Intelligent fire protection detection system for thermal power plants

Abstract

This invention relates to the field of fire safety monitoring and intelligent video analysis technology in thermal power plants, specifically to a smart fire detection system for thermal power plants. The system includes: a data acquisition module for acquiring, synchronizing, and spatially registering visible light video streams, infrared thermal imaging video streams, and operating parameters to generate a dual-spectral temporal video stream; a benchmark reconstruction module for generating a dual-spectral dynamic benchmark based on a compliant sample library and operating parameters; a simulation generation module for generating theoretically abnormal simulation states based on combustion evolution operators; a dual-track differential module for generating and extracting temperature evolution features, edge diffusion features, and temporal change features; a coupling decision module for performing trajectory alignment and calculating spatiotemporal topological similarity, outputting a fire state, a non-fire state, or a state awaiting verification; and a feedback update module for updating the compliant sample library when there is no fire and the operating conditions are stable, and stopping updating the dynamic benchmark when there is a fire state or a state awaiting verification. This invention improves the reliability of early fire identification and reduces false alarms.

Description

Technical Field

[0001] This invention relates to the field of fire safety monitoring and intelligent video analysis technology in thermal power plants, specifically to a smart fire protection intelligent detection system for thermal power plants. Background Technology

[0002] With the improvement of the intelligent operation level of thermal power plants, the requirements for early fire identification in high-risk areas such as boiler islands, main steam pipe corridors, coal conveying and transfer points, and cable interlayers are becoming increasingly higher. Due to the long-term presence of complex industrial interferences such as steam plumes, equipment heat radiation, reflection, water mist, dust, and load fluctuations at thermal power plant sites, how to accurately distinguish between actual combustion phenomena and normal operation phenomena under complex working conditions has also become a technical challenge in the field of smart fire protection. Traditional fire detection methods in thermal power plants currently mainly rely on the following approaches: smoke and fire identification based on visible light video, temperature rise alarm based on infrared thermal imaging, and linkage judgment combined with a small number of operating condition signals. However, visible light video recognition, infrared temperature rise alarm, and simple linkage judgment all have certain shortcomings. For example, visible light video recognition is prone to misjudging steam, white fog, or reflection as smoke and fire. Infrared temperature rise alarm has the problem of misjudging normal thermal radiation or short-term thermal disturbance of high-load equipment as fire. Simple linkage judgment lacks dynamic benchmark reconstruction of normal scenarios under current operating conditions and lacks effective reference for the evolution of real fires, making it difficult to adapt to the complex scenarios in thermal power plants where multiple heterogeneous information coexists and abnormal forms change significantly over time. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent fire protection detection system for thermal power plants, solving the following technical problems: This technology enables the separation of complex industrial noise from thermal power plants, distinguishing normal industrial phenomena from actual combustion phenomena, thereby improving the reliability of early fire detection and avoiding false alarms caused by normal steam, dust, or reflections.

[0004] The objective of this invention can be achieved through the following technical solutions: The intelligent fire protection and detection system for thermal power plants includes: The data acquisition module is configured to acquire visible light video streams, infrared thermal imaging video streams, and operating parameters of the target monitoring area of ​​the thermal power plant, and to perform time synchronization and spatial registration of the visible light video streams and the infrared thermal imaging video streams to generate a dual-spectrum time-series video stream, which includes visible light image features and infrared temperature field features. The benchmark reconstruction module is configured to generate a dual-spectral dynamic benchmark, including a visible light benchmark field and an infrared benchmark field, based on a pre-stored compliant sample library and the operating condition parameters. The simulation generation module is configured to perform abnormal perturbations on the dual-spectral dynamic benchmark according to a preset combustion evolution operator in order to generate a theoretical abnormal simulation state that includes thermal flash evolution and smoke diffusion evolution. The dual-track differential module is configured to perform differential analysis on the dual-spectral temporal video stream and the dual-spectral dynamic reference, as well as on the theoretical anomaly simulation state and the dual-spectral dynamic reference, to generate real residuals and theoretical residuals, and to extract temperature evolution features, edge diffusion features and temporal change features. The coupling decision module is configured to perform trajectory alignment on the temporal variation characteristics of the actual residual and the theoretical residual, and calculate the spatiotemporal topological similarity based on the alignment result and the spatial diffusion structure of the abnormal region, so as to output the state evaluation result. The feedback update module is configured to update the compliance sample library when the status assessment result is a non-fire state and the operating parameters meet the stability conditions, and to stop updating the dual-spectral dynamic benchmark when the status assessment result is a fire state or a state to be reviewed. The status assessment results include fire status, non-fire status, and status pending review; The compliance sample library is constructed by writing bispectral time-series video streams when the state assessment result is a non-fire state and the changes in wind speed, wind direction, equipment load, and valve status are all less than the corresponding thresholds within a preset time window. The preset combustion evolution operator is configured to: superimpose a periodic modulation function with a frequency range between 0.5Hz and 5Hz and an amplitude following a preset Gaussian distribution on the temperature amplitude of the corresponding reference pixel to inject thermal scintillation features; and establish a transparency mixing mask that uses a preset smoke grayscale matrix as the foreground and performs weighted fusion with the visible light reference field, with the mask transparency value gradually increasing from 0 to 1 as the simulation time progresses to inject transmittance attenuation features.

[0005] Preferably, the data acquisition module includes: a visible light acquisition unit configured to acquire visible light video streams; Infrared acquisition unit, configured to acquire infrared thermal imaging video stream; The parameter acquisition unit is configured to acquire wind speed, wind direction, equipment load, and valve status. The time alignment unit is configured to synchronize the visible light video stream, the infrared thermal imaging video stream, and the operating parameters in time. The spatial registration unit is configured to perform field-of-view calibration and spatial mapping on the visible light video stream and the infrared thermal imaging video stream to generate a spatially aligned bispectral temporal video stream.

[0006] Preferably, the baseline reconstruction module includes: The sample retrieval unit is configured to retrieve corresponding historical compliance samples from the compliance sample library based on the operating condition parameters. The visual reference generation unit is configured to generate a visible light reference field based on the historical compliance samples and the operating condition parameters. The thermal reference generation unit is configured to generate an infrared reference field based on the historical compliance samples and the operating condition parameters. The reference fusion unit is configured to fuse the visible light reference field and the infrared reference field to generate the dual-spectral dynamic reference.

[0007] Preferably, the operating parameters include valve status and equipment load; the visual reference generation unit is configured to generate a steam diffusion expected distribution based on the opening degree and opening / closing state of the valve status and combined with the steam plume area statistical results in historical compliance samples; the thermal visual reference generation unit is configured to generate a thermal radiation gradient distribution on the equipment surface based on the historical mapping relationship between equipment load and equipment surface temperature.

[0008] Preferably, the simulation generation module includes: The operator invocation unit is configured to invoke the preset combustion evolution operator to perturb the temperature field and visible light transmission characteristics in the dual-spectral dynamic reference. A scintillation injection unit is configured to execute the preset combustion evolution operator to inject the thermal scintillation feature; The smoke injection unit is configured to inject expansion features and transmittance attenuation features into the visible light reference field, and heat shielding and edge blurring features into the infrared reference field; The simulation state output unit is configured to output the theoretical anomaly simulation state based on the result after injection.

[0009] Preferably, the dual-track differential module includes: The reality difference unit is configured to perform difference analysis on the visible light image features and infrared temperature field in the dual-spectral temporal video stream and the dual-spectral dynamic reference, respectively, and fuse the difference results to generate the reality residual. The theoretical difference unit is configured to perform differences between the visible light simulation features and the infrared simulation temperature field in the theoretical anomaly simulation state and the dual-spectral dynamic reference, respectively, and fuse the difference results to generate the theoretical residual. The feature extraction unit is configured to extract temperature evolution features, edge diffusion features, and temporal change features from the actual residual and the theoretical residual.

[0010] Preferably, the coupling decision module includes: The trajectory alignment unit is configured to align the temporal variation characteristics of the actual residual and the theoretical residual using dynamic time warping; The topology calculation unit is configured to calculate the spatiotemporal topological similarity based on the aligned temporal change characteristics and the connected domain structure, boundary expansion path, and temperature gradient distribution contained in the edge diffusion characteristics of the corresponding abnormal region. The decision output unit is configured to output the fire status when the spatiotemporal topological similarity is greater than or equal to a first threshold, output the non-fire status when the spatiotemporal topological similarity is less than a second threshold, and output the status to be reviewed when the spatiotemporal topological similarity is greater than or equal to the second threshold and less than the first threshold. Wherein, the first threshold is greater than the second threshold.

