before state recursion, preventing invalid references from polluting the state cache.

[0101] When using, slow-changing components With fast-changing components The task of carrying out drift at different time scales is distributed; the freezing rule is executed before state propagation to avoid invalid references being written to the state cache; the version number and timestamp are retained with each update to reference the total drift state in step three. Provide a clear entry point.

[0102] Furthermore, the total drift state quantity The document only answers how compensation should be made currently, without addressing whether the current compensation is within a credible range. In fixed online deployments, phenomena such as water droplets adhering to the lens assembly surface, dirt film appearing on the window edge, localized dust coverage on the flip-type reference film, and transient response anomalies in a certain spectral channel detector can all increase the total drift state quantity. It can still be calculated, but that doesn't mean the result is suitable to continue passing to step three.

[0103] In order to convert whether the reference layer is self-consistent and whether the spectral layer is cooperative into stability fields, step three no longer looks at the compensation amount itself, but only at the state of the compensation amount.

[0104] The edge computing board first sets the total drift state quantity Write the data to the corresponding spectral channel reference measurement frame to obtain the fixed reference region response after state compensation. Then calculate whether the fixed reference region is consistent with the one before compensation. Instead of using the simple absolute sum of differences as the criterion, we use the residual aggregation quantity with orientation preservation capability, because some pixels in the fixed reference region may have directional shifts due to lens vignetting, uneven transmittance at the edge of the window, or local contamination on the surface of the reference film.

[0105] Based on this, the construction method can achieve the following effect in constructing residual aggregation:

[0106]

[0107] Among them, residual polymerization amount : No. The residual intensity of each spectral channel after compensation in a fixed reference region is a real number used to measure the total drift state quantity. After the baseline compensation term is applied, does the reference area return to a stable range? (Fixed reference area) Pixel Index and original pixel response The definition remains consistent with step one; total drift state quantity , used to indicate the current number The combined drift results of each spectral channel;

[0108] To ensure that slowly varying components evolve smoothly at a low absorption rate between consecutive reference sampling events, and to prevent short-term thermal disturbances from being directly written into the long-term drift baseline, the drift solver plate... The slowly varying drift state variables of each spectral channel are updated recursively. In the... When the reference sampling event arrives, with the first... The slow-varying drift state after the event is used as the historical state, with the current reference deviation as the reference value. As a new observation input, the first... The slow-change drift state quantities after the event are as follows:

[0109]

[0110] in, Indicates the first The spectral channel in the first... Slowly varying drift state quantity after the next reference sampling event; Indicates the first The spectral channel in the first... Slowly varying drift state quantity after the next reference sampling event; This represents the update coefficient of the slowly varying component, which is a real number greater than 0 and less than 1. It is used to control the absorption rate of the slowly varying drift state quantity to the current reference deviation. Indicates the first The spectral channel in the first... The reference deviation corresponding to the next reference sampling event is used to characterize the normalized deviation of the current reference response relative to the reference response of that channel.

[0111] Baseline compensation item : No. The baseline compensation values ​​for each spectral channel, obtained from static calibration during the initial installation phase, are real numbers used to offset fixed biases associated with long-term manufacturing deviations; mapping function An odd function mapping, taking real values, is applied to the residuals to preserve direction and aggregate the data. This is used to suppress the influence of a small number of sharp residual spikes on the overall decision while retaining the positive and negative direction information of the residuals. In this embodiment, the mapping function... Pick This form is approximately linear when the residuals are small, but gradually saturates when the residuals are large; the terms in the denominator Used to aggregate residual amount Normalize to a scale related to the current channel response amplitude to avoid bias caused by using the same absolute threshold for low-response and high-response channels;

[0112] Example 1: In a field implementation, the device faces the valve assembly area. At night, as air humidity increases, a thin mist forms at the edge of the window, causing some pixels in the fixed reference area to slowly decay. The edge computing board will then respond with the current original pixel. Total drift state quantity and baseline compensation item Feeding residual polymerization amount The computation chain is used if the fog only causes anomalies in a small number of edge pixels, then the mapping function... It will suppress the impact of abnormal spikes on the overall residual. If the haze has developed to the fixed reference body, the residual aggregation will be reduced. It continues to deviate from the stable value.

[0113] When used, residual aggregation amount With fixed reference area To standardize the testing methods, the total drift state quantity is directly verified. Regression capability to the reference layer; mapping function While preserving the residual direction, peaks are suppressed to reduce the impact of local contamination or single-point anomalies on the overall judgment; baseline compensation term With total drift state quantity When used in combination, long-term manufacturing deviations and operational drift are addressed in a stratified manner.

[0114] Residual polymerization amount obtained for different spectra Next, step two needs to answer another question: whether there is still a synergistic relationship between different spectral channels. Because the identification of leaky plumes requires comparison of multi-band infrared image sequences, if the residual of a certain channel is not large in a fixed reference area, but its drift trend is completely different from that of other channels, then the drift compensation characterization obtained in step three will still be biased.

