Anti-false-alarm method for detecting particle components in air

Abstract

The invention provides a false alarm prevention method for detecting particle components in air, and relates to the technical field of air component detection, and the false alarm prevention method comprises the following steps: S1, generating incident light by using a laser source; s2, light scattering signals in a plurality of preset scattering directions are collected through a scattered light collection system, and the scattered light collection system comprises a main path receiving unit and a branch path receiving unit; s3, performing signal processing on the collected light scattering signals to obtain processed scattered light data; s4, analyzing the processed scattered light data, judging whether the signal intensity of the main path receiving unit reaches a preset threshold, if so, executing the step S5, and if not, returning to the step S2 to continue to collect and detect the light scattering signal; s5, analyzing and calculating the scattered light data of the branch receiving unit by using a particle feature analysis module to obtain particle feature parameters; and S6, performing particle classification judgment based on the particle characteristic parameters, and selectively triggering an alarm signal according to a classification result. According to the invention, smoke particles and interference particles such as dust and water mist can be effectively distinguished, so that the false alarm rate is greatly reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of air component detection, and in particular to a false alarm prevention method for detecting particle components in the air. Background Art

[0002] Traditional aspirating detectors rely mainly on a single receiving sensor to detect particles in the air. The working principle of this detector is based on the scattering of light, especially the intensity change of forward scattered light. The detector emits a light beam through a laser light source. When the light beam encounters particles in the air, it is scattered. The receiving sensor then detects the intensity change of the scattered light. Since the intensity of scattered light is related to the concentration of particles, this method can infer the concentration of particles in the air. However, this single sensor system has significant limitations, mainly manifested in the inability to distinguish different types of particles, such as smoke, dust or water mist, which can easily lead to false alarm problems.

[0003] In order to deal with the shortcomings of single-sensor systems, existing technologies have taken some measures. One common method is to use mechanical filters to remove larger particles. However, this method is difficult to effectively filter out PM2.5-level interfering particles that are similar in size to smoke particles. In addition, the filter will increase airflow resistance and affect the sensitivity of the detector. At the same time, the filter itself is easily contaminated and requires frequent maintenance. Another method is to use a software delayed alarm strategy to reduce false alarms by delaying the alarm signal. However, this method may result in an untimely response, especially for short-term smoke conditions. Summary of the invention

[0004] In view of this, the present invention proposes an anti-false alarm method for detecting particle components in the air, aiming to overcome the limitations of traditional single-sensor aspirating fire detectors. By adopting multiple light sensors with different scattering directions, combined with signal processing and machine learning algorithms, the detection accuracy is significantly improved, and smoke particles are effectively distinguished from interfering particles such as dust and water mist, thereby greatly reducing the false alarm rate.

[0005] The technical solution of the present invention is implemented as follows: The present invention provides a method for preventing false alarms in detecting particle components in air, comprising:

[0006] S1 uses a laser source to generate incident light;

[0007] S2 collects light scattering signals in multiple preset scattering directions through a scattered light collection system, wherein the scattered light collection system includes a main receiving unit and a branch receiving unit;

[0008] S3 performs signal processing on the collected light scattering signal to obtain processed scattered light data;

[0009] S4 analyzes the processed scattered light data to determine whether the signal strength of the main receiving unit reaches a preset threshold value. If so, execute step S5. If not, return to step S2 to continue collecting and detecting light scattering signals.

[0010] S5 uses the particle characteristic analysis module to analyze and calculate the scattered light data of the branch receiving unit to obtain particle characteristic parameters;

[0011] S6 classifies particles based on particle characteristic parameters and selectively triggers an alarm signal according to the classification results.

[0012] On the basis of the above scheme, preferably, the number of preset scattering directions is 3, and the scattered light collection system includes a main receiving unit and two branch receiving units, wherein the main receiving unit and the two branch receiving units respectively collect light scattering signals in the 3 preset scattering directions.

[0013] Based on the above solution, preferably, step S3 includes:

[0014] S31 performs ensemble empirical mode decomposition on the collected light scattering signals to obtain a series of intrinsic mode functions IMFs;

[0015] S32 performs adaptive noise threshold screening on the obtained IMF to remove noise-related IMF;

[0016] S33 performs sparse representation optimized independent component analysis on the remaining IMFs to separate independent signal components;

[0017] S34 extracts multi-scale features of independent components based on wavelet packet transform;

[0018] S35 introduces an adaptive threshold filter, which dynamically adjusts the threshold to select the component most relevant to particle scattering based on the energy distribution, frequency characteristics and wavelet packet characteristics of the signal;

[0019] S36 uses a fuzzy inference system to evaluate and optimize the selected components;

[0020] S37 performs adaptive weight reconstruction on the optimized components to obtain the denoised and enhanced scattered light data.

