A dual-wavelength photoelectric sensing point-type fire alarm method and system

Abstract

The application relates to the technical field of fire safety monitoring, and discloses a point type fire alarm method and system based on dual-wavelength photoelectric sensing. The method comprises the following steps: obtaining original photoelectric data under irradiation of a multi-frequency band light source, and performing multi-scale decomposition and filter reconstruction to obtain pure photoelectric signals; performing blind source separation decoupling and cross interference on the pure photoelectric signals to obtain independent scattering signals; extracting envelopes and performing frequency domain analysis to obtain wavelength response characteristics representing differences in dual-wavelength responses; performing response state analysis to obtain a particle concentration level, and combining with environmental fluctuation verification to generate an alarm instruction. The method can solve the problem that weak fire characteristic signals are difficult to accurately separate under strong background noise and multi-source cross interference, leading to early fire false alarms and omissions.

Description

Technical Field

[0001] This invention relates to the field of fire safety monitoring technology, and in particular to a point-type fire alarm method and system using dual-wavelength photoelectric sensing. Background Technology

[0002] In the field of industrial fire safety monitoring, distributed point-type photoelectric smoke detection systems based on wireless networking are widely used for early fire warning in large buildings and industrial facilities. This is particularly relevant in industrial settings involving flammable and explosive materials, such as petrochemical and chemical fiber manufacturing plants, including purified terephthalic acid (PTA) production facilities and polyester fiber production workshops. These locations contain large amounts of combustible dust and volatile organic compounds, and early fires are often accompanied by the release of ultrafine aerosol particles, placing extremely high demands on the sensitivity and anti-interference capabilities of fire detection.

[0003] In existing technologies, detection equipment uses multi-band light sources to alternately illuminate suspended particles in space, utilizing the differences in the scattering characteristics of particles at different wavelengths to obtain photoelectric signals. The presence of fire risk is then determined by analyzing the changing trends of these scattered signals. However, in typical chemical environments such as PTA production plants and polyester workshops, the scattered signals from different wavelengths overlap, creating cross-interference. Simultaneously, the amplitude of background noise, including ambient light fluctuations, dust and water vapor scattering, and circuit thermal noise, is often much greater than the weak scattered signals generated by fire particles. Strong background noise and cross-interference from multiple sources prevent detectors from accurately separating the true scattering characteristics of ultrafine particles from the aliased signals, and further hinders the extraction of response patterns at specific wavelengths for subsequent analysis. Furthermore, ultrafine particles such as dust and fumes in normal industrial environments highly overlap with those generated in the early stages of a fire in terms of particle size and concentration. Relying solely on concentration thresholds can easily misinterpret normal environmental fluctuations as fire hazards. Existing signal preprocessing and matching methods are ill-suited to handle such complex signal environments.

[0004] In summary, existing technologies suffer from the problem of difficulty in accurately separating weak fire characteristic signals under strong background noise and multi-source cross-interference, leading to false alarms and missed alarms in early fires. Summary of the Invention

[0005] This invention provides a point-type fire alarm method and system with dual-wavelength photoelectric sensing to solve the problem that weak fire characteristic signals are difficult to separate accurately under strong background noise and multi-source cross interference, leading to false alarms and missed alarms in early fires.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a point-type fire alarm method based on dual-wavelength photoelectric sensing, comprising: The raw photoelectric data of alternating illumination of space by multiple frequency band light sources is acquired, and the raw photoelectric data is decomposed and filtered to reconstruct a pure photoelectric signal. The raw photoelectric data contains cross-interference of scattered signals between light sources of different wavelengths. The pure photoelectric signal is framed and windowed to construct a multi-dimensional observation matrix. Blind source separation is performed on the multi-dimensional observation matrix to obtain a separation matrix. The cross interference is decoupled based on the separation matrix to obtain the independent scattering signal corresponding to the ultrafine particles. The envelope of the independent scattering signal is extracted and differential operation is performed to obtain the photocurrent change. The photocurrent change is then subjected to frequency domain analysis and dimensionality reduction to obtain wavelength response characteristics that characterize the difference between the two wavelength responses. The wavelength response characteristics are analyzed for response status and classified to obtain particle concentration levels. Historical environmental data corresponding to the particle concentration levels are extracted, and meteorological elements are separated from the historical environmental data to obtain temperature and humidity sequences. Fluctuation analysis is performed on the temperature and humidity sequences to obtain environmental fluctuation benchmark values. If the particle concentration level is greater than a preset concentration threshold and the environmental fluctuation benchmark value is less than a preset benchmark threshold, an alarm command is generated.

[0007] Secondly, the present invention provides a point-type fire alarm system with dual-wavelength photoelectric sensing, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the point-type fire alarm method with dual-wavelength photoelectric sensing as described above.

[0008] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the point-type fire alarm method with dual-wavelength photoelectric sensing as described above.

[0009] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention avoids the limitation of insufficient signal-to-noise ratio of single-wavelength detection in strong background noise environment by collecting raw photoelectric data under alternating illumination of multiple frequency band light sources and performing multi-scale decomposition and filtering reconstruction, thus ensuring the accuracy of extracting scattering features from aliased signals. Furthermore, based on the pure photoelectric signal, a multi-dimensional observation matrix is ​​constructed by framing and windowing, and the dual-wavelength cross-interference is decoupled by blind source separation. This enables the accurate extraction of independent scattering signals corresponding to ultrafine particles, thus solving the problem that existing technologies cannot separate weak fire feature signals from strong background noise and multi-source cross-interference.

[0010] (2) This invention extracts the envelope of the independent scattering signal and performs differential operation to obtain the photocurrent change. Then, it performs frequency domain analysis and local neighborhood preservation dimensionality reduction on the photocurrent change to obtain wavelength response features that characterize the difference between the two wavelength responses. This can accurately distinguish the differentiated modes of ultrafine particle scattering response under infrared and blue light bands. Then, it uses support vector machine to classify the response state to obtain the particle concentration level and identify the combustion stage. This can realize the dynamic distinction between the early smoldering and open flame sudden stages, effectively improve the coarseness of single concentration threshold judgment, and enhance the graded response capability of fire alarm.

[0011] (3) This invention obtains the environmental fluctuation benchmark value by analyzing the temperature and humidity sequence fluctuations of the correlation particle concentration level and historical environmental data, which can determine the stability of the current environmental state in real time. Furthermore, by combining the particle concentration level with the preset concentration threshold and the environmental fluctuation benchmark value with the preset benchmark threshold, this solution can be deployed in the drying workshop of the PTA production unit. This can effectively distinguish between normal dust from PTA dust and ultrafine carbon black particles released in the early stage of smoldering, avoid false alarms caused by dust concentration fluctuations, and ensure the reliable capture of real fire signals, thereby reducing the probability of false alarms and missed alarms caused by a single threshold mechanism. Attached Figure Description

[0012] Figure 1 A flowchart illustrating a point-type fire alarm method using dual-wavelength photoelectric sensing provided in an embodiment of the present invention; Figure 2 A schematic diagram of the multidimensional observation matrix provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the SVM response state classification provided in an embodiment of the present invention. Detailed Implementation

[0013] 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.

[0014] Reference Figure 1 The first embodiment of the present invention provides a point-type fire alarm method based on dual-wavelength photoelectric sensing, comprising the following steps: S11, acquire the original photoelectric data when the space is alternately illuminated by multiple frequency band light sources, perform multi-scale decomposition and filtering reconstruction on the original photoelectric data to obtain a pure photoelectric signal, wherein the original photoelectric data contains cross interference of scattered signals between light sources of different wavelengths; S12, the pure photoelectric signal is framed and windowed to construct a multi-dimensional observation matrix. Blind source separation is performed on the multi-dimensional observation matrix to obtain a separation matrix. The cross interference is decoupled according to the separation matrix to obtain the independent scattering signal corresponding to the ultrafine particles. S13, extract the envelope of the independent scattering signal and perform differential operation to obtain the photocurrent change. Perform frequency domain analysis and dimension reduction on the photocurrent change to obtain wavelength response characteristics that characterize the difference between the two wavelength responses. S14, perform response state analysis and feature classification on the wavelength response characteristics to obtain the particle concentration level, extract historical environmental data corresponding to the particle concentration level, separate meteorological elements from the historical environmental data to obtain the temperature and humidity sequence, and perform fluctuation analysis on the temperature and humidity sequence to obtain the environmental fluctuation benchmark value. If the particle concentration level is greater than the preset concentration threshold and the environmental fluctuation benchmark value is less than the preset benchmark threshold, then generate an alarm command.

