Infrared thermal imaging-based concealed sprinkler installation quality acceptance method and system

Abstract

This invention discloses a method and system for quality acceptance of concealed sprinkler installation based on infrared thermal imaging, relating to the fields of building fire protection facility acceptance and non-destructive testing technology. The method includes: modeling and calibration, establishing a material parameter library containing thermal diffusivity, thermal conductivity, and temperature resistance limits, and completing coordinate system alignment and time synchronization; controlled excitation and triggering, applying hot / cold pulses to the pipe network or connecting metal parts where the sprinkler is located and recording the frequency band identifier; acquisition and preprocessing, obtaining infrared image sequences before, during, and after excitation and extracting temporal observation features such as isothermal contour radius, temperature difference response, temperature gradient center, and contour ellipticity; determining update weights based on signal-to-noise ratio, occlusion, and excitation effectiveness, and performing block fusion and recursion with time decay; installation geometry and fastening status estimation, outputting the sprinkler's position, tilt angle, extension amount or elevation, and fastening status indicators within the ceiling plane; compliance judgment and reporting, generating conclusions and rectification suggestions based on threshold-weighted scoring.

Description

Technical Field

[0001] This invention relates to the field of building fire protection facility acceptance and non-destructive testing technology, specifically a method and system for quality acceptance of concealed sprinkler installation based on infrared thermal imaging. Background Technology

[0002] In fire sprinkler systems of public buildings and industrial plants, concealed sprinkler heads are usually installed in the ceiling. Their location, angle, extension or elevation, and fastening condition directly affect the spray coverage and fire extinguishing effect. Because the sprinkler heads and their connecting components are covered by decorative panels, traditional acceptance methods often rely on visual sampling, random opening, or partial disassembly, which are destructive, costly, inefficient, and difficult to achieve full coverage and traceability.

[0003] Existing technologies use infrared imaging to detect defects, leaks, and electrical hotspots in building envelopes. Some technologies also indirectly reveal the location of obscured metal parts through passive or simple excitation thermal imaging. However, these technologies have certain limitations. For example, they lack quantitative criteria corresponding to the installation geometry and fastening status of fire sprinklers. They are also unable to suppress false alarms caused by environmental drift, wind disturbances, and obstruction under complex ceiling materials and thicknesses. Furthermore, they cannot quantify the position, tilt angle, extension, or elevation of sprinklers at the millimeter level without damaging the finish. Additionally, they lack automated reporting capabilities that can be integrated with engineering compliance clauses.

[0004] Therefore, there is an urgent need for a technical solution that, without damaging the ceiling, combines controlled temperature difference excitation, infrared thermal imaging time-series analysis, and structured fusion calculation to quickly, quantitatively, and traceably accept the installation geometry and fastening status of concealed sprinklers, and output compliance conclusions and rectification suggestions corresponding to the standard thresholds. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and propose a method and system for quality acceptance of concealed nozzle installation based on infrared thermal imaging, so as to solve the above-mentioned problems.

[0006] The objective of this invention is achieved through the following technical solution: a method for quality acceptance of concealed nozzle installation based on infrared thermal imaging, comprising the following numbered steps:

[0007] S1. Modeling and Calibration: Obtain information on ceiling finish material and thickness, nozzle model and its connecting components, and establish a material parameter library including thermal diffusivity, thermal conductivity and temperature resistance limits; complete the calibration of internal and external parameters of the infrared imaging equipment and align it with the field coordinate system, and use a synchronization controller to synchronize the infrared imaging equipment and the excitation device in time.

[0008] S2. Controlled excitation and triggering: A controlled thermal pulse or cold pulse excitation is applied to the pipe network or metal parts connecting the sprinklers in the area where the sprinklers are located in the ceiling. The excitation includes a settable pulse width, temperature difference amplitude and frequency band identification, and a synchronous trigger signal is sent to the infrared imaging device by the synchronous controller.

[0009] S3. Acquisition and Preprocessing: Infrared image sequences are acquired before, during and after excitation. Non-uniformity correction and background drift suppression are performed. Temporal observation features, including at least isothermal contour radius, temperature difference response, temperature gradient center and contour ellipticity, are extracted. The temporal observation features are then time-registered with the excitation parameters.

[0010] S4. Selectively update quantitative information: Adaptively compensate for different finish materials and thicknesses based on the material parameter library; based on time-series observation characteristics and excitation parameters, selectively update the quantitative information of thermal conductivity characteristics, thermal response curves, or relative position change trends corresponding to the time-series observation characteristics. Among them, the update weight is increased when the excitation is effective or the signal-to-noise ratio is high, and the update weight is decreased when environmental disturbances or occlusion occur; thereby obtaining time-series data reflecting the quantitative information of thermal conductivity characteristics, thermal response curves, or relative position change trends between the nozzle and the finish.

[0011] S5. Structured temporal mask construction: Based on the update weights obtained in S4, a temporal mask that is causal in time and decays with time intervals is constructed. The temporal mask is used to fuse features from different time points, different perspectives and different excitation frequency bands.

[0012] S6. Block Fusion and Recursion: Divide the temporal mask and its corresponding temporal observation features into multiple sequence blocks according to time. Perform mask-based secondary fusion calculation within each sequence block, perform recursive fusion update between adjacent sequence blocks, and perform numerical normalization within the block and pass the fusion state and normalization factor between blocks to obtain the numerically processed fusion result.

[0013] S7. Installation geometry and fastening status estimation: Based on the temporal status and fusion results, estimate the position of the nozzle in the ceiling plane, the tilt angle of the nozzle axis relative to the vertical direction, the nozzle extension or elevation, and the fastening status index determined by the combined temperature difference response decay rate and local micro-vibration characteristics, and give the confidence level of the estimation results.

[0014] S8. Compliance Judgment and Report: Based on the preset installation compliance threshold, calculate the normalized errors of the eccentricity, tilt angle, extension amount and fastening status indicators determined by the location. According to the preset weights of the importance of each normalized error, output the conclusion of whether it is qualified or not, the problem type and confidence level, and automatically generate an acceptance report including heat map overlay, isotherm display, coordinate and time annotation, conclusion and rectification suggestions.

[0015] The temperature difference amplitude of the hot or cold pulse in S2 is within a preset range, and the pulse width and duty cycle are adjustable. The excitation temperature rise or drop is limited by the safety threshold of the surface material, nozzle model and its connecting components. When the temperature is detected to be close to the threshold, the excitation intensity is automatically reduced or the excitation is stopped and an alarm is recorded.

[0016] In S1, time synchronization employs a dual mechanism of trigger signal and timestamp, with synchronization accuracy no less than millisecond level; in S3, preprocessing includes a combination of inter-frame differential and exponential smoothing to suppress slow changes in ambient temperature and imaging drift.

[0017] The material parameter library in S1 includes at least the thermal parameters and thickness ranges of gypsum board, mineral wool board and metal plate; the adaptive compensation in S4 selects different compensation coefficients according to the material and thickness, and allows the compensation coefficients to be corrected by a small number of field calibration samples.

[0018] In S4, the quantitative information of thermal conduction characteristics, thermal response curves, or relative position change trends corresponding to time-series observation features is selectively updated. The update weight is determined based on the signal-to-noise ratio, effective excitation identifiers, and occlusion detection results. When the update weight is continuously lower than the threshold or the temperature difference fluctuation is abnormal, a strong excitation short-sequence acquisition and intra-block fusion priority strategy is triggered to improve the reliability of estimation in high-noise scenarios.

[0019] In S5, the temporal mask is based on updating weights and introducing a decay factor with time intervals, so that the near-time features have a higher proportion in the fusion and the far-time features gradually decrease in proportion after excitation. In S6, the time length and number of each sequence block are adaptively determined according to the target sequence length and hardware resources. The mask-based secondary fusion is performed within the block, and the recursive fusion is performed between blocks. The maximum value and cumulative sum of the rows are maintained within the block to enhance the robustness of numerical processing.

[0020] In S7, the nozzle position, tilt angle, and extension amount are obtained through geometric estimation in the ceiling plane coordinate system. The fastening status index is determined by the temperature difference response decay rate and local micro-vibration characteristics. In S8, the compliance judgment is first determined by rules based on the threshold, and then the confidence of the abnormal type is corrected by the lightweight model.

