Kitchen gas stove abnormity early warning system based on infrared thermal imaging multispectrum

Abstract

The invention relates to the technical field of kitchen safety monitoring, and discloses a kitchen gas stove abnormity early warning system based on infrared thermal imaging multispectrum. The system comprises an infrared thermal imaging acquisition module, a dynamic feature extraction module and an abnormal behavior identification module. The infrared thermal imaging acquisition module synchronously captures the surface temperature field distribution of the gas cooking range through a multispectral infrared sensor array, obtains flame form data in combination with a visible light image, and generates a multi-channel fused infrared thermal imaging characteristic spectrum through processing. The dynamic feature extraction module adopts a three-dimensional convolution kernel to scan continuous time sequence frames, identifies flame morphological feature differences, calculates thermal radiation intensity energy distribution offset, and outputs dynamic feature vectors. The abnormal behavior recognition module extracts a correlation mode through a deep residual network, compares current features with a historical normal working condition feature library, and generates a preliminary early warning signal containing an abnormal type code and a risk level when the attenuation rate of flame spectral energy in a specific infrared band exceeds a first threshold value.

Description

Technical Field

[0001] This invention relates to the field of kitchen safety monitoring technology, specifically to an abnormal early warning system for kitchen gas stoves based on infrared thermal imaging multispectral imaging. Background Technology

[0002] Kitchen gas stoves are commonly used cooking appliances in homes and restaurants, and their operational safety is directly related to personal and property safety. Currently, most safety monitoring methods for gas stoves on the market focus on single-parameter detection. For example, thermocouple sensors monitor whether the flame is extinguished, thus triggering the gas valve to close; or gas sensors detect the concentration of leaked gas and issue an alarm when it exceeds the standard. However, these traditional methods have significant limitations. Thermocouple sensors can only determine whether a flame exists and cannot detect potential risks such as changes in flame shape and abnormal temperature field distribution, such as localized high temperatures caused by fluctuating flame size, or abnormal combustion states caused by an imbalance in the gas-air mixture. Gas sensors mainly provide early warnings for gas leaks and are difficult to detect problems such as incomplete combustion and flame deviation during combustion, which are often important causes of carbon monoxide poisoning and localized overheating leading to fires. While some existing monitoring systems incorporate temperature detection, they often employ single-point or limited point-type temperature sensors, which can only acquire localized temperature data and fail to reflect the overall temperature field distribution and dynamic changes of the gas stove. When the flame shifts locally or experiences a sudden temperature rise, point-type sensors may miss warning opportunities because they do not cover the abnormal area. Furthermore, traditional systems lack analysis of flame spectral characteristics. The spectral properties of a flame change with the combustion state; for example, the infrared energy distribution of a flame during incomplete combustion differs significantly from that during normal combustion, and this difference is difficult to capture directly through temperature or gas concentration detection. In practical applications, abnormal conditions of gas stoves are often the result of multiple factors working together, making it difficult to comprehensively and accurately identify risks through single-parameter monitoring. For example, fluctuations in gas pressure can lead to unstable flame patterns, accompanied by uneven temperature field distribution and changes in spectral characteristics. Relying on only one indicator for monitoring can easily result in false alarms or missed alarms. As people's requirements for kitchen safety increase, there is an urgent need for a technical solution that can monitor the operating status of gas stoves in a multi-dimensional and high-precision manner to compensate for the shortcomings of existing monitoring methods and achieve early identification and warning of various potential anomalies. Summary of the Invention

[0003] The purpose of this invention is to provide an abnormal early warning system for kitchen gas stoves based on infrared thermal imaging multispectral imaging, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides a kitchen gas stove anomaly early warning system based on infrared thermal imaging multispectral imaging, the system comprising: The infrared thermal imaging acquisition module is used to synchronously capture the surface temperature field distribution of the gas stove head through a multispectral infrared sensor array, combine it with visible light image acquisition equipment to obtain real-time changes in flame shape, perform non-uniformity correction and environmental noise filtering on the original thermal radiation signal, and generate a multi-channel fused infrared thermal imaging feature map. The dynamic feature extraction module, based on the infrared thermal imaging feature map, uses a three-dimensional convolution kernel to scan the temperature gradient change region in continuous time frames, identifies the morphological feature differences between the flame core region and the diffusion edge, calculates the energy distribution offset of thermal radiation intensity in a preset frequency band, and outputs a dynamic feature vector containing temperature change coordinates and spectral characteristics. The abnormal behavior identification module receives the dynamic feature vector, extracts the correlation pattern between the spatial distribution of the temperature field and the flame spectral features through a deep residual network, compares the matching degree of the current feature vector with the historical normal operating condition feature library, and generates a preliminary warning signal containing the abnormal type code and risk level when the attenuation rate of the flame spectral energy in a specific infrared band exceeds the first threshold.

[0005] Preferably, the system further includes: The collaborative analysis and decision-making module, based on the preliminary early warning signal, integrates real-time monitoring data from the gas flow meter with the operating parameters of the stove power control unit, constructs a multi-dimensional state matrix, infers the cause of the anomaly and the risk propagation path through a probabilistic graphical model, and outputs the anomaly location coordinates and risk diffusion trend prediction results. The graded response control module sends graded adjustment commands to the gas valve control unit based on the abnormal location coordinates and risk spread trend prediction results, sends an alarm data packet containing a risk heat map and handling suggestions to the human-machine interface terminal, and simultaneously triggers the audible and visual alarm device to execute the warning mode corresponding to the risk level.

[0006] Preferably, the infrared thermal imaging acquisition module includes: The multispectral sensing unit is equipped with four independent detector arrays covering the short-wave infrared to long-wave infrared bands. Each detector array captures the thermal radiation signal on the surface of the stove at a millisecond sampling frequency. The environmental compensation unit integrates environmental temperature and humidity sensors and air pressure monitoring elements to collect kitchen environmental parameters in real time and eliminates the interference of environmental variables on thermal radiation signals through an adaptive Kalman filter. The data fusion unit receives the output data from the multispectral sensing unit and the environmental compensation unit, and uses a pixel-level weighted fusion algorithm to align the spatial coordinates of the visible light image and the infrared channel to generate the infrared thermal imaging feature map with three-dimensional attributes of temperature, spectrum and space.

[0007] Preferably, the dynamic feature extraction module includes: The time-series analysis unit calculates the temperature change rate of each pixel between adjacent frames based on the continuous frame sequence of the infrared thermal imaging feature map, and marks abnormally active areas where the temperature change rate exceeds a preset rate. The morphological processing unit performs corrosion and expansion operations on the abnormally active region to separate the flame core region from the edge diffusion region and extracts the core region area change gradient and edge irregularity index. The frequency domain conversion unit converts the thermal radiation signal of the selected area into a frequency domain energy distribution spectrum and calculates the energy ratio offset between the mid-wave infrared band and the long-wave infrared band. The feature integration unit encapsulates the temperature change rate, core area change gradient, edge irregularity index, and energy ratio offset into the dynamic feature vector.