[0011] Preferably, the coupling decision module further includes a noise suppression and elimination unit. The noise suppression and elimination unit is configured to compare the heat dissipation rate characterized by temperature evolution features and the spatial diffusion direction characterized by edge diffusion features in the actual residual with the corresponding features in the theoretical residual when the spatiotemporal topological similarity satisfies the fire state corresponding condition. The preset consistency condition includes: the heat dissipation rate of the actual residual is not greater than the heat dissipation rate of the theoretical residual, and the angle between their spatial diffusion directions is less than a preset angle threshold. The fire state is maintained when the comparison result satisfies the preset consistency condition, and the non-fire state is output when the comparison result does not satisfy the preset consistency condition.

[0012] Preferably, the feedback update module includes: a sample update unit, configured to acquire a bispectral time-series video stream that meets the conditions for writing to construct the compliant sample library, and to perform writing thereon; The benchmark stop update unit is configured to pause the adaptive update of the bispectral dynamic benchmark based on the current bispectral time-series video stream when the status assessment result is the fire status or the pending review status.

[0013] The beneficial effects of this invention are: 1. This invention solves the problem of lack of normal scenario reference in traditional detection by introducing a dual-spectral dynamic benchmark that changes with operating conditions; the system integrates historical compliant samples and multi-dimensional operating condition parameters, and pre-exposes the inherent steam drift and equipment thermal radiation interference of thermal power plants; this effectively avoids the defect of misjudging normal white fog or high-load thermal halo as smoke and fire, and significantly improves the detection accuracy under complex operating conditions. 2. This invention innovatively constructs a theoretical anomaly simulation state that includes thermal flash and smoke diffusion evolution, making up for the lack of real fire evolution reference in traditional methods; the system generates a theoretical fire standard that is adapted to the current working conditions by actively applying perturbations oriented towards fire mechanism, realizing the transformation from passive identification to active verification, and providing a reliable reference for accurately extracting real combustion characteristics. 3. This invention uses dual-track differential and trajectory alignment technology to overcome the defect that single-frame images are easily misled by instantaneous disturbances; by comparing the spatiotemporal topological similarity between the real and theoretical residuals, and further verifying the heat dissipation rate and spatial diffusion direction, the system accurately distinguishes between brief thermal disturbances and real fires, and effectively filters out positive alarms caused by high similarity caused by steam or reflection. 4. This invention has an adaptive feedback update mechanism, which solves the problem that static models are prone to failure as equipment ages or the environment changes; the system dynamically writes compliant samples when it is confirmed that there is no fire and the operating conditions are stable, and strictly freezes the benchmark update when an anomaly is determined; this not only ensures the model's long-term adaptive and evolutionary capabilities, but also prevents abnormal features from polluting the normal background, ensuring the long-term stability of the judgment. Attached Figure Description

[0014] The invention will now be further described with reference to the accompanying drawings.

[0015] Figure 1 This is a schematic diagram of the modules of the intelligent fire protection and detection system for thermal power plants provided in the embodiments of this application. Detailed Implementation

[0016] 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.

[0017] Please see Figure 1 A smart fire protection intelligent detection system for thermal power plants includes: a data acquisition module configured to acquire visible light video streams, infrared thermal imaging video streams, and operating parameters of the target monitoring area of ​​the thermal power plant, and to perform time synchronization and spatial registration of the visible light video streams and the infrared thermal imaging video streams to generate a dual-spectrum time-series video stream, wherein the dual-spectrum time-series video stream contains visible light image features and infrared temperature field features; The benchmark reconstruction module is configured to generate a dual-spectral dynamic benchmark, including a visible light benchmark field and an infrared benchmark field, based on a pre-stored compliant sample library and the operating condition parameters. The simulation generation module is configured to perform abnormal perturbations on the dual-spectral dynamic benchmark according to a preset combustion evolution operator in order to generate a theoretical abnormal simulation state that includes thermal flash evolution and smoke diffusion evolution. The dual-track differential module is configured to perform differential analysis on the dual-spectral temporal video stream and the dual-spectral dynamic reference, as well as on the theoretical anomaly simulation state and the dual-spectral dynamic reference, to generate real residuals and theoretical residuals, and to extract temperature evolution features, edge diffusion features and temporal change features. The coupling decision module is configured to perform trajectory alignment on the temporal variation characteristics of the actual residual and the theoretical residual, and calculate the spatiotemporal topological similarity based on the alignment result and the spatial diffusion structure of the abnormal region, so as to output the state evaluation result. The feedback update module is configured to update the compliance sample library when the state assessment result is a non-fire state and the operating parameters meet the stability conditions, and to stop updating the dual-spectral dynamic benchmark when the state assessment result is a fire state or a state pending review. The state assessment result includes fire state, non-fire state, and state pending review. The compliance sample library is constructed by writing dual-spectral time-series video streams when the state assessment result is a non-fire state and the changes in wind speed, wind direction, equipment load, and valve status are all less than the corresponding thresholds within a preset time window. The preset combustion evolution operator is configured to: superimpose a periodic modulation function with a frequency range between 0.5Hz and 5Hz and whose amplitude follows a preset Gaussian distribution onto the temperature amplitude of the corresponding benchmark pixel to inject thermal scintillation features; and establish a transparency mixing mask using a preset smoke grayscale matrix as the foreground and weightedly fused with the visible light reference field, with the mask's transparency value gradually increasing from 0 to 1 as the simulation time progresses to inject transmittance attenuation features.

[0018] This embodiment provides the implementation mechanism of the intelligent fire protection detection system for thermal power plants. Specifically, the system is deployed in high-risk areas such as the main steam pipe corridor, coal transfer point, and cable interlayer between the boiler island and the turbine hall of the thermal power plant. The system is described with the main steam bypass valve operating periodically during high-load operation at night and the possibility of steam plumes, welding reflections, or early smoldering in some areas as the main scenario. The system does not simply regard bright spots or white fog in the picture as fire. Instead, it first establishes a dynamic benchmark for the target monitoring area under the current working conditions, and then actively constructs a theoretical abnormal simulation state when a real fire occurs. It compares the on-site residual with the theoretical fire process in two directions to distinguish between normal industrial phenomena and real combustion phenomena. To avoid ambiguity caused by multiple terms referring to the same technical object in subsequent descriptions, the dynamic benchmark mentioned later in this article refers to the aforementioned dual-spectral dynamic benchmark, the theoretical simulation state or simulation state refers to the aforementioned theoretical anomaly simulation state, and the topological similarity or similarity refers to the aforementioned spatiotemporal topological similarity. Unless otherwise specified, their technical meaning will not be changed. Specifically, the data acquisition module continuously acquires visible light video, infrared thermal video, and operating condition information such as wind speed, wind direction, equipment load, and valve status for the same area; the visible light channel reflects appearance changes such as smoke obstruction, reflection, and steam form, while the infrared channel reflects thermal changes such as temperature rise, heat halo, and heat obstruction. The operating condition parameters are used to explain whether these changes are inevitable phenomena caused by normal unit operation; the reference reconstruction module generates a set of visible light reference fields and infrared reference fields that vary with the operating conditions based on samples from historical compliant states and the current operating conditions. For example, when the exhaust valve is opened, the system expects a white plume to appear in a certain fixed area; when the unit load increases, the system expects that the thermal radiation level of several pipes and flange surfaces will exceed the preset background value; the resulting dual-spectrum dynamic reference is not a static background, but a fire-free reference field that changes synchronously with the equipment operating status. Based on this, the simulation generation module calls the combustion evolution operator to apply abnormal perturbations oriented towards the fire mechanism to the dual-spectral dynamic benchmark; for the initial open flame, the simulation emphasizes the flashing of local heat sources, the formation of hot spot centers, and the outward expansion of the surrounding temperature gradient. For early smoldering or smoke, the simulation emphasizes the decrease in transmittance, edge diffusion, and the shift of the plume due to wind pull in the visible light image, as well as the thermal shading and blurred boundaries in the infrared image; the theoretical abnormal simulation state obtained in this way expresses the spatiotemporal variation law that real combustion should exhibit under the current working conditions. The dual-track differential module forms two parallel analysis paths; one path compares the on-site dual-spectral time-series video with the dual-spectral dynamic benchmark to extract the actual residuals of the on-site anomalies; the other path compares the theoretical anomaly simulation state with the same dual-spectral dynamic benchmark to extract the theoretical residuals under interference-free conditions. Both methods further extract temperature evolution characteristics, edge diffusion characteristics, and temporal variation characteristics. Among them, temperature evolution characteristics correspond to whether the heat source continuously accumulates heat and whether there is a stable heat core caused by combustion; edge diffusion characteristics correspond to whether the boundary of the plume or flame is continuously expanding or randomly drifting; and temporal variation characteristics reflect whether the abnormal process is a brief flash, driven by valve action, or has a self-sustaining increasing trend. The coupling decision module does not only look at the single frame form, but also aligns the evolution trajectory of the actual residual with the theoretical residual in time, and then calculates the spatiotemporal topological similarity by combining the connectivity structure, expansion direction and thermal gradient distribution of the abnormal region. When the on-site anomaly and the theoretical fire process are consistent in the chain of heat source formation - boundary expansion - temperature outward transfer, the fire status is output. When there are changes in appearance at the scene, but the diffusion and heat dissipation methods are more consistent with non-combustion phenomena such as steam, water mist, or reflection, the non-fire status is output; when there is a partial match between the two but it is not enough to confirm, such as when there is only a short-term temperature rise and the shape of the smoke plume is not yet stable, the status pending verification is output. The feedback update module is responsible for ensuring the long-term validity of the sample library and benchmark model. If the current judgment is that it is not a fire, and the wind speed, wind direction, equipment load and valve status remain stable over a period of time, it means that the video reflects a compliant operating condition and can be written into the compliant sample library for subsequent benchmark improvement. Conversely, as soon as a fire or pending verification status occurs, the system will stop updating the dynamic baseline with the current data to prevent abnormal processes from being mistakenly absorbed into the normal model. In a further implementation, when the visible light image is missing details due to insufficient nighttime illumination, the system can still rely on the infrared channel and operating parameters to maintain the judgment; when the infrared lens is temporarily affected by high temperature heat waves and experiences local drift, the system retains the visible light residual and operating condition constraints, does not directly output the fire conclusion, but prioritizes entering the pending review state. When the wind speed sensor or valve position signal is lost, the system will no longer perform baseline adaptive updates that rely on that parameter, but will freeze the current stable baseline to avoid distortion of the operating condition interpretation. For example, during the nighttime full-load operation of a 600 MW coal-fired unit, the main steam bypass valve was intermittently opened, and a white plume continuously appeared above the pipe gallery on the north side of the boiler island. The system, in conjunction with the valve opening status and previous compliance samples, first established a dynamic benchmark for the presence of steam in the corresponding area, and therefore did not directly regard the white mist as smoke. About ten minutes later, a new weak heat source appeared under the cable tray. This heat source showed continuous heat accumulation in the infrared image and was accompanied by a translucent gray plume in the visible light spreading along the direction of the cable tray. After the actual residual and the theoretical fire residual were aligned, they showed high consistency in the three aspects of heat core formation, boundary expansion and wind drift. Therefore, the fire status was output and the dynamic reference was stopped from absorbing the current video content. The purpose of this step is to compare what is seen on-site with how actual combustion should evolve, thereby separating the complex industrial noise of thermal power plants, improving the reliability of early fire identification, and avoiding false alarms caused by normal steam, dust, or reflections.