[0021] Based on the above solution, preferably, in step S35, the calculation process of the adaptive threshold filter is:

[0022] Calculate the energy E of each independent component x :

[0023]

[0024] Where, IC x(t) is the signal value of the xth independent component, T is the signal length, and t is the time variable;

[0025] Calculate the wavelet packet eigenvalue W x :

[0026] W x =g(WPT(IC x ))

[0027] Where WPT is wavelet packet transform, g is feature extraction function;

[0028] Set the initial threshold θ0 and adjust the threshold according to the energy distribution and wavelet packet characteristics:

[0029]

[0030] Where M is the total number of independent components, θ x is the adaptive threshold of the xth independent component.

[0031] Based on the above solution, preferably, in step S37, the adaptive weight reconstruction process is:

[0032]

[0033] Where X(t) is the reconstructed scattered light data, w x (t) is the time-varying weight function of the x-th independent component, IC x (t) is the xth selected independent component, M is the total number of independent components;

[0034] Time-varying weight function w x The calculation formula of (t) is:

[0035]

[0036] In the formula, a x 、b x ,t x 、c x A parameter that is dynamically adjusted according to the output of the fuzzy inference system; a x is the amplitude coefficient, which controls the maximum value of the time-varying weight function; b x is the attenuation coefficient, which controls the rate at which the time-varying weight function changes over time; t x is the central time point of the time-varying weight function; c x is the baseline shift parameter; |tt x | represents the current time point t and the central time point t x The absolute distance between .

[0037] Based on the above solution, preferably, in step S4, a fast Fourier transform method is used to analyze the processed scattered light data, and the signal strength of the main receiving unit is extracted by calculating the amplitude spectrum of the signal.

[0038] Based on the above solution, preferably, step S5 includes:

[0039] S51 pre-processes the scattered light data of the two branch receiving units to remove background noise and irrelevant signals;

[0040] S52 extracts time domain features from the preprocessed scattered light data, including calculating the mean, standard deviation, peak value and fluctuation factor of the signal;

[0041] S53 performs frequency domain feature extraction on the preprocessed scattered light data, including calculating the power spectrum density and main frequency of the signal;

[0042] S54 calculates an asymmetry factor according to the average value of the scattered light intensities of the two branch receiving units;

[0043] S55 fuses and reduces the dimensionality of the extracted time domain features, frequency domain features, and asymmetric factors based on a deep learning model to obtain a comprehensive feature vector;

[0044] S56 uses an adaptive weight algorithm to assign weights to each feature in the comprehensive feature vector and calculates the final particle feature parameters.

[0045] Based on the above solution, preferably, the deep learning model in step S55 adopts a multi-scale convolutional neural network, and its structure is as follows:

[0046] The input layer is used to receive time domain features, frequency domain features and asymmetry factors, and combine these features into a multi-dimensional input vector;

[0047] The multi-scale convolution layer contains three parallel convolution branches, each branch uses a convolution kernel of a different size:

[0048] Small-scale branches use 1*1 convolution kernels to capture local features;

[0049] The mid-scale branch uses a 3*3 convolution kernel to capture medium-range feature associations;

[0050] Large-scale branch, using 5*5 convolution kernel to capture a wider range of feature patterns;

[0051] Each branch contains two consecutive convolutional layers, each followed by batch normalization and ReLU activation function;

[0052] The feature fusion layer uses 1*1 convolution to fuse the outputs of the three branches and integrate feature information of different scales;

[0053] Attention mechanism, which adopts a dual attention mechanism combining channel attention mechanism and spatial attention mechanism. Channel attention uses global average pooling and maximum pooling to calculate the weight of each channel through a multi-layer perceptron. Spatial attention uses convolution operations to calculate the importance of each spatial position.

[0054] The global pooling layer uses global average pooling to compress the features into a vector of fixed length to obtain global features;

[0055] Fully connected layer, including two fully connected layers, each layer is followed by a ReLU activation function. The first fully connected layer is used to reduce the feature dimension, and the second fully connected layer is used to compress and nonlinearly transform features;

[0056] The output layer uses a fully connected layer to generate the final comprehensive feature vector.

[0057] Based on the above solution, preferably, in step S56, the calculation formula of the final particle characteristic parameter is:

[0058]

[0059] Weight ω i The dynamic update formula is:

[0060]

[0061] Where P represents the final particle characteristic parameter; f i represents the i-th eigenvalue; ω i represents the weight of i features; α, β, γ, and δ are adjustment parameters, α adjusts the nonlinear effect of the weighted sum of features, β controls the power of the nonlinear term, γ adjusts the effect of feature differences on the nonlinear term, and δ adjusts the effect of signal statistical characteristics on the final parameters; σ is the standard deviation of the signal; max(F) is the maximum amplitude in the spectrum; ΔA is the fluctuation factor; η is the learning rate; L is the objective function; f mean is the average value of all features; ∈ is a regulation parameter used to control the impact of feature differences on weight update.