[0008] Preferably, the abnormal behavior recognition module includes: The feature comparison unit separates the temperature field distribution pattern and spectral feature parameters from the dynamic feature vector, calls the benchmark template in the historical normal operating condition feature library, and calculates the cosine similarity between the current feature vector and the benchmark template. The residual network unit receives the similarity data output by the feature comparison unit, extracts cross-channel correlation features through five layers of convolutional residual blocks, and identifies abnormal patterns such as flame extinguishing precursors, gas leak thermal diffusion, and oil ignition. The risk quantification unit generates a first-class warning code characterizing flame instability when the energy value of the long-wave infrared band decreases by more than 40% within three consecutive frames, and generates a second-class warning code characterizing gas leakage when a hot spot with non-flame characteristics appears in the mid-wave infrared band.

[0009] Preferably, the collaborative analysis and decision-making module includes: The data integration unit receives instantaneous flow fluctuation data uploaded by the gas flow meter and operating power parameters of the stove power control unit, and timestamps them with the preliminary warning signal. The matrix construction unit constructs row vectors from temperature field coordinates, spectral anomaly codes, gas flow deviation values, and power fluctuation values, and stacks them in a time series to form the multidimensional state matrix. The Bayesian inference unit, based on a conditional probability model trained on a historical accident database, infers the probability distribution of whether the current anomaly is caused by abnormal gas pressure or failure of stove components. The path prediction unit simulates the spread path of abnormal heat sources on the stove surface and the expected range of impact based on the correlation between heat diffusion rate and gas flow rate changes.

[0010] Preferably, the hierarchical response control module includes: The valve control unit receives the abnormal positioning coordinates and risk diffusion trend prediction results. When the risk level reaches level two, it sends a 30% opening limit command to the gas proportional valve. When it reaches level three, it sends a shut-off command. The heat map generation unit maps the anomaly location coordinates to the visible light image coordinate system and superimposes temperature gradient color scale markers to form the risk heat map spectrum. The alarm strategy unit matches the audible and visual alarm modes according to the risk level: Level 1 risk triggers a low-frequency intermittent buzzer and a yellow indicator light; Level 2 risk triggers a high-frequency continuous buzzer and a red indicator light; and Level 3 risk triggers a rotating warning light and a voice broadcast.

[0011] Preferably, the system further includes: The self-calibration module is configured during the daily system startup phase. It calibrates the response curves of each infrared detection channel using a built-in standard blackbody radiation source and updates the filtering parameters of the environmental compensation unit. The data storage module continuously records the dynamic changes of the infrared thermal imaging feature spectrum and stores the complete multispectral data sequence for 30 seconds before and after the occurrence of the abnormal event. The remote communication module uploads Level 3 risk alarm data packets to the cloud monitoring platform in real time via an encrypted transmission protocol.

[0012] Preferably, the self-calibration module includes: The calibration execution unit drives the stepper motor to move the blackbody radiation source to the center of the infrared sensor's field of view, and acquires the raw response data at three temperature reference points; The curve fitting unit uses the least squares method to fit the temperature-voltage response equation of each detector channel and generates a new correction coefficient matrix. The parameter update unit writes the correction coefficient matrix into the non-volatile memory of the environmental compensation unit.

[0013] Preferably, the alarm strategy unit further includes: The feedback learning unit collects user response delay data to alarms. If the user does not intervene within ten seconds after a single Level 3 risk alarm, the initial risk level of the same type of anomaly will be automatically increased. When a level 3 alarm is triggered, the equipment linkage unit simultaneously shuts down the kitchen exhaust system fan and sends a forced exhaust command to the intelligent ventilation device.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This kitchen gas stove anomaly early warning system based on infrared thermal imaging multispectral imaging achieves comprehensive capture of the temperature field distribution on the gas stove surface through an infrared thermal imaging acquisition module. Combined with visible light images, it acquires real-time data on flame morphology. After non-uniformity correction and noise filtering, it generates a multi-channel fused feature map, enabling the system to simultaneously grasp both temperature and morphology information. Compared to traditional single-point or localized monitoring methods, this multi-dimensional data acquisition method covers the entire area of ​​the gas stove, fully presenting the temperature field distribution characteristics and dynamic changes in the flame, avoiding the omission of anomalies due to limited monitoring range. The dynamic feature extraction module uses a 3D convolutional kernel to scan continuous temporal frames, identifying morphological differences between the flame core and its edges. It calculates the energy distribution shift of thermal radiation intensity within a preset frequency band and outputs a dynamic feature vector containing coordinates of temperature abrupt changes and spectral characteristics. This process combines the dynamic changes in the temperature field with the shift in spectral features, enabling the capture of potential risks corresponding to subtle changes in flame morphology. Examples include changes in thermal radiation energy distribution accompanying abnormal flame edge diffusion, or localized combustion anomalies indicated by the appearance of temperature abrupt changes, making the system more sensitive to abnormal states. The abnormal behavior recognition module extracts the correlation patterns between the spatial distribution of the temperature field and the spectral characteristics of the flame through a deep residual network. It compares the current characteristics with the historical normal operating condition feature database, and generates a warning signal when the energy attenuation rate in a specific infrared band exceeds a threshold. This recognition method, based on multi-feature correlation analysis, integrates information from temperature, morphology, and spectrum, enabling it to distinguish different types of abnormal states and determine risk levels, reducing misjudgments that may result from single-feature analysis. For example, when incomplete combustion occurs in the flame due to an improper fuel mixture ratio, the changes in its spectral characteristics are specifically correlated with the abnormal distribution of the temperature field. The system can accurately identify this type of anomaly through this correlation pattern, rather than making a judgment based solely on a single isolated indicator. The system combines infrared thermal imaging technology with multispectral analysis, overcoming the limitations of traditional monitoring methods that rely on single parameters, and achieving comprehensive monitoring of the combustion status of gas stoves. Through multi-channel fused feature maps, it can simultaneously reflect comprehensive information on temperature field, flame morphology, and spectral characteristics, making the judgment of abnormal states more comprehensive. Simultaneously, dynamic analysis of continuous time frames allows the system to track the development of abnormal states, issuing early warnings at the initial stage of risk, giving users time to take countermeasures. Furthermore, the establishment of a historical normal operating condition feature database allows the system to adaptively adjust to the individual differences of different gas stoves and usage environments, adapting to the operating characteristics of different brands and models of gas stoves, and improving its applicability in diverse scenarios. Attached Figure Description

[0015] Figure 1This is a schematic diagram illustrating the working principle of the kitchen gas stove abnormality early warning system based on infrared thermal imaging multispectral as described in this invention. Figure 2 A flowchart for extracting dynamic features from a 3D convolutional kernel scanning temporal frame; Figure 3 A flowchart for constructing a multidimensional state matrix and inferring risk propagation paths; Figure 4 A flowchart for calibrating and updating the response curve of a blackbody radiation source. 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 This invention provides an abnormal early warning system for kitchen gas stoves based on infrared thermal imaging multispectral imaging, the system comprising: The infrared thermal imaging acquisition module captures real-time data on the temperature field distribution and flame morphology changes on the gas stove surface. A multispectral infrared sensor array simultaneously acquires infrared radiation signals in different bands, combined with visible light image acquisition equipment to obtain flame morphological characteristics. The raw thermal radiation signals undergo non-uniformity correction and environmental noise filtering to generate a multi-channel fused infrared thermal imaging feature map. The dynamic feature extraction module uses a three-dimensional convolutional kernel to scan the temperature gradient change region in continuous time frames, identifying the morphological differences between the flame core and the diffusion edge, calculating the energy distribution offset of thermal radiation intensity in a preset frequency band, and outputting a dynamic feature vector containing coordinates of temperature abrupt changes and spectral characteristics. The abnormal behavior recognition module extracts the correlation pattern between the spatial distribution of the temperature field and the flame spectral characteristics through a deep residual network, compares the current feature vector with the historical normal operating condition feature library, and generates a preliminary warning signal when the attenuation rate of the flame spectral energy in a specific infrared band exceeds a first threshold.

