A neural network-based method and device for tracking and positioning of an aspirating smoke fire detector

Abstract

The application provides a kind of neural network-based aspirating smoke fire detector tracking positioning method and device, solve the existing aspirating smoke fire detector adopts zoning positioning, positioning range is big, low precision, cannot be positioned to specific sampling hole, difficult to meet the technical problems of fire rapid disposal and linkage fire extinguishing demand.It includes constructing the flow resistance model under the laminar flow condition of sampling pipe;Based on the flow resistance model, a neural network model is constructed with the total number of sampling holes and the total length of the sampling pipe as the input, the equivalent travel of each section as the intermediate layer, and the smoke arrival time as the output;The weight parameters of the neural network model are trained by gradient descent method, and the secondary alarm time after back blowing to clear smoke is used to realize sampling hole level positioning.The application can be widely applied to the field of fire detection technology.

Description

Technical Field

[0001] This application belongs to the field of fire detection technology, and more specifically, relates to a method and device for tracing and locating aspirating smoke detectors based on neural networks. Background Technology

[0002] Aspirating smoke detectors are an important component of the fire alarm system, primarily used in fire alarm applications in large spaces, warehouses, cold storage facilities, data centers, and other similar locations.

[0003] However, because aspirating smoke detectors can only locate specific areas within zones, covering a large area, they cannot pinpoint smaller areas, posing challenges to coordinated fire suppression and rapid fire response. Therefore, there is an urgent need to provide a tracking and location solution to address these issues. Summary of the Invention

[0004] The purpose of this application is to provide a neural network-based tracking and positioning method and device for aspirating smoke detectors, in order to solve the technical problems in the prior art where aspirating smoke detectors use zoned positioning, resulting in a large positioning range, low accuracy, inability to locate to a specific sampling hole, and difficulty in meeting the needs of rapid fire response and coordinated fire suppression.

[0005] To achieve the above objectives, a first aspect of this application provides a neural network-based tracking and localization method for aspirating smoke detectors, comprising the following steps: Construct a flow resistance model for laminar flow conditions in the sampling tube; Based on the flow resistance model, a neural network model is constructed with the total number of sampling holes and the total length of the sampling tube as inputs, the equivalent stroke of each segment as the intermediate layer, and the smoke arrival time as the output. The weight parameters of the neural network model are trained by gradient descent, and sampling hole-level positioning is achieved by utilizing the secondary alarm time after backflushing to clear smoke.

[0006] Preferably, the process of constructing the flow resistance model of the sampling tube under laminar flow conditions includes: determining the airflow state by the Reynolds number; when in laminar flow, deriving the relationship between the gas flow rate of the sampling tube and the pressure difference along the flow path based on Poiseuille's law; introducing the flow resistance along the flow path; determining the relationship between the gas flow rate of the sampling tube, the pressure difference along the flow path, and the flow resistance along the flow path; and obtaining the flow resistance model of the sampling tube along the flow path. Based on Poiseuille's law, the relationship between the pressure difference inside and outside the orifice and the orifice flow rate is derived. By introducing the orifice flow resistance, the relationship between the pressure difference inside and outside the orifice, the orifice flow rate and the orifice flow resistance is determined, and the orifice flow resistance model of the sampling orifice is obtained. The flow resistance model is obtained by combining the friction flow resistance model of the sampling tube and the orifice flow resistance model of the sampling orifice.

[0007] Preferably, the process of constructing the neural network model includes: based on the flow resistance model, the gas path model is equivalent to a circuit model, the flow rate is analogous to the current, the pressure difference is analogous to the voltage, and the flow resistance is analogous to the resistance, and the fluid equation and the iterative relationship of the flow velocity in the sampling tube are derived based on Thevenin's theorem; Based on the flow velocity iteration relationship, and according to the relationship between travel, flow velocity and time, the smoke arrival time is expressed as a weighted sum of the travel distance and the reciprocal of the flow velocity for each segment. A neural network model is constructed with the total number of sampling holes and the total length of the sampling tube as the input layer, the equivalent travel distance of each segment as the intermediate layer, and the smoke arrival time as the output layer.

[0008] Preferably, the process of training the weight parameters of the neural network model includes: using the backpropagation algorithm, taking the squared difference between the output of the neural network model and the actual smoke detection time as the error function, calculating the gradient of the error with respect to each weight through the chain rule, and iteratively updating the weights using the gradient descent method until the error meets the termination condition, thereby obtaining the optimal weight parameter array and completing the training of each weight parameter in the neural network model.

[0009] Preferably, the process of achieving sampling hole-level positioning includes: after the first alarm, the aspirating smoke detector clears residual smoke by backflushing, resumes the aspirating mode, and starts timing; when the second alarm occurs, the actual smoke detection time is recorded, the predicted time of the neural network model is output by retrieving the pre-trained weight array that matches the length and number of holes of the on-site sampling tube, and the prediction time is quickly adapted and biased by combining a small number of feature hole smoke release tests, thereby achieving sampling hole-level smoke tracking and positioning.

[0010] Preferably, the fluid equation within the sampling tube is: Q n+1 +(P a -P n ) / z=(P n -P n-1 ) / Z; In the formula, Q n+1 P represents the flow rate of the sampling tube between sampling orifice n and sampling orifice n+1. a Atmospheric pressure, P n P n-1 These represent the sampling tube pressures at sampling hole n and sampling hole n-1, respectively.

[0011] Preferably, the flow velocity iteration relationship is as follows: V n =[z / (Z+2z)](V n+1 +V n-1 ); In the formula, V n Let V be the airflow velocity in the nth segment of the sampling tube, z be the orifice resistance, Z be the friction loss, and V be the flow velocity. n+1 V is the airflow velocity in the (n+1)th segment of the sampling tube. n-1This represents the airflow velocity in the (n-1)th segment of the sampling tube.