[0062] Based on the above solution, preferably, step S6 includes:

[0063] The calculated particle characteristic parameters are compared with the pre-calibrated interference particle characteristic parameters, and the particles are judged to be smoke particles or interference particles according to the comparison results;

[0064] If the particle is determined to be a smoke particle and the signal strength exceeds the safety threshold, an alarm signal is triggered;

[0065] If the particle is determined to be an interference particle, no alarm signal is triggered and the test result is recorded to optimize subsequent tests.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] (1) Through the multi-sensor system and signal processing algorithm, it can effectively distinguish different types of particulate matter and significantly reduce the false alarm rate while maintaining high sensitivity and rapid response capabilities to abnormal situations;

[0068] (2) The light scattering signal collection method with multiple preset scattering directions can capture the scattering characteristics of particles in different directions;

[0069] (3) The signal processing technologies adopted, such as adaptive noise threshold screening, independent component analysis and wavelet packet transform, have greatly improved the quality and information content of the original signal. These technologies effectively remove background noise and interference, enhance useful signals, and provide a more reliable data basis for subsequent particle feature analysis, thereby improving the detection sensitivity and anti-interference ability of the entire system;

[0070] (4) The introduction of multi-scale convolutional neural networks for feature fusion and dimensionality reduction processing significantly improves the system's ability to identify subtle changes in complex environments. This method can automatically learn and adapt to particle characteristics under different environmental conditions, enhance the system's adaptability and generalization ability, and make the detection results more accurate and reliable;

[0071] (5) A dynamic weight update mechanism is adopted, which enables the system to adjust the importance of different features according to real-time data. This adaptive method significantly improves the system's response speed and accuracy to environmental changes, making the detection results more stable and reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0073] Figure 1 is a flow chart of the method of the present invention;

[0074] Figure 2 The technical implementation diagram of the present invention

[0075] Figure 3 It is a framework diagram of the anti-false alarm system of the present invention. DETAILED DESCRIPTION

[0076] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0077] like Figure 1 and Figure 2 As shown, the present invention provides a method for preventing false alarms in detecting particle components in air, comprising:

[0078] S1 uses a laser source to generate incident light;

[0079] S2 collects light scattering signals in multiple preset scattering directions through a scattered light collection system, wherein the scattered light collection system includes a main receiving unit and a branch receiving unit;

[0080] S3 performs signal processing on the collected light scattering signal to obtain processed scattered light data;

[0081] S4 analyzes the processed scattered light data to determine whether the signal strength of the main receiving unit reaches a preset threshold value. If so, execute step S5. If not, return to step S2 to continue collecting and detecting light scattering signals.

[0082] S5 uses the particle characteristic analysis module to analyze and calculate the scattered light data of the branch receiving unit to obtain particle characteristic parameters;

[0083] S6 classifies particles based on particle characteristic parameters and selectively triggers an alarm signal according to the classification results.

[0084] Specifically, in one embodiment of the present invention, there are three preset scattering directions, and the scattered light collection system includes a main receiving unit and two branch receiving units, wherein the main receiving unit and the two branch receiving units respectively collect light scattering signals in the three preset scattering directions.

[0085] Specifically, in one embodiment of the present invention, step S3 includes:

[0086] S31 performs ensemble empirical mode decomposition on the collected light scattering signals to obtain a series of intrinsic mode functions IMFs.

[0087] Specifically, add white noise to the original signal x(t) to generate multiple new signals, perform EMD decomposition on each new signal, and obtain an IMF set. Take the average of all IMF sets to eliminate the influence of white noise. The formula is as follows:

[0088] x(t)=∑IMF i (t)+rn (t)

[0089] Among them, IMF i (t) is the ith eigenmode function, r n (t) is the residual term.

[0090] S32 performs adaptive noise threshold screening on the obtained IMF to remove noise-related IMF.

[0091] Specifically, the energy of each IMF is calculated, an adaptive threshold is set, and based on the overall energy distribution of the signal, the IMF with energy below the threshold is removed. In this case, the threshold can be k*meanE x , k represents the adaptive coefficient.

[0092] S33 performs sparse representation optimized independent component analysis on the remaining IMFs to separate independent signal components.

[0093] Specifically, the residual IMFs are combined into a mixed signal matrix X, the ICA algorithm is applied, the separation matrix W is estimated, and W is optimized using the sparsity constraint.

[0094] During optimization, the optimization target is max{J(W)+λ||W||1}, J(W) is the ICA objective function, and λ is the sparse regularization parameter. After optimizing W, the independent component S is obtained according to S=WX.

