A point type flame detection method based on multi-infrared band multi-scale time-frequency fusion

Abstract

The application discloses a point type flame detection method based on multi-infrared waveband multi-scale time-frequency fusion, which comprises the following steps: acquiring time sequence sampling signals of multiple infrared channels to perform robust background estimation and baseline deduction, and obtaining baseline deducted signals of each channel; performing frequency domain transformation, calculating channel energy in a target flame flickering frequency band, and obtaining main frequency domain enhancement features and second candidate waveband auxiliary frequency domain information; calculating a window level flame probability through a light time-frequency fusion model; performing multi-phase fusion on flame probabilities of a main window and multiple offset windows to obtain a fusion flame probability of a current period; performing confidence mapping on the fusion flame probability, and adding a mapped evidence value into a rolling evidence integral queue; and outputting a flame alarm result according to infrared evidence integrals in the rolling evidence integral queue and an alarm state machine. The application can realize time-frequency joint identification and improve the adaptability to long-distance small fire sources and complex interference environments.

Description

Technical Field

[0001] This invention relates to the field of fire detection technology, and in particular to a point flame detection method based on multi-infrared band multi-scale time-frequency fusion. Background Technology

[0002] Point flame detectors typically use the radiation changes and flicker characteristics of flames in specific infrared bands to trigger alarms. Common technical approaches include infrared amplitude threshold detection, multi-band ratio detection, flame flicker frequency analysis, ultraviolet-infrared composite judgment, and flame classification based on machine learning.

[0003] In practical applications, flame detectors often operate in complex environments: small, distant fire sources result in weak flame characteristics; direct sunlight, foliage shadows, artificial light, heat sources, and welding can cause non-flame infrared fluctuations; the sensors themselves may also experience temperature drift, DC bias, and long-term background changes. Traditional rule-based methods rely on fixed thresholds, making it difficult to balance sensitivity and false alarm resistance. While machine learning methods can learn complex waveforms, they are also susceptible to environmental and hardware variations if the physical relationships between infrared bands are not constrained.

[0004] Meanwhile, the interference actually faced by flame detectors is not limited to a single frequency point or intensity band. Interference from sunlight, lamplight, welding, and heat source movement may exhibit fluctuations similar to flames in certain frequency bands, but their multi-band relative relationships, temporal continuity, and alarm persistence usually differ from those of a real flame. Therefore, it is necessary to unify the physical relationships between infrared bands, the temporal flicker pattern of flames, and continuous alarm evidence into a single detection chain, rather than relying solely on frequency domain thresholds, feature discrimination, or single model outputs for judgment.

[0005] Therefore, there is a need for a flame detection method that can utilize the spectral differences of multiple infrared bands, learn the temporal scintillation pattern of flames, and is suitable for low-computing deployment of point detectors. Summary of the Invention

[0006] The purpose of this invention is to provide a point flame detection method based on multi-infrared band and multi-scale time-frequency fusion, which can realize joint time-frequency identification and improve the adaptability to small fire sources at long distances and complex interference environments.

[0007] To achieve the objectives of this invention, a point flame detection method based on multi-infrared band multi-scale time-frequency fusion is provided, comprising the following steps:

[0008] S1. Acquire the timing sampling signals of multiple infrared channels, wherein the multiple infrared channels include the main three channels and the second flame candidate band;

[0009] S2. Perform robust background estimation and baseline subtraction on the time-series sampled signals of each infrared channel to obtain the baseline-subtracted signals of each channel;

[0010] S3. Perform frequency domain transformation on the baseline-subtracted signal and calculate the channel energy in the target flame flashing frequency band. The main frequency domain enhancement features are obtained based on the calculation results of the main three channels, and the auxiliary frequency domain information of the second candidate band is obtained based on the calculation results of the second flame candidate band.

[0011] S4. Input the main three-channel time-domain waveform, the main frequency-domain enhancement features, and the auxiliary frequency-domain information of the second candidate band into the same lightweight time-frequency fusion model, and calculate the window-level flame probability through the attention gating and fusion classification within the model.

[0012] S5. Perform multi-phase fusion on the flame probabilities of the main window and multiple offset windows to obtain the fused flame probability of the current period;

[0013] S6. Calculate the confidence level of the fused flame probability and add the mapped evidence value to the rolling evidence integration queue.

[0014] S7. Output the flame alarm result based on the infrared evidence points in the rolling evidence points queue and the alarm state machine.

[0015] As a preferred embodiment of the present invention, the multi-channel infrared system includes a first channel, a second channel, a third channel, and a fourth channel. The combination of the first channel, the second channel, and the third channel constitutes the main three channels. The fourth channel is the second flame candidate band. Specifically, the band of the first channel is located near 4.3 μm and serves as a conventional flame candidate channel; the band of the second channel is located near 5.0 μm and serves as a first background reference channel; the band of the third channel is located near 3.8 μm and serves as a second background reference channel; and the band of the fourth channel is located near 2.7 μm and serves as a second flame candidate channel.