[0015] In step S11, raw photoelectric data of alternating illumination of space by multiple frequency band light sources is acquired. The raw photoelectric data is then subjected to multi-scale decomposition and filtering reconstruction to obtain a clean photoelectric signal. The raw photoelectric data contains cross-interference of scattered signals between different wavelength light sources, including: Acquire raw photoelectric data when space is alternately illuminated by multiple frequency band light sources, wherein the raw photoelectric data includes cross-interference of scattered signals between light sources of different wavelengths; The original photoelectric data is decomposed into multi-scale components. The multi-scale components are subjected to noise removal and smoothing to obtain the denoised components. The denoised components are reconstructed to obtain a clean photoelectric signal.

[0016] It should be noted that the light source module in the detection device alternately emits first and second wavelengths of light according to a preset time sequence, alternately irradiating suspended particles in space. Specifically, the light source alternation period is set to 100ms, with the first 50ms emitting infrared light at a wavelength of 870nm and the last 50ms emitting blue light at a wavelength of 405nm. Infrared light has high scattering sensitivity for large dust particles, while blue light has high scattering sensitivity for ultrafine aerosol particles.

[0017] The alternation period is set based on the Mie scattering theory, which states that large dust particles with a diameter greater than 1 μm have higher scattering efficiency in the near-infrared band, while ultrafine particles with a diameter less than 1 μm have higher scattering efficiency in the blue-violet band. The selection of 870 nm infrared light and 405 nm blue light can cover the sensitive scattering bands of large dust particles and ultrafine aerosols, respectively. The 100 ms alternation period balances real-time detection with signal integrity, and the 50 ms single-wavelength illumination time is sufficient for the photodiode to collect approximately 51 sampling points at a sampling rate of 1024 Hz, meeting the statistical characterization requirements for the scattering signal at a single wavelength.

[0018] The photodiode of the detection device continuously receives light signals scattered by suspended particles in space, converts the light intensity into a continuous analog voltage signal, and samples it through an analog-to-digital converter. The sampling frequency is set to 1024Hz. The sampling frequency is set based on the fact that the detection frequency range is 10Hz to 50Hz. According to the Nyquist sampling theorem, the sampling frequency should be greater than 100Hz. Choosing 1024Hz provides an oversampling rate more than 20 times higher than the highest frequency, effectively avoiding spectral aliasing. Furthermore, 1024 is an integer power of 2, facilitating alignment in the subsequent butterfly operation of the Fast Fourier Transform. The raw photoelectric data is obtained after sampling.

[0019] The raw photoelectric data contains scattering response components corresponding to both infrared and blue light wavelengths. Since the residence time of suspended particles within the detection cavity can reach hundreds of milliseconds, far exceeding the 100-millisecond light cycle, the scattering signal generated by the same particle group during the infrared illumination period has not fully decayed before entering the blue light illumination period. Simultaneously, the photodiode's response decay exhibits a μs-level tail. Therefore, in a continuous time series, the infrared and blue light scattering signals intermingle and overlap on the time axis, creating cross-interference.

[0020] It is worth noting that the preset timing is controlled by a hardware timer inside the light source driving module. Data within 10μs of the instant the light source switches is marked as transient transition data and discarded in subsequent segmented processing. The 10μs setting is based on the fact that the response decay time constant of the silicon photodiode used in this embodiment is 2 to 3μs, and the signal amplitude decays to less than 5% of the peak value within 10μs. The 10μs discard window can cover the transient response tail caused by the light source switching and avoid transient interference introducing analysis errors.

[0021] For example, multi-scale decomposition is performed on the raw photoelectric data. The raw photoelectric data is input and processed using a discrete wavelet transform algorithm. The db6 wavelet is selected as the wavelet basis function, and the decomposition level is set to 3 levels. The discrete wavelet transform decomposes the signal layer by layer through a set of high-pass and low-pass filters. After each level of decomposition, the low-frequency components are further decomposed, ultimately outputting multi-scale components containing detailed and approximate information at different scales.

[0022] The db6 wavelet basis was chosen because its vanishing moment of 6 provides sufficient regularity to smoothly represent transient fluctuations in photoelectric scattering signals, while its compact support length of 11 sampling points results in moderate computational complexity. Comparative tests were conducted on candidate wavelet bases from db2 to db10, using the signal reconstruction signal-to-noise ratio (SNR) as the evaluation metric. db6 achieved a reconstruction SNR of 22.3 dB, superior to db4's 19.8 dB, and showed no significant difference from db8's 22.5 dB while incurring lower computational cost. Therefore, db6 was selected.

[0023] The decision to decompose the components into three layers is based on the following: the first layer, the detail component, primarily captures circuit white noise; the second layer captures rapid fluctuations caused by airflow disturbances; and the third layer and the approximate component retain slowly changing trend information such as ambient light drift. After three layers of decomposition, low-frequency ambient light drift is concentrated in the approximate component, while fluctuations caused by dust and smoke scattering are distributed in the detail component. The frequency ranges of each layer are independent and fully covered, and further increasing the number of decomposition layers will not introduce new separable frequency components.

[0024] It should be noted that noise removal and smoothing are performed on the multi-scale components. For the high-frequency fluctuation components in the multi-scale components (i.e., the first-level detail component D1 and the second-level detail component D2), their energy amplitude is evaluated to see if it exceeds a preset high-frequency energy threshold. Among them, D1 mainly captures circuit white noise, and D2 mainly captures rapid fluctuations caused by airflow disturbances. Both are noise and interference components that need to be removed or suppressed.

[0025] The preset high-frequency energy threshold is set using a statistical calibration method. Specifically, the detector operates continuously for 24 hours in clean air, collecting background signals in a particulate-free environment. The mean μ and standard deviation σ of the high-frequency component energy are calculated, and μ + 3σ is taken as the threshold. Under the photoelectric conversion circuit parameters of this embodiment (transimpedance amplification gain is...), The photodiode dark current is 2nA, and the statistical calibration value is approximately 150mV. When the energy amplitude of the initial high-frequency component is detected to be significantly greater than this threshold, it is determined that the signal contains background noise spikes caused by external electromagnetic interference or severe airflow disturbances within the detection cavity. In actual engineering deployments, this threshold should be recalibrated according to the noise floor of the specific hardware circuit and is not a universally applicable value as described in this embodiment.

[0026] At this point, the high-frequency fluctuation component is used as input, and an adaptive median filtering algorithm is employed for processing. The adaptive median filtering algorithm dynamically adjusts the filtering window size within a range of 3 to 7 sampling points. The trigger condition for window expansion is whether the current window center point is the maximum or minimum value within the window. If so, and the point remains an extreme value after filtering, it is determined that the current window fails to cover the noise region, and the window is expanded by 2 sampling points. If the center point is not an extreme value within the window, it is determined to be in the signal edge region, and the window is shrunk by 2 sampling points to preserve details. After each adjustment, the median filtering is re-executed, outputting a clean high-frequency component after removing noise spikes.

[0027] The lower limit of the window range is set at 3 sampling points because 3 sampling points is the smallest filtering window that can effectively distinguish a single noise sampling point from its neighboring normal signal. If the window is too small, it will be impossible to distinguish between noise and signal. The upper limit of 7 sampling points is set because 7 sampling points correspond to a time span of approximately 6.8ms. If noise spikes persist within a larger window range, the fluctuation is more likely to be a real airflow disturbance rather than a random noise spike, and the window should not be expanded further in this case.

[0028] Specifically, the low-frequency trend components in the multi-scale components (i.e., the third-layer detail component D3 and the approximate component A3) mainly characterize the slow accumulation process of the base concentration of smoke in space and the temperature drift phenomenon of ambient light sources. D3 retains the slowly changing trend information such as ambient background light drift, while A3 retains the overall baseline profile of the signal. To accurately extract the baseline trend, a moving average smoothing algorithm is used to process the low-frequency trend components. The sliding window length is set to 5 sampling periods. After inputting the low-frequency trend components, the arithmetic mean of 5 consecutive sampling points within the window is used to replace the value at the center point of the window. The window is slid-slid point by point to complete the smoothing process, outputting a pure low-frequency component free of minor glitches. The sliding window length is set based on the fact that 5 sampling periods correspond to a time span of approximately 4.88 ms. This time scale can effectively smooth random fluctuations caused by circuit thermal noise without smoothing out the true trend changes caused by variations in ultrafine particle concentration.