[0018] Example 1: See Figure 2The system employs a collaborative analysis and decision-making module and a tiered response control module. Upon receiving an initial warning signal from the abnormal behavior identification module, this module initiates a multi-source data fusion analysis process. The gas flow monitoring unit collects instantaneous gas flow data from the pipeline at a sampling frequency of five times per second. The flow meter uses a turbine-type sensor, and its output signal is converted from analog to digital and transmitted to the central processing unit. The stove power control unit records the burner solenoid valve opening, igniter operating status, and heat load adjustment parameters in real time. These operating parameters are uploaded to the data buffer via a serial communication interface. The time synchronization unit timestamps all input data at the millisecond level to ensure strict temporal alignment of information from different sources.

[0019] The construction of the multidimensional state matrix employs a sliding window mechanism, generating a state snapshot every 200 milliseconds. The matrix's row vectors consist of four core parameters: temperature field coordinates derived from the center points of marked abnormal regions in the infrared thermal imaging feature map, represented using the stove's planar coordinate system; spectral anomaly codes, encoded as three binary digits, representing the detection status of three abnormal modes: flame instability, gas leakage, and oil ignition; gas flow deviation values, calculating the percentage difference between the current flow rate and the set value; and power fluctuation values, recording the maximum power change of the stove within the last 500 milliseconds. The matrix has a fixed 20 columns, corresponding to continuous state records within the last 4 seconds, forming a spatiotemporally correlated data cube.

[0020] The probabilistic graphical model employs a dynamic Bayesian network structure, with nodes divided into observation and inference layers. Observation layer nodes are directly associated with parameters in the multidimensional state matrix, including nodes for temperature mutation intensity, spectral anomaly persistence, and flow fluctuation frequency. The inference layer contains twelve hidden nodes, forming a three-layer inference network corresponding to three main causes: equipment failure, operational anomaly, and environmental interference. Network parameters are obtained through training on a historical incident database containing 3,000 labeled samples covering eight typical failure modes. When new data is input, a message passing algorithm performs parallel computation on the network, updating the probability of all nodes within ten minutes and outputting the confidence distribution of each anomaly cause.

[0021] The risk propagation path prediction employs a thermodynamic simulation method. A physical model is established, incorporating the thermal conductivity of the stove surface material, the gas diffusion rate, and the air convection intensity. Based on the currently detected abnormal temperature distribution in the region, the heat source boundary conditions are initialized. The finite element analysis engine simulates the heat conduction and convection processes over the next thirty seconds with a step size of 0.1 seconds. During the simulation, gas flow rate variation data is dynamically introduced as a correction factor; when an abnormal increase in flow rate is detected, the energy input intensity of the simulated flame region is correspondingly increased. The prediction results generate a two-dimensional heat diffusion vector map, marking the coordinates of potentially affected hazardous areas and their estimated arrival times.

[0022] The graded response control module implements a multi-level interlocking strategy. The gas valve control unit receives three command modes: in monitoring mode, it maintains normal PID regulation; in level one warning, it enters flow limiting mode, controlling the gas supply within ±5% of the set value via a proportional valve; in level two warning, it activates flow suppression mode, reducing the flow to 30% of the set value within 300 milliseconds; and in level three warning, it triggers emergency shutdown, simultaneously cutting off the power to the gas main pipeline solenoid valve and igniter. Valve control signals are transmitted via industrial fieldbus, with response delay controlled within 50 milliseconds.

[0023] The alarm interface of the human-computer interaction terminal adopts a layered display design. The main view overlays visible light video streams and infrared thermal maps, with abnormal areas marked by dynamic contour lines. The auxiliary panel displays the current risk level, anomaly type icon, and a countdown progress bar. The handling suggestion library contains twenty-three preset solutions, which automatically generate guidance content based on the combination of anomaly types. When the system detects that the operator has not responded to the alarm for five consecutive seconds, it automatically enlarges the warning icon and starts a looping voice prompt.

[0024] The audible and visual alarm device adopts a modular design. The audio alarm unit includes three preset modes: a level 1 warning plays an intermittent 800Hz buzzer twice per second; a level 2 warning switches to a continuous 2000Hz tone, increasing the sound pressure level by 15 decibels; and a level 3 warning activates a composite alarm mode, alternately outputting low-frequency alarms and voice prompts. The optical warning unit consists of a ring-shaped LED array: a level 1 warning displays a constant yellow light; a level 2 warning switches to a red breathing light mode; and a level 3 warning activates a red strobe light and simultaneously illuminates a directional arrow pointing to the danger zone. All alarm modes maintain strict timing consistency across multiple devices via wireless synchronization signals.

[0025] The system maintenance interface has a built-in self-test program. Upon the first startup each day, it automatically checks communication link latency, sensor zero-point drift, and actuator travel. The verification data is recorded in non-volatile memory, forming a device health trend chart. Maintenance personnel can access the diagnostic logs using dedicated tools to view detailed parameter snapshots of all warning events over the past 72 hours. The remote diagnostic port supports real-time data stream mirroring, allowing authorized engineers to analyze the system's operating status online. Firmware updates utilize differential transmission technology, requiring only the transmission of modified data blocks for each upgrade, reducing network bandwidth consumption.