[0016] As a preferred embodiment of the present invention, the main three-channel time-domain waveform is a timing input composed of the baseline-subtracted signals of the first, second, and third channels. The second candidate band auxiliary frequency domain information includes the second candidate band time-domain auxiliary coding and the second candidate band auxiliary frequency domain evidence. The second candidate band auxiliary frequency domain evidence is obtained from the energy relationship between the fourth channel and the reference channel in the target flame flashing frequency band. The second candidate band time-domain auxiliary coding is composed of the baseline-subtracted signal of the fourth channel and its difference characteristics relative to the reference channel.

[0017] As a preferred embodiment of the present invention, the robust background estimation and baseline subtraction steps include:

[0018] S21. Perform robust statistics on the sampled values ​​of each channel in the current processing window or historical window, and obtain the estimated window background value after removing high-end and low-end abnormal samples;

[0019] S22. The estimated window background value is weighted and updated with the historical background baseline to obtain the updated background baseline.

[0020] S23. Perform baseline subtraction on the current window sample value using the updated background baseline to obtain the baseline-subtracted signal;

[0021] The specific expressions for robust background estimation and baseline subtraction are as follows:

[0022] ;

[0023] ;

[0024] ;

[0025] In the formula, For the first The first channel in the The original window for each processing cycle, For robust statistical computation, As the background baseline, These are the baseline update coefficients.

[0026] As a preferred embodiment of the present invention, when alarm evidence increases or candidate flame status improves, the baseline update coefficient of the main flame channel or candidate channel is reduced. The value of .

[0027] As a preferred technical solution of the present invention, the lightweight time-frequency fusion model includes: a time-domain feature extraction branch, a second candidate band auxiliary coding branch, a time-frequency auxiliary attention enhancement module, and a fusion classification module. The specific steps for calculating the window-level flame probability include:

[0028] S41. The temporal feature extraction branch performs channel-time convolution on the main three-channel temporal waveforms and obtains the main temporal features through the multi-scale parallel convolution branch.

[0029] S42. The second candidate band auxiliary coding branch receives the baseline subtraction signal of the fourth channel and its difference features or reference constraint features relative to the reference channel, performs lightweight one-dimensional convolutional coding to obtain the dynamic embedding of the second candidate band, and splices or gates the dynamic embedding with the auxiliary frequency domain information of the second candidate band to obtain the candidate band embedding vector.

[0030] S43. The time-frequency assisted attention enhancement module generates attention gating weights based on the global description of the main time domain features, the main frequency domain enhancement features, and the candidate band embedding vectors, and recalibrates the main time domain features.

[0031] S44. The fusion classification module takes the recalibrated main time domain features, the main frequency domain enhanced features, and the candidate band embedding vector as inputs and outputs the window-level flame probability.

[0032] As a preferred embodiment of the present invention, the multi-scale parallel convolutional branches include short-scale convolutional branches, medium-scale convolutional branches, and long-scale convolutional branches arranged in parallel. Each convolutional branch is provided with residual connections, so that shallow local fluctuation features can be transmitted to deep layers. Specifically, the short-scale convolutional branches extract the high-frequency rapid flickering features of the flame, the medium-scale convolutional branches extract the medium-frequency continuous flickering features of the flame, and the long-scale convolutional branches extract the low-frequency slow-change features of the flame.

[0033] As a preferred embodiment of the present invention, the confidence mapping adopts a segmented mapping form, including:

[0034] When the probability of fused flame is less than the low confidence suppression threshold, the evidence value is mapped to zero;

[0035] When the probability of fused flame is greater than or equal to the low confidence suppression threshold and less than the high confidence retention threshold, the evidence value is linearly mapped or weighted according to the amount by which the probability of fused flame exceeds the low confidence suppression threshold.

[0036] When the probability of fused flame is greater than or equal to the high confidence retention threshold, the evidence value is mapped to the probability of fused flame itself;

[0037] The infrared evidence integral is the sum of evidence values ​​from the most recent several periods.

[0038] As a preferred embodiment of the present invention, the alarm state machine includes a normal state, an observation state, a candidate alarm state, a formal alarm state, and a recovery state. An alarm entry threshold and a recovery threshold are set, and the alarm entry threshold is greater than the recovery threshold to form a hysteresis characteristic. When the infrared evidence integral reaches the alarm entry threshold, the machine enters the formal alarm state and outputs a flame alarm result.

[0039] As a preferred embodiment of the present invention, it further includes: S8, an auxiliary step for ultraviolet post-processing:

[0040] S81. Acquire point-type ultraviolet signals;

[0041] S82. When the infrared evidence integral is in the candidate alarm range and the ultraviolet data is valid, the ultraviolet signal is used to improve the alarm confirmation speed, increase the candidate confirmation weight, or shorten the confirmation delay.

[0042] S83. When the infrared evidence integration is insufficient, even if there is a short-term strong pulse in the ultraviolet signal, a formal alarm will not be triggered separately.

[0043] S84. When ultraviolet data is abnormal or missing, the system degenerates into pure infrared criteria.

