Dual-optical-path fire smoke detection smoke chamber optimization system and method

Abstract

The invention provides a dual-optical-path fire smoke detection smoke chamber optimization system and method, and relates to the technical field of fire detection. The method specifically comprises the following steps: synchronously acquiring light intensity signals of a main light path and an auxiliary light path, calculating an instantaneous ratio and a change rate of the main light path, and inputting into a weight analyzer to obtain a weight coefficient so as to generate a comprehensive light intensity signal; and extracting time sequence features based on the second time window, and outputting a fire probability and a confidence coefficient in combination with a pre-trained fire analysis model so as to optimize an alarm threshold value of the dual-optical-path fire smoke detection system. According to the invention, the fusion weight of the dual-optical-path signal can be dynamically adjusted, a virtual adjustable scattering angle is designed at the algorithm level, and the fire alarm fixed threshold is dynamically adjusted by comprehensively analyzing the fire probability and the detection confidence coefficient, so that the detection sensitivity of the system to different smog and the anti-interference capability in a complex environment are improved, and the system has high practicability. And meanwhile, the false alarm probability is effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of fire detection technology, specifically to an optimized system and method for dual-optical-path fire smoke detection smoke chamber. Background Technology

[0002] Traditional dual-path photoelectric smoke detectors use two optical paths with fixed scattering angles—a large scattering angle path and a small scattering angle path—and the ratio of these two scattered light signals to distinguish between fire smoke and environmental interference. Compared to single-path solutions, their anti-interference capability has been significantly improved. However, in practical applications, this technology still faces key technical bottlenecks due to the fixed principles of its optical structure.

[0003] In this field, existing technologies suffer from a contradiction between selective blind spots and sensitivity in their response to typical fire smoke. Due to the fixed scattering angle, the detector's response sensitivity to particles of a specific size is physically limited. Large scattering angle optical paths are sensitive to visible white smoke but weakly respond to fully combusted black smoke; small scattering angle optical paths, while having some response to black smoke, typically have low signal strength and are easily overwhelmed by noise. This physical limitation makes it difficult to simultaneously guarantee high sensitivity detection of both black and white smoke under fixed hardware conditions. Furthermore, static judgment logic suffers from insufficient adaptability when facing complex interference. Existing technologies often use fixed thresholds to determine fire alarms; however, the particle size distribution of interfering objects in real-world environments is continuously variable. Static criteria, which apply only one extreme, cannot accurately characterize this continuous and dynamic particulate matter feature. This leads to a frequent contradiction between "improving sensitivity" and "reducing false alarms" in complex application scenarios, hindering the improvement of system reliability.

[0004] Therefore, there is an urgent need for a method that can dynamically optimize the sensing characteristics of the detector from the signal processing level without relying on hardware modifications, in order to overcome the performance limitations imposed by fixed optical structures. Summary of the Invention

[0005] The purpose of this invention is to provide a dual-optical-path fire smoke detection smoke chamber optimization system and method to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: An optimization method for a dual-optical-path fire smoke detection chamber, comprising the following steps: Step 1: Synchronously acquire the raw light intensity signals output by the photoelectric sensors of the main optical path and the secondary optical path arranged in the smoke detection chamber, and preprocess the raw light intensity signals of the main optical path and the secondary optical path respectively to obtain the first light intensity signal and the second light intensity signal. Step 2: Calculate the instantaneous ratio of the first light intensity signal to the second light intensity signal at each measurement moment; construct a first time window of fixed length with the measurement moment as the endpoint, perform linear fitting on the first light intensity signal within the first time window, and use the slope of the fitted line as the rate of change of the first light intensity signal; Step 3: Input the instantaneous ratio and rate of change into a preset weight analyzer. The weight analyzer uses deterministic mapping rules to perform joint analysis on the input instantaneous ratio and rate of change, and outputs weight coefficients for calculating the light intensity fusion of the first light intensity signal and the second light intensity signal at the same moment. Step 4: Based on the weighting coefficients, perform weighted fusion operations on the first and second light intensity signals at each measurement time to obtain a comprehensive light intensity signal that characterizes the comprehensive scattering characteristics inside the smoke chamber; construct a second time window of fixed length with each measurement time as the endpoint, extract a continuous sequence of comprehensive light intensity signals within the second time window, and perform time-series feature analysis on it to obtain judgment feature data; Step 5: Input the comprehensive light intensity signal and judgment feature data at each measurement moment into the preset fire analysis model, output the fire probability and detection confidence at each moment, dynamically correct the fixed alarm threshold of the dual-light path fire smoke detection system based on the obtained fire probability and detection confidence, compare the comprehensive light intensity signal with the corrected alarm threshold, and determine whether a fire alarm is triggered.

[0007] Furthermore, the preprocessing method for the light intensity signal is as follows: the original light intensity signals of the main optical path and the secondary optical path are subjected to low-pass filtering, and their background values ​​calibrated in the smokeless state are subtracted to obtain the first light intensity signal and the second light intensity signal.

[0008] Furthermore, the preset weight analyzer has a built-in weight decision rule library, which is set as follows: multi-level judgment thresholds are set for the instantaneous ratio and rate of change of the light intensity signal, and the range of instantaneous ratio and rate of change is orthogonally divided into multiple discrete two-dimensional intervals using the multi-level judgment thresholds. A mapping relationship table between the instantaneous ratio interval, the rate of change interval and the weight coefficient is established, and the mapping relationship table is solidified into the weight analyzer as the weight decision rule library. For the input instantaneous ratio and rate of change, determine the interval they fall into. If they fall into the valid interval, look up the corresponding weight coefficient according to the mapping relationship table and output it. Otherwise, output the preset default weight coefficient. The effective interval refers to the set of all two-dimensional intervals that are defined by multi-level judgment thresholds and cover the response characteristics of typical fire smoke and non-fire interference events. The boundary of the effective interval is dynamically calibrated and fixed in the rule base based on the historical signal data of the smoke detection chamber and the environmental adaptability test results.

[0009] Furthermore, the composite light intensity signal is a weighted sum of the first light intensity signal and the second light intensity signal. The calculation method is as follows: multiply the first light intensity signal value at the measurement time with the weighting coefficient output by the weighting analyzer to obtain the first weighted component; multiply the second light intensity signal value at the measurement time with the difference between 1 and the weighting coefficient to obtain the second weighted component; add the first weighted component and the second weighted component, and the sum is the composite light intensity signal at the measurement time.

[0010] Furthermore, the specific method for performing time-series feature analysis on the comprehensive light intensity signal sequence is as follows: taking each measurement time as the endpoint, within the second time window, calculate the comprehensive light intensity signal at each measurement time; for the comprehensive light intensity signal sequence extracted based on the second time window, calculate its mean, variance, and first-order difference mean, and use the multidimensional array constructed from the obtained results as the judgment feature data of the comprehensive light intensity signal sequence at each measurement time.

[0011] Furthermore, the fire analysis model is a random forest model, and its pre-establishment process includes: Historical signal data of the smoke detection chamber is acquired, including the main optical path light intensity signal and the secondary optical path light intensity signal synchronously collected in the smoke detection chamber under various typical fire scenarios and non-fire interference scenarios, as well as the tag information corresponding to each sampling time. The tag information includes the fire type, the smoke development stage, and whether it is a real fire event. Historical signal data is preprocessed to calculate the comprehensive light intensity signal at each time point. Based on a second time window, the historical comprehensive light intensity signals are traversed, and temporal features of the comprehensive light intensity signal sequence ending at each time point are extracted. The comprehensive light intensity signal value at each time point is combined with the extracted temporal features to form the structured feature for that time point, and associated with the corresponding label information to construct a structured feature-label sample pair. The structured feature-label sample pair is randomly divided into a training set and a validation set, where the training set is used for model learning and the validation set is used for model performance evaluation. The random forest model is trained based on the training set.