[0115] Therefore, cross-spectral consistency is further incorporated into the stability field. The drift solver considers the total drift state variables across the spectral channels. and residual polymerization amount Check if the residual aggregation amount of a certain channel Exceeding the limit, and the total drift state quantity with adjacent channels. If the directions of change are continuously opposite, it is called an abnormal channel. When the number of abnormal channels exceeds the stability threshold, the drift solver will change the unstable field to an unstable state and expand the unupdated reason field to an anomaly source field. The stability field preferably includes five elements: status identifier, anomaly source field, version number, associated reference measurement frame index, and timestamp. The status identifier can be directly read in step three. The anomaly source field indicates whether it is a reference mismatch, cross-spectral imbalance, or frozen inheritance. The version number corresponds to the total drift state quantity. The associated reference measurement frame index is used to trace back the sampling event in step one, and the timestamp is used to align with the multispectral infrared image sequence.

[0116] For abnormal freezing, the same stable inheritance mechanism is added as for pre-freezing. When the current sampling event is in an unstable state, the edge computing board will not directly obtain a null value, but will output the stability field and the total drift state index determined to be stable in the previous version into the current record and display the inherited output. Step 3 requires a clear state entry point when processing continuous image sequences. Directly obtaining a null value will cause a breakpoint in step 3 when there is no state. After using the stable inheritance mechanism, step 3 can determine whether to continue using the previous stable state, and then determine whether to degrade to location-only output or alarm-only output based on the abnormal source field.

[0117] When in use, cross-spectral consistency integration records situations where a single channel is stable but the relationship between multiple channels is unbalanced in the stability field; the stability field fixes the status identifier, anomaly source field, version number, associated reference measurement frame index, and timestamp into a unified output object to stabilize the reading interface in step three; stability inheritance retains continuous entries in the case of short-term instability, which is convenient for degradation processing in step three.

[0118] Step 3: Under the stability field constraints, the total drift state quantity It is applied to multi-spectral infrared image sequences to form an operational envelope field, quantization intermediates, and error source decomposition results.

[0119] In this embodiment, step three does not directly search for the plume outline on the original image, but first uses the stability field and total drift state quantity output in step two. A multi-spectral radiation characterization is formed after drift compensation, and then the operating envelope field, plume region, concentration path length, velocity field, and error source decomposition are established layer by layer on this characterization. The reason for this setting is that when a fixed camera faces flanges, valve groups, elbows, and tank connection pipelines, temperature difference structures, background edges, thermal reflections, and equipment self-heating will all affect the plume display; if the input boundaries are not defined first, the subsequent quantization process may easily mistake changes that are not in line with the scene for leakage signals.

[0120] The total drift state quantity output in step two The stability field only indicates whether the channel compensation chain can continue processing, and does not indicate whether the current scene is suitable for the quantization chain. For fixed deployments, the proportion of sky at different times, heat reflection from metal surfaces, residual heat from the ground, and line-of-sight distance will affect the separability of the plume from the background. Therefore, before starting quantization, the scene conditions must be extracted from the compensated image and converted into an operation envelope field that can be called in step four.

[0121] The edge computing board reads the total drift state quantities for different versions according to the version number output in step two. The current multi-spectral infrared image sequence is compensated according to spectral channels to obtain the drift-compensated multi-spectral radiation characterization, without considering the total drift state quantity. Used as a global brightness shift, it is written into the channel-by-channel compensation chain together with the channel baseline compensation term to ensure that the relative response relationship is maintained after different spectral channels are compensated. Preferably, the edge computing board performs bad pixel shielding, row and column non-uniformity correction and channel compensation based on the total drift state quantity for each spectral channel in sequence to avoid individual bad pixels from being misaligned across channels and causing subsequent plume discrimination.

[0122] To organize the compensated image into a unified object suitable for subsequent scene parsing, the edge computing board... Each spectral channel is used to compensate for radiance:

[0123]

[0124] Among them, the compensation radiation amount : No. Each spectral channel in pixel index Total drift state quantity and baseline compensation item The corrected radiation characterization values ​​are real numbers;

[0125] Raw pixel response Total drift state quantity and baseline compensation item The definition remains consistent with the steps described above, where the original pixel response... Total drift state quantity from the current multi-band infrared image sequence From step two, baseline compensation item From initial installation calibration;

[0126] Radiative mapping function : No. A monotonic mapping function for each spectral channel is used to transform the compensated response value to a unified radiation characterization domain; in this embodiment, the radiation mapping function... A piecewise linear function is used, with the segment nodes given in the factory calibration table. The slope of each segment is positive, thus ensuring that the channel response order is not reversed; a radiation mapping function is employed. As provided by the factory calibration table, each spectral channel corresponds to a piecewise linear mapping table, which includes the input interval endpoints, output interval endpoints, and piecewise slopes. The edge computing board reads the corresponding mapping table according to the channel number and performs table lookup interpolation.