[0029] For example, signal reconstruction is performed on the pure high-frequency and pure low-frequency components. The pure high-frequency and pure low-frequency components are input, and the signal is reconstructed using an inverse wavelet transform algorithm. The inverse wavelet transform uses the same db6 wavelet basis and 3-layer structure as the decomposition stage, reconstructing layer by layer in order from low resolution to high resolution, ultimately outputting a pure photoelectric signal. Background interference such as ambient light drift, circuit thermal noise, and airflow disturbances are effectively suppressed in the pure photoelectric signal, while the scattering characteristics of ultrafine particles under illumination from different wavelength light sources are preserved, providing a high-quality input signal for subsequent blind source separation.

[0030] In step S12, the pure photoelectric signal is framed and windowed to construct a multi-dimensional observation matrix. Blind source separation is performed on the multi-dimensional observation matrix to obtain a separation matrix. The cross-interference is decoupled based on the separation matrix to obtain the independent scattering signal corresponding to the ultrafine particles.

[0031] The pure photoelectric signal is framed and windowed to construct a multi-dimensional observation matrix.

[0032] Specifically, the pure photoelectric signal is framed and windowed to construct a multi-dimensional observation matrix. Since the light source alternates illumination with a period of 100ms, the relative proportions of infrared and blue light scattering contributions within the time window spanned by each frame of data differ between frames due to the dynamic migration of particulate matter. Different time frames can be regarded as observations of the same set of physical source signals under different mixing ratios, thus satisfying the basic requirements of independent component analysis for multi-channel observation.

[0033] A clean photoelectric signal is input and subjected to frame-by-frame windowing using a Hamming window function. The frame length is set to 256 sampling points, and the frame shift is set to 128 sampling points. The Hamming window is chosen because it achieves a better balance between the main lobe width and side lobe attenuation than the rectangular window and Hanning window. The rectangular window attenuates the side lobes to only -13dB, resulting in severe inter-frame spectral leakage; the Hanning window attenuates the side lobes to -31dB, but has a wider main lobe width, leading to a decrease in frequency resolution; the Hamming window attenuates the side lobes to -41dB, and has a main lobe width comparable to the Hanning window, thus maintaining sufficient frequency resolution while suppressing spectral leakage. This makes it suitable for scenarios where the frequency components of infrared and blue light scattering signals are similar in this embodiment.

[0034] The 256-sampling-point frame length is based on the fact that, at a sampling rate of 1024Hz, 256 points correspond to a 250ms time window. This time length is sufficient to cover one complete alternation cycle of infrared and blue light, ensuring that each frame of data simultaneously contains response information of both wavelengths. The 128-point frame shift corresponds to a 50% overlap rate, which avoids information loss between frames and ensures a smooth transition of data between adjacent frames.

[0035] By continuously capturing and arranging data from each frame, a multi-dimensional observation matrix is ​​output. Each ICA process uses 64 consecutive frames of data to form a 64-row observation matrix. This number of frames is based on the fact that 64 frames cover a data span of approximately 16 seconds. Within this time window, the change in particulate matter concentration within the detection chamber can be approximated as steady-state, while the mixing ratio of the source signal exhibits sufficient inter-frame differences due to the dynamic migration of particles. Under the condition of using an embedded processor with a main frequency of at least 200MHz and optimizing the core iteration of the FastICA algorithm through iterative expansion, the ICA calculation time for a single 64-row observation matrix is ​​on the order of approximately 200ms. Combined with a frame period of 250ms and a pipelined parallel scheduling strategy, it can meet real-time processing requirements. The actual calculation time depends on the specific processor model and compilation optimization level, and is not a general limitation based on the values ​​in this embodiment.

[0036] In one implementation, blind source separation is performed on the multidimensional observation matrix to obtain a separation matrix, including: The multidimensional observation matrix is ​​subjected to decorrelation processing to obtain the decorrelation matrix; The variance of the decorrelation matrix is ​​normalized to obtain the orthogonal whitening matrix; The separation matrix is ​​obtained by iteratively solving the orthogonal whitening matrix by maximizing its non-Gaussianity.

[0037] For example, decorrelation processing is performed on a multidimensional observation matrix. The multidimensional observation matrix is ​​input, and the covariance matrix is ​​calculated using principal component analysis (PCA). Eigenvalues ​​and eigenvectors of the covariance matrix are obtained through eigenvalue decomposition. Components with eigenvalues ​​less than 0.1% of the largest eigenvalue are removed, and the eigenvectors corresponding to the remaining eigenvalues ​​are retained to form the decorrelation matrix. After decorrelation processing, the cross-covariance of data in each dimension is zero, providing conditions for subsequent variance normalization. The threshold is set based on testing the number of components retained and the subsequent FastICA separation performance at three thresholds: 0.01%, 0.1%, and 1%, in the PCA decomposition of a 64×256-dimensional observation matrix. At the 0.01% threshold, 31 principal components are retained, and the signal-to-interference ratio (SIR) after ICA separation is 28.3 dB; at the 0.1% threshold, 24 principal components are retained, and the SIR remains at 28.3 dB without substantial decrease; at the 1% threshold, only 11 principal components are retained, and the SIR drops to 22.1 dB. The 0.1% threshold reduces the number of components by 22.6% without compromising separation accuracy, so it was selected as the decorrelation cutoff threshold.

[0038] It's important to note that variance normalization is performed on the decorrelation matrix. The decorrelation matrix is ​​input, and each row is divided by its corresponding standard deviation, normalizing the variance of each dimension to 1. The output after variance normalization is an orthogonal whitening matrix. The whitening matrix is ​​orthogonal across all dimensions with equal variances, eliminating the influence of data scale differences on the convergence direction of the subsequent blind source separation algorithm. Whitening is a crucial preprocessing step in the entire blind source separation process. Unwhitened data, due to variance differences across different dimensions, can cause the independent component analysis algorithm to converge to local optima instead of the global optimum.

[0039] Specifically, the orthogonal whitening matrix is ​​solved iteratively by maximizing its non-Gaussianity. Given the orthogonal whitening matrix, a fast independent component analysis algorithm is used for blind source separation. This algorithm uses negative entropy as a measure of non-Gaussianity and continuously searches for the weight vector that maximizes the non-Gaussianity of the projection direction using Newton's iteration method.

[0040] The algorithm is set to a maximum of 500 iterations and a convergence accuracy threshold of 10. -6 The convergence precision is set based on the fact that the element values ​​of the weight vector are normalized to the range of 0 to 1, 10 -6 The convergence accuracy requirement is that the weight change between adjacent iterations is less than 10. -6 At this level of precision, each row of the separation matrix sufficiently approximates the true independent source directions. Comparative tests show that further tightening the convergence precision to 10... -7 At that time, the improvement in separation performance, measured by the signal-to-interference ratio (SIR), was less than 0.5 dB (the SIR on the test signal in this embodiment was approximately on the order of 28 dB), while the number of iterations increased significantly, therefore 10 -6 To achieve a balance between separation accuracy and computational efficiency, the above SIR values ​​are the measured results on standard test signals in this embodiment; actual values ​​may vary depending on signal characteristics. The algorithm is considered convergent when the change in the weight vector between two consecutive iterations is less than the convergence accuracy threshold. The 500-iteration upper limit is based on the fact that, for the data scale of 1024Hz sampling rate and 256-point frame length in this embodiment, a stable solution is typically reached within 200 to 300 iterations. The 500-iteration upper limit serves as a safety margin to prevent the algorithm from looping infinitely on individual abnormal data frames. The algorithm outputs a separation matrix after convergence.