[0026] Example 2: See Figure 3The infrared thermal imaging acquisition module and the dynamic feature extraction module work collaboratively. The multispectral sensing unit uses four independent detector arrays arranged 30 cm above the stove. Each array contains 256×320 sensitive elements, corresponding to four infrared bands: 1.5-2.5μm, 3-5μm, 5-7μm, and 8-12μm. The multispectral sensing unit is equipped with a foldable high-temperature resistant bracket made of 304 stainless steel. The bracket has a folding angle range of 0-90° and is fixed to the wall behind the stove with expansion screws. The spatial position of the detector array can be adjusted manually or automatically. When the system detects the presence of a cookware placed in the stove area using infrared thermal imaging feature maps (the determination is based on more than 50% of the stove surface being covered by a low-temperature obstruction, and the outline of the obstruction matching the shape of the cookware), the bracket automatically drives the detector array to fold towards the wall. After folding, the vertical distance between the lower edge of the detector array and the stove surface remains unchanged at 30cm, while the horizontal distance between the detector array and the center of the stove increases to 15cm, and the detector array forms a 60° angle with the stove surface. This angle and spacing design avoids common cookware with a height ≤35cm (such as pressure cookers and deep soup pots). At the same time, through a spatial coordinate transformation algorithm (mapping the obliquely collected thermal imaging data into vertical viewpoint data), the integrity of the temperature field distribution data on the stove surface is ensured (covering an area extending 5cm beyond the edge of the stove). When the system detects that the cookware has been removed (the obstruction in the stove area disappears, and a normal combustion temperature field or room temperature is detected), the bracket automatically resets to the initial monitoring position within 1 second. Furthermore, the detector array housing features a rounded corner design with a 5mm radius, reducing the risk of equipment damage or personal injury from accidental contact with arms or spatulas during cooking. The detector array is encased in a nitrogen-protected housing, with a rotating filter wheel at the front for alternating sampling of different wavelengths. Each detection unit captures thermal radiation signals at a rate of 50 frames per second, and an analog-to-digital converter converts the analog voltage into a 14-bit digital signal output. The visible light acquisition device uses a 2-megapixel CMOS sensor, with the frame rate synchronized with the infrared sampling, and time alignment ensured by a hardware trigger signal.

[0027] The environmental compensation unit integrates three sets of high-precision sensors: a temperature sensor measures ambient air temperature, covering a range of 0-50℃; a humidity sensor detects relative humidity changes, with a range of 10-95%RH; and a barometric pressure sensor monitors atmospheric pressure fluctuations, with a resolution better than 1 hPa. These environmental parameters are sampled 10 times per second and transmitted to the signal processor via an I2C bus. An adaptive Kalman filter establishes a state-space model including three variables: temperature, humidity, and barometric pressure, to predict the degree of influence of environmental factors on infrared radiation. The filter's covariance matrix is ​​dynamically adjusted based on the environmental change rate over the past 30 seconds, increasing the weight of historical data in stable environments and improving the confidence of new data during periods of rapid change.

[0028] The data fusion unit performs multi-source information integration processing. The spatial registration algorithm first establishes the geometric correspondence between the visible light image and the infrared image, and uses SIFT feature point matching combined with manually calibrated reference points to generate an affine transformation matrix. The pixel-level fusion process adopts a weighted superposition strategy, with the visible light channel contributing edge detail information and the four infrared bands providing temperature features at different depths. The fused data cube contains six channels: RGB three-channel visible light data and infrared radiation intensity values ​​in four bands. Each pixel is accompanied by three-dimensional spatial coordinates, with the origin of the coordinate system set at the geometric center of the stove, the XY plane parallel to the stove surface, and the Z-axis vertically upward.

[0029] The temporal analysis unit processes continuous infrared thermal imaging frame sequences. The differential calculation module compares the radiation intensity changes of corresponding pixels in adjacent frames and uses a sliding window to calculate the temperature change trend of each point over the past five frames. An abnormally active region detection algorithm labels two types of feature points: pixel groups with a temperature rise rate exceeding 20°C per second, and edge points exhibiting discontinuous temperature gradients. A region growing method expands these feature points into connected components, eliminating noise interference with areas smaller than 50 pixels. The labeling results are stored in a two-dimensional matrix, where matrix element values ​​represent the state type at that location: 0 indicates a background region, 1 indicates a region with a sharp temperature increase, and 2 indicates a region with abnormal heat conduction.

[0030] The morphological processing unit performs structural analysis on the detected active regions. An iterative erosion operation first eliminates isolated noise points at the region edges, followed by a dilation operation to restore the original dimensions of the main structure. A shape analysis algorithm calculates the geometric features of each connected region: the core region is defined as the set of pixels with radiation intensity exceeding 80% of the maximum value, and the edge region is the transition zone with intensity between 30% and 80%. The feature extraction process outputs three dynamic parameters: the core region area change rate reflects flame intensity fluctuations, the edge perimeter to area ratio characterizes flame morphological complexity, and the core region centroid displacement velocity indicates flame stability.

[0031] The frequency domain transformation unit performs time-frequency analysis. The thermal radiation signal of the selected analysis region is first processed by the Hanning window function to reduce spectral leakage. A Fast Fourier Transform converts the time-domain signal into a frequency-domain energy distribution, with a frequency resolution set to 0.5 Hz. Energy integration in the characteristic frequency bands calculates two key indicators: the cumulative energy value of the mid-wave infrared band (3-5 μm) in the 50-100 Hz band, and the cumulative energy value of the long-wave infrared band (8-12 μm) in the 10-30 Hz band. The energy ratio analysis module calculates the dynamic changes in energy in these two characteristic frequency bands, triggering a frequency domain anomaly flag when the ratio deviates from the reference range by more than 15%.

[0032] The feature integration unit encapsulates the outputs of each module into a standard data structure. The dynamic feature vector is organized in JSON format, containing four fields: timestamp, spatial coordinates, morphological parameters, and frequency domain features. The timestamp is accurate to the millisecond level, recording the absolute time of feature extraction. The spatial coordinate field stores the vertex sequence of the boundary polygons of the abnormal region, represented using a stove coordinate system. The morphological parameter field includes seven indicators such as core area, edge complexity index, and centroid movement speed. The frequency domain feature field records the energy distribution ratio of each band and the energy value of the characteristic frequency band. Data packets are transmitted to the abnormal behavior recognition module using zero-copy technology, with an average transmission latency controlled within 2 milliseconds.

[0033] A calibration and maintenance mechanism ensures the long-term stability of the system. A baseline calibration process is automatically executed weekly, controlling the blackbody radiation source to sequentially generate three standard temperature points: 100℃, 300℃, and 500℃. Each temperature point is maintained for three minutes, and the response curves of all detectors are collected. A piecewise linear approximation method is used to fit the temperature-voltage relationship, establishing independent correction equations within the characteristic temperature range. Calibration coefficients are stored in the FPGA's configuration memory, and the correction parameters for the corresponding range are automatically retrieved during real-time signal processing. The environmental compensation unit updates the noise model parameters monthly, recalculating the process noise matrix of the Kalman filter based on the statistical distribution of environmental data from the past thirty days.

[0034] The quality control module monitors the reliability of data at each stage. Infrared detector health checks include three components: dark current testing, response uniformity detection, and noise-equivalent temperature difference calculation, which are performed automatically every morning. Visible light cameras undergo regular white balance calibration and lens contamination detection; a maintenance alarm is triggered when more than 5% of pixels are found to be abnormal. The data transmission channel implements cyclic redundancy check, with each data packet appended with a CRC-32 checksum; automatic retransmission is requested when checksum fails. The system status dashboard displays real-time operating metrics for each module, including over twenty parameters such as CPU load, memory usage, and data throughput, allowing maintenance personnel to monitor the system's health.

[0035] Example 3: See Figure 4 The process involves deep interaction between the abnormal behavior identification module and the collaborative analysis and decision-making module. The feature comparison unit uses a dynamic time warping algorithm to process the input feature vector sequence, which can effectively match pattern changes over different time periods. The historical normal operating condition feature library stores typical operating scenario data collected over the past three months, including twelve benchmark templates such as regular cooking, heat adjustment, and cookware handling. Each template consists of fifty sets of feature vector sequences, recording the temperature field and spectral evolution process within a complete operating cycle. Similarity calculation uses an improved cosine similarity metric, assigning different weights to spatial distribution features and spectral features respectively.