[0021] A concealed sprinkler installation quality acceptance system based on infrared thermal imaging includes:

[0022] Thermodynamic excitation unit is used to apply controlled thermal or cold pulse excitation to the pipe network or metal parts connecting the nozzle in the area where the nozzle is located.

[0023] The synchronous control and acquisition unit is used to generate trigger signals, record excitation parameters and frequency band identifiers, and synchronize with the infrared imaging unit in time.

[0024] An infrared imaging unit is used to acquire infrared image sequences before, during, and after excitation.

[0025] The processing and judgment unit is used to perform the following tasks: establishing the material parameter library in S1, calculating the calibration of the internal and external parameters of the infrared imaging equipment and the alignment with the field coordinate system, data preprocessing from S3 to S8, extraction of temporal features, selective updating of the thermal conduction characteristics corresponding to the temporal observation features, quantitative information of thermal response curves or relative position change trends, construction and block fusion of temporal masks, estimation of geometric and fastening states, compliance judgment and report generation.

[0026] The processing and decision unit has a built-in producer and consumer parallel execution mechanism. The producer is responsible for loading the sequence block data into the shared buffer in blocks, and the consumer completes the mask-based secondary fusion within the block and the recursive update between blocks, and the two overlap in data transfer and calculation. The processing and decision unit adopts the same block quantization strategy as the sequence block for the feature tensor and mask block, and performs orthogonal perturbation on the features before quantization to optimize the distribution of the activation range. The key intermediate quantities are normalized and accumulated using a high-precision accumulator.

[0027] The infrared imaging unit's spectral range covers long-wave or medium-wave thermal imaging, with noise equivalent temperature difference not exceeding a preset threshold and frame rate not lower than a preset value. The system is equipped with a safety interlock function, which automatically stops the excitation and records an alarm when the excitation temperature rise or drop exceeds the safety threshold. The report generation includes thermal image overlay, isotherm display, coordinate and time annotations, judgment conclusions and rectification suggestions, as well as traceability information of material and excitation parameters.

[0028] The beneficial effects of this invention are:

[0029] This invention implements acquisition and preprocessing under controlled excitation and triggering constraints, ensuring a one-to-one correspondence between the infrared image sequences acquired by the infrared imaging unit before, during, and after excitation and the excitation parameters and frequency band identifiers of the temperature difference excitation unit. This guarantees consistency in time and scale between acquisition and preprocessing and the on-site working conditions, reducing the impact of environmental drift on time-series observation characteristics. This provides a stable observational basis for the subsequent selective updating of quantitative information such as thermal conductivity characteristics, thermal response curves, or relative position change trends corresponding to the time-series observation characteristics. By establishing a material parameter library including thermal diffusivity, thermal conductivity, and temperature resistance limits during the modeling and calibration stages, and completing the calibration of the internal and external parameters of the infrared imaging equipment and alignment with the on-site coordinate system, this invention can adaptively compensate for different surface materials and thicknesses, fundamentally reducing response differences between different project sites and enabling cross-material generalization capabilities for installation geometry and fastening state estimation.

[0030] By incorporating signal-to-noise ratio (SNR), effective excitation identifiers, and occlusion detection results into selective state updates to determine update weights, this invention can increase update weights during periods of effective excitation and high SNR, and decrease update weights when environmental disturbances or occlusions occur. This ensures that the time-series data of quantized information has a stronger focusing ability on the effective thermal response, significantly suppressing false alarms and misjudgments introduced by background disturbances. Based on this, a temporal mask with temporal causality that decays with time intervals is constructed. Using the update weights as a basis, features from different time points, different perspectives, and different excitation frequency bands are fused, resulting in a higher proportion of effective information from closer time points after excitation in the fusion, while the interference contribution from farther time points gradually decreases, further enhancing the ability of the fusion result to represent the real physical process.

[0031] The temporal mask and its corresponding temporal observation features are divided into multiple sequence blocks according to time. Within each sequence block, a block fusion and recursive strategy is performed, which involves performing secondary fusion calculation based on the mask and recursive fusion updates between adjacent sequence blocks. This strategy enables the numerical normalization within a block and the transfer of fusion states and normalization factors between blocks to form a stable computational link. This improves processing efficiency while ensuring numerical stability and meets the batch acceptance requirements of large-scale sites. Based on the temporal state and fusion results, installation geometry and fastening status estimation are performed. This enables the complete quantitative output of the sprinkler head's position in the ceiling plane, the sprinkler head axis tilt angle relative to the vertical direction, the sprinkler head extension or elevation, and fastening status indicators jointly determined by the temperature difference response decay rate and local micro-vibration characteristics. Through preset installation compliance thresholds and weighted scoring rules, conclusions, problem types, and confidence levels are output in the compliance judgment and report generation stages. This constructs a closed-loop process from controlled excitation, data acquisition and preprocessing, selective state updates, temporal mask construction to block fusion and recursion, installation geometry and fastening status estimation, and finally to compliance judgment and reporting. This significantly improves the non-destructive, comprehensive, and traceable nature of concealed sprinkler head installation quality acceptance.

[0032] This invention utilizes a producer-consumer parallel execution mechanism built into the processing and judgment unit, enabling the loading of sequence block data, intra-block mask-based secondary fusion, and inter-block recursive updates to overlap in time. Furthermore, it employs a block quantization strategy consistent with the sequence blocks for feature tensors and mask blocks, and performs orthogonal perturbation on features before quantization to optimize the distribution of activation ranges. Simultaneously, a high-precision accumulator is used for normalization and accumulation, further reducing storage and bandwidth requirements while maintaining numerical stability. Additionally, it incorporates constraints on the spectral range, noise equivalent temperature difference, and frame rate of the infrared imaging unit, along with safety interlocking functions. When the excitation temperature rise or fall exceeds a safety threshold, excitation is automatically stopped and an alarm is recorded, ensuring the safety and usability of this invention in actual construction environments.

[0033] Through the synergy of the above methods and systems, this invention can achieve highly reliable estimation of the position, tilt angle, extension amount or elevation and fastening status of concealed sprinklers with minimal manual intervention without damaging the ceiling. It can also output an acceptance report that includes thermal image overlay, isotherm display, coordinate and time annotation, judgment conclusions and rectification suggestions, as well as material and excitation parameter traceability information, providing sufficient technical basis for project delivery and subsequent re-inspection. It has comprehensive beneficial effects of being non-destructive, fast, anti-interference, quantitatively traceable and feasible for engineering implementation. Attached Figure Description

[0034] Figure 1 The process of this invention Figure 1 ;

[0035] Figure 2 The process of this invention Figure 2 ;

[0036] Figure 3 The process of this invention Figure 3 ;

[0037] Figure 4 This is a system architecture diagram of the present invention. Detailed Implementation

[0038] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0039] Example 1

[0040] like Figure 1 and Figure 4 As shown, this embodiment realizes the quality acceptance of concealed nozzle installation based on infrared thermal imaging, and adopts a device system composed of a temperature difference excitation unit, a synchronous controller, an infrared imaging unit, and a processing and judgment unit.

[0041] The process is carried out sequentially according to the numbered steps. Modeling and calibration are performed in step S1. Information on the ceiling finish material and thickness, nozzle model, and its connecting components is obtained, and a material parameter library including thermal diffusivity, thermal conductivity, and temperature limits is established. The internal and external parameters of the infrared imaging equipment are calibrated and aligned with the field coordinate system. A synchronization controller synchronizes the infrared imaging equipment and the excitation device in time, achieving millisecond-level synchronization accuracy.

[0042] In step S2, controlled excitation and triggering are performed. A controlled thermal pulse or cold pulse excitation is applied to the piping network or sprinkler connection metal parts in the area where the sprinklers are located within the ceiling. The excitation includes a settable pulse width, temperature difference amplitude, and frequency band identifier. The temperature rise or fall of the excitation is limited by the safety threshold of the finish material and the sprinkler connection components. When the detected temperature approaches the threshold, the excitation intensity is automatically reduced or the excitation is stopped, and an alarm is recorded. The synchronization controller sends a synchronization trigger signal to the infrared imaging device.