[0095] S34 performs multi-scale feature extraction on independent components based on wavelet packet transform.

[0096] Specifically, each independent component is decomposed by wavelet packet, features are extracted at different scales and frequency bands, and features are combined to form feature vectors. The formula is: WPT(s) = {d j ,k,n},d j ,k,n is the nth wavelet packet coefficient of the kth node in the jth layer.

[0097] S35 introduces an adaptive threshold filter, which dynamically adjusts the threshold to select the component most relevant to particle scattering based on the energy distribution, frequency characteristics and wavelet packet characteristics of the signal.

[0098] Specifically, the calculation process of the adaptive threshold filter is:

[0099] Calculate the energy E of each independent component x :

[0100]

[0101] Where, IC x (t) is the signal value of the xth independent component, T is the signal length, and t is the time variable;

[0102] Calculate the wavelet packet eigenvalue W x :

[0103] W x =g(WPT(IC x ))

[0104] Where WPT is wavelet packet transform, g is feature extraction function;

[0105] Set the initial threshold θ0 and adjust the threshold according to the energy distribution and wavelet packet characteristics:

[0106]

[0107] Where M is the total number of independent components, θ x is the adaptive threshold of the xth independent component.

[0108] S36 uses a fuzzy inference system to evaluate and optimize the selected components.

[0109] Specifically, input variables and output variables are defined: Input variables include: signal strength (S), main frequency (F), energy distribution (E), wavelet packet characteristics (W), etc. Output variable: component correlation score (R).

[0110] Fuzzy sets are defined for each variable: for example, signal strength (S) is defined as {low, medium, high}; dominant frequency (F) is defined as {low frequency, medium frequency, high frequency}; energy distribution (E) is defined as {concentrated, scattered}; wavelet packet features (W) are defined as {insignificant, significant}; and output variable correlation score (R) is defined as {low, medium, high}.

[0111] Design fuzzy rules: Rule 1: IF(S is high) AND(F is medium frequency) AND(E is concentrated) THEN(R is high); Rule 2: IF(S is low) OR(E is scattered) THEN(R is low); Rule 3: IF(W is significant) AND(F is high frequency) THEN(R is medium); ... (more rules).

[0112] Fuzzification: Convert crisp input to fuzzy set using membership function. Membership function: μ A (x):X→[0,1].

[0113] Rule evaluation: Use fuzzy operators (such as AND, OR) to evaluate the antecedents of each rule. For example: R i =min(μ A (x1),μ B (x2),...,μ N (x n )).

[0114] Rule output aggregation: Merge the outputs of all rules. For example: R aggregate =max(R1,R2,...,R m ).

[0115] Defuzzification: Convert fuzzy output to crisp values. crisp =∫(x*μ(x)dx) / ∫(μ(x)dx).

[0116] Output evaluation: The importance of each signal component is evaluated based on the defuzzified correlation score R.

[0117] S37 performs adaptive weight reconstruction on the optimized components to obtain the denoised and enhanced scattered light data.

[0118] Specifically, the adaptive weight reconstruction process is:

[0119]

[0120] Where X(t) is the reconstructed scattered light data, w x (t) is the time-varying weight function of the x-th independent component, IC x (t) is the xth selected independent component, M is the total number of independent components;

[0121] Time-varying weight function w x The calculation formula of (t) is:

[0122]

[0123] In the formula, a x 、b x ,t x 、c x A parameter that is dynamically adjusted according to the output of the fuzzy inference system; a x is the amplitude coefficient, which controls the maximum value of the time-varying weight function. It determines the overall importance of the independent component in the reconstruction process. x The value of b indicates that the component has a greater contribution to the final reconstructed signal; b x is the attenuation coefficient, which controls the rate at which the time-varying weight function changes over time. A larger b x The value will cause the weight function to decay faster, making the influence of this component more concentrated on t x Nearby, smaller b x The value of t will cause the weight function to decay more slowly, so that the influence of the component is maintained over a longer time range; x is the central time point of the time-varying weight function, which represents the best or most important position of the independent component on the time axis. The weight function is at t = t xWhen the maximum value is reached (assuming c x =0); c x is the baseline shift parameter, which ensures that even if far away from t x At this time point, the component may still have a minimum contribution, and the larger the c x A value of tt means that the component has a certain influence over the entire time range; x | represents the current time point t and the central time point t x The absolute distance between them, this term makes the weight function appear to be t x A symmetrical shape with its center at the center.

[0124] In this embodiment, a x 、b x ,t x 、c x The process of obtaining is as follows:

[0125] Map the relevance score R of the fuzzy inference system to the interval [0,1]. x =R max +(1-R max )×R,R max is the preset maximum amplitude, and R is the correlation score of the fuzzy inference output.