[0036] in, This represents the overall similarity score. and These are the weighting coefficients for spatial features and spectral features, respectively. and Represents the spatial distribution components of the current feature vector and template vector. and These represent the current spectral feature vector and the historical template spectral vector, respectively. This is the temperature normalization factor. This represents the number of feature dimensions. When the overall similarity is below 0.65, the system determines that the current state deviates from the normal range.

[0037] The residual network unit employs a five-layer convolutional structure to handle anomalous features. The first convolutional layer uses a 3×3×3 3D convolutional kernel to extract local correlation features along the temporal, spatial, and channel dimensions. The second to fourth layers introduce skip connections, each containing two sets of 3×3×1 2D convolutional kernels, coupled with batch normalization and the ReLU activation function. The final layer uses a 1×1×1 convolution to achieve channel dimensionality reduction, outputting a 128-dimensional depth feature vector. During network training, a focal loss function is used to assign higher weights to minority class anomalous samples. During online inference, the network outputs classification results every 200 milliseconds, including probability distributions for three anomalous types: flame instability, gas leaks, and oil spill ignition.

[0038] The risk quantification unit implements a multi-level threshold judgment strategy. The long-wave infrared energy monitoring window is set to five consecutive frames of data, calculating the moving average rate of change of energy values. When an energy value decreases by more than 40% within three frames, a flame instability warning is triggered. Mid-wave infrared hotspot analysis employs dual criteria: a spatial criterion requiring the abnormal area to exceed 5% of the total stove area, and a spectral criterion requiring the energy ratio in the 3-5μm band to exceed twice the standard deviation of the historical mean. Meeting either criterion generates a gas leak warning signal. Oil ignition detection is combined with visible light image analysis; when the visible light image corresponding to the infrared hotspot area shows oil stain reflective characteristics, it is determined to be an oil-related anomaly.

[0039] The data integration unit of the collaborative analysis and decision-making module implements a precise time alignment mechanism. The hardware time synchronization module uses the PTP protocol to maintain clock consistency across data sources, with errors controlled within 1 millisecond. Gas flow data and power parameters are transmitted via message queue middleware, with each message accompanied by a timestamp accurate to the microsecond level. The matrix construction process uses a sliding window technique, generating a state snapshot every 50 milliseconds. Temperature field coordinate data undergoes coordinate transformation, uniformly converting to a polar coordinate system centered on the gas nozzle. Spectral anomaly encoding uses a three-bit bitmap format, with the highest bit indicating flame instability, the middle bit indicating gas leakage, and the lowest bit indicating oil ignition. Flow deviation values ​​are calculated using a moving average filter, based on the average of the past ten seconds as a benchmark reference.

[0040] A Bayesian inference unit constructs a dynamic probabilistic network model. The observation layer contains eighteen evidence nodes, corresponding to features such as the intensity of temperature field abrupt changes, the duration of spectral anomalies, and the amplitude of flow fluctuations. The hidden layer constructs a three-layer inference structure: the first layer distinguishes between equipment failures and operational anomalies; the second layer refines the failure type (e.g., valve failure, sensor drift); and the third layer identifies specific components (main burner, ignition electrode, etc.). The network parameters are trained using the expectation-maximization algorithm, using a training set containing 5,000 labeled samples, covering fifteen typical failure scenarios. An approximate message-passing algorithm is used for real-time inference, completing the full network probability update within fifty milliseconds.

[0041] The path prediction unit implements multiphysics coupled simulation. The heat conduction model considers the anisotropic thermal conductivity of the stove's metal components and uses an unstructured mesh to divide the computational domain. The gas diffusion model is solved based on the Navier-Stokes equations, combined with a turbulence model to simulate the three-dimensional distribution of leaked gas. The simulation engine uses an explicit time-progression method, outputting prediction results every 0.5 seconds. The visualization interface generates a two-dimensional risk heat map, uses the HSV color space to encode the temperature gradient, and overlays vector arrows to indicate the direction of heat flow. The prediction results are updated every three seconds, and recalculation is triggered immediately when new abnormal features are detected.

[0042] The anomaly localization algorithm integrates multi-source information. The spatial matching module aligns infrared thermal imaging coordinates with visible light image feature points, combining this with the stove's mechanical design drawings to determine the physical location of the abnormal component. The probabilistic fusion module combines the component failure probability from a Bayesian network with the heat spread path prediction results to calculate the confidence factor for each component. The localization result is output in polar coordinates, including three parameters: radial distance, azimuth, and height, with millimeter-level accuracy. The system maintenance interface provides a historical location history query function, allowing users to trace the location information of all anomalies within the past 24 hours.

[0043] The tiered response strategy employs a dynamic adjustment mechanism. Level 1 response addresses transient flame fluctuations, maintaining monitoring without triggering active intervention, only displaying a prompt on the human-machine interface. Level 2 response addresses persistent anomalies, automatically adjusting the damper opening to optimize the air-fuel ratio in addition to limiting gas flow. Level 3 response executes a complete safety interlock: shutting off the main gas valve, cutting off power, activating the ventilation system, and simultaneously sending an alarm SMS to preset contacts. Response latency testing shows that the end-to-end latency from anomaly detection to the execution of a Level 3 response is controlled within 800 milliseconds. Strategy parameters can be adjusted through the configuration interface, supporting customized response thresholds for different kitchen environments.

[0044] Example 4: Complete operation mechanism and feedback optimization system of the graded response control module. The valve control unit adopts a hybrid PID and bang-bang control strategy. The regulation algorithm of the gas proportional valve considers the dynamic balance of current flow rate, target limit value, and rate of change.

[0045] in, This indicates the valve opening adjustment amount. The deviation between real-time traffic and the target value. , , These are the proportional, integral, and differential coefficients, respectively. For the integration time window, This is a nonlinear compensation factor. The control algorithm can adjust the flow rate to within ±3% of the target value while ensuring stability, with a response time of no more than 200 milliseconds. When an emergency shutdown command is triggered, an independent safety relay module directly cuts off the solenoid valve's power supply circuit, bypassing the main control system to achieve hardware-level protection.

[0046] The heatmap generation unit implements multi-dimensional data visualization. A temperature gradient mapping algorithm converts absolute temperature values ​​acquired by infrared sensors into relative risk levels, employing an improved HSL color space encoding scheme: hue components represent temperature from low to high (blue-green-yellow-red), saturation reflects the rate of temperature change, and brightness adjustment is automatically optimized based on ambient lighting conditions. Hazard area contour extraction uses the MarchingSquares algorithm to generate smooth isotherm boundaries. The visualization engine renders 15 frames of heatmaps per second, overlaying them onto a 1080p resolution visible light video stream, and eliminating image misalignment through edge blending technology.