[0012] Preferably, the formula for equivalent stroke is: L n =N×WN n +L×WL n ; In the formula, L n For equivalent stroke, WN n WL represents the weight of the influence of the total number of sampling holes N on the equivalent stroke of the nth segment. n The influence weight of the total length L of the sampling tube on the equivalent stroke of the nth segment.

[0013] Preferably, the formula for the error function is: σ=0.5(T) n ´-T n ) 2 =0.5(T n ´-∑ n i=1 w i L i ) 2 ; In the formula, σ is the difference of squares function, and T n ´ represents the actual counting time of the microcontroller, T n For the model output time of smoke entering at sampling hole n, w i For the i-th equivalent stroke L i Weighting of the impact on output time.

[0014] The second aspect of this application provides a neural network-based aspirating smoke detector tracing and positioning device, comprising: a sampling tube with a plurality of uniformly distributed sampling holes, one end of the sampling tube being connected to an aspirating main unit; The aspirating unit includes a photoelectric maze and an aspirating pump. One end of the photoelectric maze is connected to a sampling tube, and the other end is connected to the aspirating pump.

[0015] The beneficial effects of this application are as follows: This application provides a method and device for tracing and locating aspirating smoke detectors based on neural networks. First, a flow resistance model under laminar flow conditions in the sampling tube is constructed to accurately quantify the transmission characteristics of smoke in the pipeline, laying the foundation for subsequent calculation of smoke transmission time. Next, a neural network model is constructed with the total number of sampling holes and the total length of the sampling tube as inputs, the equivalent travel of each segment as intermediate layers, and the smoke arrival time as the output. This model has a simple structure, is suitable for the embedded operating environment of aspirating smoke detectors, and trains the model weight parameters using the gradient descent method. Simultaneously, it utilizes the secondary alarm time after backflushing to achieve sampling hole-level positioning, ensuring the model's output accuracy while significantly reducing the smoke location range. Furthermore, it can accurately locate specific sampling holes, providing reliable technical support for rapid fire response and precise coordinated fire suppression. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating a neural network-based tracking and localization method for an aspirating smoke detector, as provided in an embodiment of this application. Figure 2 A schematic diagram of the structure of a neural network-based aspirating smoke detector locating device provided in an embodiment of this application; Figure 3 This is a sampling tube gas path model provided in one embodiment of this application; Figure 4 This is a circuit equivalent model diagram of the sampling tube gas path provided in an embodiment of this application; Figure 5 This is a schematic diagram of a neural network model established by a circuit equivalent model, provided as an embodiment of this application.

[0018] In the diagram: 1. Sampling hole; 2. Sampling tube; 3. Photoelectric maze; 4. Suction pump; 5. Suction-type main unit. Detailed Implementation

[0019] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0020] Please see Figure 1The first embodiment of this application provides a neural network-based tracking and localization method for aspirating smoke detectors, comprising: S1: Construct a flow resistance model under laminar flow conditions in the sampling tube.

[0021] Please see Figure 1 The flow rate is calculated based on the pressure difference between each section of the pipeline, and then converted into the flow velocity V inside the pipe from the cross-sectional area of ​​the pipeline. n The transmission time t of the smoke can be calculated.

[0022] Specifically, the time t for the smoke to reach the detection cavity is given by the following formula: t=(L n / V n )+(L n-1 / V n-1 )+(L n-2 / V n-2 )+…+(L2 / V2)+(L1 / V1); In the formula, L n V represents the distance the smoke travels between the nth sampling holes in the sampling tube. n Let be the flow velocity of the smoke between the nth sampling holes in the sampling tube, where n is the total number of sampling holes.

[0023] To achieve the above calculations, this application constructs a fluid dynamic resistance model for the sampling pipe of an aspirating smoke detector to quantitatively analyze the flow velocity V within the pipe. n Relationship with pipeline pressure loss. Flow velocity V inside the pipe. n Related to pressure loss in the pipeline, this pipeline pressure loss consists of two parts: the pressure difference along the flow of gas in the sampling tube (i.e., pressure loss along the flow, including bends), and the pressure difference at the small orifice (sampling hole) when the gas passes through the sampling hole.

[0024] To apply laminar fluid dynamics formulas for pressure loss analysis, the first step is to determine whether the airflow is laminar or turbulent, which can be determined by the Reynolds number Re.

[0025] The Reynolds number Re can be calculated from the kinematic viscosity coefficient ν, using the following formula: Re=VD / ν; ν = µ / ρ; In the formula, µ is the dynamic viscosity coefficient, ρ is the gas density, D is the inner diameter of the sampling tube, and V is the flow velocity inside the sampling tube. After simplification, we get: Re = ρVD / µ; For aspirating sampling pipelines, the length of a single pipe typically does not exceed 120m. According to national standards, the smoke test at the furthest point must trigger an alarm within 120 seconds. Therefore, an average gas flow velocity of not less than 1m / s in the sampling pipe is sufficient to meet the requirements. In this embodiment, a PVC pipe with an outer diameter of 25mm is used as the sampling pipe, with an inner diameter D of 0.02m, a gas density ρ = 1.2kg / m³, and a dynamic viscosity coefficient of air µ = 1.8x10⁻⁶. -5 Pa.s., using the flow velocity V = 1.5 m / s obtained in the sampling tube for calculation, we can get: Re=ρVD / µ=1.2x1.5x0.02 / 1.8x10 -5 =2000; When the Reynolds number is less than 2000, the gas is in a laminar flow state. Therefore, it can be determined that the airflow in the sampling tube of this application is in a laminar flow state.