[0044] Compared with existing technologies, this invention provides a point flame detection method based on multi-infrared band multi-scale time-frequency fusion, which has the following beneficial effects:

[0045] This invention inputs the main three-channel time-domain waveforms and frequency-domain enhancement features into the same model, enabling the model to simultaneously utilize flame flicker morphology and the physical relationships of multiple infrared bands to achieve joint time-frequency identification. By using frequency-domain enhancement features in attention gating, feature recalibration is correlated with the energy relationship between flame candidate channels and reference channels, which helps suppress background disturbances. Parallel convolutional branches at short, medium, and long scales cover flicker changes at different time scales, improving adaptability to long distances, small fire sources, and different combustion states. Multi-phase window fusion and continuous evidence integration transform window-level probabilities into stable alarm states, reducing transient false alarms and missed detections at window boundaries. With its three-channel time-series input, low-dimensional frequency-domain enhancement features, and lightweight attention gating structure, it is suitable for embedded, low-computing-power deployment of point-type flame detectors. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0047] Figure 2 This is a schematic diagram of the lightweight time-frequency fusion attention convolution model of the present invention;

[0048] Figure 3 This is a flowchart of the continuous alarm integration process of the present invention. Detailed Implementation

[0049] 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.

[0050] Please see Figures 1-3 This invention provides a point flame detection method based on multi-infrared band multi-scale time-frequency fusion, comprising the following steps:

[0051] S1. Acquire timing sampling signals from multiple infrared channels, including three main channels and a second auxiliary candidate channel. The system acquires timing signals from four infrared channels.

[0052] S2. Perform robust background estimation and baseline subtraction on each channel to eliminate the effects of ambient temperature drift, slow changes in background heat sources, and DC bias of the sensor, and obtain the baseline-subtracted signal for each channel.

[0053] S3. Perform frequency domain transformation on the baseline-subtracted signal, calculate the channel energy in the target flame flashing frequency band, obtain the main frequency domain enhancement features based on the calculation results of the main three channels, and obtain the auxiliary frequency domain information of the second candidate band based on the calculation results of the second auxiliary candidate channel.

[0054] S4. Input the main three-channel time-domain waveform, the main frequency-domain enhancement features, and the second candidate band auxiliary frequency-domain information into the same lightweight time-frequency fusion model. The second candidate band auxiliary frequency-domain information includes the time-domain auxiliary coding and auxiliary frequency-domain evidence of the fourth channel. Calculate the window-level flame probability through the attention gating and fusion classification within the model.

[0055] S5. Perform multi-phase fusion of the flame probabilities of the main window and multiple offset windows to obtain the fused flame probability of the current cycle.

[0056] S6. Perform confidence mapping on the fused flame probability and add the mapped evidence value to the rolling evidence integration queue.

[0057] S7. Output the flame alarm result based on the infrared evidence points in the rolling evidence points queue and the alarm state machine.

[0058] In one embodiment of the present invention, the raw infrared acquisition includes four channels, and the bands and technical functions of each channel are defined as follows: The first channel is located around 4.3 μm and serves as a conventional flame candidate channel to characterize the main changes in flame radiation. The second channel is located around 5.0 μm and serves as a first background reference channel to characterize common background perturbations. The third channel is located around 3.8 μm and serves as a second background reference channel to characterize common background perturbations and reference variations. The fourth channel is located around 2.7 μm and serves as a second flame candidate channel to provide supplementary spectral evidence when the response in the conventional flame band is insufficient.

[0059] The main model uses three timing input channels consisting of 4.3μm, 5.0μm, and 3.8μm, and the 2.7μm channel serves as the second candidate flame band, which enters the model through the auxiliary coding branch of the second candidate band.

[0060] Let the current processing window length be... The baseline-subtracted signals of the four infrared channels are respectively The main three-channel time-domain input is represented as follows:

[0061] ;

[0062] in, This indicates a 4.3 μm conventional flame candidate channel. and Indicates two reference channels. This indicates a 2.7 μm second flame candidate channel.

[0063] The second flame candidate band auxiliary coding branch lightweightly encodes the baseline subtraction signal of the 2.7 μm channel, the difference relative to the reference channel, and the energy relationship of the target frequency band, forming a candidate band embedding vector. This embedding vector, together with the main three-channel convolutional features and frequency domain enhancement features, participates in attention gating and fusion classification, enabling the model to obtain supplementary judgment capabilities when the evidence of the conventional flame band is weak but the evidence of the second candidate band is obvious, and ensuring that the auxiliary band is always constrained by the main features and the relationship with the reference channel.

[0064] In one embodiment of the present invention, a dynamic background baseline is established for each infrared channel. When each processing window arrives, robust statistics are performed on the channel sample values ​​in that window or historical window to remove high-end and low-end abnormal samples, resulting in a window background estimate, which is then weighted and updated with the historical background baseline.

[0065] After baseline update, baseline subtraction is performed on the current window sample values ​​to obtain the AC flicker signal. This signal serves as the time-domain input to the model.

[0066] Let the first The first channel in the The original window for each processing cycle is robust statistical computing The background baseline is Then it can be expressed as:

[0067] ;

[0068] ;

[0069] .

[0070] in, This is the baseline update factor. When alarm evidence increases or candidate flame status improves, the baseline update factor for the flame candidate correlation channel is reduced. The value of is used to prevent the background baseline from absorbing the continuous flickering component of the real flame too quickly.