[0127]

[0128] Among them, segment endpoints Slope and intercept By the The factory calibration table for each spectral channel is provided; the edge computing board reads the corresponding segment table according to the channel number and interpolates the values.

[0129] Pixel Index The location of a single pixel in the current image plane, used to converge the compensating radiance of each spectral channel at the same location. ;

[0130] In one specific embodiment, the camera is mounted on a high pole in the spherical tank area, and the edge computing board obtains the total drift state quantity of the current version from the drift solver board. Then, compensation is performed channel by channel for the current frame of each spectral channel, and the compensated radiance is then calculated. The multi-spectral radiation characterization was stacked and compensated for drift. The results observed on-site showed that the brightness relationship of the metal pipeline edge in different channels tended to be consistent, and the overall brightness introduced by lens heating was pushed back into a stable range.

[0131] When in use, the total drift state quantity Baseline Compensation Item Commonly entering the compensation radiation amount The construction chain ensures that the output of step two is completely continued into step three, and the fixed compensation order ensures that the compensated representation is more suitable for subsequent plume analysis.

[0132] After obtaining the compensated radiation amount Afterward, the plume analysis board does not immediately output the plume region. Instead, it first delineates the background composition, device boundaries, and candidate flow regions. This is because plume visibility depends not only on spectral absorption characteristics but also on background contrast, line-of-sight path length, device occlusion, and field-of-view configuration. Therefore, the plume analysis board performs three layers of scene boundary recognition on the multi-spectral radiation representation after drift compensation: the first layer identifies static device boundaries, the second layer distinguishes between the sky and ground backgrounds, and the third layer filters out candidate flow regions with continuous migration characteristics. After completing the recognition, the plume analysis board writes the background contrast, line-of-sight distance grouping, field-of-sight incident angle grouping, candidate flow region area ratio, and stability field status into the operation envelope field.

[0133] In order for the operation envelope field to be directly used for gating in step four, the plume parser further constructs a background contrast quantity:

[0134]

[0135] Among them, background contrast : The degree of separability of the two selected spectral channels relative to the background within the candidate flow region, expressed as a non-negative real number; candidate flow region A set of pixels with continuous migration characteristics after scene boundary recognition, used to limit the background contrast. Molecular statistical range; background reference region : with candidate flow region The set of adjacent background pixels that do not contain strong device edges; Selected spectral channels and selected spectral channels These represent the number of participants in the background comparison. The calculated spectral channel numbers take values ​​located in the interval [missing information]. Different positive integers; compensation radiation amount and compensation radiation Representing candidate flow regions or background reference area The two selected spectral channels are at the pixel index. The compensated radiation characterization value at the location is used to characterize the degree of spectral separation of the candidate flow region relative to the background.

[0136] In a preferred embodiment, the camera faces the compressor outlet pipe section, and the plume analyzer first identifies the metal edges and platform railings, then identifies the candidate flow area that continuously oscillates behind the railings. Mark the area and select a background reference area near it. Subsequently, the background contrast was measured. The grouping of line-of-sight distance and field of view incident angle is written together into the operation envelope field. If the background contrast is... If the background is too low and the sky background ratio is too high, the operation envelope field will indicate an unfavorable scenario; if the background contrast is low... If the candidate flow region is high and located in front of the stable device boundary, then the operation envelope field points to a favorable scenario.

[0137] In use, scene boundary recognition distinguishes between device boundaries, sky background, and candidate flow regions, preventing subsequent misidentification of device thermal edges as plume boundaries; background contrast... Together with the line-of-sight distance grouping and the field of view incident angle grouping, the scene conditions are entered into the operation envelope field, so that the scene conditions are saved in a structured way. At this stage, the operation envelope field absorbs the stability field state, so the subsequent step four does not need to backtrack to step two to perform gating.

[0138] Furthermore, even if the operation envelope field indicates that the current scene meets the conditions for entering the quantization chain, the plume still needs to be separated from the background and device boundaries and transformed into the quantization intermediate quantity and error source decomposition in step four. When a fixed camera is aimed at the leakage source, the shape of the plume in the image will change with wind direction, temperature difference, and background structure. If only the unit area or unit brightness difference is calculated, the gating and feedback scheduling tasks cannot be completed.

[0139] The plume parsing board first selects the flow region based on the operating envelope field. As the originating region, the plume region is extracted from the adjacent multi-frame compensated radiance sequence. Instead of using single threshold segmentation, a series of spectral difference constraints, migration continuity constraints, and boundary shrinkage constraints are used.