[0026] Furthermore, this application uses Reynolds number Re calculation and flow rate control to ensure that the gas flow within the sampling tube is in a laminar state, satisfying the laminar fluid dynamics formula. Therefore, based on the laminar fluid dynamics formula, the formula for the gas flow rate Q in the sampling tube and the pressure difference ΔP1 along the flow path can be derived as follows: Q = ΔP1 / (128µL / πD) 4 ); In the formula, ΔP1 is the pressure difference along the pipe, which represents the pressure lost due to friction of the pipe wall as the gas flows from one end to the other in the sampling tube. L is the length of the sampling tube, D is the inner diameter of the sampling tube, and µ is the dynamic viscosity coefficient of the fluid.

[0027] Specifically, after determining the airflow state, it is confirmed that the airflow in the sampling tube is laminar, conforming to Poiseuille's law. Therefore, Poiseuille's law can be used to analyze the pressure difference along the sampling tube. Poiseuille's law describes the relationship between the gas flow rate Q and the pressure difference ΔP1 along the sampling tube when an incompressible fluid undergoes laminar flow in a horizontal circular tube, satisfying the following equation: ΔP1=128µLQ / (πD 4 ); In the formula, Q is the gas flow rate of the sampling tube, D is the inner diameter of the sampling tube, ΔP1 is the pressure difference along the pipe, and µ is the dynamic viscosity coefficient of the fluid (the dynamic viscosity of air is µ = 1.8 x 10⁻⁶). -5 Pa.s), L is the length of the sampling tube, i.e., the length along the sampling path.

[0028] Transforming the above equation, we obtain the following expression for calculating gas flow rate from pressure difference: Q = (πD) 4 / 128µL) ΔP1=ΔP1 / (128µL / πD 4 ); This application introduces the concept of flow resistance in the form of Ohm's law for circuits, defining the flow resistance per unit length as Z, as shown in the following formula: Z=128µL / πD 4 ; Therefore, the relationship between the gas flow rate Q, the friction resistance Z, and the friction pressure difference ΔP1 in the sampling tube can be simplified to: Q = ΔP1 / Z, thus obtaining the friction resistance model of the sampling tube.

[0029] Furthermore, based on the fluid formula for laminar flow, the formula for the pressure difference between the inside and outside of the sampling orifice is derived. The relationship between the gas flow rate q entering the sampling tube through the sampling orifice and the pressure difference ΔP2 between the inside and outside of the sampling orifice is: q=ΔP2 / (128µh / πd 4 ); In the formula, h is the wall thickness of the sampling tube, d is the diameter of the sampling hole, and µ is the dynamic viscosity coefficient of the fluid.

[0030] Specifically, the working principle of aspirating smoke detectors is as follows: a negative pressure is created inside the sampling tube by an air pump. The pressure difference between the atmospheric pressure outside the tube and the negative pressure inside the tube drives air to enter the sampling tube through the sampling holes. When the pressure inside the tube approaches atmospheric pressure, the sampling capacity decreases significantly. Therefore, the size and number of sampling holes directly determine the sampling capacity of the aspirating smoke detector.

[0031] The sampling orifice typically has an opening diameter of 2 mm, while the wall thickness of the sampling tube is approximately 2.5 mm. Therefore, the sampling orifice can be considered equivalent to a short circular tube with a length equal to the wall thickness h and a diameter of d. Using Poiseuille's law, the relationship between pressure difference and flow rate is analyzed as follows: ΔP2=128µhq / (πd 4 ); In the formula, q is the orifice flow rate, d is the diameter of the sampling orifice, ΔP2 is the pressure difference inside and outside the orifice, and µ is the dynamic viscosity coefficient of the fluid (the dynamic viscosity of air is µ = 1.8 x 10⁻⁶). -5 Pa), h is the pipe wall thickness.

[0032] q = (πd) 4 / 128µh) ΔP=ΔP / (128µh / πd 4 ); Wherein, the pressure difference between the inside and outside of the orifice ΔP2=P a -P, P a Where is atmospheric pressure, and P is the pressure inside the sampling tube.

[0033] Similarly, following the circuit principle, the orifice resistance is defined as z = 128µh / πd. 4 Then the flow rate q of the small orifice can be expressed as: q=ΔP2 / z, thus obtaining the flow resistance model of the sampling orifice.

[0034] Furthermore, by calculating the friction resistance of the bend as equivalent to that of a straight pipe, it can be seen that the local resistance of the 90° bend, when converted into the equivalent straight pipe length of the same diameter, is much smaller than the pipe length between sampling holes. Therefore, the influence of the sampling pipe bend on the fluid model can be ignored, thus simplifying the model.

[0035] Specifically, to verify the rationality of simplifying the sampling pipeline into a straight pipe model, this application analyzes the local resistance of the bend. In addition to friction resistance, bends also exhibit local resistance due to changes in flow direction. In engineering, the equivalent length method is commonly used to convert this local resistance into an equivalent straight pipe friction resistance. The equivalent length L... e The localized loss incurred by a pipe fitting (such as an elbow) is equivalent to the length of that straight section of pipe, and is called the equivalent length of the fitting. According to the Darcy-Weisbach formula, the friction loss h of a fluid in a circular pipe... f With local loss h j They are respectively: h f =λ(D / L)(V 2 / 2g); h j =ζ(V 2 / 2g); In the formula, λ is the friction coefficient, ζ is the local resistance coefficient, D is the inner diameter of the sampling tube, L is the length of the sampling tube, V is the flow velocity of the fluid in the sampling tube, and g is the acceleration due to gravity.

[0036] When the airflow in the pipeline is stable, the local loss h j Along the way loss h f Equivalent, i.e., h j =h f (Corresponding straight pipe length is L) e Solve the two equations simultaneously and eliminate V. 2 / 2g term, the equivalent length L can be obtained e The calculation formula is as follows: L e =ζD / λ; For laminar flow in a circular pipe, the friction factor λ = 64 / Re, where Re is taken as 2000. The inner diameter of the sampling pipe is D = 20 mm. For a smooth-walled PVC pipe, the local resistance factor of a 90° standard bend is approximately ζ ≈ 0.25, obtained from a table. Substituting these values ​​into the calculation, we can obtain: L e =ζD / λ=0.25x20x2000 / 64=156mm≈0.15m; Therefore, the local resistance of a single 90° bend is equivalent to the friction resistance of approximately 0.15m of straight pipe. Relative to the typical pipe length of several meters between sampling holes, this equivalent length L... eThe impact is negligible, so the sampling pipeline is simplified to a straight pipe model, which does not affect the accuracy of the flow resistance model.