[0071] The effects of this processing include: reducing the impact of ambient temperature drift, changes in solar irradiance, and slow changes in background heat sources; reducing the interference of sensor DC bias on model input; and making the model more attentive to dynamic changes caused by flame flicker.

[0072] The baseline update rate can be linked to the alarm evidence status. As the candidate flame evidence continues to rise, the baseline update rate of the flame candidate channel is reduced to decrease the risk of the real flame duration being quickly absorbed into the background.

[0073] In one embodiment of the present invention, frequency domain transformation is performed on each infrared channel signal within the current window to calculate the channel energy within the target flame flicker frequency band. The target frequency band can be determined based on the sampling rate and window length, preferably covering the main energy range of low-frequency flame flicker.

[0074] Set up a channel The baseline subtraction signal is Its discrete frequency domain representation is:

[0075] ;

[0076] Let the set of frequency points corresponding to the target flame flashing frequency band be . Then the channel The energy within the target frequency band is defined as:

[0077] ;

[0078] For a 4.3 μm conventional flame candidate channel and two reference channels, let's denote... , , Forming a dominant frequency domain enhanced eigenvector:

[0079] ;

[0080] in, This is a stable term. The four components mentioned above represent the energy ratio of the flame candidate channel relative to the first reference channel, the energy ratio relative to the second reference channel, the advantage quantity relative to the two reference channels as a whole, and the common perturbation quantity of the reference channels, respectively.

[0081] For the 2.7 μm second flame candidate band, let's denote... This forms an auxiliary frequency domain evidence vector:

[0082] ;

[0083] Together with the 2.7 μm time-domain auxiliary coding, a candidate band embedding vector is formed.

[0084] The enhanced features in the main frequency domain and the evidence in the auxiliary frequency domain are not used as alarm thresholds on their own, but rather as physical constraint information of the model to participate in feature recalibration and classification. This preserves the physical relationship of the flame in the frequency domain while avoiding the insufficient expressive power of pure rule thresholds.

[0085] The frequency domain enhancement features are not used to construct feature engineering or output alarm results based on the similarity comparison of pre-stored features. Instead, these features describe the dynamic advantage of candidate flame bands relative to reference bands within the current window and serve as input for attention gating and fusion classification within the model. Thus, frequency domain information and temporal convolutional features are coupled within the same model, enabling the model to simultaneously consider flame flicker morphology and multi-band physical relationships.

[0086] This design avoids limiting flame detection to a few fixed frequency points, fixed thresholds, or pre-stored sample feature sets, which helps improve adaptability to different fire source distances, combustion states, and background environments.

[0087] In one embodiment of the present invention, the time-frequency fusion model includes three mutually constrained evidence streams. The first stream is the main time-domain evidence stream, which performs channel-time convolution on the baseline subtraction waveforms of the 4.3 μm, 5.0 μm, and 3.8 μm channels, and obtains the main time-domain features through parallel convolution branches of short-scale, medium-scale, and long-scale convolutions. The second stream is the second candidate band evidence stream, which performs auxiliary encoding on the 2.7 μm waveform and its changes relative to the reference channel to form candidate band embeddings. The third stream is the frequency-domain enhancement evidence stream, which calculates the energy advantage of the main candidate band and the second candidate band relative to the reference channel within the target flame scintillation frequency band.

[0088] The three evidence streams mentioned above are not independent rule-based decision paths, but are fused within the same lightweight model. Frequency-domain enhanced features and candidate band embeddings jointly participate in the generation of compressed-excitation gating weights, recalibrating multi-scale temporal features. The recalibrated temporal features then enter the fusion classification layer together with the frequency-domain enhanced features and auxiliary candidate embeddings, outputting a window-level flame probability. This structure makes frequency-domain physical relationships a constraint on feature selection within the model, rather than an externally imposed threshold judgment.

[0089] In one embodiment of the invention, the temporal input is a three-channel window waveform after baseline subtraction. The model first extracts local cross-channel differences and short-term fluctuation patterns through channel-time two-dimensional convolution.

[0090] Let the channel-time convolution front end be... Then the initial time-domain characteristics of the main three-channel input are:

[0091] ;

[0092] This 2D convolutional front end is used to capture: the relative changes between the flame candidate channel and the reference channel; the short-term flicker of the flame candidate channel; the background disturbances that may occur when all three channels are enhanced simultaneously; and the local morphology of continuous small fluctuations in a weak fire scene.

[0093] In one embodiment of the invention, the second candidate band auxiliary coding branch receives the baseline subtracted signal from the 2.7 μm channel and introduces its difference features or reference constraint features relative to the two reference channels. The difference features may include differential sequences. This branch employs a small amount of one-dimensional convolution, pooling, and nonlinear mapping to extract short-term abrupt changes, persistent weak fluctuations, and relative enhancement information after reference suppression in the second candidate band.

[0094] This branch outputs the candidate band time-domain embedding, and combines it with the auxiliary frequency-domain evidence vector. By splicing or gated fusion, a second candidate band embedding vector is formed. . Instead of outputting alarm results separately, it participates in subsequent attention gating and fusion classification.

[0095] The reference constraint sequence for the second candidate band can be represented as:

[0096] ;

[0097] Let the auxiliary coding branch be Candidate fusion mapping is ,but:

[0098] ;

[0099] .