[0140] Spectral difference constraints ensure absorption within the selected spectral channels in the extracted region; migration continuity constraints ensure continuity in the displacement direction across adjacent frames; boundary retraction constraints ensure retraction when the region contacts strong edges of the device or sky edges, preventing high-contrast static boundaries from being drawn into the plume region. After the plume is determined, the quantization output board solves for the concentration path length along the line-of-sight direction and forms a velocity field by tracking the plume leading edge position in consecutive frames; the concentration path length is solved using a cross-sectional layering method, i.e., multiple vertical cross-sections are set along the main axis of the plume, the cumulative spectral absorption on each cross-section is obtained, and the cross-sectional level concentration path length is formed using the line-of-sight length model; the velocity field is jointly established using sparse feature tracking and block matching.

[0141] To link the concentration path length to subsequent error propagation, the quantization output board constructs a path integral:

[0142]

[0143] Among them, path integral : Cumulative concentration path length of the plume across multiple cross-sectional strata, a non-negative real number; stratum number The first one set along the main axis of the plume Each cross-sectional layer has a value located in the interval [missing information]. A positive integer used to identify the specific cross-sectional location; total number of layers. The total number of cross-sectional layers participating in the integration, taking a positive integer value, is used to control the path integral. The range of discrete summation;

[0144] Hierarchical weights : No. The path integral of each cross-section layer The weights in the quantization parameters are non-negative real numbers, used to reflect the contribution of the cross-sectional layer to the overall quantization result; absorption characterization quantity. : No. The absorption characterization results calculated from the differences in selected spectral channels on each cross-sectional layer are non-negative real numbers, used to characterize the gas absorption intensity of that cross-sectional layer.

[0145]

[0146] Among them, the cross-sectional area Indicates the first The set of pixels corresponding to each plume cross-section layer; plume membership degree Represents pixel index The degree of plume affiliation; the amount of compensation radiation and compensation radiation The two selected spectral channels are respectively located at the pixel index. The amount of compensation radiation at the location.

[0147]

[0148] Among them, the membership degree of the feather stream Represents pixel index The degree to which it belongs to the plume region, with values ​​ranging from 0 to 10. ; feather flow discriminant It is formed by the combination of cross-channel compensated radiation difference and multi-frame migration continuity. The threshold for distinguishing plumes.

[0149] If hard partitioning is used, then let Degenerates into binary form:

[0150]

[0151] Path length factor : No. The effective length of each cross-sectional layer along the line of sight is a non-negative real number, used to characterize the cross-sectional absorption. Converted into an accumulative path quantity;

[0152]

[0153] Among them, path length factor Indicates the first Effective length of each cross-section along the line of sight; cross-sectional projected width Indicates the first The pixel width of each cross-sectional layer in the image plane; distance from the group center. This indicates the target distance group center value corresponding to this cross-sectional layer; horizontal focal length. The horizontal focal length and line-of-sight angle are parameters within the camera's internal specifications. This indicates the angle between the camera's main line of sight and the normal to the target surface.

[0154] In one example, the camera is aimed at the flange connection, and the plume escapes from the lower edge of the flange and rises along the outside of the pipeline. The plume analyzer first identifies the continuously moving plume region within the candidate flow area. Several cross-sectional layers are distributed along the main axis of the plume. A quantization output plate calculates the absorption characterization at each cross-sectional layer. and path length factor Path integral The leading edge of the plume moves to the upper right between consecutive frames, and sparse feature tracking and block matching are combined to establish a local velocity vector to form a velocity field.

[0155] In practice, plume region extraction uses a triple constraint concatenation to prevent device edges and sky boundaries from accidentally entering the plume region; path integral... Characterization of cross-sectional absorption and path length factor This combination facilitates subsequent gating judgment and scheduling; the velocity field is jointly established by sparse feature tracking and block matching, taking into account both regions with clear boundaries and regions with weak texture.

[0156] Obtain the path integral. After the velocity field is applied, the quantization output board does not immediately output the final quantized value. Instead, it first identifies the sources of error affecting the result. For this purpose, the error sources are divided into three categories: the first category is the drift residual term, which originates from the residual aggregation in step two. and the current total drift state quantity The first category is the cross-channel dispersion; the second category is the segmentation uncertainty, which originates from the expansion and contraction amplitude of the plume region boundary in adjacent frames and the number of boundary retractions; the third category is the velocity uncertainty, which originates from the dispersion of vector directions within the velocity field and the block matching back-off ratio. After calculating the intensity of the three types of error sources, the quantization output board writes them into the error source decomposition field in a fixed order, and forms confidence interval index values ​​according to the fixed order, so that step four can process them according to the same protocol.