[0012] Furthermore, the specific method for training the random forest model is as follows: during the model inference phase, the decision feature data corresponding to the measurement time is input into the initial random forest model, and each decision tree outputs the category prediction results for the smoke development stage in parallel; the final decision category is determined through a majority voting mechanism; the probability of occurrence of the decision category is calculated based on the voting support rate, and the detection confidence is generated by combining the consistency of the prediction results of each decision tree. After each training iteration, the temporal features of the sample pairs in the validation set are input into the current model to predict their smoke development stage and calculate the fire probability. The prediction results are compared with the true labels to calculate the model's performance metrics on the validation set. These performance metrics include accuracy, recall, comprehensive discrimination ability, and overall distinguishing ability. If the fluctuation of the model's comprehensive discrimination ability on the validation set is less than a preset threshold for three consecutive evaluations, the model training is considered to have converged, training is stopped, and the final fire analysis model is output. Otherwise, training continues.

[0013] Furthermore, the specific method for optimizing the alarm threshold of the dual-path fire smoke detection system is as follows: multiply the product of the fire probability and the detection confidence by the sensitivity coefficient and add 1. The result is used as a dynamic correction factor. The ratio of the original alarm threshold to this dynamic correction factor is the optimized alarm threshold. The sensitivity coefficient is the product of the fire probability weighting coefficient and the detection confidence weighting coefficient; The fire probability weighting coefficient and the detection confidence weighting coefficient are determined by an offline calibration method. The calibration process is as follows: In a standard fire test environment, multiple sets of probability and confidence data under different fire conditions are collected. With the goal of minimizing the system response time and ensuring that the false alarm rate does not exceed the set upper limit, the combination of the fire probability weighting coefficient and the detection confidence weighting coefficient is adjusted within the preset coefficient range using a traversal search method. The combination value that optimizes the overall performance is recorded as the final set value.

[0014] The present invention also provides a dual-optical-path fire smoke detection smoke chamber optimization system, which is used to execute the above-described dual-optical-path fire smoke detection smoke chamber optimization method, including: Data acquisition module: used to synchronously acquire the raw light intensity signals output by the photoelectric sensors of the main optical path and the secondary optical path arranged in the smoke detection chamber, and to preprocess the raw light intensity signals of the main optical path and the secondary optical path respectively to obtain the first light intensity signal and the second light intensity signal; Parameter calculation module: used to calculate the instantaneous ratio of the first light intensity signal to the second light intensity signal at each measurement moment; construct a first time window of fixed length with the measurement moment as the endpoint, perform linear fitting on the first light intensity signal within the first time window, and use the slope of the fitted line as the rate of change of the first light intensity signal; Signal fusion module: used to input the instantaneous ratio and rate of change into a preset weight analyzer. The weight analyzer uses deterministic mapping rules to jointly analyze the input instantaneous ratio and rate of change, and outputs weight coefficients for calculating the light intensity fusion of the first light intensity signal and the second light intensity signal at the same moment. Feature extraction module: Based on the weighting coefficients, it performs weighted fusion operations on the first light intensity signal and the second light intensity signal at each measurement time to obtain a comprehensive light intensity signal that characterizes the comprehensive scattering characteristics inside the smoke chamber; it constructs a second time window of fixed length with each measurement time as the endpoint, extracts a continuous sequence of comprehensive light intensity signals within the second time window, and performs time-series feature analysis on it to obtain judgment feature data; Fire Analysis Module: This module inputs the comprehensive light intensity signal and judgment feature data at each measurement moment into a preset fire analysis model, outputs the fire probability and detection confidence level at each moment, dynamically corrects the fixed alarm threshold of the dual-path fire and smoke detection system based on the obtained fire probability and detection confidence level, and compares the comprehensive light intensity signal with the corrected alarm threshold to determine whether a fire alarm is triggered.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention dynamically generates adaptive fusion weighting coefficients by extracting the instantaneous ratio and rate of change of the dual-path light intensity signals in real time, thus constructing a "virtual adjustable scattering angle" at the algorithm level. When suspected black smoke features appear, the algorithm automatically increases the weight of the small scattering angle light path signal; conversely, it increases the dependence on the large scattering angle light path signal. This dynamic adjustment mechanism enables the system to intelligently optimize its response characteristics, breaking through the limitations of fixed optical structures on sensitivity selection, thereby significantly improving the detection coverage of various fire smoke types, such as black smoke and white smoke, without increasing hardware costs and complexity.

[0016] This invention effectively suppresses interference components that do not match the current smoke characteristics at the signal feature level through dynamic weight analysis, providing a cleaner information foundation for subsequent judgments. Furthermore, the system inputs the fused temporal features into a pre-trained random forest model, outputting a probabilistic fire judgment result and corresponding confidence level. This represents a shift from a simple judgment method based on "whether the light intensity signal is below a threshold" to a more refined decision-making approach that considers "the probability of a real fire event and the potential size of the fire," significantly reducing false alarms caused by non-fire interference such as flying insects and dust, while maintaining relatively high fire identification sensitivity.

[0017] This invention abandons the traditional static threshold determination mechanism and proposes a method for dynamically correcting alarm thresholds based on fire probability and detection confidence. By introducing a sensitivity coefficient and a dynamic correction factor, the system can adaptively adjust alarm sensitivity according to current environmental characteristics and signal quality. This mechanism can respond quickly when fire characteristics are obvious, and remain cautious when signals are ambiguous, thus effectively balancing the contradiction between "sensitivity" and "stability" in complex application scenarios and improving system reliability. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 This embodiment provides a histogram of the statistical distribution of instantaneous ratios of historical data. Figure 3 This embodiment provides a histogram of the statistical distribution of the historical data change rate. Figure 4 This is a spatial distribution diagram of the dual-optical-path signal characteristics provided in this embodiment; Figure 5 The weight coefficient grid mapping table provided in this embodiment; Figure 6 The ROC curve of the random forest model provided in this embodiment; Figure 7 This is a structural block diagram of the overall system of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0020] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0021] Example:

[0022] Please see Figures 1 to 6 The present invention provides a technical solution: An optimization method for a dual-optical-path fire smoke detection chamber, comprising the following steps: Step 1: Simultaneously acquire the raw light intensity signals output by the photoelectric sensors of the main optical path and the secondary optical path arranged in the smoke detection chamber, and preprocess the raw light intensity signals of the main optical path and the secondary optical path respectively to obtain the first light intensity signal and the second light intensity signal.

[0023] The core physical principle of a fire smoke detector is light scattering. The detector's smoke chamber contains a light source and at least one light receiver. When there is no smoke, the light propagates along a preset path, and the light signal received by the receiver is a stable background value. When smoke particles enter the smoke chamber, the light is scattered upon encountering the particles. Some of the scattered light changes direction and is captured by the receiver, causing a change in the receiver's signal strength.

[0024] Smoke particles vary in size, concentration, and refractive index, resulting in different spatial distributions of scattered light. The dual-path system is designed based on this difference in scattering characteristics. It contains two receivers positioned at different scattering angles: the primary receiver at a larger scattering angle and the secondary receiver at a smaller scattering angle. The larger scattering angle path is more sensitive to larger, lighter-colored particles (such as white smoke), while the smaller scattering angle path responds faster to smaller, darker-colored particles (such as black smoke). By simultaneously monitoring these two signals, the system can obtain richer characteristic information about the smoke particles.

[0025] The goal of Step 1 is to acquire a high-quality data foundation. To ensure that the main and secondary optical path signals reflect the state of the same smoke sample at the same time, strict synchronous acquisition is essential. Using a microcontroller (MCU) or a dedicated signal processing chip with a synchronous sample-and-hold analog-to-digital converter, driven by the same sampling clock pulse, the analog voltage signals output by the photodiodes of the main and secondary optical paths are instantaneously captured and converted into digital quantities. The sampling frequency needs to be set according to the dynamic characteristics of the fire smoke.