[0047] The alarm strategy unit uses a state machine model to manage the audible and visual alarm devices. State transition conditions include seven event types, such as changes in risk level, user response timeout, and equipment failure. The audio signal generator is programmable to output any waveform within the 20Hz-20kHz range. Level 1 alarms use a square wave with a 30% duty cycle, level 2 alarms switch to a sawtooth wave, and level 3 alarms use a composite sound effect with amplitude modulation. The LED driver circuit features 256 levels of brightness adjustment, and the breathing effect is achieved through PWM frequency modulation. The strobe mode follows the 3-5Hz flicker frequency specified by international safety standards. All alarm mode parameters are stored in an encrypted configuration file and support remote OTA updates.

[0048] The feedback learning unit constructs a user behavior analysis model. The operation response database records the user's intervention time, operation type, and handling effect for each alarm event. The feature extraction module mines typical patterns from historical data and establishes a 20-dimensional decision tree model including environmental parameters, alarm type, and time period characteristics. The adaptive adjustment algorithm dynamically updates the risk level mapping rules based on the latest 30 interaction records. When the response delay of a specific type of alarm is detected to continuously exceed a threshold, the initial warning level for that type of anomaly is automatically increased. The learning process employs an incremental update mechanism to avoid drastic fluctuations in model parameters.

[0049] The equipment linkage unit implements a multi-system collaborative control protocol. The smoke exhaust system interface adopts the Modbus RTU communication standard, supporting precise adjustment of fan speed. When the strong exhaust mode is activated, the system first detects the data from the flue pressure sensor, and then increases the speed in three stages: initially running at 50% power for 3 seconds, then linearly accelerating to 85% in the transition stage, and finally maintaining 100% power until the anomaly is resolved. Intelligent ventilation control is based on real-time monitoring of the indoor air quality index (AQI). When PM2.5 or CO concentrations exceed safety limits, the fresh air system is automatically activated to supplement clean air. All linkage commands are managed through a priority arbitration mechanism to ensure that critical safety operations receive the highest execution authority.

[0050] The human-computer interface implements a context-aware design. The touchscreen display automatically adjusts its content based on the user's distance: when the infrared proximity sensor detects someone within 1 meter, detailed parameters and operation buttons are displayed; beyond this range, the display simplifies to status icons and risk level indicators. The voice interaction module supports both offline keyword recognition and online semantic understanding modes, and includes a built-in dictionary specifically for cooking scenarios. The help system employs a context-sensitive card-based design, dynamically generating guidance content with illustrations and text based on the current anomaly type and kitchen equipment model. The interface layout is optimized using eye-tracking data, ensuring key control elements remain within the optimal visual area.

[0051] The system health monitoring implements closed-loop management. The self-diagnostic program scans the status registers of each hardware module every morning, checking for 42 indicators including memory integrity, sensor calibration status, and communication link quality. The fault recovery strategy employs a three-tiered progressive mechanism: primary errors attempt automatic repair and log recording; intermediate errors trigger a partial module restart; and critical errors execute a system safety mode switch. The maintenance reminder function intelligently predicts based on equipment operating hours and usage environment, notifying users two weeks in advance of routine maintenance items such as filter replacement and sensor cleaning. The remote diagnostic interface uses a two-way authenticated encrypted channel, allowing engineers to view equipment status and download diagnostic logs in real time.

[0052] The security protection mechanism implements a defense-in-depth strategy. Physical layer protection includes tamper-proof switches and a sealed casing design; any unauthorized opening will trigger a security lock. The data layer employs AES-256 encryption for storage and transmission, with critical parameters stored in a tamper-proof secure chip. The communication layer uses whitelist filtering and frequency hopping technology to prevent wireless signal interference and man-in-the-middle attacks. Operation permissions are managed hierarchically: routine cooking operations require no authentication, system settings require password verification, and modifications to critical security parameters require physical key authorization. All security events are recorded in a separate security log, retained for at least three years, and support forensic-level audit trails.

[0053] Version management employs a dual-backup redundancy mechanism. The main control system runs in an isolated container environment, and update packages are only installed after digital signature verification. The rollback mechanism retains the three most recent stable versions, automatically switching to the previous available version when a new version is detected to be malfunctioning. Configuration data is subject to a differential backup strategy, synchronizing to a secure storage area hourly to ensure rapid restoration of personalized settings after a system reset. Firmware integrity verification is performed at each boot, using the SHA-3 algorithm to verify the hash values ​​of each module; any mismatch triggers a security alert and enters a restricted operating mode.

[0054] Example 5: Focusing on the collaborative working mechanism between the self-calibration module and the data management system. The system enters its maintenance window at 3:00 AM daily, and the self-calibration program automatically initiates the calibration process. A stepper motor drive moves the blackbody radiation source from its standby position to directly beneath the infrared sensor array, achieving a positioning accuracy of 0.1 mm. The blackbody radiation source is externally equipped with a double-layer safety protection structure: the inner layer is a ceramic fiber heat shield (5 mm thick, temperature resistance ≥1200℃, thermal conductivity ≤0.03 W / (m²)). The blackbody radiation source has an outer frosted stainless steel shell (1.5mm thick), with a 10mm gap between the two layers forming an air insulation layer, which keeps the outer surface temperature of the insulation cover below 40℃ (at an ambient temperature of 25℃). A fiberglass fireproof cotton pad (8mm thick, temperature resistance ≥800℃) is attached to the bottom of the blackbody radiation source to prevent high-temperature components from directly contacting the stovetop or cabinets during calibration. Three NTC temperature sensors are distributed across the surface of the blackbody radiation source to monitor the temperature of the radiating surface, the inner layer of the insulation cover, and the outer layer of the insulation cover, respectively. When the temperature of the radiating surface exceeds 550℃, the temperature of the inner layer of the insulation cover exceeds 150℃, or the temperature of the outer layer exceeds 40℃, the system immediately cuts off the power supply circuit to the blackbody radiation source and simultaneously sends a temperature over-limit alarm to the human-machine interface terminal. The stepper motor drive system is equipped with a torque sensor and limit switches. When the motor drives the blackbody radiation source to move, if the detected resistance torque exceeds 0.5N... If the blackbody radiation source is detected as obstructed (m), or reaches the preset movement limit position, the motor will immediately stop running. Two infrared ranging sensors (detection distance 0-30cm) are installed on both sides of the blackbody radiation source's movement path. If a foreign object (such as a pot or kitchen utensil) is detected in the path, the system will pause the calibration process and send a foreign object alarm. The system will restart after the user removes the foreign object. The self-calibration program supports custom configuration of calibration time. Users can set the calibration time to non-cooking periods (such as 9:00-11:00 AM or 3:00-5:00 PM) through the kitchen human-machine interface terminal or the accompanying mobile APP. The system's default calibration time is 10:00 AM. Ten minutes before the calibration process starts, the system will display a text prompt that self-calibration is about to start, expected to last 15 minutes, and emit a 500Hz buzzer for 2 seconds. Within 10 minutes, users can click on the terminal to delay calibration (which can be delayed by 1-24 hours) or cancel calibration (cancellation will automatically postpone to the same time period the next day). During calibration, the system pushes calibration progress updates to the user every 3 minutes via the app (e.g., collecting 100℃ reference data, collecting 300℃ reference data). If any abnormalities occur, such as exceeding temperature limits or motor malfunction, a calibration anomaly alert is immediately sent, requesting the user to check the equipment's emergency alarm and triggering a red flashing light (3Hz) and a continuous 1000Hz beep until the user confirms the issue. The blackbody radiating surface heats up at a rate of 5 degrees Celsius per minute, stabilizing at three reference temperatures: 100℃, 300℃, and 500℃. Each temperature is maintained for three minutes to allow the system to reach thermal equilibrium. The multispectral sensor array collects 500 sets of raw response data at each temperature point, covering the pixel positions of all detection units.