[0157] In a preferred embodiment, the status indicator in the stability field is checked first. If the status indicator is unstable, it is directly written as a limited output, and the drift residue in the error source decomposition field is used as the main output. If the status indicator is stable, the background contrast in the operation envelope field is checked next. Grouping by line-of-sight distance and the percentage of candidate flow area. If the background contrast... If the value is too low or the proportion of candidate flow region area is too small, the result will be directly written as a location-only output, and the path integral will be retained. Velocity field index, but not written as an event-level quantization result; if background contrast... If the applicable range is within the continuous velocity field, then the quantization result, confidence interval index value, and error source decomposition field are generated as inputs for step four.

[0158] When the device needs to output event-level quantization results, the quantization output board uses the path integral. Construct event-level quantization results with the projected velocity of the velocity field in the plume normal direction:

[0159]

[0160] Among them, event-level quantization results Indicates the quantized output value of the current sampled event; response conversion factor. Determined by target gas type, spectral response factor, and equipment calibration table; normal velocity The mean value is obtained by projecting each effective vector in the velocity field onto the normal of the main section of the plume.

[0161] To prevent the computation chain from breaking in the event of a local failure, a solver fallback path is specified: the velocity field solution first calls the sparse feature tracking solver, and falls back to the block matching solver if there are insufficient tracking points; path length factor The solution first calls the layered section interpolation path, and then backtracks to the adjacent section extension path when the section boundary fractures. The corresponding source identifier is written in the output field for both backtracking paths.

[0162] In use, the error source decomposition field stores the drift residual term, the segmentation uncertainty term, and the velocity uncertainty term separately, enabling step four to perform gating based on the dominant source; the three paths of restricted output, location-only output, and quantization result output maintain a direct correspondence with the stability field and the operation envelope field; the solver backtracking path is the velocity field and the path length factor. The solution provides an engineering safety net, preventing partial failures from causing a break in the entire third step.

[0163] Step 4: Consolidate the state fields, intermediate quantities, and error sources formed in Steps 2 and 3 into event-level gating results, record and solidify the results, and set the scheduling parameters for the next cycle.

[0164] In this implementation, step four first determines the current output state of the result based on the stability field and the operation envelope field, then solidifies the determination result along with the source field, and finally sends the error source decomposition field back to the scheduling feedback board as the basis for the next round of reference sampling and calculation order adjustment. Fixed online deployment is not a one-time measurement process, but a continuous, cyclical observation process. If only the quantification result is output without specifying the boundary and formation path of the result, subsequent equipment maintenance, threshold correction, and task scheduling will lack clear guidelines.

[0165] Although step three outputs the quantization results, confidence interval index values, and error source decomposition fields, the conditions under which these objects are formed are not entirely the same. Some sampling events occur under favorable background conditions, resulting in path integrals. Corresponding continuously to the velocity field index; some sampling events, although identifying the plume region, are in a low background contrast or high occlusion state; and some sampling events are themselves written as unstable in step two. If explicit gating is not performed here, the backend interface will treat the output under different conditions equally. Therefore, all source fields are first uniformly pulled to the same decision plane before outputting the three-state event results.

[0166] The gating decision board first reassembles the input fields in a fixed order: stability field, operation envelope field, and path integral. The parameters include: velocity field index, quantization result, confidence interval index value, error source decomposition field, version number, and associated reference measurement frame index. The stability field is placed first because it directly follows the steps in step two regarding the total drift state quantity. and residual polymerization amount The determination result reflects not only whether the current image is suitable for calculation, but also whether the preceding reference chain is continuous. After the gating decision board completes the field reorganization, it first performs a state priority determination.

[0167] Preferably, the state priority order is as follows: reference unavailable state takes precedence over unstable state, unstable state takes precedence over envelope-restricted state, and envelope-restricted state takes precedence over quantization-available state. The reason for using priorities is that when the reference chain is interrupted, any image layer determination will fail; when the reference chain is continuous but the stability field is unstable, the total drift state quantity has not yet reached the range that can directly support quantization, and the operation envelope field only has an effect when the first two states are not triggered.

[0168] For example, the gating decision board receives event records from the flange area, inherits the output of the stability field, and shows a moderate proportion of the candidate flow region area in the operation envelope field with a small background contrast. The gating decision board reads the status in the stability field and determines that the current event is not the latest stable result but is inherited from the previous version of the stability field. It then pushes the event into the restricted decision channel, instead of directly entering the quantization available decision channel. On-site, it was observed that the interface marked the suspected leak location but did not simultaneously provide the final quantization value.

[0169] When in use, the gating input reassembly pulls the objects scattered in steps two and three onto the same decision plane, avoiding confusion in the backend's field interpretation order; state priority determination, reference chain interruption and unstable events are intercepted first, reducing the number of results outside the boundary entering the quantization output chain; fixed order writing of version number and associated reference measurement frame index, so that subsequent record solidification boards directly connect to the same object order.