[0100] With the above structure, the second candidate band retains independent spectral sensitivity while being constrained by the common perturbation of the reference channel, making it suitable for supplementing scenarios where the candidate band response is insufficient for specific combustion types or conventional flames.

[0101] In one embodiment of the invention, multiple convolutional branches with different temporal receptive fields are set after the two-dimensional convolutional front end. The different branches extract short-term rapid fluctuations, moderate continuous flickering, and weak fire variations on a longer time scale, respectively.

[0102] Residual connections are set in each branch to allow shallow local fluctuation features to be propagated to deeper layers, while reducing the training difficulty of the lightweight model. The outputs of each branch are pooled to form multi-scale temporal features.

[0103] Let the short-scale, medium-scale, and long-scale parallel convolution branches be respectively , , Then the three types of time-domain features can be expressed as:

[0104] ;

[0105] After splicing and pooling, the main temporal synthesized features are obtained:

[0106] ;

[0107] This structure is designed to adapt to differences in scintillation patterns caused by varying distances from the fire source, flame size, combustion state, and sampling window position.

[0108] In one embodiment of the invention, the model incorporates an attention enhancement module after convolutional feature extraction. This module takes a global description of the main temporal convolutional features as its primary input and introduces main frequency domain enhancement features. Second candidate band embedding vector As auxiliary gating information, feature weights are generated.

[0109] Specifically, global pooling is performed on the multi-scale convolutional features to obtain a global feature description; this description is then combined with... , The combined data is input into a lightweight gating unit to generate channel weights and branch weights; these weights are then used to recalibrate multi-scale features. The lightweight gating unit can employ a channel recalibration structure consisting of compressed description, nonlinear excitation, and weight write-back, enabling the fusion of cross-channel feature selection, frequency domain advantage information, and second candidate bands within the same weight generation process.

[0110] set up For global pooling, It is a non-linear activation function. for If the function is defined, then the gating weights can be expressed as:

[0111] ;

[0112] Recalibrate the main time-domain integrated features using gating weights:

[0113] ;

[0114] in, This indicates channel-wise or feature-wise multiplication.

[0115] The purpose of this attention enhancement method is as follows: when the flame candidate channel has a frequency domain advantage over the reference channel, increase the weights of relevant temporal features; when both reference channels are enhanced together and the flame candidate advantage is insufficient, reduce the weights of features related to background perturbation; when there is effective relative enhancement in the second candidate band, increase the weights of the backbone features consistent with the temporal embedding of the candidate band; and at different flame flicker scales, enhance the convolutional branches that better match the frequency domain features of the current window.

[0116] This module incorporates infrared frequency domain enhancement features and second candidate bands into the feature recalibration process, linking the attention weights to the physical signal relationship of point flame detection.

[0117] In one embodiment of the present invention, the attention-enhanced multi-scale temporal features are concatenated and pooled into a temporal composite feature. Subsequently, this temporal composite feature and the dominant frequency domain enhanced feature are... Auxiliary frequency domain evidence Second candidate band embedding The common input is the classification layer, and the output is the window-level flame probability.

[0118] Window-level fusion features can be represented as:

[0119] ;

[0120] in, This represents the fusion mapping layer. The window-level flame probability is:

[0121] ;

[0122] in, This represents the classification layer weight vector; Represents the weight vector transpose; This represents the category layer bias term.

[0123] In this structure, the frequency domain enhancement features play two roles: assisting in the generation of attention weights during the feature extraction stage; and participating in the probability output together with the time domain features and the second candidate band embedding during the classification stage.

[0124] This structure allows the model to utilize both flame flashing patterns and frequency domain relationships across multiple infrared bands, reducing the dependence of purely time-domain models on environmental noise.

[0125] In one embodiment of the invention, a second flame candidate band auxiliary encoding is used to enhance the model's coverage of specific combustion types or anomalous spectral response scenarios. This encoding uses the 2.7 μm channel as the candidate signal and the 5.0 μm and 3.8 μm channels as reference constraints to form auxiliary evidence in the time and frequency domains, respectively.

[0126] In the time domain, the system extracts the window waveform after baseline subtraction of the 2.7 μm channel and calculates its relative change sequence with respect to the reference channel. The auxiliary coding branch performs lightweight convolutional coding on the above sequence to obtain the dynamic embedding of the second candidate band.

[0127] In the frequency domain, the system calculates the energy advantage of the 2.7 μm channel relative to the two reference channels within the target flame scintillation frequency band, obtaining auxiliary frequency domain evidence. The dynamic embedding and the auxiliary frequency domain evidence together form the candidate band embedding vector, which participates in attention gating and final fusion classification.

[0128] This approach allows the 2.7 μm channel to participate in the same window inference as internal model evidence. It not only preserves the stable representation of the 4.3 μm conventional flame band by the main detection path, but also provides a supplement to the final confidence score when a valid response appears in the second candidate band.

[0129] The relationship between the second candidate band auxiliary coding and the main detection path can be expressed as follows:

[0130] ;

[0131] in, This represents the overall model, which includes the main temporal branch, multi-scale convolutional branch, second candidate band auxiliary coding branch, gating recalibration, and fusion classification. This expression indicates that the second candidate band participates in the unified probability output as internal evidence within the model, rather than forming an independent alarm result.