[0041] After obtaining the independent source signals for each frame through ICA decomposition, to address the inherent ambiguity of ICA—that is, the order of the same physical source may differ in the ICA decomposition results of different frames—it is necessary to first align the independent source signals of each frame. The specific method for alignment is as follows: using the amplitude variance values ​​of each independent source signal in the first frame as a reference, the independent source signals in each subsequent frame are sorted in ascending order according to their amplitude variance values, ensuring that independent source signals with the same physical source in each frame are arranged in a consistent order. After sorting, a weighted smoothing stitching method is used to overlap the parts of adjacent frames using the Hamming window coefficient as the weighting function. This eliminates amplitude jumps at frame boundaries, restoring the independent scattered signals in the continuous time domain. After alignment, independent source signals corresponding to the same physical source in different frames are continuously connected on the time axis, ensuring the physical consistency of subsequent variance determination.

[0042] In one implementation, the cross-interference is decoupled according to the separation matrix to obtain the independent scattering signal corresponding to the ultrafine particles, including: The cross-interference is decoupled based on the separation matrix to obtain multiple independent source signals; The variance of the independent source signals is calculated to obtain the amplitude variance value; If the amplitude variance value is less than the preset variance threshold, the independent source signal is determined to be an independent scattering signal corresponding to the ultrafine particles.

[0043] It should be noted that cross-interference is decoupled based on the separation matrix. The input separation matrix is ​​applied to an orthogonal whitening matrix, and a linear transformation is performed on the mixed signal using each row of the separation matrix as a projection direction, outputting multiple independent source signals. Each independent source signal corresponds to the scattering contribution from a specific physical source within the detection cavity, such as water vapor interference, large dust particles, or ultrafine particles.

[0044] A preset variance threshold is set, which is based on calibration experiments of known ultrafine particle standard samples. Ultrafine carbon black particle aerosols with particle sizes ranging from 50nm to 500nm are introduced into the standard test chamber, and 100 sets of scattering signals are collected. The amplitude variance is calculated, and the mean variance is statistically obtained. The mean plus 1.5 times the standard deviation is taken as the judgment threshold. Under the signal amplitude normalization conditions of this embodiment (the amplitude of independent source signals is normalized to the 0~1V range according to the photoelectric conversion voltage), this statistical calibration value is approximately 0.05V². When the amplitude variance value of a detected independent source signal is less than this threshold, the independent source signal is determined to be an independent scattering signal corresponding to ultrafine particles. It should be noted that the actual value of the amplitude variance threshold depends on the amplitude normalization reference of the independent source signal; in engineering deployment, the measured value under the calibration conditions of this equipment should be used as the standard.

[0045] like Figure 2As shown in the figure, the aliasing distribution of photoelectric signals in each frame under alternating illumination by dual-wavelength light sources is displayed in the time domain. The heatmap clearly shows that before blind source separation, the infrared and blue light scattering signals are distributed in an alternating red and blue stripe pattern in the multidimensional observation matrix (multiple sets of alternating positive and negative extreme values ​​appear in the horizontal axis time sampling point range of 0-30). The different wavelength responses show a gradual difference in the mixing ratio in the frame number direction (vertical axis 0-250) due to the dynamic migration of particles. This result intuitively proves that there is indeed cross-interference between different wavelength light sources in the original photoelectric data, and that the observation data of each frame can be regarded as a linear superposition of the same set of physical source signals under different mixing ratios, providing a solid data foundation and separability basis for subsequent blind source separation using independent component analysis.

[0046] In step S13, the envelope of the independent scattering signal is extracted and differential operation is performed to obtain the photocurrent change. The photocurrent change is then subjected to frequency domain analysis and dimensionality reduction to obtain wavelength response characteristics that characterize the difference between the two wavelength responses.

[0047] In one implementation, the envelope of the independent scattering signal is extracted and differentially analyzed to obtain the photocurrent change, including: Extreme points are extracted from the independent scattering signals to obtain a time-axis extreme value sequence; The fluctuation amplitude is obtained by fitting the envelope of the time axis extreme value sequence; The change in photocurrent is obtained by performing a differential operation on the fluctuation amplitude.

[0048] Extreme points are extracted from independent scattering signals. The independent scattering signals are input, and a local optimization algorithm is used to perform sliding window detection on the time axis, with the sliding window length set to 5 sampling points. The window length of 5 is based on the fact that at a sampling rate of 1024Hz, 5 sampling points correspond to approximately 4.88ms. Compared to the characteristic time scale of photoelectric signal fluctuations caused by ultrafine particle concentration fluctuations (approximately 10ms under the combined effects of airflow disturbance and Brownian motion), this time scale can stably detect 1 to 2 extreme points within a single cycle. If the window is too small, noise spikes are easily misjudged as extreme values; if the window is too large, the true extreme values ​​within the cycle may be missed. At each window position, the value of the center point of the window is compared one by one with the values ​​of the other 4 sampling points within the window. If the center point is the maximum value within the window, it is marked as a local maximum; if it is the minimum value, it is marked as a local minimum. The sliding window traverses the entire time series with a single sampling point step, outputting a time axis extreme value sequence composed of all local maxima and local minima arranged in chronological order.

[0049] For example, envelope fitting is performed on the time axis extreme value sequence. The input time axis extreme value sequence is processed using a cubic spline interpolation algorithm for both the maximum and minimum value sequences. The choice of cubic spline interpolation is based on the following: linear interpolation is not differentiable at extreme points, resulting in a sharp angle in the generated envelope at inflection points, failing to accurately reflect the continuous and gradual change in the intensity of the scattered signal; quadratic splines only guarantee the continuity of the first derivative, leading to unnatural curvature jumps in the envelope near inflection points; cubic splines simultaneously guarantee the continuity of both the first and second derivatives, generating a geometrically smoother envelope that appears most natural to the eye. Furthermore, the cubic polynomial is the lowest-order polynomial satisfying second-order smoothness, making the computational and storage overhead acceptable. This algorithm constructs a piecewise cubic polynomial, generating a smooth curve between adjacent extreme points, and requires the first and second derivatives of adjacent piecewise polynomials to be continuous at the connection points. Natural boundary conditions are applied at the endpoints other than the first and last extreme points, i.e., the second derivative at the endpoints is set to zero. After fitting the upper and lower envelopes separately, the amplitude difference between the upper and lower envelopes at the same moment is calculated, and the fluctuation amplitude, reflecting the overall fluctuation state of the signal, is output. The fluctuation amplitude directly corresponds to the range of scattering intensity variation of ultrafine particles under light irradiation.

[0050] Specifically, the fluctuation amplitude is differentially calculated. The fluctuation amplitude is input, and a first-order forward differencing algorithm is used to calculate the difference between fluctuation amplitude values ​​at adjacent time points by dividing by the time step, outputting a sequence of photocurrent changes. The photocurrent changes are quantized in μA, characterizing the transient response rate of the scattered photocurrent caused by minute fluctuations in ultrafine particle concentration. It should be noted that positive values ​​in the photocurrent change sequence indicate enhanced scattering, while negative values ​​indicate weakened scattering; the absolute value reflects the speed of the response.

[0051] In one implementation, frequency domain analysis and dimensionality reduction are performed on the photocurrent change to obtain wavelength response characteristics characterizing the difference in dual-wavelength responses, including: If the absolute value of the change in photocurrent is greater than a preset change threshold, then the time domain segment of the change in photocurrent is extracted to obtain the sequence to be analyzed; Perform a Fast Fourier Transform on the sequence to be analyzed to obtain the frequency domain energy distribution; By performing local neighborhood-preserving dimensionality reduction on the frequency domain energy distribution, wavelength response characteristics that characterize the difference in dual-wavelength responses are obtained.

[0052] It should be noted that, in order to extract spectral response features with significant physical meaning, the photocurrent changes need to be screened and analyzed in the frequency domain. A preset change threshold is set, which is based on the response calibration of standard ultrafine particle samples under different concentration gradients: in the concentration step experiment, during the transient process of the ultrafine particle concentration suddenly increasing from the baseline level to twice the baseline concentration, the amplitude distribution of the photocurrent change is recorded, and 50% of the smallest effective change in the amplitude distribution is taken as the threshold value to trigger the frequency domain analysis. Under the conditions of light source power (infrared LED driving current 50mA, blue LED driving current 30mA) and photoelectric detection gain in this embodiment, this calibration threshold is approximately on the order of 10μA. When the absolute value of a photocurrent change is detected to be greater than this threshold, the time domain segment of the photocurrent change is extracted and output as the sequence to be analyzed. During engineering deployment, this threshold should be recalibrated according to the light source power and detection gain. The length of the sequence to be analyzed is set to 256 sampling points, consistent with the frame length of the frame-segmented windowing, and the frequency resolution corresponding to the 256-point FFT is 4Hz.