[0043] In step S3, data acquisition and preprocessing are performed. Infrared image sequences are acquired before, during, and after excitation. Non-uniformity correction and background drift suppression are performed. Temporal observation features, including isothermal contour radius, temperature difference response, temperature gradient center, and contour ellipticity, are extracted and time-registered with the excitation parameters.

[0044] In step S4, the quantization information is selectively updated. Adaptive compensation is performed for different finish materials and thicknesses based on the material parameter library. Based on the time-series observation characteristics and excitation parameters, the quantization information corresponding to the time-series observation characteristics, such as thermal conductivity characteristics, thermal response curves, or relative position change trends, is selectively updated. Specifically, the update weight is increased when the excitation is effective or the signal-to-noise ratio is high, and the update weight is decreased when environmental disturbances or occlusion occur.

[0045] In step S5, a structured temporal mask is constructed. Based on the update weights obtained in S4, a temporal mask that is causal in time and decays with time intervals is constructed. The temporal mask is used to fuse features from different time points, different perspectives, and different excitation frequency bands.

[0046] In step S6, block fusion and recursion are performed. The temporal mask and its corresponding temporal observation features are divided into multiple sequence blocks according to time. Within each sequence block, a mask-based secondary fusion calculation is performed. Recursive fusion updates are performed between adjacent sequence blocks. Numerical normalization is performed within the block, and the fusion state and normalization factor are passed between blocks to obtain the numerically processed fusion result.

[0047] In step S7, the installation geometry and fastening status are estimated. Based on the temporal status and fusion results, the position of the nozzle in the ceiling plane, the tilt angle of the nozzle axis relative to the vertical direction, the nozzle extension or elevation, and the fastening status index determined by the combined temperature difference response decay rate and local micro-vibration characteristics are estimated, and the confidence level of the estimation results is given.

[0048] In step S8, compliance determination and reporting are performed. Based on the preset installation compliance threshold, the normalized errors of the eccentricity, tilt angle, extension amount and fastening status indicators determined by the location are calculated. According to the weighted scoring rules based on the importance of each normalized error and preset weights, the conclusion of whether it is qualified or not, the problem type and confidence level are output, and an acceptance report including heat map overlay, isotherm display, coordinate and time annotation, conclusion and rectification suggestions are automatically generated.

[0049] This embodiment enables the location of concealed sprinklers and the quantification of installation geometry and fastening status without damaging the ceiling. Controlled excitation and synchronous triggering ensure consistency between excitation and observation; selective state updates and structured temporal masks enhance the utilization of effective thermal response and suppress environmental drift; block fusion and recursion ensure numerical stability while maintaining computational efficiency. The final output includes position, tilt angle, extension amount, and fastening status indicators, performs compliance judgment, and generates a traceable acceptance report containing thermal maps and isotherms.

[0050] The processing and decision unit performs time registration, associating each frame of image with the excitation parameters via timestamps. Temporal observation features are extracted, including the isothermal contour radius. Temperature difference response Temperature gradient center With profile ellipticity .

[0051] The formula for calculating the center of the temperature gradient is:

[0052]

[0053] in For pixels gradient magnitude at that point The values ​​are the row and column indices in the image coordinate system, and the summation range is all pixels within the region of interest.

[0054] The processing and judgment unit determines the thermal diffusivity based on the material parameter library. Thermal conductivity Adaptive compensation is applied to the thickness of time-series observations. The compensation coefficient is selected by defining a time-scale normalization factor. The isothermal profile radius is determined according to Normalize the temperature difference response according to Normalize.

[0055] The updated weight calculation formula is as follows:

[0056]

[0057] in Normalized signal-to-noise ratio (the ratio of peak temperature response to standard deviation of background noise). This is the occlusion probability (the occlusion detection logic is described in detail in Example 2, and is set to 0 in Example 1). To incentivize valid identifiers (confirmed based on excitation amplitude, frequency band, and safety threshold). These are weighting coefficients (typical values ​​are 0.4, 0.4, and 0.2). To ensure the truncation function .

[0058] A weighted recursive update mechanism is adopted:

[0059]

[0060] in It is a time-series state vector. Here is the state transition matrix. To control the input matrix, To control the input vector.

[0061] The core physical relationships are as follows:

[0062] Relationship between diffusion radius and time:

[0063] in The thermal diffusivity of the material (unit: mm² / s). The time elapsed after the stimulus (in seconds). , where is the initial boundary constant (a geometric factor reflecting the initial contact characteristics between the nozzle and the finish, in mm²).

[0064] Temperature response time decay:

[0065] in To stimulate the initial temperature difference amplitude (unit: °C). The time elapsed after the stimulus (in seconds). The decay time constant (reflecting the combined effect of heat conduction and heat dissipation, in seconds).

[0066] The core form of a timing mask is:

[0067] in The first time sequence mask matrix Line number Column elements, For the first The update weight at each moment (calculated from the previous step S4). For from the first Time to the The product of all updated weights at any given time. The attenuation coefficient is related to the material (unit: s⁻¹). For time indexing. The relationship between the attenuation coefficient and material parameters is as follows: ,in This is the base value for the attenuation coefficient (typical range 0.1~1.0 s⁻¹·mm⁻², selected according to material type). For the thermal diffusivity of the material, The thickness is the material thickness.

[0068] Perform a secondary fusion calculation within each sequence block:

[0069] in For the characteristic matrix, For the time sequence mask submatrix, This is the feature matrix after fusion.

[0070] Location estimation: Extract the time average of the temperature gradient center from the fusion results and convert it into ceiling plane coordinates. .

[0071] Tilt Angle Estimation: Based on the relationship of profile ellipticity, when the nozzle axis is tilted relative to the ceiling plane, the heat diffusion area will exhibit an elliptical shape, and its ellipticity is correlated with the tilt angle. The tilt angle estimation formula is as follows: ,in For the ellipticity of the profile ( (These are the radii of the major and minor axes of the profile, respectively). The tilt angle (in rad or °) is the angle between the nozzle axis and the vertical direction. This relationship is based on the geometric relationship between the heat source point and the surface plane, and has good linear approximation within the tilt angle range of no more than 15°.

[0072] Elongation estimation: The initial boundary constant is obtained by fitting the diffusion radius relationship. Isothermal profile radius data at multiple times With physical models Perform least squares fitting to obtain Value. The inverse formula is: ,in This refers to the nozzle extension distance or elevation (in mm). This is the standard installation reference height (unit: mm). These are the initial boundary constants obtained from the current fitting. The initial boundary constants are the standard reference. The conversion factor (unit: mm² / mm) is used to determine the deviation between the extension amount and the initial boundary constant.

[0073] Tightness index: Combining the decay rate of temperature difference response and local micro-vibration characteristics, the calculation formula is as follows:

[0074] in This is an indicator of tightness (the higher the value, the better the tightness). Weighting coefficients (satisfying) Typical values ​​are 0.6 and 0.4. The decay time constant (in seconds) is obtained by fitting the temperature difference response decay curve. The mapping function from decay time constant to fastening index (monotonically increasing, When smaller Larger size indicates better tightness. The standard deviation of temperature fluctuation in the excited sequence (in mK). The mapping function from micro-vibration characteristics to fastening parameters (monotonically decreasing, the stronger the vibration) The smaller the value, the worse the tightness.

[0075] Calculate the normalization error:

[0076] Eccentricity error: ,in For the estimated planar position, For standard reference position, Euclidean distance. The position error tolerance is in mm.

[0077] Tilt angle error: ,in For the estimated dip angle, This refers to the tolerance for the tilt angle (in degrees or rad).

[0078] Extension error: ,in For the estimated extension amount, For standard extension amount, For the allowable extension (in mm). Tightening condition error: ,in For the estimated fastening index, The tolerance for fastening parameters is 0.1~0.2 (typical value).

[0079] Overall rating:

[0080] in Weights for importance of each item (satisfying) Typical values ​​are 0.25, 0.25, 0.25, 0.25, or adjusted according to project specifications.

[0081] Pass / Fail Judgment Rules: When It was judged as qualified at that time. It was judged as unqualified at that time.