[0132] In this invention, the second candidate band auxiliary encoding serves as internal model-level supplementary evidence. Its purpose is to supplement spectral information in scenarios where the response of conventional flame candidate bands is insufficient for specific combustion types or situations. This supplementary evidence does not alter the structure of the main detection path based on the 4.3 μm flame candidate channel and two reference channels, nor does it independently generate an alarm conclusion. The effectiveness of the second candidate band needs to be constrained simultaneously by the differences in reference channels, supplementary frequency domain evidence, and main time domain characteristics, thereby preventing transient fluctuations in a single supplementary band from being directly amplified into alarm results.

[0133] In one embodiment of the present invention, in order to reduce the influence of window boundaries on the detection results, the system constructs a main window and several offset windows. Each window undergoes baseline subtraction, frequency domain enhancement feature calculation, and model inference to obtain the window flame probability.

[0134] The flame probability for the current period is obtained by fusing the probabilities of multiple windows. The fusing method can be determined based on window integrity, window time position, and probability magnitude. If multiple adjacent offset windows all support the presence of flames, the current period probability is more stable; if only an isolated window has a high probability, its contribution to subsequent alarm integration is reduced.

[0135] Assume the current period includes The phase window, the first The model output probability for each window is The corresponding fusion weight is And satisfy , Then the fusion probability can be expressed as:

[0136] ;

[0137] in, It can be determined by window integrity, window time position, consistency of adjacent windows and historical evidence status, so that the consistent high response of continuous windows can be given a higher contribution, while the short-term high response of isolated windows is reduced in contribution.

[0138] In one embodiment of the present invention, the current period flame probability output by the model is first subjected to segmented confidence mapping: low confidence probabilities are set to zero or reduced in weight; medium confidence probabilities are proportionally entered into the integration queue; and high confidence probabilities retain their main contribution.

[0139] The system maintains a rolling evidence queue, summing the evidence values ​​from the most recent several periods to obtain the infrared evidence integral. The alarm state machine includes normal, observation, candidate alarm, formal alarm, and recovery states. Different alarm entry and recovery thresholds create hysteresis, reducing alarm jitter.

[0140] This process transforms window-level model probabilities into device-level alarm evidence over continuous time, which can suppress single-point interference while retaining the ability to accumulate continuous weak fire signals.

[0141] Set the confidence mapping function to The evidence value for entering the points queue in the current period is:

[0142] ;

[0143] Segmented mapping can be used:

[0144] ;

[0145] in, The low-confidence suppression threshold. To retain the threshold for high confidence, Let be the intermediate confidence level mapping coefficients. Let the length of the integration queue be... The infrared evidence integral is:

[0146] ;

[0147] Alarm status can be represented as:

[0148] ;

[0149] in, To reach the alarm threshold, To recover the threshold, and This results in an alarm delay.

[0150] In one embodiment of the present invention, the point-type ultraviolet signal is used as a post-processing auxiliary signal, which is not input into the infrared model and does not trigger the formal flame alarm separately.

[0151] Ultraviolet (UV) auxiliary information includes cumulative UV intensity, cumulative valid UV events, and UV data validity. When the infrared evidence integral is within the candidate alarm range, UV auxiliary information can improve alarm confirmation speed, increase candidate confirmation weight, or shorten confirmation delay. When infrared evidence is insufficient, short-duration strong UV pulses do not trigger formal alarms independently. When UV data is abnormal or missing, the system degenerates into pure infrared criteria.

[0152] Let ultraviolet auxiliary evidence be... The effective indicator of ultraviolet data is The infrared candidate alarm range is Then, UV-assisted selection will only participate in candidate confirmation if the following condition is met:

[0153] ;

[0154] This condition is used to accelerate the confirmation of infrared candidate states or improve the candidate ranking; when At that time, ultraviolet auxiliary evidence does not trigger a formal alarm on its own.

[0155] In the complete embodiment of conventional flame detection of the present invention, the system acquires data from four infrared channels. 4.3μm, 5.0μm, and 3.8μm are selected as the inputs for the three-channel model.

[0156] Robust background estimation and baseline subtraction are performed on the three-channel data. A frequency domain transformation is applied to the windowed data, and frequency domain enhanced feature vectors are calculated.

[0157] The three-channel time-domain waveform is fed into the channel-time two-dimensional convolution front end. Scintillation features at different time scales are extracted through multi-scale residual convolution branches.

[0158] Auxiliary time-domain coding and auxiliary frequency-domain evidence calculation are performed on the second candidate band at 2.7μm to form the candidate band embedding.

[0159] The convolutional global description, frequency domain enhancement features, and candidate bands are embedded into the input attention gating unit to obtain feature weights.

[0160] The attention-enhanced temporal features, frequency-enhanced features, and candidate bands are embedded into a common input classification layer to obtain the window flame probability.

[0161] The probabilities of the main window and the offset window are fused. The fused probability is added to the continuous evidence integration queue, and the alarm status is output.