[0053] For example, a Fast Fourier Transform (FFT) is performed on the sequence to be analyzed. The input sequence is converted from the time domain to the frequency domain using the FFT algorithm. The FFT reduces the computational complexity of the Discrete Fourier Transform (DFT) from O(N²) to O(NlogN) through a butterfly operation structure, outputting a frequency domain energy distribution composed of discrete frequency components and their corresponding amplitudes. The low-frequency amplitudes in the frequency domain energy distribution reflect the slow accumulation trend of ultrafine particle concentration, while the mid-to-high frequency amplitudes reflect the random fluctuation characteristics of the Brownian motion of the particles. Under different wavelengths of illumination, the frequency domain energy distribution of the same ultrafine particle sample exhibits identifiable patterns of difference.

[0054] Specifically, local neighborhood-preserving dimensionality reduction is performed on the frequency domain energy distribution. The input frequency domain energy distribution is then subjected to nonlinear dimensionality reduction using a locally linear embedding algorithm.

[0055] The selection of the locally linear embedding algorithm is based on the fact that principal component analysis (PCA) is a linear dimensionality reduction method, which cannot capture the nonlinear coupling relationship between different frequency components in the frequency domain energy distribution. In this embodiment, the separability of infrared and blue light response modes after PCA dimensionality reduction on the frequency domain data is insufficient. The equidistant mapping algorithm is sensitive to neighborhood parameters and the topology estimation is unstable under noisy conditions. The locally linear embedding algorithm achieves dimensionality reduction by maintaining the linear reconstruction weights in the local neighborhood of each data point. It has a better ability to preserve the local geometric structure than the above methods. Moreover, the algorithm only involves sparse matrix eigenvalue decomposition, and the computational complexity is on the order of O(pn²), where p is the number of neighborhood points and n is the number of data points. It can be completed in real time on the processor of the detection device with a cumulative scale of n approximately 500 frames.

[0056] The 500 frames of data are accumulated by successively storing the FFT results of the sequence of events to be analyzed into a frequency domain data buffer. The buffer is managed using a first-in, first-out (FIFO) sliding window strategy. When the buffer is full of 500 frames, the newest data replaces the oldest data. Standard LLE is a batch method, requiring all sample points to participate in the calculation at once to obtain the low-dimensional embedding. To meet the real-time requirements of streaming processing, this embodiment adopts a strategy combining batch initialization and incremental expansion. When the buffer first accumulates to 500 frames, a complete LLE dimensionality reduction calculation is performed on all 500 frames to obtain the initial low-dimensional embedding space and the 2D wavelength response features of each frame. Subsequently, whenever a new frame of data enters the buffer and replaces an old frame of data, the out-of-sample expansion method of LLE is used to calculate the 2D coordinates of the new frame. That is, based on the k nearest neighbors of the new frame in the original high-dimensional space and their known coordinates in the initial embedding space, the coordinates of the new frame are linearly interpolated using the local reconstruction weights determined in the initial training phase, without having to re-execute the complete LLE calculation. The single-frame computation time for out-of-sample expansion is approximately 2ms, far less than the frame data accumulation interval of approximately 250ms, meeting the requirements of embedded real-time processing. When the buffer has accumulated more than 100 frames (approximately 20% data update ratio), a complete LLE calculation is re-executed for all 500 frames in the current buffer to update the embedding space, preventing a decrease in expansion accuracy due to data distribution drift. If the analysis results need to be output before the buffer reaches 500 frames, LLE dimensionality reduction is performed using all currently accumulated frames. Because LLE is robust to changes in the number of samples, incremental verification experiments of more than 50 frames have confirmed that the angle between the two principal directions of the dimensionality reduction result does not exceed 5 degrees within the range of 100 to 500 frames.

[0057] The locally linear embedding algorithm first determines a local neighborhood of k nearest neighbors for each data point in the high-dimensional frequency domain space, setting the number of neighborhood points k to 12. Then, by minimizing the reconstruction error, it calculates the optimal weight coefficients for linear reconstruction of each data point from its neighborhood points. Finally, it searches for a distribution of data points in the low-dimensional space that maintains the same local reconstruction weight relationship. The choice of 12 neighborhood points k is based on testing the dimensionality reduction effect of k=8 to 20 on test data with a total sample size of approximately 500 frames. When k=12, the low-dimensional features have the best discriminative power for infrared and blue light response patterns, and the ratio of inter-class dispersion to intra-class dispersion reaches a maximum of 2.8.

[0058] The dimension of the reduced low-dimensional feature space is set to 2. It should be noted that Locally Linear Embedding (LLE), as a nonlinear manifold learning method, does not inherently correspond to specific physical components of infrared or blue light scattering intensity in its output low-dimensional embedding coordinates. To further define the physical meaning of the two-dimensional embedding directions, after LLE dimensionality reduction, canonical correlation analysis (CCA) is performed on the two-dimensional embedding coordinates and the reference samples, using synchronously acquired known single-wavelength response samples (frequency domain features of scattering signals acquired only under infrared light illumination and only under blue light illumination, respectively, processed using the same procedure). CCA is used to obtain two linear combination directions in the two-dimensional embedding space, maximizing their canonical correlation coefficients with the infrared and blue light reference features, respectively.

[0059] Under the calibration conditions of this embodiment, the typical correlation coefficient between the first calibration direction and the infrared light scattering intensity is 0.89, and the typical correlation coefficient between the second calibration direction and the blue light scattering intensity is 0.91, indicating that the two-dimensional embedding, after CCA calibration, can effectively encode the intrinsic scattering property differences between the two wavelengths. When retaining higher dimensions, the typical correlation coefficient between the newly added dimension and any reference feature is less than 0.5, losing its clear physical calibration significance; therefore, the dimension is set to 2. In actual deployment, due to the randomness of LLE initialization, the embedding coordinate direction may be rotated or flipped. CCA calibration needs to be re-performed using standard particle samples to determine the physical correspondence between the two embedding directions. The final output 2D feature vector after CCA calibration is the wavelength response feature.

[0060] In step S14, the wavelength response characteristics are analyzed for response status and classified for particle concentration level. Historical environmental data corresponding to the particle concentration level is extracted, and meteorological elements are separated from the historical environmental data to obtain temperature and humidity sequence. Fluctuation analysis is performed on the temperature and humidity sequence to obtain environmental fluctuation benchmark value. If the particle concentration level is greater than a preset concentration threshold and the environmental fluctuation benchmark value is less than a preset benchmark threshold, an alarm command is generated.

[0061] In one implementation, response state analysis and feature classification are performed on the wavelength response characteristics to obtain particle concentration levels, including: If the wavelength response feature is greater than a preset feature threshold, then the response feature parameters of the wavelength response feature are extracted, and the response feature parameters are classified and mapped to obtain the sensitive response state. The feature parameters contained in the sensitive response state are normalized to obtain a multidimensional feature vector; Calculate the Euclidean distance between the multidimensional feature vector and the preset standard feature library to obtain the feature matching degree; The particle concentration levels are obtained by classifying the particles based on the feature matching degree.

[0062] It should be noted that after obtaining the wavelength response characteristics, the first step is to determine whether they reach the minimum threshold for triggering fire analysis. A preset feature threshold is set, which is based on comparative calibration experiments of various interference sources and real fire particles. 200 sets of wavelength response characteristics are collected for each of six scenarios: dust rising, oil fume diffusion, water mist generation, airflow disturbance, smoldering cotton rope, and smoldering wood. The first four scenarios are non-fire interference scenarios (negative), and the last two are early-stage real fire scenarios (positive). Receiver operating characteristic (ROC) curves of the wavelength response characteristics are plotted for each scenario. The threshold point that results in a true positive rate of over 95% and a false positive rate of less than 10% is selected as the feature threshold. Under the LLE-CCA joint calibration conditions in this embodiment, the L2 norm value corresponding to this threshold is approximately 50. This value is determined by ROC curve analysis and varies with different feature dimensionality reduction parameters (LLE neighborhood number k, CCA calibration sample set). The L2 norm of the two-dimensional vector of wavelength response characteristics is taken as the wavelength response characteristic value. When the wavelength response characteristic value is greater than the characteristic threshold calibrated by this device, the response status analysis process is initiated.