[0126] Define the fuzzy set {fast decay, medium decay, slow decay}. Use fuzzy rules to determine the decay type based on the persistence and rate of change of the signal. x =b min +(b max -b min )×μ b , b min and b max is the preset attenuation coefficient range, μ b is the degree of membership of the decay type.

[0127] Analyze the peak position and energy distribution of the signal. Use fuzzy rules to evaluate the important time periods of the signal. x =t start +(t end -t start )×μ t , t start and t end are the start and end time of the signal, μ t is the fuzzy evaluation result of the center position.

[0128] Evaluate the background noise level of the signal. Use fuzzy rules to determine the appropriate baseline shift. x =c min +(c max -c min)×(1 - R), c min and c max is a preset baseline range, and R is a correlation score.

[0129] Specifically, in one embodiment of the present invention, step S4 includes:

[0130] Perform a fast Fourier transform FFT on the processed scattered light data x(t) of the main road receiving unit: X(f) = FFT(x(t)).

[0131] Calculate the absolute value |X(f)| of the FFT result. Normalize the amplitude spectrum: A(f) = |X(f)| / max(|X(f)|).

[0132] Extract the signal intensity of the main road receiving unit:

[0133] Find the maximum value of the amplitude spectrum: S max = max(A(f)). Calculate the average amplitude in a specific frequency band: where [f1, f2] is the frequency range of interest, and M is the number of frequency points in this range.

[0134] Calculate the comprehensive signal intensity index: S = w1 * S max + w2 * S avg , where w1 and w2 are weight coefficients.

[0135] Set a preset threshold v, compare S and v. If S ≥ v, the signal intensity reaches the threshold; if S < v, the signal intensity does not reach the threshold.

[0136] If the signal intensity reaches the threshold: Record the current timestamp and signal characteristics, and trigger subsequent particle characteristic analysis (perform step S5). If the signal intensity does not reach the threshold: Return to step S2 and continue to collect and detect the light scattering signal.

[0137] Specifically, in one embodiment of the present invention, step S5 includes:

[0138] S51 Preprocess the scattered light data of the two branch receiving units to remove background noise and irrelevant signals.

[0139] Specifically, background noise removal: Use a moving average filter to remove high-frequency noise, and apply an adaptive median filter to remove impulse noise.

[0140] Signal normalization: Normalize the signal amplitude to the range [0, 1], formula: x norm = (x - min(x)) / (max(x) - min(x)).

[0141] Trend removal: Use polynomial fitting to estimate the signal trend and subtract the trend component from the original signal.

[0142] Bandpass filtering: A bandpass filter is applied to retain the frequency components related to particle scattering, and the filter parameters are determined based on prior knowledge.

[0143] S52 performs time domain feature extraction on the preprocessed scattered light data, including calculating the mean, standard deviation, peak value and fluctuation factor of the signal.

[0144] Specifically, calculate the signal mean:

[0145] Calculate the standard deviation:

[0146] Calculate the peak value: Peak = max(|x i |).

[0147] Calculate the fluctuation factor: CF = Peak / RMS,

[0148] S53 performs frequency domain feature extraction on the preprocessed scattered light data, including calculating the power spectrum density and main frequency of the signal.

[0149] Specifically, the power spectral density PSD is calculated: PSD(f)=|FFT(x(t)*w(t))| 2 / (fs*N), w(t) is the window function, fs is the sampling frequency, and N is the number of data points.

[0150] Extract the main frequency: f main =argmax(PSD(f)).

[0151] Calculate the band energy ratio: E ratio =∑ f in band PSD(f) / ∑ all f PSD(f).

[0152] S54 calculates an asymmetry factor according to the average value of the scattered light intensities of the two branch receiving units.

[0153] Specifically, the average values ​​of the scattered light intensities of the two branch receiving units are calculated: I1=mean(x1), I2=mean(x2).

[0154] Calculate the asymmetry factor: AF = (|I1-I2|) / (I1+I2).

[0155] Based on the deep learning model, S55 fuses and reduces the extracted time domain features, frequency domain features, and asymmetric factors to obtain a comprehensive feature vector.