[0026] Furthermore, the preprocessing method for the light intensity signal is as follows: the obtained light intensity signal is converted from analog to digital to obtain an initial signal sequence; the initial signal sequence is low-pass filtered and its background value calibrated under smokeless conditions is subtracted to obtain a pure optical path signal sequence.

[0027] Raw photoelectric signals acquired directly from sensors are typically mixed with various noises, such as circuit thermal noise, power supply ripple, and ambient light transient interference, and also contain the detector's own background offset. Preprocessing aims to remove unwanted information and inherent system biases to obtain a pure sequence of light intensity changes caused only by smoke.

[0028] The number of bits in an analog-to-digital converter (ADC) determines the quantization accuracy of the signal amplitude. In this field, 12-bit to 16-bit ADCs are typically used. Considering that signal changes caused by smoke can be very weak, especially for black smoke, to obtain sufficient dynamic range and sensitivity to distinguish minute changes in light intensity, this embodiment selects a 14-bit ADC to perform analog-to-digital conversion on the acquired raw signal. A 14-bit ADC can accurately capture the full range of signal changes from weak black smoke to dense white smoke, meeting the wide dynamic range requirements of a dual-optical-path system. Compared to a 12-bit ADC, it has higher signal-to-noise discrimination capability, and it is easier to integrate than a 16-bit ADC. Therefore, this accuracy setting can effectively support the accuracy requirements of subsequent calculations of signal ratios, rates of change, and other characteristics.

[0029] The discrete digital signal sequence obtained after analog-to-digital conversion is used as the initial signal sequence. Low-pass filtering is applied to the initial signal sequence to remove high-frequency noise. This embodiment uses a first-order low-pass filter for noise reduction, and the specific formula is as follows:

[0030] in, Represents the original signal at time t. This represents the denoised signal at time t. This represents the denoised signal from the previous sampling time. The filter coefficients determine the cutoff frequency and response speed of the filter, and their values ​​range from (0,1]. The larger the value, the smaller the filter inertia, the faster it follows signal changes, but the weaker its noise reduction capability. The smaller the value, the better the smoothing effect, but the phase lag will also increase; therefore, a trade-off needs to be made between the filter's response speed and signal smoothness. Based on the spectral characteristics analysis of historical signal data, this embodiment sets the filter's cutoff frequency at a position much lower than half the sampling frequency but higher than the highest effective frequency of smoke change, and derives calculations accordingly. value.

[0031] After filtering and noise reduction, the signal value at each sampling time is subtracted from its optical path calibration background value to achieve signal correction. The background value is the stable non-zero signal output by the smoke chamber detection system in a smoke-free state. By subtracting the background value, the absolute measurement value of the signal can be converted into a relative change, thus making the signal sequence more realistic and direct in reflecting the light intensity changes caused by particulate matter entering the smoke chamber.

[0032] The background value is calculated as follows: After the smoke detection chamber is installed, under the condition that its optical channel is clean and the operating environment is smoke-free, the light intensity signal of the main optical path and the light intensity signal of the secondary optical path are collected over a period of time and low-pass filtered. The arithmetic mean of the filtered main and secondary optical path signal sequences is calculated respectively. The arithmetic mean obtained is the background value calibrated by the main and secondary optical paths respectively.

[0033] After the above-mentioned analog-to-digital conversion, low-pass filtering, and background subtraction operations, the preprocessing of the original signal is completed, thereby obtaining a pure main and secondary optical path light intensity signal sequence.

[0034] Step 2: Calculate the instantaneous ratio of the first light intensity signal to the second light intensity signal at each measurement moment; construct a first time window of fixed length with the measurement moment as the endpoint, perform linear fitting on the first light intensity signal within the first time window, and use the slope of the fitted line as the rate of change of the first light intensity signal.

[0035] The core of this step lies in the shift from "fixed dual-path observation" to "dynamic intelligent sensing." Traditional dual-path detection systems have a fundamental limitation: once the optical scattering angle is fixed, the system's response sensitivity to smoke of specific particle sizes and colors is also fixed, making it unable to dynamically adapt to complex and ever-changing real fire and interference environments.

[0036] To enable the detection system to flexibly adjust the scattering angle without altering the hardware, a signal processing-based "virtual scattering angle" needs to be designed. This involves adaptively fusing the light intensity signals of the main and secondary optical paths, which have fixed physical angles, using real-time calculated weighting coefficients. This simulates the dynamic adjustment effect of the scattering angle at the signal level. These weighting coefficients determine the contribution ratio of different optical path signals to the final synthesized signal, allowing the system to dynamically adjust its equivalent detection angle for different smoke and interference characteristics, thereby intelligently optimizing its sensing performance.

[0037] Although the pre-processed dual-path pure signals reflect the changes caused by smoke, their individual characteristics are isolated. In order to overcome the physical limitations of the fixed optical structure, it is necessary to extract composite features from these two signal sequences that can profoundly reflect the physical nature and dynamic changes of smoke particles.

[0038] This embodiment employs a combination of static and dynamic features to derive the weighting coefficient at the target time: calculating the instantaneous ratio of the light intensity signals at the target time as a static snapshot feature reflecting the smoke type; and calculating the rate of change of light intensity in the signal sequence ending at the target time as a temporal trend feature reflecting the dynamic change pattern of the signal. These two cross-dimensional features are input into a preset weighted analyzer, which comprehensively determines which signal at the target time is more reliable and richer in fire information, and outputs a quantified weighting coefficient. This allows the system to no longer treat the two signals equally or use them fixedly, but rather dynamically and selectively fuses the two light intensity signals based on the real-time sensed physical characteristics, thereby constructing a new and more adaptable "virtual sensing channel" at the information level.

[0039] Furthermore, the instantaneous ratio of the light intensity signals is the ratio of the first light intensity signal to the second light intensity signal at the same moment.

[0040] Particles of different sizes and materials exhibit characteristic ratios in the intensity of scattered light at small forward angles to large lateral angles, and these ratios directly correlate with the physicochemical properties of smoke particles. A high ratio indicates that the smoke is dominated by large-diameter, light-colored particles, such as white smoke; a low ratio indicates that the smoke is dominated by small-diameter, dark-colored particles, such as black smoke. Therefore, the ratio of the dual-path light intensity signals provides the most basic basis for type determination in weighted analysis.

[0041] The formula for calculating the instantaneous ratio of light intensity signal at the target moment is:

[0042] in, This represents the instantaneous ratio of the light intensity signals at time t. This represents the main optical path signal value at time t. This represents the signal value of the secondary optical path at time t.

[0043] The entry of smoke into a smoke chamber is a dynamic, time-series process, and its concentration changes are reflected in the evolution of the light intensity signal over time. Different types of interference events (such as flying insects and steam) and real fire events (such as smoldering and open flames) exhibit distinctly different temporal dynamic characteristics in the time-series data of light intensity signals. The rate of change of light intensity is a quantitative description of the speed and trend of signal change, reflecting the dynamic properties of the event.

[0044] Therefore, this embodiment establishes a first time window and performs linear trend analysis on a segment of main optical path signal data ending at the target time. A main optical path signal sequence of equal length to the first time window is extracted, and a scalar value representing the average rate of change of the main optical path signal within that time period, i.e., the rate of change, is calculated. This feature is used to distinguish between rapid transient events and slow accumulation processes, providing crucial dynamic behavioral information for subsequent weighted analysis.