[0082] Automatically generate acceptance reports, including heatmap overlays, isotherm displays, coordinate and time annotations, conclusions and rectification suggestions, material and excitation parameter traceability information, equipment pose and timestamps. Also export data packages, packaged in CSV or JSON format and including MD5 verification.

[0083] The entire data processing process is executed sequentially within the same task pipeline, with no manual interruptions and log traceability capabilities.

[0084] It employs non-contact infrared thermal imaging technology, eliminating the need to damage the ceiling structure; it achieves cross-material generalization through a material parameter library and adaptive compensation; the estimation results based on the physical model have interpretability and confidence intervals; and complete safety interlocks and audit records ensure the safety and traceability of the testing process.

[0085] Those skilled in the art can directly implement the above disclosure and adjust the relevant parameters according to different project requirements to adapt to specific engineering conditions.

[0086] Example 2

[0087] like Figure 1 , Figure 2 As shown, this embodiment, based on the device and process of embodiment 1, achieves robust enhancement for complex working conditions such as high noise, obstruction, and weak response through multi-view / multi-frequency configuration and high noise backoff strategy, so as to support stable acquisition of installation geometry and fastening status estimation under the conditions of personnel movement, local wind field disturbance, uneven ceiling material, and frame loss on site.

[0088] During on-site deployment, the infrared imaging unit employs a dual-view camera to acquire data from the same target area. The two views are calibrated using extrinsic parameters and projected onto a unified ceiling-mounted coordinate system, enabling cross-view data alignment. The temperature difference excitation unit retains dual channels for both thermal and cold pulses, with each excitation accompanied by a frequency band identifier to distinguish amplitude and thermal type, facilitating layered processing in subsequent timing masking and block fusion. The synchronization controller simultaneously sends triggers and timestamps to both views, ensuring consistency of the time base across multiple views and frequency bands.

[0089] In step S4, the quantitative information corresponding to the time-series observation features, such as thermal conduction characteristics, thermal response curves, or relative position change trends, is selectively updated. An update weight calculation and anomaly backoff mechanism for multi-view / multi-frequency observations are introduced. The processing and judgment unit calculates indicators such as signal-to-noise ratio, occlusion detection results, and temperature difference variance for each viewpoint and frequency band of the time-series observation features. Based on these indicators, update weights are determined to ensure that observations with effective excitation, high signal-to-noise ratio, and low occlusion probability contribute more to the quantitative information.

[0090] Signal-to-noise ratio (SNR) is the ratio of the peak value of the temperature difference response to the standard deviation of the background noise. The calculation formula is:

[0091]

[0092] in For the first Perspective 1 Signal-to-noise ratio at any given time (dimensionless). This represents the maximum temperature difference response (in °C) at this point in time from this viewpoint. The spatial standard deviation of the image during the baseline period before excitation from this perspective represents the background noise level (in °C or mK). Index the viewpoint (first or second viewpoint).

[0093] Occlusion detection results refer to indicators used to determine whether the current frame is occluded. Occlusion detection is measured by changes in local contrast and edge integrity. The occlusion determination threshold is adaptively set according to background noise: the gradient center displacement threshold is taken as a percentage of the standard deviation of the background pixel position. The temperature difference abrupt change threshold is taken as a multiple of the background temperature standard deviation. times, of which This is an adaptive coefficient (ranging from 2 to 4, automatically estimated from the statistical data collected in the pre-collection window). The minimum number of frames (preferably 2 to 3 frames, to suppress glitches) required to continuously meet the conditions for occlusion. Local contrast is defined as the ratio of the temperature difference between the peak temperature response region and the surrounding region; edge integrity is evaluated by the continuity and closure of the isotherm contours. Occlusion probability. The calculation takes into account both the decrease in contrast and the degree of edge breakage; the larger the value, the higher the probability of occlusion.

[0094] Temperature variance refers to the time variance of the temperature response within the excitation window, and the calculation formula is:

[0095]

[0096] in For the first Temperature variance at different angles (unit: °C or mK). For the set of frame indices of the post-excitation analysis window, The number of frames within the window. This represents the time average of the temperature difference response within the window. The temperature difference variance reflects the stability of the thermal response; an excessively large variance indicates abnormal fluctuations.

[0097] The updated weights are determined based on the above indicators, ensuring that observations with effective excitation, high signal-to-noise ratio, and low occlusion probability contribute more to the quantification information. The core calculation relationship for the updated weights is as follows:

[0098]

[0099] in For the first Perspective 1 Weights are updated in real time. This represents a cutoff function that restricts the value to between 0 and 1. Weighting coefficients (satisfying) (The sum of the three is 1, with no requirement) For the first The excitation validity is determined by the excitation amplitude, frequency band, and safety threshold. For complex multi-view / multi-frequency conditions, the weighting coefficients require a larger dynamic range to accommodate drastic changes in signal-to-noise ratio and occlusion probability; typical values ​​are... , , Compared to 0.4, 0.4, and 0.2 in Example 1, this example reduces the weight of signal-to-noise ratio to avoid high noise dominance in a single viewpoint, and significantly increases the weight of occlusion detection and excitation effectiveness to enhance robustness.

[0100] For data from two perspectives, the processing and decision-making units first calculate update weights within their respective perspectives during the update process, and then merge them according to reliability in a unified coordinate system. The merging form is a confidence-based weighted average, and the core cross-perspective fusion relationship is...

[0101]

[0102] in This is the time-series observation feature vector after cross-view fusion (including the fused isothermal profile radius, temperature difference response, gradient center, ellipticity, etc.). For the first Temporal observation feature vectors of perspective, For the first The fusion weights of the viewpoints (determined by the product of the update weights and the signal-to-noise ratio). For the first The update weight of the viewpoint (from the previous formula). For the first The signal-to-noise ratio (SNR) of the viewpoint. Through this weighted averaging method, views with higher SNR and larger update weights contribute more to the fusion result, thus improving the reliability of the fused features. To avoid the degradation situation where the weights of both views simultaneously approach zero, resulting in a zero denominator, when... hour( (Using the minimum threshold) the joint features of the most recent valid moment are used to maintain and mark the frame as low confidence.

[0103] Temporal observation features after cross-view fusion This is used to selectively update the quantitative information corresponding to the time-series observation features, such as thermal conductivity characteristics, thermal response curves, or relative position change trends. Specifically, the method involves fusing the time-series observation features... As control input vector A portion of the data is substituted into the recursive update formula in Example 1 to perform state updates, thereby obtaining time-series data that reflects quantitative information such as the thermal conduction characteristics, thermal response curve, or relative position change trend between the nozzle and the finish.

[0104] If the update weight remains below the threshold or the temperature difference variance is abnormal, the processing and judgment unit automatically triggers a strong-excitation short-sequence acquisition and intra-block fusion priority strategy. A continuously low update weight means that the update weights of both views are below a preset threshold for several consecutive frames. The preset threshold typically ranges from 0.3 to 0.5, and the number of consecutive frames is typically five to ten frames. An abnormal temperature difference variance means that the temperature difference variance exceeds a preset upper limit. The preset upper limit is determined based on the material type and excitation amplitude, and typically ranges from 30% to 50% of the initial temperature difference amplitude.

[0105] Strong excitation short sequence refers to increasing the temperature difference amplitude and shortening the pulse width for a short period of time, without exceeding the safety threshold, to significantly improve the peak response. The specific implementation method is as follows: when the trigger condition is detected, the synchronous controller sends a strong excitation command to the temperature difference excitation unit, which increases the temperature difference amplitude by 20% to 50% and simultaneously increases the pulse width... The excitation duration is typically shortened by 30% to 50%, with a typical duration of 0.3 to 1 second. During strong excitation, the synchronous controller continuously monitors the temperature sensor readings to ensure that the temperature rise or drop does not exceed the safety threshold of the finish material. If the temperature approaches the safety threshold, the excitation is immediately stopped and an alarm is recorded. To prevent heat accumulation, a minimum cooling interval is set between two consecutive strong excitations. The method for determining the minimum cooling interval is as follows:

[0106] Based on the heat capacity of the finishing material With heat dissipation rate Calculate the cooling time constant ;

[0107] in Specific heat capacity (unit: J / (kg·K)) Density (unit: kg / m³) Thickness (in meters). The convective heat transfer coefficient is expressed in W / (m²·K). This refers to the heat dissipation area (unit: m²). The temperature difference with the environment (in K).