[0063] The L2 norm is calculated as the square root of the sum of the squares of the components of a two-dimensional vector. The "wavelength response characteristic value" used in the ROC curve analysis during the calibration experiment is this L2 norm value. The 200 samples from each of the six scenarios described above, along with the samples used in the SVM training set, standard feature library, and concentration level calibration experiment described later, were collected at different time periods, with no overlap between the datasets.

[0064] For example, response feature parameters of the wavelength response are extracted and classified and mapped. The wavelength response feature is a 2D vector containing two components: a first principal direction (corresponding to infrared light scattering response) and a second principal direction (corresponding to blue light scattering response). The wavelength response feature sequence is scalarized. First, the L2 norm of the 2D vector, i.e., the square root of the sum of the squares of the two components, is calculated frame by frame to obtain the time series of wavelength response intensity. Then, the first derivative is calculated using the central difference algorithm on this time series to calculate the rate of change between adjacent data points, and the rising slope is extracted as the first dimension of the response feature parameter. The rising slope is expressed in μA per ms and characterizes the transient change rate of the comprehensive scattering response of ultrafine particles at different wavelengths.

[0065] The central difference algorithm was chosen because forward and backward differences only use adjacent points on one side to calculate the derivative, resulting in a truncation error on the order of O(h), where h is the sampling interval. Central difference, on the other hand, uses two symmetrical adjacent points to calculate the derivative, also with a truncation error on the order of O(h²), and significantly higher derivative accuracy at the same sampling interval. In this embodiment, h corresponds to approximately 0.98 ms for a 1024 Hz sampling rate, and the derivative estimation error of central difference is approximately 1% of that of forward difference. Simultaneously, time periods where the wavelength response features are continuously greater than a preset feature threshold are extracted as response periods. The difference between the start and end timestamps of the response period is calculated, and the duration is extracted as the second dimension of the response feature parameters. The rising slope and duration are combined into a two-dimensional input vector, which is then input into a pre-trained support vector machine classifier for classification mapping.

[0066] The Support Vector Machine (SVM) classifier uses a radial basis function kernel for nonlinear mapping. The training process involves collecting 500 sets of standard sample data with known combustion states, of which 250 sets correspond to the early smoldering stage and 250 sets correspond to the sudden open flame stage. These are randomly divided into training and validation sets in a 7:3 ratio. On the training set, a grid search method is used to select the parameter combination with the highest validation accuracy in the parameter space γ∈[0.1,2.0] and C∈[0.1,10.0] using 5-fold cross-validation. In this embodiment, the preferred parameters for the training set are γ=0.5 and C=1.0. In actual deployment, these parameters should be re-optimized through grid search as the training samples are expanded; the values ​​in this embodiment are not universally applicable.

[0067] The classifier outputs a sensitive response state, which includes a combustion stage label output by the classifier and two feature parameters that constitute the classifier's input: the rising slope and the duration. The combustion stage label has a value of 0 indicating the early accumulation stage of smoldering and a value of 1 indicating the sudden release stage of open flame. This label is used to distinguish alarm levels in subsequent joint determination, with the smoldering stage corresponding to the warning level and the open flame stage corresponding to the alarm level.

[0068] It should be noted that the feature parameters of the sensitive response state are normalized. The rising slope and duration values ​​are extracted from the sensitive response state. The unit of the rising slope is μA per ms, and the unit of the duration is ms; their numerical ranges and physical dimensions differ significantly. A max-min normalization algorithm is used for processing, with the scaling target range set to 0 to 1. The maximum and minimum values ​​are derived from the statistical extreme values ​​of the corresponding dimensions in the standard feature library. After normalization, a multi-dimensional feature vector is output, where each component of the multi-dimensional feature vector lies within the range of 0 to 1.

[0069] For example, the Euclidean distance between the multidimensional feature vector and the preset standard feature library is calculated.

[0070] The construction process of the pre-set standard feature library is as follows: In a standard combustion laboratory, 200 sets of scattering signal data were collected for each of the four common combustibles—cotton rope, wood, polyurethane foam, and polyvinyl chloride cable—under both smoldering and open flame combustion conditions. These 1600 sets of scattering signal data were processed by the same S11 to S13 preprocessing chain as the online detection process. After obtaining the wavelength response characteristics of each set of data, the rise slope and duration were extracted, and after max-min normalization, they were saved as standard feature vectors. A total of 1600 sets of standard feature vectors were stored in the standard feature library. These 1600 sets of sample data were taken from different batches of combustion experiments and there was no overlap between the samples and the SVM classifier training set and the samples used in the ROC calibration experiment.

[0071] The system inputs the current measured multidimensional feature vector, traverses each set of standard feature vectors in the standard feature library, calculates the Euclidean distance between them in multidimensional space, and outputs a distance set. The minimum Euclidean distance in the distance set reflects the physical similarity between the current particulate matter and the most similar standard combustion product. A preset distance threshold is set, which is based on the intra-class distance statistical analysis of the standard feature library. The Euclidean distance between different samples of the same combustion product in the standard feature library is calculated, and the 95th percentile of the intra-class distance is obtained. In the standard feature library of this embodiment, the 95th percentile is approximately 5.5. Considering engineering margin, it is rounded up to 6 as the maximum allowable upper limit of the intra-class distance. When the minimum Euclidean distance is less than this threshold, it indicates that the current particulate matter is highly similar to a certain standard combustion product. In actual deployment, this threshold should be recalculated as the standard feature library is expanded and updated.

[0072] A nonlinear mapping using an inverse proportional function is applied to the minimum Euclidean distance, with the mapping function being d'=1 / (1+d), where d is the minimum Euclidean distance and d' is the feature matching degree. The inverse proportional function is chosen because the decay rate of the exponential decay mapping d'=exp(-αd) is controlled by the parameter α; an inappropriate choice of α can lead to the matching degree being either too sensitive or too sluggish to changes in distance. The linear mapping d'=1-d / D requires a preset maximum distance D, but the selection of D lacks an objective standard. The inverse proportional function d'=1 / (1+d) has no free parameters; the matching degree is 1 when the distance d=0, and approaches 0 as d approaches infinity. At d=5, i.e., the distance threshold, the matching degree is 1 / 6≈0.17. This decay curve is highly consistent with the scores given by human experts for the matching degree at different distances, with a Pearson correlation coefficient of 0.96. The feature matching degree ranges from 0 to 1, with a larger value indicating a higher matching degree.

[0073] Specifically, classification is performed based on feature matching scores. A preset matching threshold is set to 0.5. This threshold is based on the fact that when the feature matching score is greater than 0.5, the corresponding Euclidean distance is less than 1, indicating that the similarity between the current particulate matter feature and a known combustion product has entered the high confidence interval, providing sufficient evidence to trigger further concentration grading. When the feature matching score is greater than the preset matching threshold, the concentration classification process is triggered. When the feature matching score is less than or equal to the preset matching threshold, it indicates that the similarity between the current particulate matter feature and a known combustion product is insufficient, the concentration classification process is not triggered, and the process returns to collect the next frame of data for analysis.

[0074] The rising slope and duration of the multidimensional feature vectors are extracted to construct the feature matrix to be classified. The feature matrix is ​​then decorrelated to eliminate the linear correlation between the rising slope and duration dimensions, resulting in a decorrelated feature matrix. This decorrelated feature matrix is ​​input into a trained support vector machine (SVM) classification model, which uses a radial basis function kernel for nonlinear mapping. The training process for this model is as follows: 1200 sets of feature data collected during the construction of the concentration level mapping table are randomly divided into training and validation sets in a 7:3 ratio. A one-to-one multi-class strategy is used to construct a 6-class SVM combination. A grid search method is used to select the parameter combination with the highest validation accuracy in the parameter space γ∈[0.1,2.0] and C∈[0.1,10.0] using 5-fold cross-validation. In this embodiment, the preferred parameters for the training set are γ=0.8 and C=2.0. In actual deployment, these parameters should be re-grid searched as the training samples are expanded.

[0075] The classification model divides the feature space into multiple regions corresponding to different concentration levels based on multiple classification hyperplanes determined during the training phase, and outputs a set of distances between each data point and each classification hyperplane. For the distance set generated by multi-classification, the category corresponding to the classification hyperplane with the smallest distance value is selected as the decision category, and this smallest distance value is used as the input for table lookup matching.