[0045] Furthermore, the length of the first time window is sufficient to cover the complete duration of a typical instantaneous interference event passing through the detection smoke chamber, and the length of the first time window is adaptively adjusted according to the statistical distribution characteristics of various events in historical signal data; The design of the first time window is based on the temporal physical characteristics of interfering events: transient interferences such as flying insects and steam plumes exist for a short time in the smoke chamber, and the resulting changes in light intensity signals exhibit pulse characteristics of rapid abrupt changes followed by rapid recovery. To accurately capture the complete dynamic process of such events, the length of the first time window needs to be long enough to cover the entire process of the interference from its entry into the detection area to its complete departure, avoiding signal fragmentation due to an excessively short window, which would lead to distortion in feature extraction. At the same time, the window should not be too long to prevent the introduction of excessive background noise or interference from adjacent events, which would affect the sensitivity of the rate of change calculation. Therefore, the window length is set to be slightly longer than the maximum duration of most typical interference events to ensure stable reflection of their transient characteristics. This length is predetermined based on extensive experimental testing and statistical analysis, taking into account the smoke chamber structure, airflow velocity, and motion characteristics of common interference sources for different detector models.

[0046] Regarding the specific time-series feature analysis method, this embodiment employs linear regression to fit the extracted main optical path signal sequence. Compared to the simple difference method, linear regression performs an overall fit based on all data points within the window, effectively smoothing out the interference of single-point random noise and obtaining a more robust trend estimate. The slope of the fitted line intuitively represents the average rise or fall rate of the signal within the window period, which better reflects the overall behavior within the window than simple difference, which relies solely on the difference between the endpoint and the starting point.

[0047] The reason for analyzing only the signal timing characteristics of the main optical path is that the main optical path signal usually has a wider dynamic range and can more clearly reflect the gradient of concentration change. Its rate of change is used as a general representative of the system's dynamic characteristics and has good versatility. In addition, the signal rate of change is only part of the overall system architecture; the characteristic information of the secondary optical path has been incorporated into the decision-making process through the "instantaneous ratio".

[0048] The formula for calculating the rate of change of light intensity signal is as follows:

[0049] Where k is the slope of the fitted line, i.e. the rate of change of the light intensity signal; n is the number of signal data points, i.e., the first time window contains n time points; This represents the numerical value of the data point's sequence number. This represents the signal value at time i.

[0050] After the above steps, a signal feature pair of "instantaneous ratio - rate of change" can be constructed for each target time. In order to comprehensively consider the information of these two dimensions and to make differentiated responses to different types of smoke according to the preset logic, a weight analyzer also needs to be established. Based on the input instantaneous ratio and signal rate of change, after comprehensive analysis, the weight coefficient of the light intensity signal at the target time is output to determine which optical path contains more reliable information features in the currently received signal data.

[0051] Furthermore, the preset weight analyzer has a built-in weight decision rule library, which is set as follows: multi-level judgment thresholds are set for the instantaneous ratio and rate of change of the light intensity signal, and the range of instantaneous ratio and rate of change is orthogonally divided into multiple discrete two-dimensional intervals using the multi-level judgment thresholds. A mapping relationship table between the instantaneous ratio interval, the rate of change interval and the weight coefficient is established, and the mapping relationship table is solidified into the weight analyzer as the weight decision rule library. For the input instantaneous ratio and rate of change, determine the interval they fall into. If they fall into the valid interval, look up the corresponding weight coefficient according to the mapping relationship table and output it. Otherwise, output the preset default weight coefficient. The effective interval refers to the set of all two-dimensional intervals that are defined by multi-level judgment thresholds and cover the response characteristics of typical fire smoke and non-fire interference events. The boundary of the effective interval is dynamically calibrated and fixed in the rule base based on the historical signal data of the smoke detection chamber and the environmental adaptability test results.

[0052] This part of the task involves transforming signal patterns summarized from expert experience and historical data into a set of structured execution rules. This means establishing a weighted decision rule base, and the weight analyzer, based on this rule base, outputs a signal weight coefficient for the input "instantaneous ratio - rate of change". In a dual-path detection system, fire smoke and non-fire interference typically exhibit different, regular distribution patterns in the two-dimensional feature space of "instantaneous ratio - rate of change". The weight analyzer, essentially a feature space mapper, rapidly locates and matches pairs of real-time input feature data, and the output weight coefficient determines the trust ratio of the main and secondary optical path signals.

[0053] The establishment of a weighted decision rule base is based on the extremely high security and reliability requirements of fire alarm systems. Compared to using real-time computation models, table lookup operations do not suffer from model convergence or randomness issues, and offer high interpretability, with each mapping having a clear physical meaning. Furthermore, table lookup is one of the least computationally complex operations, consuming almost no microcontroller computing resources, and boasts extremely fast response times, meeting the requirements of real-time detection. The simple rule base structure also makes it easier to implement and store in resource-constrained embedded systems.

[0054] The weighted decision rule base is established as follows: First, historical light intensity signal data covering various typical scenarios are collected and preprocessed to form a historical sample set. Each sample contains a time series segment and a corresponding scene label. For each sampling moment, its signal feature pair of "instantaneous ratio - rate of change" is calculated and associated with the scene label corresponding to its time series segment. The label information includes fire type, smoke development stage, and whether it is a real fire event.

[0055] See Figure 2 and Figure 3 As shown, statistical analysis was performed on all instantaneous ratios and rates of change, and their respective probability distribution histograms were plotted. Based on the statistical analysis results, multi-level judgment thresholds that can effectively distinguish different event types were set.

[0056] See Figure 4 As shown, the numerical intervals defined by the multi-level threshold for instantaneous ratios are combined with the numerical intervals defined by the multi-level threshold for rate of change through a Cartesian product, and a grid is drawn on a two-dimensional plane. Each grid cell is uniquely determined by a pair of numerical intervals. A two-dimensional array is created to store the weight coefficients corresponding to each grid cell.

[0057] For each grid cell, if a certain number of signal feature data pairs fall into it, the grid cell is considered to belong to the valid interval; if the number of signal feature data pairs falling into the grid cell is lower than the threshold of explicit statistics, the grid cell is considered to belong to the invalid interval, and a status flag is directly assigned to it and stored in a two-dimensional array.

[0058] See Figure 5 As shown, for each effective interval, its corresponding weight coefficient is determined. First, the label information distribution of signal feature data pairs falling into the effective interval is statistically analyzed. An initial weight range is set based on physical principles: for example, if there are more data pairs representing black smoke, a lower weight coefficient is tended to be assigned to enhance the confidence ratio of the secondary optical path signal; if there are more data pairs representing white smoke, a higher weight coefficient is tended to be assigned to enhance the confidence ratio of the primary optical path signal; if there are more data pairs representing interference events, the weight coefficient tends to be a middle value. Within the initial weight range, with the optimization goal of improving the overall alarm performance of the system, the optimal value of the weight coefficient for each effective interval is determined through simulation experiments. In addition, a default weight coefficient needs to be set; when the system encounters unrecognizable feature combinations, a conservative fusion strategy of equal weighting of the primary and secondary optical path signals is adopted.

[0059] The completed two-dimensional array is the mapping table between the instantaneous ratio interval, the rate of change interval and the weight coefficient. This table is then embedded into the weight analyzer as a weight decision rule library.

[0060] When an instantaneous ratio and rate of change are input, the corresponding weight coefficient is read from the mapping table based on the signal feature data pair they constitute. If a valid value is read, it means that the signal feature data pair falls within the valid range, and the corresponding weight coefficient is output. If a status label is read, it means that the signal feature data pair falls within the invalid range, and the preset default weight coefficient is output.

[0061] Step 3: Input the instantaneous ratio and rate of change into a preset weight analyzer. The weight analyzer uses deterministic mapping rules to perform joint analysis on the input instantaneous ratio and rate of change, and outputs weight coefficients for calculating the light intensity fusion of the first light intensity signal and the second light intensity signal at the same moment. After establishing and solidifying the weighted decision rule base, the system possesses the ability to intelligently identify real-time signals. In the aforementioned steps, a weighted mapping relationship table based on instantaneous ratios and rates of change was constructed through historical data statistical analysis and expert experience calibration, and this table was solidified in the weighted analyzer. During actual detection, the corresponding weight coefficients are output from the table based on the signal characteristics at the current test moment, thus achieving a quantitative assessment of the signal confidence level.