[0170] After completing the state priority determination, the gating decision board further aggregates the operation envelope field, quantization result, and error source decomposition field into a gating score, which is used to distinguish between quantized output state, suggested output state, and invalid output state.

[0171] A comprehensive gating system with suppression terms is adopted. In actual field scenarios, there are often intersections where the background contrast is high but the drift residue term is large, or the velocity field is continuous but the segmentation uncertainty term is large. If only a step-by-step passage mechanism is relied upon, the gating logic will be rigid. Therefore, the gating decision board constructs a gating score:

[0172]

[0173] Among them, the gating score : A comprehensive judgment quantity for the current sampling event entering the quantization output state. It takes a non-negative real number and is used to distinguish between the quantization output state, suggested output state, and invalid output state when the state priority has not triggered an interception; Background comparison quantity. The definition remains consistent with step three, used to characterize the spectral separability of the candidate flow region relative to the background reference region; path integral. The definition is consistent with step three, and is used to characterize the cumulative path length of the plume across multiple cross-sectional layers;

[0174] velocity continuous quantity The continuity quantity, derived from the velocity field index, is a non-negative real number used to characterize whether the migration direction and amplitude of the plume are continuous between adjacent frames; in this embodiment, the velocity continuity quantity... It is obtained by combining the proportion of available vectors in the velocity field index and the direction consistency table;

[0175] Drift residue Error source decomposition field and total drift state quantity and residual polymerization amount The relevant sub-term strengths are non-negative real numbers, used to suppress drift non-convergence events from entering the quantization output state; segmenting uncertain terms. The intensity of the sub-terms related to the number of times the plume region boundary expands and contracts in the error source decomposition field are non-negative real numbers; the velocity uncertainty term... The intensity of the sub-item in the error source decomposition field that is related to the dispersion of the velocity field direction and the back-off solution ratio is a non-negative real number, used to suppress velocity field discontinuity events from entering the quantization output state;

[0176] Gain Weight Gain weight and gain weight These represent the background contrast. Path integral and velocity continuity The weights in the numerator are all real numbers greater than or equal to 0, used to adjust the gating score based on the three types of effective information. Contributions;

[0177] Suppress weights Suppression weights and suppression weights These represent drift residues. , segmenting uncertain terms and velocity uncertainty The weights in the denominator are all real numbers greater than or equal to 0, used to adjust the impact of the three types of adverse information on the gating score. The degree of inhibition;

[0178] The gate control decision board obtains the gate control score. Then, a three-state output is generated based on the aforementioned state priorities. If the state priority does not trigger interception and the gating score is achieved... If the value falls within the quantization output range, a quantization output status is generated. The output object includes the quantization result, confidence interval index value, operation envelope field, stability field, version number, and record index reserved space. If the status priority does not trigger interception but the gating score is achieved... If the value falls within the suggested output range, a suggested output status is generated. The output object includes the alarm identifier, location area index, operation envelope field, stability field, and error source decomposition field as the main items. If the status priority has already triggered interception or gating score... If the value falls into the invalid range, an invalid output state is generated. The output object includes a maintenance prompt, an exception source field, and the reason for the non-quantification. With this design, the gating result of step four is not only a state name, but also an event-level composite object with a source field.

[0179] When in use, the gating score By merging the effective and unfavorable quantities in step three into the same decision formula, the cross-state conditions in the field can be handled; the three-state output is generated in a fixed field order, which makes it easy to directly solidify the latter half of step four; the quantized output state, suggested output state, and invalid output state correspond to different combinations of source fields, so the subsequent interface does not need to infer the applicable boundary of the current result.

[0180] Furthermore, only when the gating results are entered into the continuous recording chain can they support subsequent equipment maintenance, manual review, and scheduling parameter correction. If the error source decomposition field is not sent back to the scheduling feedback board, the reference sampling in step one and the solution order in step three cannot be adjusted to adapt to current on-site changes. Therefore, it is necessary to both solidify the records and implement feedback scheduling.

[0181] After recording the three-state objects received and output by the recording board, regardless of whether they belong to the quantized output state, suggested output state, or invalid output state, a unified event record header is constructed first. The unified event record header contains, in sequence, the device serial number, event timestamp, version number, stability field index, operation envelope field index, gating status identifier, and anomaly source field index. After the unified event record header, different data segments are attached according to the state type. The quantized output state is attached with the quantization result segment, confidence interval index value segment, and error source decomposition field segment; the suggested output state is attached with the positioning area index segment, dominant error term segment, and limiting cause segment; and the invalid output state is attached with the maintenance prompt segment and unquantized cause segment. After the recording board completes the assembly, it is first written to the high-speed circular cache, then to the non-volatile memory array, and then the record index is sent to the upper interface buffer.