[0156] Specifically, the deep learning model uses a multi-scale convolutional neural network with the following structure:

[0157] The input layer is used to receive time domain features, frequency domain features and asymmetry factors, and combine these features into a multi-dimensional input vector;

[0158] The multi-scale convolution layer contains three parallel convolution branches, each branch uses a convolution kernel of a different size:

[0159] Small-scale branches use 1*1 convolution kernels to capture local features;

[0160] The mid-scale branch uses a 3*3 convolution kernel to capture medium-range feature associations;

[0161] Large-scale branch, using 5*5 convolution kernel to capture a wider range of feature patterns;

[0162] Each branch contains two consecutive convolutional layers, each followed by batch normalization and ReLU activation function;

[0163] The feature fusion layer uses 1*1 convolution to fuse the outputs of the three branches and integrate feature information of different scales;

[0164] Attention mechanism, which adopts a dual attention mechanism combining channel attention mechanism and spatial attention mechanism. Channel attention uses global average pooling and maximum pooling to calculate the weight of each channel through a multi-layer perceptron. Spatial attention uses convolution operations to calculate the importance of each spatial position.

[0165] The global pooling layer uses global average pooling to compress the features into a vector of fixed length to obtain global features;

[0166] Fully connected layer, including two fully connected layers, each layer is followed by a ReLU activation function. The first fully connected layer is used to reduce the feature dimension, and the second fully connected layer is used to compress and nonlinearly transform features;

[0167] The output layer uses a fully connected layer to generate the final comprehensive feature vector.

[0168] S56 uses an adaptive weight algorithm to assign weights to each feature in the comprehensive feature vector and calculates the final particle feature parameters.

[0169] Specifically, step S56 includes:

[0170] Step 1: Initialize feature weights: ω i =1 / n, where n is the number of features. Set the adjustment parameters α, β, γ, δ, ∈, and initialize the number of iterations t=0.

[0171] Step 2: Feature preparation: Collect all relevant features f i , including features output from deep learning models and features extracted directly.

[0172] Step 3: Calculate particle characteristic parameters:

[0173]

[0174] Where P represents the final particle characteristic parameter; f i represents the i-th eigenvalue; ω i represents the weight of i features; α, β, γ, and δ are adjustment parameters, α adjusts the nonlinear effect of the weighted sum of features, β controls the power of the nonlinear term, γ adjusts the effect of feature differences on the nonlinear term, and δ adjusts the effect of signal statistical characteristics on the final parameters; σ is the standard deviation of the signal; max(F) is the maximum amplitude in the spectrum; ΔA is the fluctuation factor.

[0175] Step 4: Evaluate the objective function L:

[0176] L=L(P,{f i}); L is an objective function defined based on the physical model and is used to evaluate the rationality of P.

[0177] Step 5: Calculate the gradient: That is, L for each weight ω i The partial derivative of .

[0178] Step 6: Update weights:

[0179]

[0180] Among them, η is a small learning rate used to control the update step size; L is the objective function; f mean is the average value of all features; ∈ is a regulation parameter used to control the impact of feature differences on weight update.

[0181] Step 7: Set t=t+1 and return to step 3.

[0182] Step 8: Repeat steps 3 to 7 until the maximum number of iterations is reached or the change in L is less than a preset threshold.

[0183] Step 9: Calculate the particle characteristic parameter P using the final optimized weight.

[0184] Specifically, in one embodiment of the present invention, step S6 includes:

[0185] The calculated particle characteristic parameters are compared with the pre-calibrated interference particle characteristic parameters, and based on the comparison results, the particles are judged to be smoke particles or interference particles.

[0186] In this embodiment, an interference particle feature database is pre-established, which contains feature parameters of various known interference sources (such as dust, water mist, etc.). A similarity measurement method (such as Euclidean distance, cosine similarity, etc.) is used to calculate the similarity between the current particle feature and each item in the database. A similarity threshold is set to determine whether it is an interference particle.

[0187] If the similarity with any known interference source is higher than the threshold, it is determined to be an interference particle. If the similarity with all known interference sources is lower than the threshold, it is determined to be a smoke particle.

[0188] If the particle is determined to be a smoke particle and the signal strength exceeds the safety threshold, an alarm signal is triggered.

[0189] In this embodiment, for the case where the particle is determined to be a smoke particle: the system will also check whether the signal strength exceeds the safety threshold. The safety threshold may be a preset fixed value or may be dynamically adjusted according to the environment. Only when the particle is determined to be a smoke particle and the signal strength exceeds the safety threshold, the alarm signal will be triggered.

[0190] If the particle is determined to be an interference particle, no alarm signal is triggered and the test result is recorded to optimize subsequent tests.

[0191] In this embodiment, if the particle is determined to be an interfering particle, the system will not trigger an alarm signal, but will record the detection result.

[0192] Specifically, the system mentioned in the present invention refers to an anti-false alarm system including a laser source, a scattered light collection system, a signal processing unit, a particle feature analysis module, and a control and decision-making unit.