[0162] Based on the standard flame detection implementation, the system establishes a second candidate band auxiliary coding for the 2.7μm channel. Specifically, this includes: baseline subtraction of the 2.7μm channel; calculation of its temporal difference sequence relative to the 5.0μm and 3.8μm reference channels; calculation of its energy advantage relative to the reference channels within the target frequency band; and fusion of the temporal auxiliary coding and auxiliary frequency domain evidence into a candidate band embedding. This candidate band embedding participates in the model's internal attention gating and fusion classification to improve supplementary identification capabilities for specific flame types.

[0163] In the ultraviolet-assisted embodiment of the present invention, ultraviolet data forms an accumulation of intensity and a accumulation of valid events. When the infrared evidence score is in a candidate alarm state, if the ultraviolet-assisted information is valid, the alarm confirmation level is increased; if the ultraviolet is abnormal or the infrared evidence is insufficient, the ultraviolet does not change the formal alarm result.

[0164] The present invention also provides a point-type infrared flame detection device, comprising:

[0165] The signal acquisition module is used to acquire the timing sampling signals of multiple infrared channels;

[0166] The preprocessing module is used to perform robust background estimation and baseline subtraction on the time-series sampled signals of each infrared channel;

[0167] The frequency domain feature calculation module is used to perform frequency domain transformation on the baseline-subtracted signal, calculate the channel energy in the target flame flashing frequency band, and form the main frequency domain enhanced feature vector and the second candidate band auxiliary frequency domain information.

[0168] The time-frequency fusion model module includes a main time-domain feature extraction unit, a multi-scale residual convolution unit, a second candidate band auxiliary coding unit, an attention enhancement unit, and a fusion classification unit;

[0169] The multi-phase window fusion module is used to fuse the flame probabilities of the main window and multiple offset windows;

[0170] The evidence integration and alarm module is used to perform confidence mapping on the fused flame probability, add it to the rolling evidence integration queue, and output the flame alarm result based on the infrared evidence integration and alarm state machine.

[0171] Optionally, it also includes an ultraviolet post-processing auxiliary module for using ultraviolet signals to assist in confirming alarms within the infrared candidate alarm range.

[0172] In the time-frequency fusion model module: the main time-domain feature extraction unit performs channel-time two-dimensional convolution on the main three-channel time-domain waveforms; the multi-scale residual convolution unit includes parallel short-scale, medium-scale, and long-scale convolution branches, each branch having residual connections; the second candidate band auxiliary coding unit performs lightweight coding on the baseline-subtracted signal of the fourth channel and its difference sequence relative to the reference channel to form the second candidate band embedding vector; the attention enhancement unit combines the global description of the main time-domain convolution features with the main frequency domain enhancement feature vector and the second candidate band embedding vector to generate gating weights, and recalibrates the multi-scale time-domain features; the fusion classification unit receives the recalibrated time-domain features, the main frequency domain enhancement feature vector, the auxiliary frequency domain evidence, and the second candidate band embedding vector, and outputs the window-level flame probability.

[0173] The present invention also provides an electronic device, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the point flame detection method described in the present invention is implemented.

[0174] The present invention also provides a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the point flame detection method described in the present invention.

[0175] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A point-type flame detection method based on multi-infrared band multi-scale time-frequency fusion, characterized in that, Includes the following steps: S1. Acquire the timing sampling signals of multiple infrared channels, wherein the multiple infrared channels include the main three channels and the second flame candidate band; S2. Perform robust background estimation and baseline subtraction on the time-series sampled signals of each infrared channel to obtain the baseline-subtracted signals of each channel; S3. Perform frequency domain transformation on the baseline-subtracted signal and calculate the channel energy in the target flame flashing frequency band. The main frequency domain enhancement features are obtained based on the calculation results of the main three channels, and the auxiliary frequency domain information of the second candidate band is obtained based on the calculation results of the second flame candidate band. S4. Input the main three-channel time-domain waveform, the main frequency-domain enhancement features, and the auxiliary frequency-domain information of the second candidate band into the same lightweight time-frequency fusion model, and calculate the window-level flame probability through the attention gating and fusion classification within the model. S5. Perform multi-phase fusion on the flame probabilities of the main window and multiple offset windows to obtain the fused flame probability of the current period; S6. Perform confidence mapping on the fused flame probability and add the mapped evidence value to the rolling evidence integration queue; S7. Output the flame alarm result based on the infrared evidence points in the rolling evidence points queue and the alarm state machine.

2. The point flame detection method based on multi-infrared band multi-scale time-frequency fusion according to claim 1, characterized in that: The multi-channel infrared system includes a first channel, a second channel, a third channel, and a fourth channel. The combination of the first, second, and third channels constitutes the main three channels. The fourth channel is the second flame candidate band. Specifically, the first channel's band is located around 4.3 μm and serves as a conventional flame candidate channel; the second channel's band is located around 5.0 μm and serves as a first background reference channel; the third channel's band is located around 3.8 μm and serves as a second background reference channel; and the fourth channel's band is located around 2.7 μm and serves as a second flame candidate channel.