[0182] To prevent result loss due to storage path failures, the recording and solidification board employs a dual-path exception write mechanism. The first path is the primary write path, targeting the non-volatile storage array; the second path is a degraded write path, targeting only the append-only log area. When the primary write path returns a verification failure, the recording and solidification board immediately switches the current event to the degraded write path and rewrites the gating status to restricted record status. Simultaneously, it retains the original quantization result segment but prevents it from entering the higher-level reporting interface. The immediate result of this write action is that the complete source chain of the event is still preserved locally, but the higher-level interface only receives the restricted record notification and will not propagate the incompletely solidified quantization value as a formal result.

[0183] In a practical example, when a write delay in a certain block of the non-volatile storage array triggers a timeout feedback from the record persistence board during nighttime operation, the current event record is rewritten to the append-only log area, with the record index set to degraded write. Then, only the restricted record status, location area index, and anomaly source fields are written to the upper-level interface buffer, without writing the quantization result segment. This way, while on-site maintenance personnel can see the event trigger location and the source of the write anomaly, they will not see the incompletely persisted quantization value in the next day's daily report.

[0184] In use, the unified event record header is pushed into the core fields of the same step and is no longer arbitrarily arranged in the index system, so the retrieval order remains unchanged; dual-path exception writes maintain the source chain and do not lose events when the main write fails; restricted record status prevents incomplete solidified results from spreading to upstream interfaces and keeps the output chain boundaries clear.

[0185] After the records are solidified, the scheduling feedback board decomposes the dominant items of the error source decomposition field and the residual aggregation amount. The current level, gating status indicator, and continuity of the velocity field index are used to redetermine the reference sampling period, threshold table, and solution order for the next cycle. Instead of directly using the single-condition rule of shortening the reference sampling period when the error is large, a feedback priority is first established, and then the feedback priority drives the updates of different types of parameters.

[0186] The reason is that some events are dominated by drift residue terms, so the reference sampling period in step one should be reduced first; some events are dominated by velocity uncertainties, where the reference sampling period is not a primary issue, and the calling order of the velocity field solver in step three should be adjusted first; some events are dominated by segmentation uncertainties, so the boundary shrinkage threshold of the candidate flow region and the update interval of the background reference region should be adjusted first. Therefore, the feedback priority of the scheduling feedback board is as follows:

[0187]

[0188] Among them, feedback priority : The push strength of the current event on the scheduling parameters of the next cycle, taking a non-negative real number, used to determine whether to update the reference sampling period, threshold table, and solution order; drift residue term , segmenting uncertain terms and velocity uncertainty The definitions remain consistent with those mentioned above, reflecting the driving effect of the three types of error sources on scheduling adjustments; residual convergence. : The residual polymerization amount of each spectral channel The converged event-level residuals are non-negative real numbers used to characterize the overall residual level of the current drift chain; in this embodiment, the converged residual amount... Using the residual aggregation amount of each channel The absolute value weighted sum form;

[0189] Gated holding quantity : The duration of the gated state remaining unchanged across multiple consecutive loops, which is a non-negative real number; feedback weight Feedback weights Feedback weights Feedback weights and feedback weights These represent drift residues. , segmenting uncertain terms speed uncertainty term Residual convergence and gate holding quantity In feedback priority The strength of the effect is determined by real numbers greater than or equal to 0, which are used to adjust the impact of different information on scheduling updates.

[0190] The scheduling feedback board receives feedback priority. Then, update the parameters according to the dominant error term. If there is a drift residual term... As the leading item and feedback priority If the drift scheduling threshold is exceeded, the reference sampling period in step one is shortened, and the release interval of the freeze threshold in step two is simultaneously reduced; if the uncertainty term is segmented... As the leading item and feedback priority If the segmentation scheduling threshold is exceeded, the boundary shrinkage threshold of the candidate flow region in step three is tightened, and the background reference region is extended. The refresh interval; if the speed is uncertain. As the leading item and feedback priority If the velocity scheduling threshold is exceeded, the block matching solver is moved before the sparse feature tracking solver, and the time window length of the velocity field index is shortened. After the above updates are completed, the scheduling feedback board writes the new reference sampling period, threshold table version number, and solution sequence number into the next loop configuration register for direct reading by steps one and three.

[0191] Feedback priority during use Decompose the error source fields and the residual aggregation amount A scheduling layer is introduced to avoid bias caused by single-condition adjustment; the reference sampling period, boundary shrinkage threshold and solution order are updated according to the dominant error term, so that the adjustment direction of each step directly corresponds to the source of the current problem; the new configuration register is written at the end of this step, and the updated parameter set can be read in the next loop without manual intervention.