[0193] like Figure 3As shown, the laser source emits a stable laser beam for detecting particulate matter in the air. The scattered light collection system includes a main receiving unit and multiple (at least two) branch receiving units, which collect light scattering signals in multiple preset scattering directions and transmit the collected original scattered signals to the signal processing unit. The signal processing unit performs complex processing and analysis on the original scattered signals. The processing flow includes: ensemble empirical mode decomposition (EEMD), adaptive noise threshold screening, independent component analysis (ICA), wavelet packet transform, adaptive threshold filtering, fuzzy inference system, adaptive weight reconstruction, and then outputs the processed high-quality scattered light data and passes it to the particle feature analysis module. The particle feature analysis module extracts and analyzes particle features from the processed scattered light data, including time domain feature extraction (mean, standard deviation, peak value, fluctuation factor, etc.), frequency domain feature extraction (power spectral density, main frequency, etc.), asymmetry factor calculation, deep learning model (multi-scale convolutional neural network) for feature fusion and dimensionality reduction, adaptive weight algorithm, and passes the calculated particle feature parameters to the control and decision unit. The control and decision-making unit classifies and determines based on the particle characteristic parameters to decide whether to trigger an alarm. The decision-making process includes: comparing the calculated particle characteristic parameters with the pre-calibrated interference particle characteristic parameters, judging whether the particles are smoke particles or interference particles, and checking whether the signal strength exceeds the safety threshold. If the unit receives an optimization instruction, it updates the interference particle characteristic database, adjusts the classification algorithm parameters, optimizes the signal processing method, and adjusts the alarm threshold. The alarm system triggers the alarm according to the instructions of the control and decision-making unit, including a multi-level alarm mechanism, such as early warning, mild alarm, severe alarm, etc. The interference particle characteristic database stores the characteristic parameters of known interference sources (such as dust, water mist, etc.), which is continuously updated and optimized according to the system's detection results and operator feedback. The operator feedback mechanism allows the operator to confirm or correct the system's classification results. Environmental data input is an optional component, not a required component. It collects environmental factor data that may affect the decision, such as temperature, humidity, etc., which provides additional decision-making basis for the control and decision-making unit and improves the environmental adaptability of the system.

[0194] The control and decision-making unit of the system continuously optimizes system performance based on the detection results and operator feedback. Optimization measures include updating the database, adjusting algorithm parameters, improving signal processing methods, etc. Through this combination of self-optimization and manual feedback, the system continuously improves detection accuracy and reduces false alarm rates.

[0195] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for preventing false alarms in detecting particle components in the air, characterized in that: include: S1 uses a laser source to generate incident light; S2 collects light scattering signals in multiple preset scattering directions through a scattered light collection system, wherein the scattered light collection system includes a main receiving unit and a branch receiving unit; S3 performs signal processing on the collected light scattering signal to obtain processed scattered light data; S4 analyzes the processed scattered light data to determine whether the signal strength of the main receiving unit reaches a preset threshold value. If so, execute step S5. If not, return to step S2 to continue collecting and detecting light scattering signals. S5 uses the particle characteristic analysis module to analyze and calculate the scattered light data of the branch receiving unit to obtain particle characteristic parameters; S6 classifies particles based on particle characteristic parameters and selectively triggers an alarm signal according to the classification results.

2. The method for preventing false alarms of detecting particulate components in air according to claim 1, characterized in that: There are three preset scattering directions, and the scattered light collection system includes a main receiving unit and two branch receiving units, wherein the main receiving unit and the two branch receiving units respectively collect light scattering signals in the three preset scattering directions.

3. A method for preventing false alarms in detecting particulate components in air as claimed in claim 2, characterized in that: Step S3 includes: S31 performs ensemble empirical mode decomposition on the collected light scattering signals to obtain a series of intrinsic mode functions IMFs; S32 performs adaptive noise threshold screening on the obtained IMF to remove noise-related IMF; S33 performs sparse representation optimized independent component analysis on the remaining IMFs to separate independent signal components; S34 extracts multi-scale features of independent components based on wavelet packet transform; S35 introduces an adaptive threshold filter, which dynamically adjusts the threshold to select the component most relevant to particle scattering based on the energy distribution, frequency characteristics and wavelet packet characteristics of the signal; S36 uses a fuzzy inference system to evaluate and optimize the selected components; S37 performs adaptive weight reconstruction on the optimized components to obtain the denoised and enhanced scattered light data.

4. A method for preventing false alarms in detecting particulate components in air as claimed in claim 3, characterized in that: In step S35, the calculation process of the adaptive threshold filter is: Calculate the energy E of each independent component x : Where, IC x (t) is the signal value of the xth independent component, T is the signal length, and t is the time variable; Calculate the wavelet packet eigenvalue W x : W x =g(WPT(IC x )) Where WPT is wavelet packet transform, g is feature extraction function; Set the initial threshold θ0 and adjust the threshold according to the energy distribution and wavelet packet characteristics: Where M is the total number of independent components, θ x is the adaptive threshold of the xth independent component.