[0062] Next, the comprehensive light intensity signal at the measured time needs to be calculated based on the obtained weighting coefficients, which is to achieve data fusion. This step introduces the weighting coefficients as adaptive adjustment factors into the synthesis calculation of the light intensity signal, and performs weighted fusion of the original forward and backscattered light intensities to generate a comprehensive light intensity signal with strong anti-interference ability and sensitive fire response.

[0063] Furthermore, the composite light intensity signal is a weighted sum of the first light intensity signal and the second light intensity signal. The calculation method is as follows: multiply the first light intensity signal value at the measurement time with the weighting coefficient output by the weighting analyzer to obtain the first weighted component; multiply the second light intensity signal value at the measurement time with the difference between 1 and the weighting coefficient to obtain the second weighted component; add the first weighted component and the second weighted component, and the sum is the composite light intensity signal at the measurement time.

[0064] By using a weighted summation, a new signal channel is created at the algorithmic level. The characteristics of this channel are no longer fixed by the physical structure, but are dynamically adjusted by the weighting coefficients, equivalent to a virtual scattering angle that can be adjusted in real time, thus overcoming the limitation of fixed optical structures in terms of specific smoke sensitivity.

[0065] The formula for calculating the combined light intensity signal is:

[0066] in, This represents the combined light intensity signal value at time t. This represents the weighting coefficient of the light intensity signal at time t. This represents the main optical path signal value at time t. This represents the signal value of the secondary optical path at time t.

[0067] When the weighting coefficient approaches 1, the combined signal is almost equivalent to the main optical path signal, and the system is equivalent to a large scattering angle detector, which has the best response to white smoke and large-diameter particles. When the weighting coefficient approaches 0, the combined signal is almost equivalent to the secondary optical path signal, and the system is equivalent to a small scattering angle detector, which has the best response to black smoke and small-diameter particles. When the weighting coefficient approaches the middle value, the system is an equal weighted average of the two signals, which is equivalent to a compromise scattering angle and is often used as a conservative strategy or default state to suppress unknown interference.

[0068] After the above steps, the system successfully transformed from a fixed dual physical optical path into a virtual single optical path with a dynamically adjustable scattering angle.

[0069] Step 4: Based on the weighting coefficients, perform weighted fusion operations on the first and second light intensity signals at each measurement time to obtain a comprehensive light intensity signal that characterizes the comprehensive scattering characteristics inside the smoke chamber; construct a second time window of fixed length with each measurement time as the endpoint, extract a continuous sequence of comprehensive light intensity signals within the second time window, and perform time-series feature analysis on it to obtain judgment feature data; The occurrence of a fire is a dynamic process that evolves over time, and its smoke concentration typically exhibits specific temporal patterns such as accumulation and increase. Conversely, many transient disturbances, such as flying insects and short-lived air currents, may cause signal abrupt changes, but often lack persistence. Therefore, relying solely on the fused signal at a single moment is insufficient for reliable fire identification. To capture the dynamic development characteristics of smoke, it is necessary to observe the development pattern of the integrated light intensity signal over a continuous period ending at the time of measurement, providing crucial dynamic discrimination criteria for subsequent fire analysis models. Therefore, a second time window is established to observe the temporal characteristics of the integrated light intensity signal sequence.

[0070] Furthermore, the length of the second time window is designed to reflect the typical dynamic accumulation time of fire smoke in the initial stage, and its length is adaptively adjusted based on the statistical distribution characteristics of various events in historical signal data. The length of the second time window is a key timescale parameter, and its physical basis is the dynamic characteristics of smoke in the initial stage of a fire. Typically, in the initial stage, the concentration of fire smoke does not reach its peak instantaneously, but rather undergoes a gradual accumulation and increase from nothing to something, a process that usually lasts from several seconds to tens of seconds. Therefore, the value of the second time window length needs to cover the main period of this typical accumulation process to ensure that the temporal features extracted from this window can fully capture the continuous dynamic evolution pattern unique to the initial stage of a fire, thereby effectively distinguishing it from changes in instantaneous interference signals.

[0071] Furthermore, the specific method for performing time-series feature analysis on the comprehensive light intensity signal sequence is as follows: taking each measurement time as the endpoint, within the second time window, calculate the comprehensive light intensity signal at each measurement time; for the comprehensive light intensity signal sequence extracted based on the second time window, calculate its mean, variance, and first-order difference mean, and use the multidimensional array constructed from the obtained results as the judgment feature data of the comprehensive light intensity signal sequence at each measurement time.

[0072] The use of mean, variance, and first-order difference mean as parameters for time-series features is based on the following considerations: the selected features should have significant statistical distinguishing power for different types of events; they should be insensitive to noise and minor perturbations, and stably reflect the essential pattern of the signal; they should have clear physical and statistical significance; and they should be easy to compute in resource-constrained embedded systems. Among these: The average value characterizes the energy level of the composite signal sequence and the persistence of smoke. In the initial stage of a fire, smoke concentration generally rises or remains at a high level, resulting in an average scattered light intensity significantly higher than the background value of the light intensity signal in clean air. Although many transient interferences may also cause transient high signal values, their short duration means that the influence of these transient high signal values ​​on the average value will be diluted after averaging over the entire second time window, thus helping to suppress interference from such transient events. Therefore, a higher average value is a necessary condition for the presence of suspected smoke. The calculation formula is:

[0073] in, This represents the average value of the composite light intensity signal sequence within the second time window, where m represents the number of composite signal data points within the second time window, i.e., the second time window contains m time points. This represents the composite signal value at time i within the second time window. The average value of the composite signal directly reflects the average intensity of the smoke-scattered light during that time period. A high average value indicates the possible presence of persistent smoke, while a low average value indicates the possible presence of weak interference.

[0074] Variance characterizes the intensity and stability of fluctuations in a comprehensive signal sequence. In a typical smoldering fire or a steadily developing open flame, the increase in smoke concentration is relatively gradual and continuous. Reflected in the light intensity signal, the changes are trend-like, rather than characterized by violent, irregular fluctuations. Therefore, real fire smoke signals often exhibit low to medium levels of variance within the observation window, indicating that the signal sequence develops with a certain trend and relatively smooth fluctuations. Conversely, various non-fire interferences can cause sudden and abrupt changes in the signal, leading to a sharp increase in variance. Therefore, when a signal sequence with suspected smoke characteristics appears, variance is also one of the important criteria for determining whether it is a real fire event. The calculation formula is:

[0075] variance The larger the variance, the more drastic the fluctuations in the overall light intensity signal sequence, which may be caused by transient interference, such as flying insects or unstable air currents; conversely, the larger the variance, the more drastic the fluctuations. The smaller the value, the more stable the overall light intensity signal sequence, corresponding to smokeless conditions, stable smoke, or other slowly changing interference.

[0076] The first-order difference mean characterizes the monotonic trend and cumulative direction of the signal. This is a key feature distinguishing the fire growth process from trendless fluctuations. The essence of a fire is the expansion of the combustion scale, causing the smoke concentration to typically exhibit a monotonic or quasi-monotonic increasing trend in the initial stage. This means that positive values ​​dominate in the first-order difference sequence of the signal, making the first-order difference mean significantly positive. This is one of the most crucial dynamic fingerprints of a fire signal. The calculation method is as follows: The first-order difference sequence of the composite light intensity signal sequence, i.e., the change between two adjacent time points, is calculated using the following formula:

[0077] Calculate the first difference mean The formula is:

[0078] If it is positive, it means that the overall light intensity signal sequence shows an upward trend in the second time window, which is one of the typical characteristics of fire smoke concentration accumulation. If it approaches 0, it indicates that the signal sequence has no significant trend and may be in a stable state or experiencing fluctuations or disturbances. If the value is negative, it indicates that the signal sequence is generally trending downwards, possibly indicating that the smoke is dissipating.