[0019] In this embodiment of the invention, the data acquisition module includes: a visible light acquisition unit configured to acquire a visible light video stream; an infrared acquisition unit configured to acquire an infrared thermal imaging video stream; a parameter acquisition unit configured to acquire wind speed, wind direction, equipment load, and valve status; and a time alignment unit configured to synchronize the visible light video stream, the infrared thermal imaging video stream, and the operating parameters in time. The spatial registration unit is configured to perform field-of-view calibration and spatial mapping on the visible light video stream and the infrared thermal imaging video stream to generate a spatially aligned bispectral temporal video stream.

[0020] This embodiment provides a mechanism for dual-spectral data acquisition and alignment. Specifically, in the above-mentioned main steam pipe gallery scenario, video acquisition without strict alignment will cause the position of white fog in visible light to be misaligned with the position of hot spot in infrared light, thus misinterpreting the same industrial phenomenon as two independent anomalies. Therefore, this embodiment further decomposes the acquisition process into five parts: visible light acquisition, infrared acquisition, parameter acquisition, time alignment, and spatial registration, so that subsequent analysis is based on an objective correspondence at the same time and in the same space. Specifically, the visible light acquisition unit is preferably deployed on the side wall of the corridor or on the steel structure platform to acquire smoke morphology, degree of obstruction, reflective edges and on-site operations; the infrared acquisition unit faces the same monitoring area to focus on observing the temperature rise of equipment surface, local hot spots and heat diffusion profile. The parameter acquisition unit connects to the anemometer, load signals in the distributed control system, and valve opening / closing status interface to provide environmental driving force and equipment status information. The time alignment unit unifies the three types of information onto the same time axis. Its purpose is not to pursue abstract data neatness, but to ensure that a valve opening at a certain moment, a steam plume appearing at a certain moment, and a local heat core rising at a certain moment belong to the same event chain that can be interpreted from each other. The spatial registration unit uses fixed reference points, equipment structure outlines, or preset calibration targets to perform field calibration and mapping on visible light and infrared images, so that the same pipe, the same valve, and the same section of cable tray fall into the same area in the dual channels. To clearly illustrate the data flow, the following simplified data model is used: it is assumed that the visible light channel captures a white plume in the upper left corner of the pipe gallery at time T1, the infrared channel captures a temperature rise zone in the upper middle part near time T1, and the valve signal shows that the bypass valve is open at time T1. If not aligned, the three signals may be interpreted as independent anomalies; after time alignment, the three are merged into the same event window; after spatial registration, the plume region in visible light and the hot zone in infrared light fall together at the physical location near the bypass valve outlet, and subsequent modules can determine that it is more likely valve steam release than fire. In a further implementation, when a channel is briefly interrupted, the system retains the remaining channels to continue acquiring data, but will reduce the automatic decision level during that time period; for example, when visible light is overexposed due to backlighting or lighting switching, infrared and operating parameters can still maintain basic monitoring, but the output results will be prioritized for review. If infrared temperature distortion occurs in local areas due to lens contamination, the system will remove the contaminated area during the spatial registration stage and will not use that area for high-confidence judgment; if the timestamp jitter exceeds the limit, the system will not force stitching, but will mark the time segment as a low-confidence window. For example, in the aforementioned nighttime operation scenario, the visible light image of monitoring point A shows a continuously drifting white fog above the main steam pipe gallery, while the infrared image simultaneously shows that the surface temperature near the valve flange is higher than that of the surrounding structural components, and the parameter channel records that the valve opening changes from closed to 30%. After time alignment and spatial registration, the system confirmed that the white fog, hot spots and valve position changes corresponded to the same location at the same time, so the phenomenon would not be misjudged as an independent fire source in the future. The purpose of this step is to establish the spatiotemporal identity among multi-source heterogeneous information, thereby providing a reliable input basis for subsequent benchmark reconstruction, differentiation, and decision-making.

[0021] In this embodiment of the invention, the benchmark reconstruction module includes: a sample retrieval unit configured to retrieve corresponding historical compliant samples from the compliant sample library based on the operating condition parameters; and a visual benchmark generation unit configured to generate a visible light benchmark field based on the historical compliant samples and the operating condition parameters. The thermal reference generation unit is configured to generate an infrared reference field based on the historical compliance samples and the operating parameters; the reference fusion unit is configured to fuse the visible light reference field and the infrared reference field to generate the dual-spectral dynamic reference.

[0022] This embodiment provides a mechanism for dynamic baseline reconstruction; specifically, when relying solely on the current screen for judgment, the system is prone to mistaking the thermal radiation, steam drift, and periodic changes in illumination that are inherent in thermal power plants as abnormalities. To address this issue, this embodiment introduces four steps: sample retrieval, visual benchmark generation, thermal benchmark generation, and benchmark fusion, enabling the system to possess the expected visual and thermal distribution characteristics of the current unit under specific operating conditions before analyzing the site. Specifically, the sample retrieval unit selects historical samples from the compliant sample library based on the similarity of operating conditions. The preferred retrieval conditions include unit load range, valve opening and closing combination, wind direction quadrant, and wind speed level. The historical samples selected in this way are not random images, but normal segments that are comparable to the current operating status. The visual reference generation unit restores the visible light reference field, focusing on reflecting normal steam plumes, common shadows, fixed reflective areas, and equipment outlines; the thermal reference generation unit restores the infrared reference field, focusing on reflecting the temperature distribution that the equipment should have under the current load, the location of normal thermal bridges, and the range of thermal corona; the reference fusion unit further combines the two in the same space to form a dual-spectral dynamic reference that includes both appearance information and thermal information. For ease of understanding, a simplified sample retrieval illustration can be used; assume that there are three sets of segments in the compliant sample library: segment A corresponds to low load and valve closed, segment B corresponds to high load and bypass valve open, and segment C corresponds to high load but wind direction changes abruptly; In the current scenario of high load, valve open, and stable wind direction, the system prioritizes segment B as the main reference and then supplements it with segments with similar wind direction to form a dynamic baseline. The baseline generated in this way will naturally include the white steam area near the valve outlet and the corresponding heat distribution, and will not regard such phenomena that should naturally occur as abnormal. In a further implementation, if no historical fragment completely consistent with the current working condition is found in the compliant sample library, the system can retreat to using the closest multiple sets of samples to construct a conservative benchmark, while raising the threshold for subsequent judgments to avoid misjudgment due to benchmark mismatch. If a certain operating condition combination occurs for the first time and the sample is scarce, the system can generate a basic benchmark using only the fixed area of ​​the equipment structure and the stable temperature zone, and will not make a strong interpretation of the dynamic steam region for the time being; if the current load fluctuates rapidly, making it difficult for the benchmark to converge in time, the system will temporarily freeze the benchmark under the most recent stable operating condition and bias the output of the results towards the one to be verified. For example, in the above main steam pipe gallery scenario, the unit is operating at more than 90% load, the bypass valve is periodically opened, and the wind direction is from west to east; the system retrieves multiple similar stable operation videos from the sample library and restores the visible light reference and infrared reference that there is a fixed steam plume near the valve outlet, a high temperature radiation zone near the pipe bend, and the night lighting forms a stable reflective visible light reference at the edge of the support. The resulting dual-spectral dynamic benchmark can cover most normal industrial phenomena, providing a reference for the subsequent stripping of truly abnormal phenomena. The purpose of this step is to pre-deduct the complex background that should appear under normal operating conditions from the anomaly detection, thereby achieving adaptation to dynamic backgrounds in industrial scenarios.