[0037] S2: Based on the flow resistance model, a neural network model is constructed with the total number of sampling holes and the total length of the sampling tube as inputs, the equivalent stroke of each segment as the intermediate layer, and the smoke arrival time as the output.

[0038] Based on the flow resistance model, this application equates the gas path model to a circuit model: replacing the current in the circuit with gas flow rate, replacing the voltage difference in the circuit with pressure difference, and replacing the flow resistance in the circuit with flow resistance. Therefore, Figure 3 The gas path model is equivalent to Figure 4 The circuit model shown is used to derive the fluid equations within the sampling tube based on Thevenin's theorem: Q n+1 +(P a -P n ) / z=(P n -P n-1 ) / Z; In the formula, Q n+1 P represents the flow rate of the sampling tube between sampling orifice n and sampling orifice n+1. a Atmospheric pressure, P n P n-1 These represent the sampling tube pressures at sampling hole n and sampling hole n-1, respectively.

[0039] Based on Thevenin's theorem, the gas path is equivalent to a circuit model. After establishing the fluid equation in the sampling tube, the iterative relationship of the flow rate is obtained through algebraic derivation. Finally, combined with the relationship between flow rate and velocity Q=VS, the iterative relationship of the velocity is derived.

[0040] Specifically, the flow resistance model along the sampling tube satisfies Q=ΔP1 / Z, and the flow resistance model of the sampling orifice satisfies q=ΔP2 / z. Combined with the variation patterns of flow rate, pressure, and flow resistance, this perfectly conforms to Ohm's law regarding the variation patterns of current, voltage, and resistance. Therefore, using... Figure 4 Circuit model analysis shown Figure 3 The gas path parameters are shown.

[0041] Figure 4 In the diagram, the red arrows represent the direction of gas flow, z1, z2, ... z n+1 For the sampling orifice flow resistance, Z1, Z2, ... Z n+1 Let Q1, Q2, ... Q be the flow resistance per meter along the friction length. n+1 Let q1, q2, ... q be the flow rates in the sampling pipe. n+1 Let z be the flow rate of the sampling orifice. Since the flow resistance of the sampling orifice is only related to the characteristics of the sampling orifice, it can be approximated as a constant, then: z1=z2=z3=…=z n=z; Q1=Q2+q1=Q3+q2+q1=Q4+q3+q2+q1=Q n+1 +q n +...+q3+q2+q1. Based on q=ΔP2 / z and Thevenin's theorem, the relationship between the sampling orifice flow rate and pressure can be obtained as follows: q n =(P a -P n ) / z=P a / zP n / z; In the formula, q n Let P be the airflow rate of the nth sampling orifice. a Atmospheric pressure, P n Let z be the pressure inside the tube on the inside of the nth sampling hole, and z be the flow resistance of a single sampling hole.

[0042] Substituting the relationship between pipeline flow rate and pressure, we get: Q n =Q n+1 +q n =Q n+1 +P a / zP n / z; In the formula, Q n Let Q be the flow rate in the nth sampling tube. n+1 The flow rate in the sampling tube of the (n+1)th segment is denoted as .

[0043] The iterative relationship of pressure can be obtained by rearranging: P n =z(Q n+1 -Q n )+P a ; P n-1 =z(Q n -Q n-1 )+P a ; In the formula, P n Let P be the pressure inside the tube at the nth sampling hole. n-1 The pressure inside the tube at the (n-1)th sampling hole.

[0044] According to the friction loss formula Q=ΔP1 / Z, we can obtain: Q n =(P n -P n-1 ) / Z; Q n =[z / (Z+2z)](Q n+1 +Q n-1 ); In the formula, Z represents the flow resistance per meter.

[0045] If the distance between the two sampling apertures is close, Z can be considered a constant. Then: Q n =(P n -P n-1 ) / Z=[z(Q n+1 -Q n )+P a -z(Q n -Q n-1 )-P a ] / Z; =[zQ n+1 -2zQ n +zQ n-1 )] / Z; ZQ n =zQ n+1 -2zQ n +zQ n-1 ; ZQ n +2zQ n =z(Q n+1 +Q n-1 ); (Z+2z)Q n =z(Q n+1 +Q n-1 ); Q n =[z / (Z+2z)](Q n+1 +Q n-1 ); Combining the relationship between flow rate and velocity, Q=VS (where S is the cross-sectional area of ​​the sampling tube), the iterative relationship of the flow velocity can be obtained: V n =[z / (Z+2z)](V n+1 +V n-1 ); In the formula, V n Let V be the gas flow rate in the nth sampling tube. n+1 V is the gas flow rate in the (n+1)th sampling tube. n-1 The gas flow rate in the (n-1)th sampling tube is denoted as .

[0046] Therefore, the airflow velocity V in the nth segment of the sampling tube can be determined. n There is an iterative correlation between the gas velocity and the velocity of the preceding and following pipe sections, and the overall trend is linear. Therefore, the gas velocity distribution inside the pipe exhibits a linear broken line characteristic.

[0047] Furthermore, from the relationship between time T, distance L, and flow velocity V, T=L / V, we can obtain: The time T for detecting smoke entering through the nth hole is...n =∑L i / V i =∑w i L i Let i = 1, 2, 3, ..., n. Where w i The weight is the reciprocal of the flow velocity w. i =1 / V i .