[0108] Minimum cooling interval set to ,in For a safety margin (typically 3 to 5), ensure that the temperature drops to below 95% of its initial value.

[0109] The intra-block fusion priority strategy refers to prioritizing the use of high-confidence information within the current sequence block for estimation during block fusion of the current sequence, reducing reliance on information from distant time points. Specifically, when constructing the time-series mask, for time periods triggering strong excitation short sequences, the attenuation coefficient of the time-series mask is adjusted. Temporarily increasing the mask weights to two to three times the normal value causes the mask weights to decay more rapidly with respect to information from further time points within that time period, thus making the fusion result more reliant on new observations within the current sequence block. It only takes effect within the frame interval covered by the strong excitation short sequence, and a transition band is set before and after this interval with a slow start and slow exit. The length of the transition band is two to four frames to avoid numerical oscillations caused by sudden weight changes. By combining the strong excitation short sequence with the intra-block fusion priority strategy, the effective thermal response ratio can be quickly recovered and stably estimated in high-noise scenarios.

[0110] In step S5, a structured temporal mask is constructed. To adapt to the dynamic changes in multi-view / multi-frequency data and update weights, a temporal mask with causality and decay over time is adopted. The mask weights are based on the update weights, and a frequency-band-related attenuation factor is introduced to differentiate the time windows of the fusion impact of different frequency bands. The frequency-band-related attenuation factor refers to the different attenuation coefficients corresponding to different excitation frequency bands, reflecting the time-scale differences in the thermal diffusion process under different excitation conditions.

[0111] To clarify the core components of a mask, the key relationship of a mask is given, the formula is as follows:

[0112]

[0113] in frequency band The corresponding timing mask matrix of the first Line number Column elements, The updated weights after cross-view fusion (obtained from the cross-view fusion calculation in step S4). Indicates from the first Time to the The product of all updated weights after cross-view fusion at any given moment. In order to be compatible with frequency bands The relevant attenuation factor (unit: s⁻¹ or frame⁻¹). For time index (unit: frames or time in seconds). is the base of the natural logarithm. When A time mask weight of 0 ensures causality. The time mask weights are determined by the product of consecutive updates (reflecting the cumulative weights of historical updates) and the exponential decay (reflecting the decay due to thermal diffusion).

[0114] Frequency band related attenuation coefficient The attenuation coefficient is determined as follows: For thermal pulse excitation, the attenuation coefficient is determined based on the temperature difference amplitude and the material's thermal diffusivity; the larger the temperature difference amplitude or the higher the thermal diffusivity, the larger the attenuation coefficient. The determination method for cold pulse excitation is similar. Typical values ​​range from 0.1 to 1.0 per second. The quantitative relationship between the attenuation coefficient and the material's thermal diffusivity, thickness, excitation amplitude, and pulse width is as follows:

[0115]

[0116] in The fundamental coefficients related to the frequency band (typically 1.5 to 3.0 for thermal pulse excitation and 1.0 to 2.0 for cold pulse excitation, in s⁻¹·mm⁻²). The thermal diffusivity of the material (unit: mm² / s). Material thickness (unit: mm). This is the temperature difference amplitude correction factor (typically 0.1 to 0.3). The actual excitation temperature difference amplitude (unit: °C). This is a reference temperature difference range (typically 10℃). This is the pulse width correction factor (typically 0.1 to 0.2). Pulse width (in seconds). This is a reference pulse width (typically 1 second). During field deployment, The values ​​are obtained by referring to a table based on the material's thermal diffusivity, thickness, excitation amplitude, and pulse width ternary set.

[0117] The mask ensures that near-time features account for a higher proportion in the fusion after excitation, while the proportion of far-time features gradually decreases, and allows different attenuation rates to be set for different frequency bands, thereby suppressing the interference of low-amplitude or hysteresis responses on the estimation.

[0118] In step S6, block fusion and recursion are performed. To adapt to inspection sequences of different lengths and terminal computing power limitations, the sequence is divided into multiple sequence blocks according to time. The block length and number of blocks are adaptively determined according to the target sequence length and hardware resources. Within each sequence block, a mask-based secondary fusion calculation is performed, and recursive fusion updates are performed between adjacent sequence blocks. The maximum row value and cumulative sum are maintained within the block for numerical normalization, and the fusion state and normalization factor are passed between blocks to maintain numerical continuity and stability.

[0119] In multi-view / multi-frequency scenarios, intra-block secondary fusion first performs hierarchical weighting by frequency band, then performs confidence fusion by viewpoint, ultimately obtaining joint features under the same time index. Hierarchical weighting by frequency band means using the corresponding time mask for each frequency band. A secondary fusion calculation is performed to obtain the fused features for that frequency band. Confidence fusion by viewpoint refers to weighting the fused features from two viewpoints within the same frequency band according to their respective confidence levels. The confidence level is determined by the product of the time-averaged weights and the time-averaged signal-to-noise ratio. The resulting joint features integrate information from multiple viewpoints and frequency bands, exhibiting higher robustness.

[0120] Inter-block recursion only transmits joint features and normalization factors, avoiding the spread of irrelevant information over time. Irrelevant information refers to intermediate calculation results and temporary variables from various perspectives and frequency bands; this information does not need to be passed to the next sequence block after fusion within the block. By only transmitting joint features and normalization factors, the amount of data transmitted between blocks is reduced, improving computational efficiency.

[0121] To ensure the feasibility of adaptive selection of block length and number of blocks, the core determining relationship of block length is given, and the formula is as follows:

[0122]

[0123] in The duration of each block (in frames). This represents the total number of frames in the sequence (in units of frames). This is the block length coefficient (dimensionless, a parameter set according to hardware resources and real-time requirements). The block length index (dimensionless, 0 < 0) <1), This represents the rounding function. Parameters Typical values ​​range from 5 to 20 for the parameter. The typical value ranges from 0.3 to 0.6. In actual deployment, and The frame rate, video memory, and throughput targets obtained from offline calibration can be jointly determined. The block implementation adopts a double-buffered or circular buffer structure, and performs overlapping operations between block loading and computation; when the video memory pressure is detected to exceed the upper limit, the block length is automatically shortened and the number of blocks is increased to avoid overflow, while maintaining the continuous transfer of normalization factors and fusion states between blocks.

[0124] Number of sequence blocks for ,in This indicates rounding up. By adaptively selecting the sequence block length, computational efficiency is ensured while meeting hardware resource constraints.

[0125] In steps S7 and S8, installation geometry and fastening status estimation, compliance determination, and report generation are performed, using the same processing method as in Example 1. For details, please refer to the description of steps S7 and S8 in Example 1. The fusion result of this example is directly input into the estimation algorithm of Example 1, and the output data structure remains consistent with that of Example 1.

[0126] The report additionally records traceability information specific to this embodiment:

[0127] Cross-view configuration: camera pose, extrinsic calibration parameters, and excitation parameters for each frequency band from two viewpoints;

[0128] Strong stimulus short sequence event: trigger time, strong stimulus temperature difference amplitude, pulse width, duration, trigger cause and safety interlock status.

[0129] Cross-view fusion significantly reduces misjudgments caused by environmental disturbances; the strong excitation short sequence and intra-block fusion priority strategy maintains stable detection performance in high-noise scenarios; the frequency band correlation attenuation factor enables reasonable allocation of time windows for different excitation conditions; and adaptive sequence block length selection ensures numerical stability and computational efficiency under different hardware conditions.

[0130] Those skilled in the art can implement this enhancement scheme by adding a second-view camera and a multi-band excitation device, and the parameters can be adjusted according to project requirements.

[0131] Example 3

[0132] like Figures 1 to 4 As shown, this embodiment is geared towards engineering deployment scenarios. It refines the system component division and execution mechanism of the device. Through the parallel execution mechanism of producers and consumers, block quantization and orthogonal disturbance, as well as safety interlocking and reporting traceability, the methodology formed in Embodiments 1 and 2 can achieve real-time, low power consumption and traceability requirements in actual construction sites or building operation and maintenance sites.