[0023] In this embodiment of the invention, the operating parameters include valve status and equipment load; the visual reference generation unit is configured to generate a steam diffusion expected distribution based on the opening degree and opening / closing state of the valve status and combined with the steam plume region statistical results in historical compliance samples; the thermal visual reference generation unit is configured to generate a thermal radiation gradient distribution on the equipment surface based on the historical mapping relationship between equipment load and equipment surface temperature.

[0024] This embodiment provides a refined reference generation mechanism for steam plume and thermal radiation gradient; specifically, although the previous scheme can reconstruct a general dual-spectral reference, in thermal power plants, the most likely causes of false alarms are steam diffusion driven by valve action and equipment thermal radiation changes driven by load changes. If only coarse-grained historical image averaging is used, normal steam drift is often mistaken for smoke, and infrared thermal halos under high load are often mistaken for smoldering fire. Therefore, this embodiment further explicitly introduces the relationship between valve status and expected steam distribution, and between equipment load and thermal radiation gradient into the baseline model. Specifically, the visual benchmark generation unit generates the expected distribution of steam diffusion based on whether the valve is closed, slightly open, or fully open, and the duration of its opening, combined with the common diffusion areas of steam plumes under different wind directions in historical compliance samples. Essentially, it pre-marks in the visible light reference which locations might appear with white plumes, which side the plumes usually shift to, and whether the boundaries are compact or loose; the thermal reference generation unit then recovers the thermal radiation gradient distribution on the equipment surface based on the historical mapping relationship between equipment load and equipment surface temperature. This gradient is not a simple ranking of high and low temperatures, but rather a characterization of the normal attenuation pattern of heat transfer from the center of the heat source outward. For example, under high load, the area near the flange and support of the main steam pipe will naturally be hotter than the surrounding steel structure, which is a normal thermal state. However, if a new heat core unrelated to the load appears under a cable tray, it is more likely to indicate an anomaly. The simplified illustration is as follows: Assume there are three valve states: closed (V0), partially open (V1), and fully open (V2); In historical compliance samples, V1 usually corresponds to the steam plume region of the first range, and V2 corresponds to the diffusion region of the second range, with the second range being larger than the first range. If the valve is currently in position V2 and the wind direction is east, then a large area of ​​expected steam will be generated on the east side of the valve outlet in the visual reference. Assuming that the equipment load is in the high load range, the main steam pipe and the adjacent bend will form a high gradient zone in the thermal reference, while the cable tray far from the equipment should still remain in the low temperature zone. In this way, if a localized persistent hot spot appears under the cable tray later, it will not be masked by the overall thermal background of the increased load. In a further embodiment, if the valve position signal is abnormal but the valve can be determined from historical trends to be in a stable open or closed state, the system can use the most recent reliable valve position state to maintain the expected steam distribution. If drastic changes in wind direction cause instability in the direction of the steam plume, the boundary of the steam prediction zone will be widened accordingly, and the affirmative judgment made solely based on the appearance will be reduced. If equipment load data is temporarily missing, the thermal reference will revert to the thermal radiation gradient corresponding to the most recent stable load, and a more cautious waiting-for-verification strategy will be adopted for newly emerging hotspots. For example, in the aforementioned high-load nighttime scenario, after the bypass valve is switched from closed to a larger opening, the system knows from previous compliance samples that an expanding white steam plume will form in the area above the valve outlet to the east side. Therefore, the expected steam distribution is given in advance in the visible light reference. Meanwhile, based on the mapping relationship between the unit's current load and the equipment's surface temperature, the system establishes a continuous thermal radiation gradient from the outer wall of the main steam pipe to the support in the infrared reference; thus, the presence of steam and high temperature near the valve will no longer directly trigger an alarm, while isolated new hot spots under the cable tray will be more clearly identified. The purpose of this step is to incorporate the most deceptive normal phenomena of thermal power plants into the prior model, thereby achieving targeted suppression of false steam alarms and false high-load thermal corona alarms.

[0025] In this embodiment of the invention, the simulation generation module includes: an operator invocation unit configured to invoke the preset combustion evolution operator to perturb the temperature field and visible light transmission characteristics in the dual-spectral dynamic reference; and a scintillation injection unit configured to execute the preset combustion evolution operator to inject thermal scintillation characteristics. The smoke injection unit is configured to inject expansion features and transmittance attenuation features into the visible light reference field, and heat shielding and edge blurring features into the infrared reference field; the simulation state output unit is configured to output the theoretical anomaly simulation state based on the injection results.

[0026] This embodiment provides a theoretical abnormal simulation state generation mechanism; specifically, dynamic benchmarks alone are insufficient to complete high-confidence determination, because although the system knows what the normal state is like, it also needs to know how a real fire should develop in the current scenario; Without this reference, many new anomalies not covered by the benchmark may still be incorrectly classified as fires; therefore, this embodiment introduces combustion evolution operators, thermal flash injection and smoke injection processes to actively construct a theoretical fire evolution state that is adapted to the current operating conditions. Specifically, the operator calling unit uses a dual-spectral dynamic reference as the base map and applies combustion mechanism perturbation to the temperature field and visible light transmission characteristics. The initial stage of combustion is usually accompanied by local heat release instability, heat nucleus formation, heating of the surrounding medium and accumulation of smoke particles. Therefore, the simulation not only makes a certain place hotter, but also reflects how heat is transferred outward, how smoke obscures the background and how the boundary changes from clear to diffuse. The scintillation injection unit superimposes thermal scintillation features in the range of 0.5Hz to 5Hz onto the infrared reference. This frequency band is closer to the low-frequency thermal fluctuations of initial open or smoldering flames in industrial scenarios, rather than stable heating of equipment or daily thermal inertia changes. The smoke injection unit adds changes such as plume expansion and decreased transmittance to the visible light reference, and adds thermal shielding and edge blurring to the infrared reference, thereby forming a theoretically abnormal simulation state that combines visible smoke effects and infrared thermal effects. In terms of specific algorithm implementation, injecting thermal scintillation features into the infrared reference field is achieved by superimposing a periodic modulation function with a frequency range between 0.5Hz and 5Hz and an amplitude following a preset Gaussian distribution onto the temperature amplitude of the corresponding reference pixel; injecting transmittance attenuation features into the visible light reference field is achieved by establishing a transparency mixing mask, using a preset smoke grayscale matrix as the foreground and performing weighted fusion with the visible light reference field, and the transparency value of the mask gradually increases from 0 to 1 as the simulation time progresses, to characterize the accumulation of smoke concentration and the decrease in background visibility; As a simplified application scenario example; suppose that a certain region originally had a clearly visible cable tray outline and a stable low temperature background; after the combustion evolution operator is applied, the region exhibits three changes in the simulation state: First, a hot spot is formed in the infrared with a central temperature higher than the surrounding preset threshold and a temperature gradient transition to the surrounding area; Second, the outline of the bridge in visible light is obscured by a semi-transparent gray cloud, and the obscuration range expands over time; third, the intensity of the hot spot does not rise steadily, but is accompanied by slow flashing; the theoretical anomaly simulation state formed by this is not a random generation of bright areas, but rather an attempt to reproduce the evolution process that real combustion may present under the current wind direction, current load and current background. Furthermore, to make the aforementioned simulation process have a clearer data flow relationship, the operator calling unit preferably first determines the disturbance region based on the current operating parameters, and then performs time-series injection on the infrared temperature field and visible light transmission characteristics respectively; for the infrared channel, it is preferable to first determine the potential thermonuclear center and its outward expansion range, and then superimpose the thermal scintillation and heat transfer changes. For the visible light channel, it is preferable to first determine the visibility of the initial occlusion area and background structure, and then superimpose the changes in transmittance, edge diffusion, and expansion direction. The simulation output unit combines the injection results at each time point into a continuous simulation sequence according to a unified time axis, so that when comparing with the actual residual, it can not only compare the differences of a single frame, but also compare the continuous process from the formation to the expansion of the anomaly. In a further implementation, if a certain type of area is not suitable for injecting a typical flame model, such as the area near a strong ventilation opening where smoke plumes are more likely to elongate rather than heat nuclei accumulate, the system can prioritize a simulation method that emphasizes smoke diffusion. If the monitoring area is affected by highly reflective materials, the simple high brightness feature can be weakened in the simulation to avoid mistaking specular reflection as a flame reference; if the current working conditions change too quickly, making it difficult to establish a stable theoretical simulation, the system will prioritize outputting a more conservative simulation state, retaining only core fire features such as thermal flash and changes in transmittance. For example, at the intersection of the main steam pipe gallery and the cable tray, the system found that white steam and equipment heat radiation existed in the area even under normal operating conditions. Therefore, direct detection is prone to feature overlap, which can lead to misjudgment. Therefore, the system injects a set of theoretical perturbations for smoldering cable trays onto the corresponding dynamic benchmark: low-frequency thermal flashes and continuous heat accumulation appear in the infrared, and semi-transparent smoke slowly expands along the direction of the cable tray in the visible light, with the background structure inside the smoke gradually fading; this theoretical abnormal simulation state is used as a standard reference and coupled with the field residuals for verification. The purpose of this step is to establish an evolutionary template for real fires for subsequent judgments, thereby achieving a shift from passive identification to active verification.