[0048] Specifically, the time taken for the smoke to be detected in the first hole is: T1 = L1 / V1 = w1L1, where w1 is the weight and L1 is the variable; Time taken for smoke to be detected in the second hole: T2 = L1 / V1 + L2 / V2 = w1L1 + w2L2; The time taken for smoke to be detected in the third hole is: T3 = L1 / V1 + L2 / V2 + L3 / V3 = w1L1 + w2L2 + w3L3; … Time taken for smoke to be detected at the nth hole: T n =∑L i / V i =∑w i L i , i=1,2,3,…,n; Based on the iterative relationship of flow velocity, the airflow velocity V in the nth segment of the sampling tube can be determined. n There is a linear relationship between the flow resistance along the sampling tube and the flow resistance of the sampling orifice, and there is a forward V... n-1 and backward V n+1 The superposition effect is due to the flow velocity V between any two sampling holes. n It will be affected by the total number of sampling holes N, and also by the total length of the pipeline L. Therefore, it can be constructed... Figure 4 The neural network model shown is used to quickly calculate the smoke arrival time. The neural network model includes: The input layer uses the total number of sampling holes N and the total length of the sampling tube L as input parameters, ignoring the influence of the bend in the tube on the model.

[0049] The intermediate layer (hidden layer) consists of N neural units, each corresponding to the equivalent travel L between the 1st and Nth sampling holes. n Equivalent stroke L n The equivalent distance, not the physical distance, is affected by the total number of sampling holes N and the total length of the pipeline L. Its expression is: L n =N×WN n +L×WL n ; In the formula, WN n WL represents the weight of the influence of the total number of sampling holes N on the equivalent stroke of the nth segment. nThe influence weight of the total pipeline length L on the equivalent stroke of the nth segment.

[0050] Output layer, smoke arrival time T at the nth sampling hole n For output, its expression is: T n =∑ n i=1 w i L i ; In the formula, w i For the i-th equivalent stroke L i The weight of the impact on output time is taken as the reciprocal of the flow rate, i.e., w. i =1 / V i The output layer uses the ReLU activation function. When smoke enters the nth sampling hole, only the first n neural units participate in the operation, and the output of subsequent units is 0, so as to realize the time calculation for different smoke entry positions.

[0051] S3: The weight parameters of the neural network model are trained by gradient descent, and the sampling hole level positioning is achieved by utilizing the secondary alarm time after backflushing to clear the smoke.

[0052] After constructing a neural network model, the smoke inlet hole can be accurately located by learning the smoke transmission time characteristics of different sampling holes.

[0053] Specifically, this application uses a tooling system for weight learning training to obtain weight parameters WN for different combinations of sampling tube lengths and the total number of sampling holes. n WL n and w n . Figure 5 The rectangular box at the bottom center represents the backpropagation function, used to optimize the weights WN through error feedback. n WL n and w n This minimizes the deviation between the smoke transmission time output by the model and the actual smoke detection time.

[0054] This application uses the squared difference function σ=0.5(T) n ´-T n ) 2 As the error function between the model output and the actual value, through T´ and T n To train WN using the error n WL n and w n Among them, T n The output time T is the time for the model to output smoke at hole n. n ´ represents the actual counting time of the microcontroller.

[0055] The error function is obtained as follows: σ=0.5(T) n ´-T n ) 2 =0.5(T n ´-∑ n i=1 w i L i ) 2 ; In the formula, σ represents the error.

[0056] Furthermore, σ versus w i The gradient is as follows: ▽σw i =L n (∑ n i=1 w i L i -T n ´); σ is calculated using backpropagation via the chain rule for WN. n WL n The gradient is obtained by the following equation: ▽σWN n =N w n (∑ n i=1 w i L i -T n ´); ▽σWL n =L w n (∑ n i=1 w i L i -T n ´); Then, the gradient descent method is used to update the weights in reverse and iterate, with the following formula: w i+1 =w i -η▽σw i , i=1,2,3,...,n; w i+1 =w i -ηL i (∑w i L i -T n ´), i=1,2,3,...,n; In the formula, η is the learning rate, and ▽σ is the gradient of the error function.

[0057] Therefore, we obtain w i A set of weight arrays, w i =[w1, w2, w3, ..., w n ].

[0058] Using the chain rule to analyze WN n WL n Updates and iterations: WN i+1 =WN i -η▽σWN i , i=1,2,3,...,n; WL i+1 =WL i -η▽σWL i , i=1,2,3,...,n; Get WN i A set of weight arrays, WN i =[WN1, WN2, WN3, ..., WN n ] and WL i A set of weight arrays, WL i =[WL1, WL2, WL3, ..., WL n The left-hand subscript i+1 of the above equal sign indicates iteration over the parameter at the right-hand subscript i.

[0059] Substitute the updated weights back into the equation. Figure 5 Model calculation output value T n Then check if the error function σ decreases and approaches 0. If it approaches 0, record and save this set of updated weight values.

[0060] Output T based on the recorded and saved weight values. n Through T n By establishing a correspondence between the sampling port and the smoke inlet, it can be determined which sampling port the smoke entered, thus achieving smoke localization.

[0061] In an optional embodiment, an error termination value is defined, and the iteration ends when σ is less than or equal to the termination value, thereby obtaining a weight parameter array WN for different combinations of the total number of holes N and the total length L. n Pipe length weight WL n (n=1,2,3,...,N), and the weight w of the smoke travel time between each hole. i (i=1,2,3,...,n; n ≤ N). Based on this weight array, the model outputs the smoke transmission time value T. n The actual smoke detection time value T recorded by the microcontroller n The model closely matches the data, providing a reliable model basis for on-site positioning.