[0079] Based on the above temporal feature extraction steps, by calculating the mean, variance, and first-order difference mean of the comprehensive light intensity signal sequence within the second time window, a continuous comprehensive light intensity signal sequence is refined into a low-dimensional feature vector rich in dynamic information. This step realizes the transformation from signal data to abstract features, providing a standardized and computable key input for fire probability judgment based on machine learning.

[0080] Step 5: Input the comprehensive light intensity signal and judgment feature data at each measurement moment into the preset fire analysis model, output the fire probability and detection confidence at each moment, dynamically correct the fixed alarm threshold of the dual-light path fire smoke detection system based on the obtained fire probability and detection confidence, compare the comprehensive light intensity signal with the corrected alarm threshold, and determine whether a fire alarm is triggered.

[0081] This step is the final decision-making and optimization stage of the information processing flow. Its core principle is probabilistic inference and adaptive system adjustment based on machine learning. Two types of data generated in the preceding steps—the comprehensive light intensity signal at the time of measurement and the temporal characteristics of the comprehensive light intensity signal sequence ending at the time of measurement—are input into a pre-trained intelligent model that incorporates fire and interference discrimination knowledge; this is the fire analysis model. This model does not perform simple binary judgment but outputs two key probabilistic indicators: fire probability and detection confidence. Based on these two indicators, the system dynamically adjusts the sensitivity of its final decision, achieving adaptive adjustment to traditional fixed alarm thresholds.

[0082] Furthermore, the fire analysis model is a random forest model. As a model for fire analysis, the random forest model has the following advantages: strong nonlinear discrimination ability, effectively learning the complex nonlinear boundaries of fire and disturbance in the four-dimensional feature space; resistance to overfitting and robustness, the ensemble learning mechanism of the random forest model makes it insensitive to noise and outliers, and has good generalization ability; natural output probability and confidence, the majority voting mechanism of the random forest model can be directly converted into fire probability parameters, and the consistency among its decision trees can be used as a detection confidence parameter; high inference efficiency, the inference process of the trained model is a parallel lookup and voting of multiple decision trees, without the need for complex calculations, making it easy to embed into a limited-scale smoke detection system.

[0083] Furthermore, the pre-establishment process of the fire analysis model includes: Historical signal data of the smoke detection chamber is acquired, including the main optical path light intensity signal and the secondary optical path light intensity signal synchronously collected in the smoke detection chamber under various typical fire scenarios and non-fire interference scenarios, as well as the tag information corresponding to each sampling time. The tag information includes the fire type, the smoke development stage, and whether it is a real fire event.

[0084] Historical signal data is preprocessed to calculate the comprehensive light intensity signal at each time point. Based on a second time window, the historical comprehensive light intensity signals are traversed, and temporal features of the comprehensive light intensity signal sequence ending at each time point are extracted. The comprehensive light intensity signal value at each time point is combined with the extracted temporal features to form the structured feature for that time point, and associated with the corresponding label information to construct a structured feature-label sample pair. The structured feature-label sample pair is randomly divided into a training set and a validation set, where the training set is used for model learning and the validation set is used for model performance evaluation. The random forest model is trained based on the training set.

[0085] Furthermore, the specific method for training the random forest model is as follows: during the model inference phase, the decision feature data corresponding to the measurement time is input into the initial random forest model, and each decision tree outputs the category prediction results for the smoke development stage in parallel; the final decision category is determined through a majority voting mechanism; the probability of occurrence of the decision category is calculated based on the voting support rate, and the detection confidence is generated by combining the consistency of the prediction results of each decision tree.

[0086] For an initial random forest model, all sample pairs in the training set are at the root node of the tree. The model will find an optimal split point as a split threshold based on the structured features of all inputs. After dividing all samples into two groups according to this threshold, the consistency of the data within the two groups can be maximized. After finding the optimal split point, the parameter types in the current structured features and the split threshold of the optimal split point are used together as a split rule and stored in the tree structure.

[0087] Repeat the above splitting process, continuously searching for new optimal split points for other feature parameters, until the stopping condition is met: the number of samples in the current node is lower than the minimum threshold of the dominant statistical distribution. When a node stops splitting, it forms a leaf node. The class of samples that constitutes the absolute majority in each leaf node is recorded as the prediction conclusion for that leaf node. At the same time, the proportion of different samples within that leaf node is also recorded and used to calculate the confidence score. When all nodes on a tree have finished splitting, a decision tree with complete splitting rules is formed. After all decision trees are generated, one round of training is completed, and the random forest model is built.

[0088] After each training iteration, the temporal features of the sample pairs in the validation set are input into the current model to predict their fire probability, label information, and detection confidence.

[0089] The model's performance metrics on the validation set are calculated by comparing the degree of agreement between the predicted results and the true labels. These performance metrics include accuracy, recall, comprehensive discrimination ability, and overall distinguishing ability. If the fluctuation of the model's comprehensive discrimination ability on the validation set is less than a preset threshold for three consecutive evaluations, the model training is considered to have converged, training is stopped, and the final fire analysis model is output. Otherwise, training continues.

[0090] The structured features of the sample pairs in the validation set are input into the random forest model after one round of training. Each tree independently judges each structured feature, and all trees vote to determine the event type, i.e., the label corresponding to the current structured feature. Based on the proportion of votes supporting the obtained label, the fire probability at the corresponding time of the structured feature is output. This enables the current random forest model to complete the label information and fire probability prediction for the structured features of all sample pairs in the validation set. After completing one round of prediction, compare the degree of agreement between the predicted label information and the true label information, and calculate the performance metrics of the current model on the validation set: Associate each sample pair in the validation set with its corresponding predicted label to form a new set, denoted as A. Each sample in set A contains a structured feature, a true label, and a predicted label from a historical sampling time.

[0091] In set A, samples whose predicted labels are the same as their true labels are denoted as set a. The ratio of the number of samples in set a to the total number of samples in set A is the accuracy (ACC) of this validation.

[0092] in, The number of samples in set A represents the total number of structured feature-label sample pairs contained in the validation set; The number of samples in set a is the total number of samples with correctly predicted labels.

[0093] In set A, find all sample pairs whose true label is "fire," denoted as set FR. In set FR, count the samples whose predicted label is also "fire," denoted as set FR_CO. Calculate the ratio of the number of samples in FR_CO to the number of samples in FR. The result is the recall rate RC.

[0094] in, This represents the total number of samples in the validation set that represent real fires. This represents the number of samples that correctly predict a real fire. Recall indicates how many real fire events can be accurately predicted out of all real fire events. In other words, in the set FR, samples that do not belong to the set FR_CO are samples that did not correctly predict a real fire event, i.e., a missed report. A higher recall means a lower probability that a real fire event is missed.

[0095] In set A, find all sample pairs whose predicted label is "fire," denoted as set PFR. Then, count the samples in set PFR whose true label is also "fire," denoted as set PFR_CO. Calculate the ratio of the number of samples in PFR_CO to the number of samples in PFR. The result is the precision COO, i.e.:

[0096] in, This represents the total number of samples that are predicted as fire events by the current model. This indicates how many of the fire event predictions made by the current model are actual fire events. Recall alone cannot accurately measure the predictive ability of the current model. In extreme cases, if the model predicts all sample pairs in the validation set as fires, the recall rate would be 100%, which is clearly illogical. Therefore, it is also necessary to measure the model's prediction precision, i.e., how many of the samples that the model classifies as fire events are actually fires. In other words, in the set PFR, samples that do not belong to PFR_CO are samples that falsely report non-fire interference events as fires.