[0076] The construction process of the concentration level mapping table is as follows: The carbon black particle samples used in the above concentration level calibration experiment are independent of the preceding SVM classifier training samples and standard feature library samples. Under standard experimental conditions, ultrafine carbon black particles are used as samples. The particle mass concentration in the test chamber is precisely controlled using an aerosol generator, and six concentration gradients are set as follows: 10 μg / m³, 50 μg / m³, 100 μg / m³, 200 μg / m³, 350 μg / m³, and 500 μg / m³. 200 sets of feature data are collected under each concentration gradient, and after decorrelation processing, they are input into the support vector machine classification model. The hyperplane distance value corresponding to each data point is recorded. The arithmetic mean of the 200 hyperplane distance values ​​under each concentration gradient is taken as the baseline distance corresponding to that concentration level.

[0077] Using the midpoint of the baseline distance between adjacent concentration gradients as the interval boundary, six concentration level intervals are divided. The particle concentration value corresponding to each level is the average concentration of the adjacent interval boundaries. The specific correspondence in the mapping table is as follows: a hyperplane distance greater than 20 corresponds to a concentration value of 10 μg / m³, 10 to 20 corresponds to a concentration value of 50 μg / m³, 6 to 10 corresponds to a concentration value of 100 μg / m³, 4 to 6 corresponds to a concentration value of 200 μg / m³, 2 to 4 corresponds to a concentration value of 350 μg / m³, and less than 2 corresponds to a concentration value of 500 μg / m³. For example, when the classification hyperplane distance of a data point is 8, it matches the concentration level interval of 100 μg / m³ in the mapping table, and the output particle concentration level is 100.

[0078] It should be noted that the above concentration mapping relationship uses ultrafine carbon black particles as the standard calibration material. The particle size distribution (50nm to 500nm) and optical scattering characteristics of carbon black particles are highly similar to those of ultrafine particles produced by smoldering of common combustibles. Verification tests using smoldering ultrafine particles from two typical combustibles, cotton rope and wood, showed that the SVM hyperplane distance deviation at the same actual concentration did not exceed 15%. Therefore, the mapping table can be directly applied to cellulosic combustibles such as cotton rope and wood. Preferably, for halogenated and nitrogen-containing combustibles with significantly different optical properties (such as polyurethane foam and polyvinyl chloride cables), the mapping table can be specifically calibrated using a standard sample of that combustible before leaving the factory.

[0079] Furthermore, this embodiment employs a lookup table mapping method for SVM hyperplane distance and concentration level. The effectiveness of this method is predicated on the premise that the hyperplane distance exhibits a monotonic relationship with particle concentration. This monotonicity has been verified in standard calibration experiments of ultrafine carbon black particles (within a concentration gradient range of 10~500 μg / m³, the Pearson correlation coefficient between hyperplane distance and the logarithm of concentration is -0.93).

[0080] In one implementation, the step of extracting historical environmental data corresponding to the particle concentration level, separating meteorological elements from the historical environmental data to obtain a temperature and humidity sequence, and performing fluctuation analysis on the temperature and humidity sequence to obtain an environmental fluctuation benchmark value includes: Based on the particle concentration level, the sampling timestamp is retrieved to obtain the target time interval; Extract historical environmental data within the target time interval; Meteorological elements were separated from the historical environmental data to obtain temperature and humidity sequences; Fluctuation analysis and weighted calculations were performed on the temperature and humidity series to obtain the environmental fluctuation baseline value.

[0081] It should be noted that after obtaining the particle concentration level, the environmental fluctuation verification process begins. Based on the detection time corresponding to the particle concentration level, the sampling timestamps for the period during which that concentration level continuously occurred are retrieved from the local time-series database. The target time interval is determined by comparing the continuity of the timestamps. For example, the time interval from 8:00 to 10:00 on the current day is extracted as a particle concentration level of 100 that continuously occurred.

[0082] Specifically, historical environmental data within the target time interval is extracted and meteorological elements are separated from the historical environmental data. The target time interval is input, and wind speed, air pressure, temperature, and humidity data recorded at 1-minute sampling intervals are read from the temperature and humidity composite sensor integrated into the detection device. The environmental data sampling interval is set to 1 minute because significant changes in temperature and humidity in industrial environments typically occur on timescales exceeding 10 minutes. A 1-minute sampling interval is sufficient to capture slow temperature drift and gradual humidity trends, while keeping the 24-hour data volume within 1440 records, allowing for the retention of at least 30 days of historical data in the 512KB environmental data cache of the detection device's local memory. Principal component analysis (PCA) is performed on the four-channel data (wind speed, air pressure, temperature, and humidity) to eliminate cross-coupling between channels. PCA calculates the covariance matrix of the four-channel data and performs eigenvalue decomposition, projecting the original coupled data into a mutually orthogonal principal component space, eliminating cross-thermal coupling between the temperature and humidity channels and common-mode interference caused by wind speed fluctuations. Based on the physical law that the condensation and sedimentation behavior of ultrafine particles in the environment is mainly affected by temperature and relative humidity, among the four principal components after decorrelation, the principal component with the largest correlation coefficient with the original temperature sensor channel Pearson is selected as the temperature sequence, and the principal component with the largest correlation coefficient with the humidity sensor channel Pearson is selected as the humidity sequence, and the temperature and humidity sequences are combined to obtain the temperature and humidity sequences.

[0083] For example, fluctuation analysis and weighted calculation are performed on a temperature and humidity series. The temperature and humidity series are input, and the proportion of invalid or blank data is calculated. A preset completeness threshold of 90% is set. This threshold is based on general requirements for environmental monitoring data quality control. When the missing rate of a time series exceeds 10%, the reliability of statistics based on incomplete data decreases significantly. Therefore, 90% completeness is set as the minimum quality threshold for data analysis. It should be noted that the specific numerical relationship between the missing rate and the statistical deviation depends on the missing pattern (random missing / continuous missing / periodic missing). The 20% mentioned here is an estimate under the random missing pattern. In practical applications, if there is continuous missing data, imputation or removal processing is required.

[0084] If the missing rate of the temperature and humidity series is lower than the preset integrity threshold, it indicates that the data quality meets the analysis conditions, and subsequent fluctuation analysis continues. If the missing rate of the temperature and humidity series is greater than or equal to the preset integrity threshold, it indicates that the data quality does not meet the analysis conditions, and this round of environmental fluctuation verification is abandoned. The environmental fluctuation benchmark value of the previous effective period is used for joint judgment.

[0085] The temperature and humidity series was truncated using a sliding truncation method with a window length of 15 sampling points and a step size of 5 sampling points to obtain multiple subsequences. Each subsequence covers a 15-minute time span; this window length allows for capturing slow environmental changes while maintaining sufficient data points for statistical calculations. Zero-mean subtraction was applied to the data points in each subsequence, i.e., the arithmetic mean of the subsequence was subtracted to eliminate baseline offset for temperature and humidity. The time series variance was obtained by calculating the degree of deviation of each data point from the mean in the zero-mean series.

[0086] The weighting coefficient for temperature variance is set to 0.6, and the weighting coefficient for humidity variance is set to 0.4. This weighting is based on logistic regression analysis of 200 independently collected historical fire data sets and 200 sets of normal environmental data sets (these 400 sets of environmental data do not overlap with the samples used for training and calibration of the aforementioned machine learning models). The standardized regression coefficient for temperature fluctuation is approximately 0.62, and the standardized regression coefficient for humidity fluctuation is approximately 0.38. Approximate values ​​of 0.6 and 0.4 are used as weighting coefficients. It should be noted that these weighting coefficients are derived from field-collected data in the region (temperate monsoon climate) where this embodiment is located. When deploying in different climate zones (such as high-temperature and high-humidity or dry and cold regions), it is preferable to re-collect local environmental data and perform logistic regression analysis to update the weighting coefficients. The temperature variance and humidity variance are weighted and summed to output the environmental fluctuation baseline value. For example, if the temperature variance is 12 and the humidity variance is 8, the weighted summed environmental fluctuation baseline value is 12 × 0.6 + 8 × 0.4 = 10.4.