[0062] Aspirating smoke detectors achieve smoke tracking and location through secondary detection. Specifically, after the aspirating smoke detector detects smoke and triggers an alarm for the first time, the controller reverses the airflow in the sampling tube to clear residual smoke through a backflushing function. Then, it resumes normal aspirating mode and simultaneously starts a time counter. When the detector detects smoke again and triggers an alarm, the time count value for that alarm is recorded; this value is the actual smoke detection time T. n ´.

[0063] The sampling tube length L and the total number of sampling holes N at the engineering site are fixed parameters. The operator can input the total sampling tube length L and the total number of sampling holes N into the microcontroller through the panel of the aspirating smoke detector. The microcontroller then retrieves the WN value corresponding to the total sampling tube length L and the total number of sampling holes N, which was pre-trained in the laboratory. n WL n w n Weighted data, through Figure 5 The neural network model outputs T n The value is combined with smoke tracking and positioning to achieve smoke localization. To eliminate positioning deviations caused by changes in atmospheric pressure, temperature, humidity, and construction errors at the engineering site, the T value output by the model can be adjusted. n Offset correction is performed to improve positioning accuracy.

[0064] Specifically, since the weights for different L and N combinations have already been pre-trained in the laboratory, it is not necessary to perform smoke testing on all sampling holes individually on-site. Only five characteristic sampling holes need to be selected for smoke testing to match the optimal weight array. To ensure uniform coverage of pipeline features, the total number of sampling holes must be odd, preferably no less than 5 and no more than 21. The sampling hole positions for the five smoke tests are selected according to the following rules: the first test is the end sampling hole (numbered N), the second test is the beginning sampling hole (numbered 1), the third test is the middle sampling hole (numbered (N+1) / 2), the fourth test is the 1 / 4 position sampling hole (numbered (N+3) / 4), and the fifth test is the 3 / 4 position sampling hole (numbered (3N+1) / 4). The T obtained through the above five smoke tests... n By matching the group with the smallest error in the pre-trained weight array, the model can be quickly adapted to the field, achieving accurate positioning without full-hole testing.

[0065] Please see Figure 2 The neural network-based smoke detector locating device provided in the second embodiment of this application includes: a sampling tube 2, which has a plurality of uniformly distributed sampling holes 1, and one end of the sampling tube 2 is connected to the aspirating host 5. The suction-type main unit 5 includes a photoelectric maze 3 and a suction pump 4. One end of the photoelectric maze 3 is connected to the sampling tube 2, and the other end is connected to the suction pump 4.

[0066] Working principle: When the suction pump 4 in the suction unit 5 is working, a negative pressure is formed in the sampling tube 2. Outside air is drawn into the sampling tube 2 through the sampling hole 1 and flows along the pipeline to the photoelectric labyrinth 3 in the suction unit 5. When the air contains smoke particles, the smoke enters the photoelectric labyrinth 3 with the airflow. The optical sensor inside the photoelectric labyrinth 3 detects the change in smoke concentration, triggers an alarm signal, and completes the basic fire detection.

[0067] Example 1: Verification example of sampling hole positioning. On-site sampling pipes are typically no longer than 120m, and are made of flame-retardant PVC pipe with an outer diameter of 25mm. The sampling orifice diameter is usually 2mm (d), while the pipe wall thickness (h) is approximately 2.5mm. The inner diameter of the sampling pipe is D = 20mm = 0.02m, the air density is ρ = 1.2kg / m³, and the air dynamic viscosity is µ = 1.8 x 10⁻⁶. -5 Pa, the distance between the two sampling holes is 5 meters.

[0068] First, calculate the friction loss Z = 128µL / πD. 4 =128x5x1.8x10 -5 / π0.02 4 =22.93x10 3 .

[0069] The friction loss Z between sampling holes every 5 meters is 23 x 10. 3 Dimensional unit, P a s / m 3 .

[0070] Then, calculate the orifice flow resistance z = 128µh / πd 4 =128 x 1.8 x 10 - 5 x 2.5 x 10 -3 / (3.14x2) 4 x10 -12 ) = 11.5 x 10 4 In this example, the orifice resistance of the sampling hole is z = 115 x 10³, dimensionless, unit: P. a s / m 3 .

[0071] This embodiment uses a sampling tube with a length L = 60 meters (which is actually the sampling tube length in most projects), a sampling hole spacing of 5 meters, and 10 sampling holes as an example for smoke tracking and localization. Figure 5 In the model, L=60 and N=10.

[0072] Taking the smoke entering through the 5th sampling hole as an example, the program calculation process is as follows: Step 1, for Figure 5The weights are given initial values, which can be random or derived from direct smoke testing. Since this embodiment is learning the smoke input from the 5th hole, only 5 numbers are needed, so 5 initial weight values ​​are randomly assigned: w i =[0.2, 0.3, 0.2, 0.1, 0.2]; WN i =[0.7, 0.7, 0.6, 0.8, 0.6]; WL i =[0.5, 0.5, 0.4, 0.6, 0.6]; Next, perform the forward operation: L1 = 10 × 0.7 + 60 × 0.5 = 37; L2 = 10 × 0.7 + 60 × 0.5 = 37; L3 = 10 × 0.6 + 60 × 0.4 = 30; L4 = 10 × 0.8 + 60 × 0.6 = 44; L5 = 10 × 0.6 + 60 × 0.6 = 42; Model output T5=∑w n L n =L1w1+L2w2+L3w3+L4w4+L5w5 =37×0.2+37×0.3+30×0.2+44×0.1+42×0.2=18.8.

[0073] If the actual alarm time detected by the microcontroller is T5´ = 38.8 seconds, substituting it into the error function: σ = 0.5(T5´-T5) 2 =0.5(T5´-∑w n L n ) 2 =0.5 (38.8-18.8) 2 =200, which shows a large error.

[0074] Step 2: Backpropagation iterates through the weights.