[0097] Recall measures the probability of a false negative by the current model, while precision measures the probability of a false positive. The harmonic mean of recall and precision is used to calculate the overall discriminative power (F1) of the current model. The formula is as follows:

[0098] The F1 score ranges from 0 to 1. A higher F1 score indicates that the model has achieved a good balance between detecting as many real fires as possible and minimizing false alarms. Therefore, F1 is a core comprehensive indicator for evaluating the performance of fire detection models.

[0099] When the model finishes a round of validation, in addition to outputting the predicted label information, it will also output a predicted fire probability value for all sample pairs in the validation set. Therefore, it is also necessary to evaluate the fire probability prediction capability of the model.

[0100] The fire probability output by the current model is associated with the corresponding real label to construct a probability-label sample pair; the complement of set FR relative to set A is denoted as FFR, which represents all non-fire events in the validation set; Define a fire event counter M1 and a non-fire event counter M2, both with an initial value of 0; Sort all probability-label sample pairs according to the fire probability from high to low, process each probability-label sample pair in turn, and generate a coordinate point for each processed pair: If the true label of the current sample pair is fire, then the value of M1 increases by 1; otherwise, the value of M2 increases by 1. Generate coordinate points based on the current counter value:

[0101] in, This indicates the number of samples in the validation set that characterize non-fire events; This indicates the number of samples in the validation set that characterize fire events.

[0102] After processing, draw a curve based on all generated coordinate points and calculate the area of ​​the region below the curve.

[0103] See Figure 6 As shown, the curve reflects the prediction performance of the current model under different fire probabilities. The closer the curve is to the upper left corner of the image, the higher the performance can be obtained under the same false alarm rate.

[0104] The area below the curve reflects the overall distinguishing ability of the current model, i.e., the AUC value. The closer the value is to 1, the stronger the model's ability to distinguish between fire and non-fire events.

[0105] The F1 score obtained after each round of model validation is used as the main monitoring indicator. However, if the F1 score obtained after three consecutive training sessions remains consistent within a certain range, that is, the difference between the F1 score obtained in each round and the F1 score obtained in the previous round is less than the preset threshold, it indicates that the predictive performance of the model tends to be stable and the performance indicator does not fluctuate significantly. At this time, it can be determined that the model training has converged; the training ends and the final fire analysis model is output.

[0106] When predicting new input samples, the model processes all decision trees in parallel, and each decision tree independently outputs a category prediction based on its internal rules. By statistically analyzing the prediction results of all decision trees, the category with the highest frequency is taken as the model's final prediction label. The proportion of votes for this category is the fire probability output by the model. At the same time, the model generates detection confidence by evaluating the degree of concentration of the prediction results of all decision trees: if all decision trees are highly consistent, the confidence is high; if there is a large divergence between decision trees, the confidence is low.

[0107] Furthermore, the specific method for optimizing the alarm threshold of the dual-path fire smoke detection system is as follows: multiply the product of the fire probability and the detection confidence level by the sensitivity coefficient, and then add 1. The result is used as a dynamic correction factor; where the sensitivity coefficient is the product of the fire probability weighting coefficient and the detection confidence level weighting coefficient; the calculation formula is:

[0108] This is the sensitivity coefficient. This is a fire probability weighting coefficient. The confidence weighting coefficient is used to detect the probability.

[0109] The fire probability weighting coefficient and the detection confidence weighting coefficient are determined by an offline calibration method. The calibration process is as follows: In a standard fire test environment, multiple sets of probability and confidence data under different fire conditions are collected. With the goal of minimizing the system response time and ensuring that the false alarm rate does not exceed the set upper limit, the combination of the fire probability weighting coefficient and the detection confidence weighting coefficient is adjusted within the preset coefficient range using a traversal search method. The combination value that optimizes the overall performance is recorded as the final set value.

[0110] The formula for calculating the dynamic correction factor is:

[0111] PRO is the dynamic correction factor, PRO is the fire probability output by the fire analysis model at the target time, and CFD is the detection confidence output by the fire analysis model at the target time.

[0112] The ratio of the original alarm threshold to the dynamic correction factor is the optimized dynamic alarm threshold.

[0113] The optimized dynamic alarm threshold is TH, which is the original fixed alarm threshold.

[0114] Therefore, for each moment when suspected fire smoke is detected, the system can adjust the alarm threshold in real time based on the fire probability and detection confidence level output by the fire analysis model at that moment. The higher the fire probability and the higher the detection confidence level, the more likely a fire event will occur. In this case, appropriately lowering the alarm threshold can trigger the alarm command more quickly.

[0115] Please see Figure 7 The present invention also provides a dual-optical-path fire smoke detection smoke chamber optimization system, which is used to execute the above-mentioned dual-optical-path fire smoke detection smoke chamber optimization method, including: Data acquisition module: used to synchronously acquire the raw light intensity signals output by the photoelectric sensors of the main optical path and the secondary optical path arranged in the smoke detection chamber, and to preprocess the raw light intensity signals of the main optical path and the secondary optical path respectively to obtain the first light intensity signal and the second light intensity signal; Parameter calculation module: used to calculate the instantaneous ratio of the first light intensity signal to the second light intensity signal at each measurement moment; construct a first time window of fixed length with the measurement moment as the endpoint, perform linear fitting on the first light intensity signal within the first time window, and use the slope of the fitted line as the rate of change of the first light intensity signal; Signal fusion module: used to input the instantaneous ratio and rate of change into a preset weight analyzer. The weight analyzer uses deterministic mapping rules to jointly analyze the input instantaneous ratio and rate of change, and outputs weight coefficients for calculating the light intensity fusion of the first light intensity signal and the second light intensity signal at the same moment. Feature extraction module: Based on the weighting coefficients, it performs weighted fusion operations on the first light intensity signal and the second light intensity signal at each measurement time to obtain a comprehensive light intensity signal that characterizes the comprehensive scattering characteristics inside the smoke chamber; it constructs a second time window of fixed length with each measurement time as the endpoint, extracts a continuous sequence of comprehensive light intensity signals within the second time window, and performs time-series feature analysis on it to obtain judgment feature data; Fire Analysis Module: This module inputs the comprehensive light intensity signal and judgment feature data at each measurement moment into a preset fire analysis model, outputs the fire probability and detection confidence level at each moment, dynamically corrects the fixed alarm threshold of the dual-path fire and smoke detection system based on the obtained fire probability and detection confidence level, and compares the comprehensive light intensity signal with the corrected alarm threshold to determine whether a fire alarm is triggered.

[0116] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0117] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0118] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0119] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An optimization method for a dual-optical-path fire smoke detection chamber, characterized in that, The specific steps include: Step 1: Synchronously acquire the raw light intensity signals output by the photoelectric sensors of the main optical path and the secondary optical path arranged in the smoke detection chamber, and preprocess the raw light intensity signals of the main optical path and the secondary optical path respectively to obtain the first light intensity signal and the second light intensity signal. Step 2: Calculate the instantaneous ratio of the first light intensity signal to the second light intensity signal at each measurement moment; construct a first time window of fixed length with the measurement moment as the endpoint, perform linear fitting on the first light intensity signal within the first time window, and use the slope of the fitted line as the rate of change of the first light intensity signal; Step 3: Input the instantaneous ratio and rate of change into a preset weight analyzer. The weight analyzer uses deterministic mapping rules to perform joint analysis on the input instantaneous ratio and rate of change, and outputs weight coefficients for calculating the light intensity fusion of the first light intensity signal and the second light intensity signal at the same moment. Step 4: Based on the weighting coefficients, perform weighted fusion operations on the first and second light intensity signals at each measurement time to obtain a comprehensive light intensity signal that characterizes the comprehensive scattering characteristics inside the smoke chamber; construct a second time window of fixed length with each measurement time as the endpoint, extract a continuous sequence of comprehensive light intensity signals within the second time window, and perform time-series feature analysis on it to obtain judgment feature data; Step 5: Input the comprehensive light intensity signal and judgment feature data at each measurement moment into the preset fire analysis model, output the fire probability and detection confidence at each moment, dynamically correct the fixed alarm threshold of the dual-light path fire smoke detection system based on the obtained fire probability and detection confidence, compare the comprehensive light intensity signal with the corrected alarm threshold, and determine whether a fire alarm is triggered.