[0133] The device consists of a temperature difference excitation unit, a synchronization controller, an infrared imaging unit, and a processing and decision unit. The processing and decision unit has a built-in producer-consumer parallel execution mechanism. The producer is responsible for loading the sequence block data into a shared buffer block by block, while the consumer completes the mask-based secondary fusion within the block and the recursive update between blocks, and overlaps the data transfer and calculation between the two.

[0134] Producers are responsible for loading data corresponding to sequence blocks into a shared buffer. Consumers, accelerated by hardware, perform mask-based secondary fusion within blocks and recursive updates between blocks. Producers and consumers communicate through a circular shared buffer and overlap their computation and data transfer operations to avoid waiting.

[0135] The processing and decision unit employs the same block quantization strategy for feature tensors and mask blocks as for sequence blocks. Before quantization, orthogonal perturbations are applied to the features to optimize the distribution of activation ranges. Key intermediate quantities are normalized and accumulated using a high-precision accumulator. Orthogonal perturbations are applied to the features before quantization to balance the distribution of activation ranges. To ensure computational stability, the consumer maintains the row maximum value and accumulated sum within each block to perform numerical normalization, and writes the normalization factor along with the fusion state into the context of the next block.

[0136] The infrared imaging unit's spectral range covers long-wave or mid-wave thermal imaging, with the noise equivalent temperature difference not exceeding a preset threshold and the frame rate not falling below a preset value. The system is equipped with a safety interlock function; when the excitation temperature rise or fall exceeds the safety threshold, excitation automatically stops and an alarm is recorded. The temperature difference excitation unit and the synchronization controller work together to achieve the safety interlock.

[0137] After completing the intra-block secondary fusion and inter-block recursion, the processing and judgment unit directly calls the installation geometry and fastening status estimation module to output the position of the nozzle in the ceiling plane, the tilt angle of the nozzle axis relative to the vertical direction, the nozzle extension amount or elevation, and the fastening status indicators and their confidence levels. This output is then connected with the compliance judgment module to generate a conclusion on whether the product is qualified or not, the problem type, and the confidence level.

[0138] The report is generated using the method described in Example 1, and additionally includes engineering information specific to this example: technical parameters such as quantization scale, orthogonal matrix identifier, and normalization factor, as well as performance statistics of producers and consumers executing in parallel.

[0139] The entire process is initiated at the level of inspection tasks. The data sources are the excitation parameters and frequency band identifiers recorded by the synchronous controller and the infrared image sequence output by the infrared imaging unit. The processing includes sequence block loading, intra-block secondary fusion and maximum value normalization, inter-block recursion and normalization factor transfer, installation geometry and fastening status estimation, compliance judgment and report generation. The output is installation geometry and fastening status indicators, qualification conclusion and report for single point.

[0140] The beneficial effects of this embodiment are as follows: First, by implementing a parallel execution mechanism of producers and consumers and cross-block transfer of maximum value normalization and normalization factors, efficient coupling of intra-block computation and inter-block scheduling is achieved, improving throughput while ensuring numerical stability. Second, block quantization and orthogonal perturbation reduce bandwidth and storage consumption without changing the numerical equivalence of the fusion operator, enabling edge devices to operate stably under limited computing power and storage conditions. Third, the coordinated control of imaging unit indicators and safety interlocks ensures that controlled excitation and infrared acquisition are always within the safety boundary, and anomalies are immediately stopped and an auditable record is formed. Fourth, the reporting and traceability mechanism records material, thickness, excitation parameters, camera pose, and time information in a comprehensive manner, facilitating re-inspection and comparison, and project delivery.

[0141] To clarify the core numerical stability relationship, maximum value normalization is adopted in the intra-block secondary fusion stage, and the formula is as follows:

[0142]

[0143] in It is an intermediate quantity of similarity or weight calculated within the block by the feature tensor and the temporal mask (usually derived from the product of the mask and the feature vector). For the first The maximum value (scalar) of a row. The result after normalization. For row index (sequence position) This is the column index (feature dimension). This process ensures that weighted summation and recursive accumulation within a block do not overflow or underflow within a finite bit width. Normalization is performed in a fixed order: first, find the maximum row value; then, shift; finally, perform weighted summation and normalization. When the sum of the weights of a row is below a very small threshold (typically...), ... to When the normalization factor of the previous valid frame is used to backfill and the confidence level of the line is lowered, division by zero and amplitude drift are avoided.

[0144] Subsequently, during inter-block recursion, only the fusion state and normalization factor are transmitted to reduce cross-block dependencies and bandwidth consumption. Before cross-block transmission, the hash check value of the fusion state and normalization factor is calculated and written to the context; when the read side finds inconsistencies or missing check values, a safe rollback to the previous consistent block is triggered and the current block is recalculated, while recording the alarm and reason code.

[0145] Regarding block quantization and orthogonal perturbation, the processing and decision units employ the same block quantization strategy for feature tensors and mask blocks as for sequence blocks. Block quantization refers to quantizing feature tensors on a sequence block basis, with each sequence block independently selecting quantization parameters to adapt to the numerical distribution characteristics of different sequence blocks. Before quantization, orthogonal perturbation is applied to the features to optimize the distribution of activation ranges. Orthogonal perturbation involves left-multiplying the feature vectors by a random orthogonal matrix before quantization to balance the distribution of feature vectors in all directions and avoid quantization saturation caused by large components in certain directions.

[0146] The specific implementation method of orthogonal perturbation is as follows: for the characteristic matrix , , Before quantization, multiply by random orthogonal matrices on the left respectively. , , The perturbated feature matrix is ​​obtained. , , Due to attention calculation ,when From time to time The numerical values ​​of the attention matrix are not changed, ensuring the numerical equivalence of the fusion operator. The orthogonal matrix is ​​generated using a fixed random seed and remains fixed in-situ, thus guaranteeing that the merged result is numerically equivalent to the unperturbed case. To ensure reproducibility, the orthogonal matrix is ​​generated using a fixed random seed and archived along with the software version, timestamp, and device serial number; the orthogonal matrix is ​​column-normalized (i.e., satisfies...). ,in (As an identity matrix) and verified by checksums. The decoding end performs inverse orthogonal transformation and inverse quantization in a fixed order: inverse quantization followed by inverse orthogonal transformation, i.e., first quantized... Perform dequantization to obtain a floating-point value, then multiply by the left. Restore the original values ​​and maintain the compatibility table during version upgrades to ensure that historical data is reversible.

[0147] Regarding safety interlocking and imaging unit specifications, the infrared imaging unit's spectral range covers long-wave or mid-wave thermal imaging, with the noise equivalent temperature difference (NETD) not exceeding a preset threshold and the frame rate not falling below a preset value. Spectral range coverage for long-wave thermal imaging refers to an operating wavelength of 8 to 14 micrometers (center wavelength approximately 10 micrometers), while spectral range coverage for mid-wave thermal imaging refers to an operating wavelength of 3 to 5 micrometers (center wavelength approximately 4 micrometers). The noise equivalent temperature difference (NETD) refers to the smallest temperature difference that the infrared imaging unit can resolve (equivalent to the noise standard deviation), with a typical preset threshold of 50 millikrvin (mK). The frame rate refers to the number of image frames acquired per second (in fps or Hz) by the infrared imaging unit, with a typical preset value of 30 frames per second. These specifications ensure that the infrared imaging unit can capture sufficiently clear thermal response signals to meet the accuracy requirements of subsequent processing.