[0087] Specifically, after the environmental fluctuation verification is passed, the joint judgment stage begins.

[0088] The preset concentration threshold is set based on a correlation analysis between ultrafine particle concentration levels and the probability of fire occurrence: particle concentration level distributions of 200 early-stage fire samples and 500 normal environmental samples were statistically analyzed, ROC curves were plotted, and the concentration level corresponding to the point with the maximum Youden index was selected as the threshold. In the concentration level mapping system of this embodiment, the concentration level label corresponding to this threshold is approximately 75. A preset baseline threshold is set, which is based on long-term statistics of environmental fluctuation baseline values ​​under fire-free conditions: daily changes in environmental fluctuation baseline values ​​were recorded in an industrial environment without fire occurrence for 90 consecutive days, and the upper limit of three times the standard deviation of the statistical mean was taken as the upper limit of stable fluctuation. Under the field environmental conditions of this embodiment, this upper limit is approximately 12 (dimensionless baseline value). In actual deployment, both of the above thresholds should be recalibrated periodically as sensors age and environmental background changes occur, and the values ​​in this embodiment are not universally applicable.

[0089] When the particle concentration level (100) is greater than the preset concentration threshold (75) and the environmental fluctuation baseline value (10.4) is less than the preset baseline threshold (12), both conditions are met simultaneously, and an alarm command is generated. The alarm level is determined based on the combustion stage label. If the combustion stage label is the early accumulation stage of smoldering, a warning command is generated; if the combustion stage label is the sudden release stage of open flame, an alarm command is generated. The alarm command is generated using a threshold triggering method. The warning command corresponds to code 0x01, and the alarm command corresponds to code 0x02, and is output to the alarm output interface of the detection device.

[0090] like Figure 3 As shown in the figure, the distribution and classification of samples at different combustion stages in the feature space during response state analysis are illustrated. A comparison clearly shows that after wavelength response feature extraction and SVM classification mapping, the early smoldering samples (blue dots) and sudden open flame samples (red dots), which were originally randomly scattered on the two-dimensional feature plane of rising slope (0.2-1.0 μA / ms) and duration (0-200 ms), are effectively converged to both sides of their respective decision region boundaries. The current sample to be tested (green pentagram) is accurately classified into the neighborhood of the sudden open flame release stage. This result intuitively demonstrates that this scheme can effectively distinguish the response state differences between different combustion stages, providing a solid classification basis for generating warning or alarm commands based on combustion stage labels.

[0091] In summary, this invention discloses a point-type fire alarm method based on dual-wavelength photoelectric sensing. By performing cascaded processing of the raw photoelectric data through multi-scale decomposition and blind source separation, this invention decouples the cross-interference of multi-band light sources, accurately separates the independent scattering signals of ultrafine particles, and combines wavelength response characteristic analysis with environmental fluctuation verification to achieve reliable early fire detection, effectively reducing the false alarm and missed alarm rates in complex industrial environments.

[0092] The second embodiment of the present invention provides a point-type fire alarm system with dual-wavelength photoelectric sensing, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the point-type fire alarm method with dual-wavelength photoelectric sensing as described above.

[0093] It should be noted that the dual-wavelength photoelectric sensing point fire alarm system provided in this embodiment of the invention is used to execute all the process steps of the dual-wavelength photoelectric sensing point fire alarm method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0094] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A point-type fire alarm method using dual-wavelength photoelectric sensing, characterized in that, include: The raw photoelectric data of alternating illumination of space by multiple frequency band light sources is acquired, and the raw photoelectric data is decomposed and filtered to reconstruct a pure photoelectric signal. The raw photoelectric data contains cross-interference of scattered signals between light sources of different wavelengths. The pure photoelectric signal is framed and windowed to construct a multi-dimensional observation matrix; Blind source separation is performed on the multidimensional observation matrix to obtain a separation matrix, including: decorrelation processing of the multidimensional observation matrix to obtain a decorrelation matrix; variance normalization of the decorrelation matrix to obtain an orthogonal whitening matrix; iterative solution of the orthogonal whitening matrix to maximize non-Gaussianity to obtain a separation matrix; and decoupling the cross-interference based on the separation matrix to obtain the independent scattering signal corresponding to the ultrafine particles. The envelope of the independent scattering signal is extracted and differential operation is performed to obtain the photocurrent change. The photocurrent change is then subjected to frequency domain analysis and dimensionality reduction to obtain wavelength response characteristics that characterize the difference between the two wavelength responses. The process of analyzing the response state and classifying the wavelength response features to obtain particle concentration levels includes: if the wavelength response feature is greater than a preset feature threshold, extracting the response feature parameters of the wavelength response feature and classifying and mapping the response feature parameters to obtain a sensitive response state; normalizing the feature parameters contained in the sensitive response state to obtain a multidimensional feature vector; calculating the Euclidean distance between the multidimensional feature vector and a preset standard feature library to obtain a feature matching degree; and performing classification processing based on the feature matching degree to obtain particle concentration levels. Historical environmental data corresponding to the particle concentration level is extracted, meteorological elements are separated from the historical environmental data to obtain a temperature and humidity sequence, and fluctuation analysis is performed on the temperature and humidity sequence to obtain an environmental fluctuation benchmark value. If the particle concentration level is greater than a preset concentration threshold and the environmental fluctuation benchmark value is less than a preset benchmark threshold, an alarm command is generated.

2. The point-type fire alarm method using dual-wavelength photoelectric sensing according to claim 1, characterized in that, The step of performing multi-scale decomposition and filtering reconstruction on the original photoelectric data to obtain a clean photoelectric signal includes: The original photoelectric data is decomposed into multi-scale components. The multi-scale components are subjected to noise removal and smoothing to obtain the denoised components. The denoised components are reconstructed to obtain a clean photoelectric signal.

3. The point-type fire alarm method using dual-wavelength photoelectric sensing according to claim 1, characterized in that, The step of decoupling the cross-interference according to the separation matrix to obtain the independent scattering signal corresponding to the ultrafine particles includes: The cross-interference is decoupled based on the separation matrix to obtain multiple independent source signals; The variance of the independent source signals is calculated to obtain the amplitude variance value; If the amplitude variance value is less than the preset variance threshold, the independent source signal is determined to be an independent scattering signal corresponding to the ultrafine particles.

4. The point-type fire alarm method using dual-wavelength photoelectric sensing according to claim 1, characterized in that, The step of extracting the envelope of the independent scattering signal and performing a difference operation to obtain the photocurrent change includes: Extreme points are extracted from the independent scattering signals to obtain a time-axis extreme value sequence; The fluctuation amplitude is obtained by fitting the envelope of the time axis extreme value sequence; The change in photocurrent is obtained by performing a differential operation on the fluctuation amplitude.

5. The point-type fire alarm method using dual-wavelength photoelectric sensing according to claim 1, characterized in that, The frequency domain analysis and dimensionality reduction of the photocurrent change yields wavelength response characteristics characterizing the difference in dual-wavelength responses, including: If the absolute value of the change in photocurrent is greater than a preset change threshold, then the time domain segment of the change in photocurrent is extracted to obtain the sequence to be analyzed; Perform a Fast Fourier Transform on the sequence to be analyzed to obtain the frequency domain energy distribution; By performing local neighborhood-preserving dimensionality reduction on the frequency domain energy distribution, wavelength response characteristics that characterize the difference in dual-wavelength responses are obtained.

6. The point-type fire alarm method using dual-wavelength photoelectric sensing according to claim 1, characterized in that, The process of extracting historical environmental data corresponding to the particle concentration level, separating meteorological elements from the historical environmental data to obtain a temperature and humidity sequence, and performing fluctuation analysis on the temperature and humidity sequence to obtain an environmental fluctuation benchmark value includes: Based on the particle concentration level, the sampling timestamp is retrieved to obtain the target time interval; Extract historical environmental data within the target time interval; Meteorological elements were separated from the historical environmental data to obtain temperature and humidity sequences; Fluctuation analysis and weighted calculations were performed on the temperature and humidity series to obtain the environmental fluctuation baseline value.

7. A point-type fire alarm system with dual-wavelength photoelectric sensing, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the point-type fire alarm method with dual-wavelength photoelectric sensing as described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the point-type fire alarm method with dual-wavelength photoelectric sensing as described in any one of claims 1 to 6.

Images