[0027] In this embodiment of the invention, the dual-track differential module includes: a real differential unit, configured to perform differential analysis on the visible light image features and infrared temperature field in the dual-spectral temporal video stream and the dual-spectral dynamic reference, respectively, and fuse the differential results to generate the real residual; The theoretical difference unit is configured to perform difference analysis on the visible light simulation features and infrared simulation temperature field in the theoretical anomaly simulation state and the dual-spectral dynamic benchmark, respectively, and fuse the difference results to generate the theoretical residual; the feature extraction unit is configured to extract temperature evolution features, edge diffusion features and temporal change features from the actual residual and the theoretical residual.

[0028] This embodiment provides a dual-track differential and feature extraction mechanism. Specifically, even if a dynamic benchmark and a theoretical anomaly simulation state are already available, if the two are not differentially compared with the same reference benchmark, the on-site anomaly and the theoretical anomaly will still be under different background conditions and will be difficult to compare directly. Therefore, this embodiment projects both the on-site anomaly and the theoretical fire into the same residual space and then extracts the key features that can represent the combustion mechanism. Specifically, the reality differential unit compares the current bispectral time-series video with the bispectral dynamic benchmark to obtain the reality residual; this means that after removing the equipment outline, steam area and thermal background that should already exist, the newly added or abnormally enhanced parts of the actual scene are retained; the theoretical differential unit compares the theoretical abnormal simulation state with the same dynamic benchmark to obtain the theoretical residual; this means that if there is a fire, what changes will be added compared to the normal operating conditions; Since both paths share the same dynamic benchmark, the content they extract is comparable. The feature extraction unit further extracts temperature evolution features, edge diffusion features, and temporal variation features from the two types of residuals. Temperature evolution features reflect whether a thermonuclear formation occurs and whether heat accumulates or dissipates rapidly. Edge diffusion features reflect whether the anomalous boundary advances continuously or drifts randomly. Temporal variation features distinguish whether the anomalous event is instantaneously driven by external operations or has a self-sustaining enhancement trend. A simplified illustration can be provided; suppose a new white fog and a slight hot spot appear in the real image of a certain monitoring area; after being differentiated from the dynamic baseline, the real residual retains a semi-transparent area that moves eastward in the visible light and a hot spot that heats up briefly and then falls back in the infrared. On the other hand, after the theoretical simulation state and the dynamic benchmark are differentiated, the theoretical residual is manifested as a plume that gradually expands along the bridge direction in visible light, and as a persistent and slowly expanding thermal core in infrared light. After feature extraction, it can be seen that the temperature evolution of the real residual is closer to rapid dissipation, and the edge changes are more affected by instantaneous wind direction disturbances. The theoretical residual, on the other hand, shows strong self-sustainability and stable diffusion. Subsequent decisions can identify that the two are not the same based on this. In a further implementation, when a frame is blocked or interfered with by strong light, causing the instantaneous difference result to be distorted, the system does not make a final judgment based on a single frame, but rather on the continuous temporal change characteristics; if the actual residual only appears in a single channel, for example, there is white fog in the visible light but no thermal anomaly in the infrared for a long time, the system retains the information but reduces its fire directionality. If there are hot spots in the infrared but no smoke in the visible light, it is not directly ruled out. Instead, the region type is considered to determine whether it may be a concealed smoldering. If the differential region is too small and the duration is too short, it is preferentially regarded as an occasional disturbance and enters the strategy of waiting for verification or ignoring. For example, in the aforementioned main steam pipe gallery scenario, after the white plume formed after the valve is opened is subtracted from the dynamic baseline, only a small amount of edge oscillation residual remains, which is mainly manifested as instantaneous thermal disturbance in the infrared. Therefore, the feature value of the actual residual is lower than the feature extraction threshold. However, when smoldering occurs in the cable under the cable tray, after subtracting from the dynamic baseline, a continuously expanding gray occlusion area appears in the visible light residual, and a heat source that expands from point to sheet appears in the infrared residual. The temperature evolution, edge diffusion, and temporal change features extracted by the system are significantly close to the theoretical residual. The purpose of this step is to unify on-site anomalies and theoretical fires into comparable anomaly representations, thereby achieving the characteristic basis required for subsequent coupled decision-making.

[0029] In this embodiment of the invention, the coupling decision module includes: a trajectory alignment unit configured to perform trajectory alignment on the temporal variation characteristics of the actual residual and the theoretical residual using dynamic time warping; and a topology calculation unit configured to calculate spatiotemporal topological similarity based on the aligned temporal variation characteristics and the connected domain structure, boundary expansion path, and temperature gradient distribution contained in the edge diffusion characteristics of the corresponding abnormal region and the temperature evolution characteristics. The decision output unit is configured to output the fire status when the spatiotemporal topological similarity is greater than or equal to a first threshold, output the non-fire status when the spatiotemporal topological similarity is less than a second threshold, and output the status to be reviewed when the spatiotemporal topological similarity is greater than or equal to the second threshold and less than the first threshold; wherein, the first threshold is greater than the second threshold.

[0030] This embodiment provides a coupled decision mechanism; specifically, the previous scheme has obtained the actual residual and the theoretical residual, but if only the shape at a certain moment is compared, it may still mistake a short-term steam cloud for smoke, or mistake the local temperature rise caused by thermal inertia for a thermonuclear mass. Therefore, this embodiment further compares the overall evolution consistency of the two types of residuals in time and space through trajectory alignment, topology calculation and hierarchical decision; Specifically, the trajectory alignment unit uses the idea of ​​dynamic time warping to handle situations where the development speed of similar phenomena varies; in real fires, the rate of temperature rise and spread may vary under different ventilation conditions, but the evolution sequence usually still follows the process of local formation - continuous enhancement - expansion to the surrounding area. Steam or reflections often appear suddenly, drift rapidly, and dissipate quickly, or change abruptly in sync with external operations. Through trajectory alignment, the system can compare whether the phase sequence of two time-series curves is consistent, rather than mechanically requiring them to change at exactly the same time. The topology calculation unit further examines whether the abnormal region forms a stable connected domain, whether the boundary expands along a fixed path or is instantaneously broken by the wind, and whether the temperature gradient shows a combustion pattern that decreases from the center to the outside. The decision output unit outputs the fire status, non-fire status, or pending review status according to the similarity classification. In order to maintain consistency with the description of the embodiment, the higher thresholds that appear in the following text all correspond to the first threshold, and the lower thresholds that appear all correspond to the second threshold. They are only used for textual explanation and do not introduce new threshold objects. A simplified illustration can be provided; assuming the time segments of the theoretical residual are A1, A2, and A3, corresponding to the initial formation of hotspots, plume expansion, and thermal gradient expansion, respectively; The temporal segments of the real residuals are B1, B2, and B3. If, after alignment, B1 corresponds to A1, B2 corresponds to A2, and B3 corresponds to A3, and the abnormal region maintains the same center and expands continuously outward in all three segments, then the similarity is high. Conversely, if the real residuals exhibit white fog in B1, rapid shift in B2, and complete disappearance in B3, and the temperature gradient is unstable, then even if the appearance of a certain frame is close to smoke, its overall topological similarity is low. For cases between the two, such as local heating but unclear expansion path, the output is "to be verified". Furthermore, to avoid spatiotemporal topological similarity remaining merely a conceptual description, in this embodiment, the topology calculation unit preferably decomposes it into three parts: temporal alignment similarity terms, spatial connectivity structure similarity terms, and thermal gradient distribution similarity terms; among which, the temporal alignment similarity terms are... This indicates the degree of consistency between the actual residual and the theoretical residual in the sequence of formation, enhancement, expansion, and decay stages; Specifically, time-series alignment of similar items It can be obtained by calculating the normalized path cost after aligning the actual residuals and theoretical residuals through dynamic time warping. The smaller the path cost, the better. The closer it is to 1; Spatial Connectivity Structure Similarity Term The intersection-union ratio (CUIR) of the anomalous connected regions in the corresponding time segments of the two types of residuals and the Hausdorff distance of the main boundary profile can be calculated. A higher CUIR and a smaller distance indicate better performance. The closer it is to 1; Thermal gradient distribution similarity term This can be obtained by extracting the temperature gradient direction histograms of the two types of residuals and calculating the cosine similarity. Spatial connected structure similarity terms This indicates that the number of connected components, the stability of the main connected components, and the consistency of boundary expansion paths are used to characterize the anomaly region; the thermal gradient distribution similarity term is... This indicates whether the temperature decreases from the center outwards around the same thermal core; Based on this, spatiotemporal topological similarity is denoted as And can be set according to preset weights , and Perform weighted calculation:

[0031] Wherein, the preset weights satisfy:

[0032] The purpose of this decomposition method is to ensure that the judgment is based on specific characteristic items, rather than being directly determined by a single comprehensive score; In specific implementation, the preset weight The configuration can be adaptively adjusted based on the interference characteristics of the target monitoring area. For example, in areas where steam drift interference is common, such as the main steam bypass valve, the weight of the spatial connectivity structure similarity term should be appropriately increased. Weights of similarity terms to thermal gradient distribution The value of is chosen to highlight the essential differences in spatial diffusion and heat conduction in actual combustion; Furthermore, in actual operation, the decision output unit preferentially checks whether the three types of sub-similar items are below the corresponding valid thresholds first, and then comprehensively considers... Perform a grading determination; for example, if higher and The value is significantly lower, indicating that although the on-site anomalies are similar to fire in terms of time sequence, they have not formed a stable expansion center in terms of space. Therefore, the direct output fire is suppressed first. like and Higher but If the duration is insufficient, it will be prioritized for review to wait for more time segments; correspondingly, the first and second thresholds can be set in layers according to the regional risk level. For high-risk areas such as cable tunnels and coal transfer points, the threshold for entering the review phase can be appropriately lowered while keeping the fire confirmation threshold unchanged; it should be noted that the spatiotemporal topological similarity calculated through the above steps It is mapped and normalized to the [0,1] interval; For example, in a normal monitoring area, the first threshold can be set to a value between 0.80 and 0.85, and the second threshold can be set to a value between 0.50 and 0.60; when When the value is [0.85,1], the output shows a fire status; when it is [0,0.50), the output shows a non-fire status; and when it is [0.50,0.85), the output shows a status pending review. For areas with frequent steam interference, such as the main steam bypass valve, it is preferable to increase the first threshold or increase... , The proportion in the overall calculation is used to reduce false alarms caused by similar steam appearance; Furthermore, to improve the clarity of the decision-making process, the trajectory alignment unit preferably first establishes time-series feature sequences of uniform length for both the actual residual and the theoretical residual, and then performs dynamic time warping; the topology calculation unit preferably calculates the number of connected components, the outer distance of the main boundary, and the position of the thermal gradient center in each alignment segment, and summarizes the results of multiple segments to avoid the abnormal fluctuations at a single moment having too much impact on the final decision; thus, the decision output unit is based not only on the similarity of a certain instantaneous image, but also on the stage consistency and spatial expansion consistency of multiple consecutive time segments; In a further implementation, when the similarity is near the threshold, the system does not make an overconfident conclusion, but instead prioritizes entering the pending review state and keeping the baseline frozen; when the time sequence length is insufficient, for example, when the anomaly only occurs for a very short time and the trajectory alignment foundation is weak, the system does not directly give a fire conclusion. If a region is affected by human operations, such as welding operations causing short-term bright spots and heat sources, the system can reduce the interpretation weight of its topological similarity by using the operation time window or region whitelist; if an abnormal region has multiple independent connected components at the same time, the system can calculate them separately and then merge them for judgment, to avoid multiple irrelevant small disturbances from being superimposed into false high similarity. For example, in the aforementioned scenario, although the white plume that appears during the release of steam from the bypass valve is locally similar to smoke in visible light, its actual residual, after alignment, shows that it rises and falls synchronously with the valve position, the center of the connected domain is unstable, and the temperature gradient does not form a continuous heat core. Therefore, the similarity is lower than the second threshold, and the output is a non-fire state. The anomaly caused by smoldering in cable trays maintains a fixed starting point in multiple consecutive time periods. The smoke plume expands along the direction of the cable tray, the infrared hotspots continuously increase and transfer heat outwards, and finally the similarity reaches the first threshold or above, outputting a fire status. The purpose of this step is to avoid biased judgments based on the appearance of a single frame, thereby achieving consistent verification of the entire evolution process of a real fire.

[0033] In this embodiment of the invention, the coupling decision module further includes a noise suppression and elimination unit. The noise suppression and elimination unit is configured to, when the spatiotemporal topological similarity satisfies the fire state correspondence condition, compare the heat dissipation rate characterized by temperature evolution features and the spatial diffusion direction characterized by edge diffusion features in the actual residual with the corresponding features in the theoretical residual. The preset consistency condition includes: The heat dissipation rate of the actual residual is not greater than the heat dissipation rate of the theoretical residual, and the angle between their spatial diffusion directions is less than a preset angle threshold; and the fire state is maintained when the comparison result meets the preset consistency condition, and the non-fire state is output when the comparison result does not meet the preset consistency condition.

[0034] This embodiment provides a secondary noise suppression and elimination mechanism. Specifically, under certain extreme conditions, phenomena such as steam leakage, water mist heating, and strong reflective heating of air disturbance may show a high topological similarity to the theoretical fire residual in a short period of time. If only the similarity classification of the previous layer is used, there will still be a small number of false positives. Therefore, this embodiment adds a consistency verification of heat dissipation rate and spatial diffusion direction on the basis of meeting the fire conditions, in order to eliminate anomalies that resemble fire but do not have the thermal continuity of combustion. The preset angle threshold is usually set according to the stability of natural ventilation or forced exhaust at the thermal power plant site, preferably between 15 degrees and 30 degrees, in order to accommodate the reasonable drift of the real plume under the environmental wind field. Specifically, real combustion usually has a relatively stable heat release center. Even when affected by wind, the heat transfer from the center to the outside still has a certain continuity, and the heat dissipation rate is less than the preset rate threshold. In contrast, anomalies caused by steam, water mist, or heat reflection often lack a continuous heat source. Once the external driving force weakens, its heat characteristics will drop rapidly. On the other hand, although the actual smoke plume or flame spread is affected by the wind direction, the spread direction usually has the characteristics of both upward drive and propagation along the path of combustibles; steam plumes are more likely to be completely dominated by the valve injection direction and the instantaneous wind direction; therefore, after initially judging it as a fire, the noise suppression and elimination unit continues to compare the consistency between the actual residual and the theoretical residual in terms of heat dissipation rate and spatial spread direction. Only when the two are consistent is the fire state retained; otherwise, it is judged as a non-fire state. A simplified illustration can be provided; assume that a certain anomaly achieves a high similarity after prior coupling calculation; the system observes two supplementary features: first, whether the real hot spot still retains a significant thermal core after the external operation stops; Secondly, whether the spread of the abnormal area continues to advance along the cable tray, the gap in the insulation layer, or the rising airflow; if the theoretical residual corresponds to continuous heat accumulation and expansion along a fixed path, while the actual residual is manifested as rapid cooling and dispersion with the wind after the valve is closed, then even if the previous similarity is high, it will still be eliminated. Furthermore, to ensure that the heat dissipation rate has an executable discrimination rule, this embodiment can record the time corresponding to the peak temperature of the hot spot and the subsequent temperature drop time within the same anomaly window, and record the heat dissipation rate as... Its expression is:

[0035] in, This indicates the peak temperature of the hot spot within the abnormal window. This indicates the temperature of the hot spot at the end of the preset observation period. This represents the time interval between the two; for the theoretical residual, It reflects the heat drop characteristics of simulated fires under the same observation window; for real-world residuals... It reflects the rate at which abnormal hot spots on site cool down after the external disturbance weakens; if the actual residual heat dissipation rate is significantly greater than the corresponding rate of the theoretical residual, it indicates that the heat source on site is more likely to be a short-term thermal disturbance rather than continuous combustion. Furthermore, to avoid relying solely on subjective descriptions for directional consistency, this embodiment preferably represents the main diffusion direction of the abnormal region as a direction vector pointing from the thermonuclear center to the main expansion boundary, and compares the degree of consistency between the directional angles of the actual residual and the theoretical residual; for example, the diffusion direction consistency is denoted as... Its expression is:

[0036] in, This represents the principal diffusion direction vector of the actual residual. This represents the theoretical residual principal diffusion direction vector. Represents the dot product of vectors. Represents the magnitude of a vector; when When the value is close to 1, it indicates that the diffusion directions of the two are similar; when When the value is significantly lower or even negative, it indicates that the two diffusion directions are inconsistent. In actual judgment, the noise suppression and elimination unit prefers to take the difference in heat dissipation rate within the allowable range and the consistency of diffusion direction above the direction threshold as conditions for maintaining the fire status, thereby avoiding maintaining the fire conclusion simply because of local appearance similarity. Furthermore, for areas with clear structural constraints, such as smoldering paths developing along cable trays, cable troughs, or wall corners, the noise suppression and elimination unit preferably uses the structural main axis direction as an auxiliary reference when comparing the theoretical residual direction; in other words, if the actual residual is slightly swayed by local wind disturbances, but its whole still revolves around the same heat core and extends along the structural main axis, then the fire will not be directly rejected due to short-term directional deviation. Conversely, if the anomaly always completely follows the valve spray direction or instantaneous wind direction and does not develop around a fixed heat core, it is more likely to be judged as not a fire; thus, this embodiment can satisfy the secondary comparison logic in the embodiment, and avoid new false alarms caused by rigid comparison in complex factory flow fields. Furthermore, in order to maintain consistency with the aforementioned judgment process, the noise suppression and elimination unit in this embodiment only performs the binary output of maintaining the fire state or changing the judgment to a non-fire state when both the heat dissipation rate and the spatial diffusion direction can be reliably obtained and the comparison conditions are met. If the above comparison cannot be reliably completed due to infrared image contamination, unstable extraction of direction vectors, or insufficient observation window, this situation is not considered as failing to meet the preset consistency conditions. Instead, the previous fire conclusion is rolled back to a pending review state and the baseline is frozen. The consistency verification will continue after the subsequent data is recovered. The purpose of this approach is to clearly distinguish between the comparison result failing to meet the consistency conditions and the comparison conditions not being met, thus avoiding the misclassification of a suspicious fire as a non-fire due to missing features. In a further implementation, if the heat dissipation rate is difficult to obtain stably, for example, if the infrared image is temporarily affected by lens contamination, the system will not force the previous conclusion to be rejected using this feature, but will instead downgrade the output for verification. If the diffusion direction is strongly constrained by structural components within a narrow channel, causing a deviation between the on-site directionality and the theoretical directionality, the system can prioritize examining whether it still develops around the same heat source center, rather than directly negating it based solely on the difference in direction; if there are forced ventilation facilities in a certain area, the system can also incorporate the equipment's operating status to correct and interpret the diffusion direction. For example, in an anomaly near the main steam bypass valve, the visible light morphology of the actual residual is quite similar to that of the theoretical plume, and local high temperature also appeared in the infrared. The preceding similarity has reached the fire judgment condition. However, the noise suppression and elimination unit further discovered that the high temperature quickly returned to the background level after the valve was closed, and the white mist was completely dispersed along the injection direction and did not continue to spread along the cable tray or the path of surrounding combustibles. Therefore, it did not meet the consistency requirements of real combustion and was finally judged as a non-fire state. Conversely, in the case of smoldering cable trays, the thermonuclear flammability does not dissipate in a short period of time, and the plume continues to extend along the top of the cable trays, thus maintaining the fire state. The purpose of this step is to screen out industrial interference that is highly similar but not inherently flammable, thereby further reducing the false alarm rate.

[0037] In this embodiment of the invention, the feedback update module includes: a sample update unit configured to acquire a bispectral time-series video stream that meets the conditions for writing into a compliant sample library and to write it; and a benchmark stop update unit configured to pause adaptive updates of the bispectral dynamic benchmark based on the current bispectral time-series video stream when the status assessment result is the fire status or the status to be reviewed.

[0038] This embodiment provides a feedback update and abnormal freezing mechanism; specifically, the aforementioned solution can already identify fire and non-fire, but if the sample library and dynamic benchmark are not updated for a long time, the system will gradually fall behind on-site changes such as equipment aging, lighting adjustment, and insulation layer replacement. Conversely, blindly updating during abnormal periods may result in the inclusion of fire processes, suspected anomalies, or even occasional noises in the normal model, leading to subsequent missed detections. Therefore, this embodiment sets up a feedback strategy of absorbing during stable periods and freezing during abnormal periods. Specifically, the sample update unit will only write the corresponding dual-spectrum time-series video into the compliant sample library if the system has determined that it is not a fire and the changes in wind speed, wind direction, equipment load and valve status within a preset time window are all less than the corresponding thresholds. The physical basis for this is that the images presented in the video have repeatable normal reference significance only when the environmental driving force and equipment status are stable enough; if the external conditions change drastically, even if there is no fire, the images may not be suitable as standard samples; the reference stop update unit suspends the use of the current video to correct the dynamic reference in the fire state or the state of waiting for verification, to prevent abnormal smoke, real thermonuclear or suspicious changes to be confirmed from being absorbed into the normal background. This can be illustrated in a very simplified way; assume that three video segments are obtained consecutively within a certain time window. , , ; If the period is determined to be non-fire-related, and the wind direction, load, and valve position are stable, then it can be written into the sample library; Even if it is not judged as a fire during this period, if the wind direction fluctuates continuously and the steam plume shape changes drastically, it will not be written as a compliant sample; if it is judged to be pending review during P3, it will not only not be written into the sample library, but the current benchmark will also be frozen and the updating of the suspected smoke plume or hot spot features to the bispectral dynamic benchmark will be suspended; this can avoid the model being contaminated. Furthermore, in order to ensure that the update has a clear execution boundary when the stability condition is met, the preset time window in this embodiment is preferably set to a short stable segment covered by multiple consecutive sampling periods, rather than a single frame determination; Specifically, observation windows ranging from 30 to 300 seconds can be selected based on the characteristics of the monitoring points. For areas with frequent steam fluctuations, a first preset time window is preferred to confirm whether the valve status and wind direction meet the stability threshold. For cable interlayers with enclosed structures and environmental disturbance parameters less than the set value, a second preset time window can be used to accelerate sample accumulation, wherein the first preset time window is larger than the second preset time window. Correspondingly, the threshold values ​​for changes in wind speed, wind direction, equipment load, and valve status are also preferably set separately according to the area type, rather than using a fixed set of thresholds uniform across the entire plant, thereby avoiding the incorrect exclusion of already highly volatile areas from sample updates. The variation amplitude refers to the standard deviation or maximum absolute deviation of the time series of the corresponding operating condition parameter within the preset time window. For example, for continuous parameters such as wind speed and equipment load, the variation amplitude is its standard deviation within the window; for wind direction angle, the variation amplitude is the maximum deflection angle within the window; for discrete parameters such as valve status, the variation amplitude is the number of status switches within the window, and the corresponding threshold can be set to 0, that is, requiring the valve status to remain unique within the window. Furthermore, the sample update unit preferably adopts a two-level strategy of caching first and then writing; that is, within the window, the candidate bispectral time-series video is first written to the temporary buffer area, and the corresponding working condition label, status assessment result and timestamp are recorded; only when the non-fire state is still met after the end of the window and the change of working condition parameters is less than the corresponding threshold, the candidate data is officially written to the compliant sample library. If the system transitions to a pending review or fire state at any point during the caching period, the current cached segment is immediately discarded and the baseline is kept frozen. The purpose of this approach is to prevent suspicious segments from being mistakenly absorbed into normal samples in the first few seconds when the system initially detects abnormal features. Furthermore, when the benchmark stop update unit pauses adaptive updates, it preferably retains the dual-spectral dynamic benchmark under the most recent stable operating condition before the anomaly occurred as the frozen benchmark, and during the freezing period, only detection and judgment are allowed and background learning is not allowed. Once manual confirmation is completed or the system continuously meets the non-fire and stable conditions again, the normal update process will resume from the frozen baseline. This ensures that the comparison baseline before and after the abnormal window remains consistent, avoiding the impact of background model drift on subsequent retrospective analysis due to the freezing period. Furthermore, to make the sample writing logic clearer, the sample update unit preferably performs a consistency check on the operating parameters at each sampling time within the window before candidate writing, and cross-confirms the status evaluation results at the beginning and end of the time window; only when there is no fire status or pending review status within the window, and the operating condition label is consistent with the retrieved sample category, will the formal writing be performed. If no alarm is detected inside the window but the operating condition changes, such as from valve closed to valve open, the video corresponding to that window should be split and then each video should be judged separately to determine whether it meets the writing conditions, so as to avoid different operating condition segments from being mixed into the same compliant sample. In a further implementation, if a region is under frequent fluctuations for a long period of time, making it difficult to meet the stable writing conditions, the system can only perform local updates on the stable sub-periods; if the sample library capacity reaches the preset limit, a method of retaining recent representative samples by operating conditions can be adopted to prevent excessive accumulation of historical redundancy. If the heat distribution of the equipment changes significantly after on-site maintenance, the system can prioritize the addition of new samples within the stable operation window after maintenance acceptance, so that the benchmark can adapt to the new normal state as soon as possible; if the status assessment fluctuates repeatedly between fire and pending review, it will remain frozen until manual confirmation or the system re-enters a stable non-fire state. For example, in the aforementioned main scenario, the steam release from the bypass valve is a long-term normal operating condition; the system repeatedly observed during multiple night shift cycles that when the unit is operating at high load and the wind direction is stable, a nearly fixed steam plume will form near the valve outlet, accompanied by a stable thermal radiation zone. Since these segments were all determined to be non-fire and the changes in operating conditions were small, they were continuously written into the compliance sample library to make the subsequent steam benchmark more accurate. On the night the cable tray smoldering occurred, once the system output a status pending review or fire, it immediately stopped updating the dynamic benchmark with the current segment to avoid solidifying the smoke and hot spots into the normal background. The purpose of this step is to maintain the system's adaptive capability during long-term operation, while preventing abnormal data from polluting the normal model, thereby achieving the sustainable evolution and discrimination stability of the benchmark library.