5. The method for preventing false alarms of detecting particulate components in air according to claim 3, characterized in that: In step S37, the adaptive weight reconstruction process is: Where X(t) is the reconstructed scattered light data, w x (t) is the time-varying weight function of the x-th independent component, IC x (t) is the xth selected independent component, M is the total number of independent components; Time-varying weight function w x The calculation formula of (t) is: In the formula, a x 、b x ,t x 、c x A parameter that is dynamically adjusted according to the output of the fuzzy inference system; a x is the amplitude coefficient, which controls the maximum value of the time-varying weight function; b x is the attenuation coefficient, which controls the rate at which the time-varying weight function changes over time; t x is the central time point of the time-varying weight function; c x is the baseline shift parameter; |tt x | represents the current time point t and the central time point t x The absolute distance between .

6. A method for preventing false alarms in detecting particulate components in air as claimed in claim 3, characterized in that: In step S4, the fast Fourier transform method is used to analyze the processed scattered light data, and the signal strength of the main receiving unit is extracted by calculating the amplitude spectrum of the signal.

7. The method for preventing false alarms of detecting particulate components in air according to claim 2, characterized in that: Step S5 includes: S51 pre-processes the scattered light data of the two branch receiving units to remove background noise and irrelevant signals; S52 extracts time domain features from the preprocessed scattered light data, including calculating the mean, standard deviation, peak value and fluctuation factor of the signal; S53 performs frequency domain feature extraction on the preprocessed scattered light data, including calculating the power spectrum density and main frequency of the signal; S54 calculates an asymmetry factor according to the average value of the scattered light intensities of the two branch receiving units; S55 fuses and reduces the dimensionality of the extracted time domain features, frequency domain features, and asymmetric factors based on a deep learning model to obtain a comprehensive feature vector; S56 uses an adaptive weight algorithm to assign weights to each feature in the comprehensive feature vector and calculates the final particle feature parameters.

8. A method for preventing false alarms in detecting particulate components in air as claimed in claim 7, characterized in that: The deep learning model in step S55 adopts a multi-scale convolutional neural network, and its structure is as follows: The input layer is used to receive time domain features, frequency domain features and asymmetry factors, and combine these features into a multi-dimensional input vector; The multi-scale convolution layer contains three parallel convolution branches, each branch uses a convolution kernel of a different size: Small-scale branches use 1*1 convolution kernels to capture local features; The mid-scale branch uses a 3*3 convolution kernel to capture medium-range feature associations; Large-scale branch, using 5*5 convolution kernel to capture a wider range of feature patterns; Each branch contains two consecutive convolutional layers, each followed by batch normalization and ReLU activation function; The feature fusion layer uses 1*1 convolution to fuse the outputs of the three branches and integrate feature information of different scales; Attention mechanism, which adopts a dual attention mechanism combining channel attention mechanism and spatial attention mechanism. Channel attention uses global average pooling and maximum pooling to calculate the weight of each channel through a multi-layer perceptron. Spatial attention uses convolution operations to calculate the importance of each spatial position. The global pooling layer uses global average pooling to compress the features into a vector of fixed length to obtain global features; Fully connected layer, including two fully connected layers, each layer is followed by a ReLU activation function. The first fully connected layer is used to reduce the feature dimension, and the second fully connected layer is used to compress and nonlinearly transform features; The output layer uses a fully connected layer to generate the final comprehensive feature vector.

9. The method for preventing false alarms of detecting particulate components in air according to claim 7, characterized in that: In step S56, the final particle characteristic parameter calculation formula is: Weight ω i The dynamic update formula is: Where P represents the final particle characteristic parameter; f i represents the i-th eigenvalue; ω i represents the weight of i features; α, β, γ, and δ are adjustment parameters, α adjusts the nonlinear effect of the weighted sum of features, β controls the power of the nonlinear term, γ adjusts the effect of feature differences on the nonlinear term, and δ adjusts the effect of signal statistical characteristics on the final parameters; σ is the standard deviation of the signal; max(F) is the maximum amplitude in the spectrum; ΔA is the fluctuation factor; η is the learning rate; L is the objective function; f mean is the average value of all features; ∈ is a regulation parameter used to control the impact of feature differences on weight update.

10. The method for preventing false alarms of detecting particulate components in air according to claim 7, characterized in that: Step S6 includes: The calculated particle characteristic parameters are compared with the pre-calibrated interference particle characteristic parameters, and the particles are judged to be smoke particles or interference particles according to the comparison results; If the particle is determined to be a smoke particle and the signal strength exceeds the safety threshold, an alarm signal is triggered; If the particle is determined to be an interference particle, no alarm signal is triggered and the test result is recorded to optimize subsequent tests.

Images