[0075] With a learning rate η = 0.0001, the weights are updated according to the iterative formula obtained from the gradient descent method described above. ∑w i L i =L1w1+L2w2+L3w3+L4w4+L5w5=18.8; W 1+1 =w1-ηL1(∑w i L i -T5´)=0.2-0.0001×37(18.8-38.8)=0.274; W 2+1 =w2-ηL2(∑w i L i -T5´)=0.3-0.0001×37(18.8-38.8)=0.374; W 3+1 =w3-ηL3(∑w i L i -T5´)=0.2-0.0001×30(18.8-38.8)=0.260; W 4+1 =w4-ηL4(∑w i L i -T5´)=0.1-0.0001×44(18.8-38.8)=0.188; W 5+1 =w5-ηL5(∑w i L i -T5´)=0.2-0.0001×42(18.8-38.8)=0.284; That is, the first round w i Updated to: w i =[0.274, 0.374, 0.260, 0.188, 0.284].

[0076] Continue with WN i WL i renew: WN i+1 =WN i -ηN w i (∑w i L i -T i ´), i=1,2,3,...,n; WN before the update i The initial value is WN i =[0.7, 0.7, 0.6, 0.8, 0.6]; Specifically, WN 1+1 =WN1-ηNw1(∑w i L i -T5´)=0.7-0.0001×10×0.2(18.8-38.8)=0.704; WN 2+1 =WN2-ηN w2(∑w i L i -T5´)=0.7-0.001×10×0.3(18.8-38.8)=0.706; WN 3+1=WN3-ηN w3(∑w i L i -T5´)=0.6-0.001×10×0.2(18.8-38.8)=0.602; WN 4+1 =WN4-ηN w4(∑w i L i -T5´)=0.8-0.001×10×0.1(18.8-38.8)=0.802; WN 5+1 =WN5-ηN w5(∑w i L i -T5´)=0.6-0.001×10×0.2(18.8-38.8)=0.604; That is, the first round of WN i Updated to: WN i =[0.704, 0.706, 0.604, 0.802, 0.604].

[0077] WL i+1 =WL i -ηL w i (∑w i L i -T n ´), i=1,2,3,...,n.

[0078] WL before the update i The initial value is WL i =[0.5, 0.5, 0.4, 0.6, 0.6].

[0079] WL 1+1 =WL1-ηL w1(∑w i L i -T5´)=0.5-0.0001×60×0.2(18.8-38.8)=0.5024; WL 2+1 =WL2-ηL w2(∑w i L i -T5´)=0.5-0.0001×60×0.3(18.8-38.8)=0.5036; WL 3+1 =WL3-ηL w3(∑w i L i -T5´)=0.4-0.0001×60×0.2(18.8-38.8)=0.4024; WL 4+1 =WL4-ηL w4(∑w iL i -T5´)=0.6-0.0001×60×0.1(18.8-38.8)=0.6012; WL 5+1 =WL5-ηL w5(∑w i L i -T5´)=0.6-0.0001×60×0.2(18.8-38.8)=0.6024; That is, the first round of WL i Updated to: WL i =[0.5024, 0.5036, 0.4024, 0.6012, 0.6024].

[0080] The first round of weight updates is performed using gradient descent and the chain rule: w i =[0.274, 0.374, 0.260, 0.188, 0.284]; WN i =[0.704, 0.706, 0.604, 0.802, 0.604]; WL i =[0.5024, 0.5036, 0.4024, 0.6012, 0.6024]; Then, use these three sets of weights to perform forward calculations on the model output values ​​and verify the reduction in the error function.

[0081] L1=10×0.704+60×0.5024=37.18; L2=10×0.706+60×0.5036=37.27; L3=10×0.604+60×0.4024=30.18; L4=10×0.802+60×0.6012=44.09; L5=10×0.604+60×0.6024=42.18; Model output T5=∑w n L n =L1w1+L2w2+L3w3+L4w4+L5w5 =37.18×0.274+37.27×0.374+30.18×0.260+44.09×0.188+42.18×0.284=10.19+13.94+7.84+8.29+11.98=52.24.

[0082] The error function is verified again: σ = 0.5(T5´-T5) 2=0.5(T5´-∑w n L n ) 2 =0.5(38.8-52.24)2=90.32.

[0083] As can be seen, the error function decreased from 200 to 90.32, indicating that after the above optimization, the model output corresponding to the new weight array is closer to the real situation.

[0084] Furthermore, this error value is compared with the target error value. If it is greater than the target error value, a second and third round of iterations can be performed to obtain a more ideal weight array.

[0085] If the defined target error value is ≤100, it can be seen that after the first round of iteration, the error value of 90.32 has reached the target error value. Therefore, the iteration is stopped and this set of weight values ​​is saved.

[0086] The above routine optimizes the model output value obtained from the smoke intake at the 5th sampling hole. The same method is used to intake smoke at the other 1-10 sampling holes to obtain the optimized weight value array. This optimized weight array will be used for actual engineering positioning.

[0087] The method of positioning is to compare the time count value from smoke entry to alarm with the model value. The model output value of the sampling hole that enters smoke is closest to the count value, and it is determined that smoke has entered the sampling hole, thus realizing the positioning of the smoke entry sampling hole.

[0088] In this example, assuming the actual model output values ​​T6, T5, and T4 in the engineering project are T6=90, T5=52.24, and T4=10.5 respectively, and the alarm count value is T´=38.8, the difference of squares is also used for judgment: (T´-T6) 2 = (38.8-90) 2 =2621.44; (T´-T5) 2 = (38.8 - 52.24) 2 =180.6; (T´-T4) 2 = (38.8 - 10.5) 2 =800.89; Obviously, the alarm count value T´=38.8 is closest to the model value of T5, so the smoke entry position is T5, which means the smoke entered through the 5th sampling hole.