[0039] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A smart fire protection intelligent detection system for thermal power plants, characterized in that: The system includes: The data acquisition module is configured to acquire visible light video streams, infrared thermal imaging video streams, and operating parameters of the target monitoring area of ​​the thermal power plant, and to perform time synchronization and spatial registration of the visible light video streams and the infrared thermal imaging video streams to generate a dual-spectrum time-series video stream, which includes visible light image features and infrared temperature field features. The benchmark reconstruction module is configured to generate a dual-spectral dynamic benchmark, including a visible light benchmark field and an infrared benchmark field, based on a pre-stored compliant sample library and the operating condition parameters. The simulation generation module is configured to perform abnormal perturbations on the dual-spectral dynamic benchmark according to a preset combustion evolution operator in order to generate a theoretical abnormal simulation state that includes thermal flash evolution and smoke diffusion evolution. The dual-track differential module is configured to perform differential analysis on the dual-spectral temporal video stream and the dual-spectral dynamic reference, as well as on the theoretical anomaly simulation state and the dual-spectral dynamic reference, to generate real residuals and theoretical residuals, and to extract temperature evolution features, edge diffusion features and temporal change features. The coupling decision module is configured to perform trajectory alignment on the temporal variation characteristics of the actual residual and the theoretical residual, and calculate the spatiotemporal topological similarity based on the alignment result and the spatial diffusion structure of the abnormal region, so as to output the state evaluation result. The feedback update module is configured to update the compliance sample library when the status assessment result is a non-fire state and the operating parameters meet the stability conditions, and to stop updating the dual-spectral dynamic benchmark when the status assessment result is a fire state or a state to be reviewed. The status assessment results include fire status, non-fire status, and status pending review; The compliance sample library is constructed by writing bispectral time-series video streams when the state assessment result is a non-fire state and the changes in wind speed, wind direction, equipment load, and valve status are all less than the corresponding thresholds within a preset time window. The preset combustion evolution operator is configured to: superimpose a periodic modulation function with a frequency range between 0.5Hz and 5Hz and an amplitude following a preset Gaussian distribution on the temperature amplitude of the corresponding reference pixel to inject thermal scintillation features; and establish a transparency mixing mask that uses a preset smoke grayscale matrix as the foreground and performs weighted fusion with the visible light reference field, with the mask transparency value gradually increasing from 0 to 1 as the simulation time progresses to inject transmittance attenuation features.

2. The intelligent fire protection detection system for thermal power plants according to claim 1, characterized in that, The data acquisition module includes: Visible light acquisition unit, configured to acquire visible light video streams. Infrared acquisition unit, configured to acquire infrared thermal imaging video stream; The parameter acquisition unit is configured to acquire wind speed, wind direction, equipment load, and valve status. The time alignment unit is configured to synchronize the visible light video stream, the infrared thermal imaging video stream, and the operating parameters in time. The spatial registration unit is configured to perform field-of-view calibration and spatial mapping on the visible light video stream and the infrared thermal imaging video stream to generate a spatially aligned bispectral temporal video stream.

3. The intelligent fire protection detection system for thermal power plants according to claim 1, characterized in that, The benchmark reconstruction module includes: The sample retrieval unit is configured to retrieve corresponding historical compliance samples from the compliance sample library based on the operating condition parameters. The visual reference generation unit is configured to generate a visible light reference field based on the historical compliance samples and the operating condition parameters. The thermal reference generation unit is configured to generate an infrared reference field based on the historical compliance samples and the operating condition parameters. The reference fusion unit is configured to fuse the visible light reference field and the infrared reference field to generate the dual-spectral dynamic reference.

4. The intelligent fire protection detection system for thermal power plants according to claim 3, characterized in that, The operating parameters include valve status and equipment load; the visual reference generation unit is configured to generate a steam diffusion expected distribution based on the opening and closing status of the valve status and the statistical results of the steam plume area in historical compliance samples; the thermal reference generation unit is configured to generate a thermal radiation gradient distribution on the equipment surface based on the historical mapping relationship between equipment load and equipment surface temperature.

5. The intelligent fire protection detection system for thermal power plants according to claim 1, characterized in that, The simulation generation module includes: The operator invocation unit is configured to invoke the preset combustion evolution operator to perturb the temperature field and visible light transmission characteristics in the dual-spectral dynamic reference. A scintillation injection unit is configured to execute the preset combustion evolution operator to inject the thermal scintillation feature; The smoke injection unit is configured to inject expansion features and transmittance attenuation features into the visible light reference field, and heat shielding and edge blurring features into the infrared reference field; The simulation state output unit is configured to output the theoretical anomaly simulation state based on the result after injection.

6. The intelligent fire protection detection system for thermal power plants according to claim 1, characterized in that, The dual-track differential module includes: The reality difference unit is configured to perform difference analysis on the visible light image features and infrared temperature field in the dual-spectral temporal video stream and the dual-spectral dynamic reference, respectively, and fuse the difference results to generate the reality residual. The theoretical difference unit is configured to perform differences between the visible light simulation features and the infrared simulation temperature field in the theoretical anomaly simulation state and the dual-spectral dynamic reference, respectively, and fuse the difference results to generate the theoretical residual. The feature extraction unit is configured to extract temperature evolution features, edge diffusion features, and temporal change features from the actual residual and the theoretical residual.

7. The intelligent fire protection detection system for thermal power plants according to claim 6, characterized in that, The coupling decision module includes: The trajectory alignment unit is configured to align the temporal variation characteristics of the actual residual and the theoretical residual using dynamic time warping; The topology calculation unit is configured to calculate the spatiotemporal topological similarity based on the aligned temporal change characteristics and the connected domain structure, boundary expansion path, and temperature gradient distribution contained in the edge diffusion characteristics of the corresponding abnormal region. The decision output unit is configured to output the fire status when the spatiotemporal topological similarity is greater than or equal to a first threshold, output the non-fire status when the spatiotemporal topological similarity is less than a second threshold, and output the status to be reviewed when the spatiotemporal topological similarity is greater than or equal to the second threshold and less than the first threshold. Wherein, the first threshold is greater than the second threshold.

8. The intelligent fire protection detection system for thermal power plants according to claim 7, characterized in that, The coupling decision module further includes a noise suppression and elimination unit. The noise suppression and elimination unit is configured to compare the heat dissipation rate characterized by temperature evolution features and the spatial diffusion direction characterized by edge diffusion features in the actual residual with the corresponding features in the theoretical residual when the spatiotemporal topological similarity satisfies the fire state corresponding condition. The preset consistency condition includes: the heat dissipation rate of the actual residual is not greater than the heat dissipation rate of the theoretical residual, and the angle between the spatial diffusion directions of the two is less than a preset angle threshold. The fire state is maintained when the comparison result satisfies the preset consistency condition, and the non-fire state is output when the comparison result does not satisfy the preset consistency condition.

9. The intelligent fire protection detection system for thermal power plants according to claim 1, characterized in that, The feedback update module includes: The sample update unit is configured to acquire a bispectral temporal video stream that meets the conditions for writing to construct the compliant sample library, and to perform writing thereon. The benchmark stop update unit is configured to pause the adaptive update of the bispectral dynamic benchmark based on the current bispectral time-series video stream when the status assessment result is the fire status or the pending review status.

Images