[0192] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0193] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0194] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0195] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0196] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A combustible gas non-contact detection camera system integrating a spectral sensor, comprising acquiring infrared image sequences of at least two spectral bands and performing gas detection, characterized in that, include: Trigger reference sampling to generate a reference measurement frame that shares at least part of the optical link with the main imaging optical path, and extract a statistical summary of the fixed reference area, temperature observation vector, and valid reference data flags; Based on the statistical summary, temperature observation vector, and valid reference data flags, update the drift state variables including channel gain drift and channel bias drift and generate a stability field. Freeze the update when the valid reference data flags are invalid. Drift compensation is performed on infrared image sequences based on drift state variables and stability fields, and operation envelope fields, quantization results, confidence intervals, and error source decomposition are generated. The gating output is decomposed based on the operation envelope field, stability field, confidence interval, and error source, and the updated scheduling parameters are used as the input for the next loop.

2. The combustible gas non-contact detection camera system according to claim 1, characterized in that: A reference measurement frame is generated in response to trigger events such as reaching a preset reference sampling time, detecting that the rate of change of ambient temperature exceeds a threshold, and switching of device operating status. A fixed reference region is set in the reference measurement frame, and the pixel position of the fixed reference region remains unchanged between consecutive reference samples. When saturation, occlusion or insufficient number of pixels in the fixed reference region occurs, the current reference sample is marked as invalid, and delayed resampling, switching to the backup reference path and outputting the reference unavailable state are performed in sequence.

3. The combustible gas non-contact detection camera system according to claim 2, characterized in that: When generating the reference measurement frame, the reference measurement frame and the corresponding infrared image are kept to have the same integration time, the same analog gain and the same non-uniform correction version. The reference measurement frame index, fixed reference area statistical summary, temperature observation vector, and reference data validity flag are then assembled into a unified input record according to a fixed field order.

4. The combustible gas non-contact detection camera system according to claim 3, characterized in that: Based on the reference measurement frame index, the statistical summary of the fixed reference area and the temperature observation vector are aligned, and a drift state variable containing channel gain drift and channel bias drift is constructed. The drift state variable is then decomposed into slow variables corresponding to changes in environmental thermal equilibrium and fast variables corresponding to short-term disturbances. Then, a stability field is generated based on the updated drift state variables. The stability field includes at least the residual statistics of the fixed reference area and the cross-spectral consistency index.

5. The combustible gas non-contact detection camera system according to claim 4, characterized in that: When the reference data validity flag is invalid, or the stability field exceeds the preset threshold, the drift state value is frozen and updated, while the channel gain drift and channel offset drift corresponding to the previous valid version remain unchanged; at the same time, the version identifier and statistical summary corresponding to the current drift state value are output, and the reason for this freeze is recorded.

6. The combustible gas non-contact detection camera system according to claim 5, characterized in that: Drift compensation is performed on the infrared image sequence based on the drift state quantity to form a multi-spectral radiometric characterization; then, background contrast, distance and field of view configuration, as well as plume motion proxy quantity are extracted from the multi-spectral radiometric characterization to generate the operation envelope field. When the stability field indicates instability, the instability information is written into the operation envelope field, and then the plume region is identified and segmented based on the operation envelope field.

7. The combustible gas non-contact detection camera system according to claim 6, characterized in that: The concentration path length and velocity field are calculated based on the plume region, and the boundary flux is determined based on the velocity field when the leakage rate is required; then the drift residual statistics, segmentation uncertainty and velocity field uncertainty are written into the error source decomposition, and the quantification results and confidence intervals are formed. When the confidence level of the plume segmentation is insufficient, the velocity field does not converge, or the operation envelope field exceeds the valid boundary, only qualitative alarms and positioning information are output, and the current dominant error term is output simultaneously.

8. The combustible gas non-contact detection camera system according to claim 7, characterized in that: Gating decisions are performed based on the stability field, operation envelope field, confidence interval, and error source decomposition, and quantized output state, restricted output state, and invalid output state are generated. In quantized output mode, the system outputs quantization results, confidence intervals, operation envelope fields, stability fields, and event record indexes. In restricted output mode, the system outputs qualitative alarms and location information. In invalid output mode, the system outputs alarms and maintenance prompts.

9. The combustible gas non-contact detection camera system according to claim 8, characterized in that: Write the current sampled event to the event log. The event log shall include at least the device serial number, event timestamp, version number, stability field index, operation envelope field index, gating status identifier, and anomaly source field index. If the event log writing fails, switch the current sampled event to a restricted output state and prevent the quantization result from entering the report link, while retaining the corresponding location information and anomaly source field.

10. The combustible gas non-contact detection camera system according to claim 9, characterized in that: The scheduling parameters are updated based on the dominant error term and drift residual level in the error source decomposition. Specifically, the reference sampling period is shortened when the drift residual increases, the execution order of uncertainty propagation is adjusted when the computational load exceeds the threshold, and the updated scheduling parameters are written back, while keeping the field order of the gating decision chain and event record chain unchanged.

Images