[0148] The temperature difference excitation unit and the synchronization controller work together to achieve a safety interlock. The safety interlock refers to a protection mechanism that automatically stops excitation and records an alarm when the detected excitation temperature rise or fall exceeds a safety threshold. Specifically, the processing and judgment unit continuously reads the peak temperature difference observed by the temperature sensor and camera. When the detected excitation temperature rise or fall approaches the safety threshold, it immediately reduces the excitation intensity or stops excitation and records an alarm. To prevent heat accumulation, a minimum cooling interval is set between two consecutive excitations. When multiple points operate in parallel, a throttling strategy is implemented for the same local area to ensure that the total heat load per unit time does not exceed the upper limit. Once a warning zone is triggered, the system enters a degraded mode within that area, completing the remaining process only with low-amplitude excitation. Minimum Cooling Interval The time in seconds is determined based on the heat capacity and heat dissipation rate of the surface material. A localized area refers to a radius of 1 to 3 meters centered on the current detection point. - (meters). A throttling strategy refers to limiting the number of points that can be simultaneously stimulated within a given area (typically no more than 2 to 3 points). Total heat load. This refers to the total heat input (in W / m²) of all excitation devices in the area per unit time. The upper limit is determined based on the temperature resistance and heat dissipation capacity of the finish material, typically ranging from 100 W / m² to 500 W / m². The warning zone refers to the range where the temperature rise or fall reaches 70% to 80% of the safe threshold. Degradation mode refers to a protective operating mode that reduces the excitation amplitude and frequency. Low-amplitude excitation refers to reducing the temperature difference amplitude to 1 / 2 to 1 / 3 of the normal value.

[0149] Simultaneously, corresponding alarm event entries are generated in the report, including trigger time, excitation parameters, camera pose, and acquisition frame index, facilitating review. An alarm event entry is a data record that documents detailed information about an abnormal event, including the timestamp of the alarm trigger, the excitation parameters at the time, the pose parameters of the infrared imaging unit, and the image frame index at the time of the alarm. This information is recorded in the appendix of the acceptance report or in the log file, facilitating subsequent analysis of the alarm causes or verification through review.

[0150] The imaging unit provides a sequence of infrared images before, during, and after excitation according to the inspection cycle. All frames are accompanied by metadata such as timestamps, camera pose, and exposure parameters to ensure temporal registration and geometric alignment within the processing and decision unit. The inspection cycle refers to the rhythm (in seconds or frames per second) of detection performed according to a predetermined time interval or detection point sequence. Camera pose refers to the position and orientation of the infrared imaging unit, including three-dimensional coordinates. Related to rotation angles (roll, pitch, yaw). Exposure parameters refer to the integration time of the infrared imaging unit. Gain Configure parameters. Metadata is appended to the image file in a structured format (such as JSON or XML) or stored as a separate metadata file, corresponding one-to-one with the image frame (i.e., each frame image ID uniquely corresponds to one metadata record).

[0151] After completing the block fusion and recursion, the processing and judgment unit calls the installation geometry and fastening state estimation and compliance judgment module in steps S7 and S8 of Example 1 to output the test results and qualification conclusions. Modules exchange data through function calls or message passing.

[0152] The data package is in CSV or JSON format and contains the basic data from Example 1, as well as engineering parameters specific to this example, such as quantization scales, orthogonal matrix identifiers, and normalization factors, and generates an MD5 checksum. Reports and data packages are archived using a unified naming convention for easy auditing and recalculation.

[0153] The entire process adopts the data processing link of Example 1, and is deployed in an engineered manner through a producer-consumer parallel mechanism. The processing includes sequence block loading, block fusion and recursion, geometric estimation, compliance judgment and report generation.

[0154] To ensure throughput under different computing power conditions, this embodiment presents the core balancing relationship for block-level throughput, as shown in the formula:

[0155]

[0156] in This represents block-level throughput (units per second or features per second). The duration of the sequence block (in frames). The feature dimension or equivalent number of channels (units). The time (in seconds) for the producer to load the block into the shared buffer. The time (in seconds) for the consumer to complete the secondary fusion and recursion of this block. This is determined by selecting an appropriate block length. By sharing the number of buffer slots, loading and computation overlap in time, thereby maximizing... .

[0157] The specific optimization method is as follows: when At this point, producers and consumers can work continuously without waiting for each other, and throughput reaches its maximum. If If this happens, the consumer will wait for the producer; in this case, the number of shared buffer slots should be increased or the data loading speed should be optimized. If the producer waits for the consumer, then computation speed should be optimized or hardware acceleration should be adopted. This relationship provides a direct basis for engineering parameter selection. Those skilled in the art can select the parameter configuration that maximizes throughput by testing different combinations of block lengths and slot numbers based on actual hardware configuration and performance requirements. The system updates after processing each sequence block. and The sliding estimate (such as the average of the last 10 blocks); when Time-triggered adaptive, where This is the imbalance threshold (typically 0.2 to 0.3); if If the number of slots is increased or the block length is decreased, then the block length is increased or the parallelism is improved. The parameter adjustment process records the values ​​before and after the adjustment and the time of the effect to ensure traceability.

[0158] During operation, the system first initializes the material parameter library, camera calibration parameters, and safety thresholds according to the inspection plan. The inspection plan refers to a pre-defined list of inspection points, inspection sequence, inspection parameters, etc. (usually stored in a table or configuration file). Initialization refers to loading the necessary configuration parameters and calibration data (including internal and external parameters) before starting inspection. Ceiling Plan Material library (etc.). Subsequently, at each detection point, the synchronous controller issues a trigger, the temperature difference excitation unit executes controlled excitation, and the infrared imaging unit acquires image sequences before, during, and after the excitation, along with metadata.

[0159] The producer process of the processing and judgment unit loads feature and temporal mask blocks into a shared buffer in units of sequence blocks. A producer process refers to a software process or thread that runs the producer function. The specific execution steps are as follows: the producer process extracts temporal observation features (including...) from the infrared image sequence. According to the division of sequence blocks (block length is...), Organize the features into a feature matrix Read the corresponding timing mask block The feature matrix is ​​subjected to orthogonal perturbation and block quantization. The quantized feature matrix and timing mask block are written into the free slots of the shared buffer, and the slot status flag is updated.

[0160] The consumer process sequentially reads and completes intra-block secondary fusion and maximum value normalization, inter-block recursion and normalization factor passing, and checks safety interlock conditions in real time. A consumer process refers to a software process or thread that runs the consumer function. The specific execution steps are: the consumer process reads the feature matrix from the filled slots of the shared buffer. With timing mask block Perform inverse quantization and inverse orthogonal transformation on the quantized data ( (etc.) Restore the original numerical range and perform matrix multiplication to calculate the secondary fusion result. Maintain the maximum row value Perform numerical normalization with the summation (as in the formula) The fusion results will be combined with the normalization factor. The data is passed to the next sequence block or written to the final output buffer, and the slot status flag is updated to release the slot for reuse by the producer. During processing, the consumer process continuously monitors the peak temperature difference between the temperature sensor reading and the infrared image. If the difference approaches the safety threshold, a safety interlock mechanism is triggered.

[0161] After all sequence blocks have been processed, the installation geometry and fastening state estimation module is invoked to output the position of the nozzle within the ceiling plane. The tilt angle of the nozzle axis relative to the vertical direction Nozzle extension or elevation and fastening condition indicators Based on confidence level and compliance assessment, a pass / fail rating and issue type are generated, along with heatmap overlay, isotherm display, coordinate and time annotations, to produce an acceptance report, which is then archived. "All sequence blocks processed" means the consumer process has completed processing the last sequence block (the...). (Block) and output the final fusion result. Calling refers to initiating the corresponding software module through function calls or message passing. Combining refers to combining multiple pieces of information or images to form a complete report.

[0162] The entire link contains no manual breakpoints. Abnormal events are uniformly recorded through logs and alarms, and are included in the report as source entries. Abnormalities are uniformly categorized as: excessive temperature rise, frame loss, buffer congestion, hash inconsistency, accelerator rollback, and packet verification failure. Each type of abnormality includes a cause code, start and end frames, affected blocks, handling strategies, and an identifier indicating whether it affects the acceptance conclusion, all listed in tabular form in the report appendix. The cause code is a numerical code identifying the specific cause of the abnormality, typically three to five digits, facilitating log analysis and fault diagnosis. Start and end frames refer to the starting and ending frame indices of the abnormality. Affected blocks are a list of sequence block numbers affected by the abnormality. Handling strategies include safe rollback, degradation mode, skipping the block, and using default values. The identifier indicating whether it affects the acceptance conclusion is a Boolean value indicating whether the abnormality causes the final acceptance conclusion to change from acceptable to unacceptable or significantly decrease in confidence. The report appendix is ​​an additional section of the acceptance report used to record detailed technical information and abnormal events.