[0055] The environmental compensation unit operates synchronously during calibration, recording instantaneous values ​​of laboratory environmental parameters. A temperature sensor monitors the gradient changes in the air layer surrounding the blackbody radiation source, a humidity sensor detects the dew point within the calibration chamber, and a barometer records atmospheric pressure fluctuations. This environmental data, along with the raw output from the infrared sensor, constitutes the calibration dataset, used to eliminate interference from environmental factors on the measurement results. An adaptive filtering algorithm analyzes the changing trends of historical calibration data and dynamically adjusts the compensation coefficients of each environmental parameter, enabling the system to adapt to seasonal climate changes.

[0056] When processing the collected calibration data, the curve fitting unit uses a piecewise linear interpolation method to establish a temperature-voltage response model. Each detection unit independently calculates the slope parameters for three temperature ranges, and the best-fit line is determined through minimum residual optimization. An abnormal pixel detection algorithm identifies detection units whose response curves deviate from the group characteristics; these units are marked as faulty pixels and excluded in subsequent processing. The correction coefficient matrix is ​​stored according to the physical layout of the detector array. During real-time signal processing, the reading of each pixel is used to call the corresponding correction parameters for compensation.

[0057] The parameter update process employs a dual verification mechanism. Newly generated correction coefficients are first written to a temporary storage area. The verification program compares the changes between the old and new parameters. When any pixel's correction value abruptly exceeds a safety threshold, a manual review process is triggered. Verified parameters are encrypted and written to the non-volatile memory of the environmental compensation unit, while simultaneously being backed up to the system's secure area. The update log meticulously records the time of each calibration, operator identification, environmental conditions, and a summary of key parameters, forming a complete audit trail.

[0058] The data storage module implements an intelligent recording strategy. Under normal operating conditions, the system saves a compressed snapshot of the infrared thermal imaging feature spectrum every five minutes, retaining only the core features of the temperature distribution. When the anomaly detection module triggers an alert, it automatically switches to a high-density storage mode, recording a complete multispectral data sequence of 30 seconds before and after the event, increasing the sampling rate to 10 frames per second. The storage format adopts a layered design: the base layer contains basic information on the temperature field and spectrum, while the extended layer stores raw sensor data and intermediate processing results, meeting the analytical needs of different levels.

[0059] Storage management employs a circular buffer mechanism, retaining ordinary monitoring data for seven days, early warning event data for thirty days, and major accident data for permanent archiving. Data compression algorithms select the optimal strategy based on information type: temperature matrices use lossy compression to retain key features, visible light images use lossless compression to ensure detail integrity, and spectral data utilizes feature extraction methods to reduce storage space. The search interface supports multi-dimensional queries by time range, anomaly type, risk level, etc., enabling rapid location of historical event records.

[0060] The remote communication module establishes a secure data transmission channel. When a Level 3 risk alarm is triggered, the system immediately packages key information such as the current heat map, sensor readings, and handling records, encrypts them using AES-256, and uploads them to the cloud monitoring platform. The transmission protocol employs adaptive bitrate technology, dynamically adjusting data fragment size and retry strategies based on network conditions to ensure critical data transmission can still be completed even in weak network environments. A heartbeat mechanism checks the connection status every five minutes and automatically switches to a backup communication path in case of abnormal disconnection.

[0061] The cloud interface enables bidirectional data synchronization. Configuration updates and algorithm upgrade packages issued by the monitoring platform are stored in an isolated area after digital signature verification and are installed according to a predetermined schedule during system idle periods. Remote diagnostic sessions employ end-to-end encryption, allowing engineers to view device status in real time through a secure tunnel; however, all operation commands require local physical confirmation before execution. Data synchronization logs record metadata for each transmission, including data volume, transmission time, and integrity checksums, for network performance analysis and optimization.

[0062] The self-test program runs a complete hardware diagnostic every time the system starts. The infrared detector performs dark current testing and uniformity checks, the visible light camera performs focus and white balance calibration, and the communication module verifies signal strength and transmission error rate. Diagnostic results are displayed as status codes on the human-machine interface: green indicates normal operation, yellow indicates caution, and red indicates a fault. The remaining lifespan of critical components is calculated based on operating hours and conditions, and a replacement reminder is issued two weeks in advance.

[0063] The maintenance support system provides comprehensive technical support. The knowledge base includes equipment structure diagrams, fault code manuals, and maintenance guides, supporting keyword search and fault tree navigation. Augmented reality assistance uses cameras to identify equipment components and overlays disassembly and assembly steps and precautions. The spare parts management system tracks the inventory status of critical components and automatically generates purchase requests when replacement is imminent. A service appointment interface connects to the manufacturer's support system, allowing for one-click initiation of remote assistance or on-site service requests.

[0064] Version control systems maintain the complete evolution history of the software. Each update generates a differentiated upgrade package, transmitting only modified data blocks to reduce bandwidth consumption. A rollback mechanism retains the three most recent stable versions, allowing for quick restoration to a previous state if a new version malfunctions. Configuration data and user settings are stored independently of the main program, ensuring no personalized information is lost during updates. Version compatibility checks verify hardware and software compatibility before installation, preventing system failures caused by incompatible versions.

[0065] The security monitoring system provides 24 / 7 protection. The intrusion detection system analyzes network traffic patterns to identify and block abnormal access attempts. Data integrity checks regularly scan system files, comparing hash values ​​to detect potential tampering. Physical security sensors monitor the status of device casings; any unauthorized opening triggers a security lock and uploads an alarm. All security events generate encrypted logs, stored in a tamper-proof dedicated chip, meeting industry security audit requirements. Access control implements the principle of least privilege, with different roles having strictly differentiated operating permissions; critical security functions require multi-factor authentication for access.

[0066] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0067] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A kitchen gas stove anomaly early warning system based on infrared thermal imaging multispectral imaging, characterized in that, The system includes: The infrared thermal imaging acquisition module is used to synchronously capture the surface temperature field distribution of the gas stove head through a multispectral infrared sensor array, combine it with visible light image acquisition equipment to obtain real-time changes in flame shape, perform non-uniformity correction and environmental noise filtering on the original thermal radiation signal, and generate a multi-channel fused infrared thermal imaging feature map. The dynamic feature extraction module, based on the infrared thermal imaging feature map, uses a three-dimensional convolution kernel to scan the temperature gradient change region in continuous time frames, identifies the morphological feature differences between the flame core region and the diffusion edge, calculates the energy distribution offset of thermal radiation intensity in a preset frequency band, and outputs a dynamic feature vector containing temperature change coordinates and spectral characteristics. The abnormal behavior identification module receives the dynamic feature vector, extracts the correlation pattern between the spatial distribution of the temperature field and the flame spectral features through a deep residual network, compares the matching degree of the current feature vector with the historical normal operating condition feature library, and generates a preliminary warning signal containing the abnormal type code and risk level when the attenuation rate of the flame spectral energy in a specific infrared band exceeds the first threshold.