2. The method for optimizing a dual-optical-path fire smoke detection chamber according to claim 1, characterized in that, The specific method for preprocessing the light intensity signal is as follows: the original light intensity signals of the main optical path and the secondary optical path are subjected to low-pass filtering, and their background values ​​calibrated in the smokeless state are subtracted to obtain the first light intensity signal and the second light intensity signal.

3. The method for optimizing a dual-optical-path fire smoke detection chamber according to claim 2, characterized in that, The preset weight analyzer has a built-in weight decision rule library, which is set as follows: multi-level judgment thresholds are set for the instantaneous ratio and rate of change of the light intensity signal, and the range of instantaneous ratio and rate of change is orthogonally divided into multiple discrete two-dimensional intervals using the multi-level judgment thresholds. A mapping relationship table between the instantaneous ratio interval, the rate of change interval and the weight coefficient is established, and the mapping relationship table is solidified into the weight analyzer as the weight decision rule library. For the input instantaneous ratio and rate of change, determine the interval they fall into. If they fall into the valid interval, look up the corresponding weight coefficient according to the mapping relationship table and output it. Otherwise, output the preset default weight coefficient. The effective interval refers to the set of all two-dimensional intervals that are defined by multi-level judgment thresholds and cover the response characteristics of typical fire smoke and non-fire interference events. The boundary of the effective interval is dynamically calibrated and fixed in the rule base based on the historical signal data of the smoke detection chamber and the environmental adaptability test results.

4. The method for optimizing a dual-optical-path fire smoke detection chamber according to claim 3, characterized in that, The composite light intensity signal is a weighted sum of the first light intensity signal and the second light intensity signal. It is calculated as follows: the first light intensity signal value at the measurement time is multiplied by the weighting coefficient output by the weighting analyzer to obtain the first weighted component; the second light intensity signal value at the measurement time is multiplied by the difference between 1 and the weighting coefficient to obtain the second weighted component; the first weighted component and the second weighted component are added together to obtain the composite light intensity signal at the measurement time.

5. The method for optimizing a dual-optical-path fire smoke detection chamber according to claim 4, characterized in that, The specific method for performing time series feature analysis on the comprehensive light intensity signal sequence is as follows: taking each measurement time as the endpoint, within the second time window, calculate the comprehensive light intensity signal at each measurement time; for the comprehensive light intensity signal sequence extracted based on the second time window, calculate its mean, variance, and first-order difference mean, and use the multidimensional array constructed from the obtained results as the judgment feature data of the comprehensive light intensity signal sequence at each measurement time.

6. The method for optimizing a dual-optical-path fire smoke detection chamber according to claim 5, characterized in that, The fire analysis model is a random forest model, and its pre-establishment process includes: Historical signal data of the smoke detection chamber is acquired, including the main optical path light intensity signal and the secondary optical path light intensity signal synchronously collected in the smoke detection chamber under various typical fire scenarios and non-fire interference scenarios, as well as the tag information corresponding to each sampling time. The tag information includes the fire type, the smoke development stage, and whether it is a real fire event. Historical signal data is preprocessed to calculate the comprehensive light intensity signal at each time point. Based on a second time window, the historical comprehensive light intensity signals are traversed, and temporal features of the comprehensive light intensity signal sequence ending at each time point are extracted. The comprehensive light intensity signal value at each time point is combined with the extracted temporal features to form the structured feature for that time point, and associated with the corresponding label information to construct a structured feature-label sample pair. The structured feature-label sample pair is randomly divided into a training set and a validation set, where the training set is used for model learning and the validation set is used for model performance evaluation. The random forest model is trained based on the training set.

7. The method for optimizing a dual-optical-path fire smoke detection chamber according to claim 6, characterized in that, The specific method for training the random forest model is as follows: during the model inference phase, the decision feature data corresponding to the measurement time is input into the initial random forest model, and each decision tree outputs the category prediction results for the smoke development stage in parallel. The final judgment category is determined through a majority voting mechanism; The probability of this decision category is calculated based on the vote support rate, and the probe confidence is generated by combining the consistency of the prediction results of each decision tree. After each training iteration, the temporal features of the sample pairs in the validation set are input into the current model to predict their smoke development stage and calculate the fire probability; the prediction results are compared with the true labels to calculate the model's performance metrics on the validation set, including accuracy, recall, comprehensive discrimination ability, and overall distinguishability. If the model's overall discriminative ability on the validation set fluctuates less than a preset threshold for three consecutive evaluations, the model training is considered to have converged, training is stopped, and the final fire analysis model is output; otherwise, training continues.

8. The method for optimizing a dual-optical-path fire smoke detection chamber according to claim 7, characterized in that, The specific method for optimizing the alarm threshold of the dual-path fire smoke detection system is as follows: multiply the product of the fire probability and the detection confidence by the sensitivity coefficient and add 1. The result is used as a dynamic correction factor. The ratio of the original alarm threshold to this dynamic correction factor is the optimized alarm threshold. The sensitivity coefficient is the product of the fire probability weighting coefficient and the detection confidence weighting coefficient; The fire probability weighting coefficient and the detection confidence weighting coefficient are determined by an offline calibration method. The calibration process is as follows: In a standard fire test environment, multiple sets of probability and confidence data under different fire conditions are collected. With the goal of minimizing the system response time and ensuring that the false alarm rate does not exceed the set upper limit, the combination of the fire probability weighting coefficient and the detection confidence weighting coefficient is adjusted within the preset coefficient range using a traversal search method. The combination value that optimizes the overall performance is recorded as the final set value.

9. A dual-optical-path fire smoke detection smoke chamber optimization system, characterized in that: The dual-optical-path fire smoke detection smoke chamber optimization system is used to execute the dual-optical-path fire smoke detection smoke chamber optimization method according to any one of claims 1-9, including: Data acquisition module: used to synchronously acquire the raw light intensity signals output by the photoelectric sensors of the main optical path and the secondary optical path arranged in the smoke detection chamber, and to preprocess the raw light intensity signals of the main optical path and the secondary optical path respectively to obtain the first light intensity signal and the second light intensity signal; Parameter calculation module: used to calculate the instantaneous ratio of the first light intensity signal to the second light intensity signal at each measurement moment; construct a first time window of fixed length with the measurement moment as the endpoint, perform linear fitting on the first light intensity signal within the first time window, and use the slope of the fitted line as the rate of change of the first light intensity signal; Signal fusion module: used to input the instantaneous ratio and rate of change into a preset weight analyzer. The weight analyzer uses deterministic mapping rules to jointly analyze the input instantaneous ratio and rate of change, and outputs weight coefficients for calculating the light intensity fusion of the first light intensity signal and the second light intensity signal at the same moment. Feature extraction module: Based on the weighting coefficients, it performs weighted fusion operations on the first light intensity signal and the second light intensity signal at each measurement time to obtain a comprehensive light intensity signal that characterizes the comprehensive scattering characteristics inside the smoke chamber; it constructs a second time window of fixed length with each measurement time as the endpoint, extracts a continuous sequence of comprehensive light intensity signals within the second time window, and performs time-series feature analysis on it to obtain judgment feature data; Fire Analysis Module: This module inputs the comprehensive light intensity signal and judgment feature data at each measurement moment into a preset fire analysis model, outputs the fire probability and detection confidence level at each moment, dynamically corrects the fixed alarm threshold of the dual-path fire and smoke detection system based on the obtained fire probability and detection confidence level, and compares the comprehensive light intensity signal with the corrected alarm threshold to determine whether a fire alarm is triggered.

Images