[0163] In terms of engineering deployment, those skilled in the art can implement this system on industrial control platforms or edge computing devices, supporting x86 / ARM architecture and equipped with 4-16 core processors and 8-32GB of memory. The system supports stand-alone or distributed deployment modes and has the capability for online parameter adjustment and remote software upgrades.

[0164] In terms of performance metrics, this embodiment maintains the accuracy metrics of Embodiment 1 while achieving engineering performance improvements: Processing time: 10-15 seconds for single-point processing, linear scaling for multi-point processing. Resource consumption: Reduces memory and bandwidth requirements by 75% through block quantization, with accuracy loss of <1%. Hardware compatibility: Supports deployment on edge devices with 4-16 core processors and 8-32GB of memory.

[0165] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for quality acceptance of concealed sprinkler installation based on infrared thermal imaging, characterized in that, Includes the following steps: S1. Modeling and Calibration: Obtain information on ceiling finish material and thickness, nozzle model and its connecting components, and establish a material parameter library including thermal diffusivity, thermal conductivity and temperature resistance limits; complete the calibration of the internal and external parameters of the infrared imaging equipment and align it with the field coordinate system, and use a synchronization controller to synchronize the infrared imaging equipment and the excitation device in time; S2. Controlled excitation and triggering: A controlled thermal pulse or cold pulse excitation is applied to the pipe network or metal parts connecting the nozzles in the area where the nozzles are located in the ceiling. The excitation includes a settable pulse width, temperature difference amplitude and frequency band identifier, and the synchronization controller sends a synchronization trigger signal to the infrared imaging device. S3. Acquisition and Preprocessing: Infrared image sequences are acquired before, during and after excitation, non-uniformity correction and background drift suppression are performed, and temporal observation features including at least isothermal contour radius, temperature difference response, temperature gradient center and contour ellipticity are extracted. The temporal observation features are then time-registered with the excitation parameters in a one-to-one correspondence. S4. Selectively update quantization information: Adaptively compensate for different finish materials and thicknesses based on the material parameter library; based on the time-series observation features and excitation parameters, selectively update the quantization information of thermal conductivity characteristics, thermal response curves, or relative position change trends corresponding to the time-series observation features, wherein the update weight is increased when the excitation is effective or the signal-to-noise ratio is high, and the update weight is decreased when environmental disturbances or occlusion occur; thereby obtaining time-series data reflecting the quantization information of thermal conductivity characteristics, thermal response curves, or relative position change trends between the nozzle and the finish. S5. Structured temporal mask construction: Based on the update weights obtained in S4, a temporal mask that is causal in time and decays with time intervals is constructed. The temporal mask is used to fuse features from different time points, different perspectives and different excitation frequency bands. S6. Block fusion and recursion: The temporal mask and the corresponding temporal observation features are divided into multiple sequence blocks according to time. A secondary fusion calculation based on the mask is performed in each sequence block. A recursive fusion update is performed between adjacent sequence blocks. Numerical normalization is performed within the block, and the fusion state and normalization factor are passed between blocks to obtain the fusion result after numerical processing. S7. Installation geometry and fastening status estimation: Based on the temporal state and the fusion result, estimate the position of the nozzle in the ceiling plane, the tilt angle of the nozzle axis relative to the vertical direction, the nozzle extension amount or elevation, and the fastening status index determined by the temperature difference response decay rate and local micro-vibration characteristics, and give the confidence level of the estimation results. S8. Compliance Judgment and Report: Based on the preset installation compliance threshold, calculate the normalized errors of the eccentricity, tilt angle, extension amount and fastening status indicators determined by the location. According to the weighted scoring rules based on the importance of each normalized error and preset weights, output the conclusion of whether it is qualified or not, the problem type and confidence level, and automatically generate an acceptance report including heat map overlay, isotherm display, coordinate and time annotation, conclusion and rectification suggestions.

2. The method according to claim 1, characterized in that, The temperature difference amplitude of the thermal pulse or cold pulse mentioned in S2 is within a preset range, and the pulse width and duty cycle are adjustable. The excitation temperature rise or drop is limited by the safety threshold of the surface material and the nozzle model and its connecting components. When the temperature is detected to be close to the threshold, the excitation intensity is automatically reduced or the excitation is stopped and an alarm is recorded.

3. The method according to claim 1, characterized in that, The time synchronization described in S1 adopts a dual mechanism of trigger signal and timestamp, with a synchronization accuracy of no less than millisecond level; the preprocessing described in S3 includes a combination of inter-frame difference and exponential smoothing to suppress slow changes in ambient temperature and imaging drift.

4. The method according to claim 1, characterized in that, The material parameter library mentioned in S1 includes at least the thermal parameters and thickness ranges of gypsum board, mineral wool board and metal plate; the adaptive compensation in S4 selects different compensation coefficients according to the material and thickness, and allows the compensation coefficients to be corrected by a small number of field calibration samples.

5. The method according to claim 1, characterized in that, The selective updating of the quantification information of the thermal conduction characteristics, thermal response curves or relative position change trends corresponding to the time-series observation features described in S4 is based on the signal-to-noise ratio, effective excitation identifiers and occlusion detection results to determine the update weight. When the update weight is continuously lower than the threshold or the temperature difference fluctuation is abnormal, a strong excitation short-sequence acquisition and intra-block fusion priority strategy is triggered to improve the reliability of estimation in high-noise scenarios.

6. The method according to claim 1, characterized in that, The temporal mask described in S5 is based on the updated weights and introduces a decay factor with time intervals, so that the near-time features have a higher proportion in the fusion after excitation and the far-time features have a gradually decreasing proportion. In S6, the time length and number of each sequence block are adaptively determined according to the target sequence length and hardware resources. The secondary fusion based on the mask is performed within the block, and the recursive fusion is performed between blocks. The maximum value and the cumulative sum are maintained within the block to enhance the robustness of numerical processing.

7. The method according to claim 1, characterized in that, The nozzle position, tilt angle, and extension amount mentioned in S7 are obtained through geometric estimation in the ceiling plane coordinate system. The fastening status index is determined by the temperature difference response decay rate and local micro-vibration characteristics. The compliance judgment in S8 is first judged according to the threshold, and then the confidence of the abnormal type is corrected by the lightweight model.

8. A concealed sprinkler installation quality acceptance system based on infrared thermal imaging, characterized in that, include: Thermodynamic excitation unit is used to apply controlled thermal or cold pulse excitation to the pipe network or metal parts connecting the nozzle in the area where the nozzle is located. The synchronous control and acquisition unit is used to generate trigger signals, record excitation parameters and frequency band identifiers, and synchronize with the infrared imaging unit in time. An infrared imaging unit is used to acquire infrared image sequences before, during, and after excitation. The processing and judgment unit is used to perform the following tasks: establishing the material parameter library as described in S1 of claim 1, calculating the calibration of the internal and external parameters of the infrared imaging device and the alignment with the field coordinate system, acquiring and preprocessing as described in S3 to S8, extracting time-series observation features, selectively updating the thermal conductivity characteristics, thermal response curves or relative position change trends corresponding to the time-series observation features, constructing and fusion of time-series masks, estimating geometric and fastening states, and determining compliance and generating reports.

9. The system according to claim 8, characterized in that, The processing and decision unit incorporates a producer-consumer parallel execution mechanism. The producer is responsible for loading the sequence block data into the shared buffer in blocks, while the consumer completes the mask-based secondary fusion within the block and the recursive update between blocks, and performs overlapping data transfer and calculation between the two. The processing and decision unit adopts the same block quantization strategy as the sequence block for the feature tensor and mask block, and performs orthogonal perturbation on the features before quantization to optimize the distribution of the activation range. Key intermediate quantities are normalized and accumulated using a high-precision accumulator.

10. The system according to claim 8, characterized in that, The infrared imaging unit's spectral range covers long-wave or medium-wave thermal imaging, with a noise equivalent temperature difference not exceeding a preset threshold and a frame rate not lower than a preset value. The system is equipped with a safety interlock function, which automatically stops the excitation and records an alarm when the excitation temperature rise or drop exceeds the safety threshold. The generated report includes thermal image overlay, isotherm display, coordinate and time annotations, judgment conclusions and rectification suggestions, as well as traceability information of material and excitation parameters.

Images