3. The point flame detection method based on multi-infrared band multi-scale time-frequency fusion according to claim 2, characterized in that: The main three-channel time-domain waveform is a timing input composed of the baseline-subtracted signals of the first, second, and third channels. The second candidate band auxiliary frequency domain information includes the second candidate band time-domain auxiliary coding and the second candidate band auxiliary frequency domain evidence. The second candidate band auxiliary frequency domain evidence is obtained from the energy relationship between the fourth channel and the reference channel in the target flame flashing frequency band. The second candidate band time-domain auxiliary coding is composed of the baseline-subtracted signal of the fourth channel and its difference characteristics relative to the reference channel.

4. The point flame detection method based on multi-infrared band multi-scale time-frequency fusion according to claim 1, characterized in that: The robust background estimation and baseline subtraction steps include: S21. Perform robust statistics on the sampled values ​​of each channel in the current processing window or historical window, and obtain the estimated window background value after removing high-end and low-end abnormal samples; S22. The estimated window background value is weighted and updated with the historical background baseline to obtain the updated background baseline. S23. Perform baseline subtraction on the current window sample value using the updated background baseline to obtain the baseline-subtracted signal; The specific expressions for robust background estimation and baseline subtraction are as follows: ； ； ； In the formula, For the first The first channel in the The original window for each processing cycle, For robust statistical calculation, As the background baseline, These are the baseline update coefficients.

5. The point flame detection method based on multi-infrared band multi-scale time-frequency fusion according to claim 4, characterized in that: When alarm evidence increases or candidate flame status improves, reduce the baseline update factor of the main flame channel or candidate channel. The value of .

6. The point flame detection method based on multi-infrared band multi-scale time-frequency fusion according to claim 1, characterized in that: The lightweight time-frequency fusion model includes: a time-domain feature extraction branch, a second candidate band auxiliary coding branch, a time-frequency auxiliary attention enhancement module, and a fusion classification module. The specific steps for calculating the window-level flame probability include: S41. The temporal feature extraction branch performs channel-time convolution on the main three-channel temporal waveforms and obtains the main temporal features through the multi-scale parallel convolution branch. S42. The second candidate band auxiliary coding branch receives the baseline subtraction signal of the fourth channel and its difference features or reference constraint features relative to the reference channel, performs lightweight one-dimensional convolutional coding to obtain the dynamic embedding of the second candidate band, and splices or gates the dynamic embedding with the auxiliary frequency domain information of the second candidate band to obtain the candidate band embedding vector. S43. The time-frequency assisted attention enhancement module generates attention gating weights based on the global description of the main time domain features, the main frequency domain enhancement features, and the candidate band embedding vectors, and recalibrates the main time domain features. S44. The fusion classification module takes the recalibrated main time domain features, the main frequency domain enhanced features, and the candidate band embedding vector as inputs and outputs the window-level flame probability.

7. The point flame detection method based on multi-infrared band multi-scale time-frequency fusion according to claim 6, characterized in that: The multi-scale parallel convolutional branches include short-scale convolutional branches, medium-scale convolutional branches, and long-scale convolutional branches set in parallel. Each convolutional branch is provided with residual connections, so that shallow local fluctuation features can be transmitted to deep layers. Specifically, the short-scale convolutional branches extract the high-frequency rapid flickering features of the flame, the medium-scale convolutional branches extract the medium-frequency continuous flickering features of the flame, and the long-scale convolutional branches extract the low-frequency slow-change features of the flame.

8. The point flame detection method based on multi-infrared band multi-scale time-frequency fusion according to claim 1, characterized in that: The confidence mapping adopts a piecewise mapping form, including: When the probability of fused flame is less than the low confidence suppression threshold, the evidence value is mapped to zero; When the probability of fused flame is greater than or equal to the low confidence suppression threshold and less than the high confidence retention threshold, the evidence value is linearly mapped or weighted according to the amount by which the probability of fused flame exceeds the low confidence suppression threshold. When the probability of fused flame is greater than or equal to the high confidence retention threshold, the evidence value is mapped to the probability of fused flame itself; The infrared evidence integral is the sum of evidence values ​​from the most recent several periods.

9. The point flame detection method based on multi-infrared band multi-scale time-frequency fusion according to claim 1, characterized in that: The alarm state machine includes a normal state, an observation state, a candidate alarm state, a formal alarm state, and a recovery state. It is equipped with an alarm entry threshold and a recovery threshold, and the alarm entry threshold is greater than the recovery threshold to form a hysteresis characteristic. When the infrared evidence integral reaches the alarm entry threshold, it enters the formal alarm state and outputs a flame alarm result.

10. The point flame detection method based on multi-infrared band multi-scale time-frequency fusion according to claim 1, characterized in that, Also includes: S8. Auxiliary steps for UV post-processing: S81. Acquire point-type ultraviolet signals; S82. When the infrared evidence integral is in the candidate alarm range and the ultraviolet data is valid, the ultraviolet signal is used to improve the alarm confirmation speed, increase the candidate confirmation weight, or shorten the confirmation delay. S83. When the infrared evidence integration is insufficient, even if there is a short-term strong pulse in the ultraviolet signal, a formal alarm will not be triggered separately. S84. When ultraviolet data is abnormal or missing, the system degenerates into pure infrared criteria.

Images