2. The kitchen gas stove anomaly early warning system based on infrared thermal imaging multispectral as described in claim 1, characterized in that, Also includes: The collaborative analysis and decision-making module, based on the preliminary early warning signal, integrates real-time monitoring data from the gas flow meter with the operating parameters of the stove power control unit, constructs a multi-dimensional state matrix, infers the cause of the anomaly and the risk propagation path through a probabilistic graphical model, and outputs the anomaly location coordinates and risk diffusion trend prediction results. The graded response control module sends graded adjustment commands to the gas valve control unit based on the abnormal location coordinates and risk spread trend prediction results, sends an alarm data packet containing a risk heat map and handling suggestions to the human-machine interface terminal, and simultaneously triggers the audible and visual alarm device to execute the warning mode corresponding to the risk level.

3. The kitchen gas stove anomaly early warning system based on infrared thermal imaging multispectral as described in claim 1, characterized in that, The infrared thermal imaging acquisition module includes: The multispectral sensing unit is equipped with four independent detector arrays covering the short-wave infrared to long-wave infrared bands. Each detector array captures the thermal radiation signal on the surface of the stove at a millisecond sampling frequency. The environmental compensation unit integrates environmental temperature and humidity sensors and air pressure monitoring elements to collect kitchen environmental parameters in real time and eliminates the interference of environmental variables on thermal radiation signals through an adaptive Kalman filter. The data fusion unit receives the output data from the multispectral sensing unit and the environmental compensation unit, and uses a pixel-level weighted fusion algorithm to align the spatial coordinates of the visible light image and the infrared channel to generate the infrared thermal imaging feature map with three-dimensional attributes of temperature, spectrum and space.

4. The kitchen gas stove anomaly early warning system based on infrared thermal imaging multispectral as described in claim 3, characterized in that, The dynamic feature extraction module includes: The time-series analysis unit calculates the temperature change rate of each pixel between adjacent frames based on the continuous frame sequence of the infrared thermal imaging feature map, and marks abnormally active areas where the temperature change rate exceeds a preset rate. The morphological processing unit performs corrosion and expansion operations on the abnormally active region to separate the flame core region from the edge diffusion region and extracts the core region area change gradient and edge irregularity index. The frequency domain conversion unit converts the thermal radiation signal of the selected area into a frequency domain energy distribution spectrum and calculates the energy ratio offset between the mid-wave infrared band and the long-wave infrared band. The feature integration unit encapsulates the temperature change rate, core area change gradient, edge irregularity index, and energy ratio offset into the dynamic feature vector.

5. The kitchen gas stove anomaly early warning system based on infrared thermal imaging multispectral as described in claim 4, characterized in that, The abnormal behavior identification module includes: The feature comparison unit separates the temperature field distribution pattern and spectral feature parameters from the dynamic feature vector, calls the benchmark template in the historical normal operating condition feature library, and calculates the cosine similarity between the current feature vector and the benchmark template. The residual network unit receives the similarity data output by the feature comparison unit, extracts cross-channel correlation features through five layers of convolutional residual blocks, and identifies abnormal patterns such as flame extinguishing precursors, gas leak thermal diffusion, and oil ignition. The risk quantification unit generates a first-class warning code characterizing flame instability when the energy value of the long-wave infrared band decreases by more than 40% within three consecutive frames, and generates a second-class warning code characterizing gas leakage when a hot spot with non-flame characteristics appears in the mid-wave infrared band.

6. The kitchen gas stove anomaly early warning system based on infrared thermal imaging multispectral as described in claim 5, characterized in that, The collaborative analysis and decision-making module includes: The data integration unit receives instantaneous flow fluctuation data uploaded by the gas flow meter and operating power parameters of the stove power control unit, and timestamps them with the preliminary warning signal. The matrix construction unit constructs row vectors from temperature field coordinates, spectral anomaly codes, gas flow deviation values, and power fluctuation values, and stacks them in a time series to form the multidimensional state matrix. The Bayesian inference unit, based on a conditional probability model trained on a historical accident database, infers the probability distribution of whether the current anomaly is caused by abnormal gas pressure or failure of stove components. The path prediction unit simulates the spread path of abnormal heat sources on the stove surface and the expected range of impact based on the correlation between heat diffusion rate and gas flow rate changes.

7. The kitchen gas stove anomaly early warning system based on infrared thermal imaging multispectral as described in claim 6, characterized in that, The hierarchical response control module includes: The valve control unit receives the abnormal positioning coordinates and risk diffusion trend prediction results. When the risk level reaches level two, it sends a 30% opening limit command to the gas proportional valve. When it reaches level three, it sends a shut-off command. The heat map generation unit maps the anomaly location coordinates to the visible light image coordinate system and superimposes temperature gradient color scale markers to form the risk heat map spectrum. The alarm strategy unit matches the audible and visual alarm modes according to the risk level: Level 1 risk triggers a low-frequency intermittent buzzer and a yellow indicator light; Level 2 risk triggers a high-frequency continuous buzzer and a red indicator light; and Level 3 risk triggers a rotating warning light and a voice broadcast.

8. The kitchen gas stove anomaly early warning system based on infrared thermal imaging multispectral as described in claim 1, characterized in that, Also includes: The self-calibration module is configured during the daily system startup phase. It calibrates the response curves of each infrared detection channel using a built-in standard blackbody radiation source and updates the filtering parameters of the environmental compensation unit. The data storage module continuously records the dynamic changes of the infrared thermal imaging feature spectrum and stores the complete multispectral data sequence for 30 seconds before and after the occurrence of the abnormal event. The remote communication module uploads Level 3 risk alarm data packets to the cloud monitoring platform in real time via an encrypted transmission protocol.

9. The kitchen gas stove anomaly early warning system based on infrared thermal imaging multispectral as described in claim 8, characterized in that, The self-calibration module includes: The calibration execution unit drives the stepper motor to move the blackbody radiation source to the center of the infrared sensor's field of view, and acquires the raw response data at three temperature reference points; The curve fitting unit uses the least squares method to fit the temperature-voltage response equation of each detector channel and generates a new correction coefficient matrix. The parameter update unit writes the correction coefficient matrix into the non-volatile memory of the environmental compensation unit.

10. The kitchen gas stove anomaly early warning system based on infrared thermal imaging multispectral as described in claim 7, characterized in that, The alarm strategy unit also includes: The feedback learning unit collects user response delay data to alarms. If the user does not intervene within ten seconds after a single Level 3 risk alarm, the initial risk level of the same type of anomaly will be automatically increased. When a level 3 alarm is triggered, the equipment linkage unit simultaneously shuts down the kitchen exhaust system fan and sends a forced exhaust command to the intelligent ventilation device.

Images