[0089] 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 in 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. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0090] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for tracing and locating aspirating smoke detectors based on neural networks, characterized in that, include: Construct a flow resistance model for laminar flow conditions in the sampling tube; Based on the aforementioned flow resistance model, a neural network model is constructed with the total number of sampling holes and the total length of the sampling tube as inputs, the equivalent stroke of each segment as the intermediate layer, and the smoke arrival time as the output. The weight parameters of the neural network model are trained by gradient descent, and sampling hole-level positioning is achieved by utilizing the secondary alarm time after backflushing and smoke clearing. The process of achieving the sampling hole-level positioning includes: after the first alarm, the aspirating smoke detector clears the residual smoke by backflushing, restores the aspirating mode and starts timing; when the second alarm occurs, the actual smoke detection time is recorded, the pre-trained weight array matching the length and number of holes of the on-site sampling tube is retrieved and the prediction time of the neural network model is output, and the prediction time is quickly adapted and biased by combining a small number of feature hole smoke release tests, so as to achieve sampling hole-level smoke tracking and positioning. The method of positioning is to compare the time count value from smoke entry to alarm with the model value. The model output value of the sampling hole that enters smoke is closest to the count value, and it is determined that smoke has entered the sampling hole, thus realizing the positioning of the smoke entry sampling hole.

2. The neural network-based tracking and localization method for aspirating smoke detectors as described in claim 1, characterized in that, The process of constructing the flow resistance model of the sampling tube under laminar flow conditions includes: determining the airflow state by Reynolds number; when in laminar flow state, deriving the relationship between the gas flow rate of the sampling tube and the pressure difference along the flow path based on Poiseuille's law; introducing the flow resistance along the flow path; determining the relationship between the gas flow rate of the sampling tube, the pressure difference along the flow path, and the flow resistance along the flow path; and obtaining the flow resistance model of the sampling tube along the flow path. Based on Poiseuille's law, the relationship between the pressure difference inside and outside the orifice and the orifice flow rate is derived. By introducing the orifice flow resistance, the relationship between the pressure difference inside and outside the orifice, the orifice flow rate, and the orifice flow resistance is determined, thus obtaining the orifice flow resistance model. The flow resistance model is obtained by combining the friction flow resistance model of the sampling tube and the orifice flow resistance model.

3. The neural network-based tracking and localization method for aspirating smoke detectors as described in claim 2, characterized in that, The process of constructing the neural network model includes: based on the flow resistance model, the gas path model is equivalent to a circuit model, the flow rate is analogous to the current, the pressure difference is analogous to the voltage, and the flow resistance is analogous to the resistance. Based on Thevenin's theorem, the fluid equation and the iterative relationship of the flow velocity in the sampling tube are derived. Based on the aforementioned flow velocity iteration relationship, and according to the relationship between travel distance, flow velocity, and time, the smoke arrival time is expressed as a weighted sum of the travel distance of each segment and the reciprocal of the flow velocity. The neural network model is constructed with the total number of sampling holes and the total length of the sampling tube as the input layer, the equivalent travel distance of each segment as the intermediate layer, and the smoke arrival time as the output layer.

4. The neural network-based tracking and localization method for aspirating smoke detectors as described in claim 1, characterized in that, The process of training the weight parameters of the neural network model includes: using the backpropagation algorithm, taking the squared difference between the output of the neural network model and the actual smoke detection time as the error function, calculating the gradient of the error with respect to each weight using the chain rule, and iteratively updating the weights using the gradient descent method until the error meets the termination condition, obtaining the optimal weight parameter array, and completing the training of each weight parameter in the neural network model.

5. The neural network-based tracking and localization method for aspirating smoke detectors as described in claim 3, characterized in that, The fluid equation within the sampling tube is: Q n+1 +(P a -P n ) / z=(P n -P n-1 ) / Z; In the formula, Q n+1 P represents the flow rate of the sampling tube between sampling orifice n and sampling orifice n+1. a Atmospheric pressure, P n P n-1 Let be the sampling tube pressure at sampling hole n and sampling hole n-1, respectively; z be the flow resistance at the small hole; and Z be the friction resistance along the flow path.

6. The neural network-based tracking and localization method for aspirating smoke detectors as described in claim 3, characterized in that, The velocity iteration relationship is as follows: In n =[z / (Z+2z)](V n+1 +V n-1 ); In the formula, V n Let V be the airflow velocity in the nth segment of the sampling tube, z be the orifice resistance, Z be the friction loss, and V be the flow velocity. n+1 V is the airflow velocity in the (n+1)th segment of the sampling tube. n-1 This represents the airflow velocity in the (n-1)th segment of the sampling tube.

7. The method for tracing and locating aspirating smoke detectors based on neural networks as described in claim 1, characterized in that, The formula for the equivalent stroke is: L n =N×WN n +L×WL n ; In the formula, L n For equivalent stroke, WN n WL represents the weight of the influence of the total number of sampling holes N on the equivalent stroke of the nth segment. n The influence weight of the total length L of the sampling tube on the equivalent stroke of the nth segment.

8. The neural network-based tracking and localization method for aspirating smoke detectors as described in claim 4, characterized in that, The formula for the error function is: σ=0.5(T n ´-T n ) 2 =0.5(T n ´-∑ n i=1 w i L i ) 2 ; In the formula, σ is the difference of squares function, and T n ´ represents the actual counting time of the microcontroller, T n For the model output time of smoke entering at sampling hole n, w i For the i-th equivalent stroke L i Weighting of the impact on output time.

9. A neural network-based tracking and positioning device for aspirating smoke detectors, applied to the neural network-based tracking and positioning method for aspirating smoke detectors as described in any one of claims 1-8, characterized in that, It includes a sampling tube with several evenly distributed sampling holes, and one end of the sampling tube is connected to the air-suction main unit; The air-suction main unit includes a photoelectric maze and an air-suction pump. One end of the photoelectric maze is connected to the sampling tube, and the other end is connected to the air-suction pump.

Images