EP3916692B1 - Fire detector; fire detection method with fire detector, computer program and machine-readable storage medium - Google Patents

Description

State of the art

The present invention relates to a method for fire detection with a fire detector, as well as a fire detector set up according to the invention, a computer program and a machine-readable storage medium.

Fire detectors are often designed as scattered light fire detectors, which have a scattered light path, a light source (e.g. LED) and a light sensor (e.g. photodiode). The light source emits light in the IR or VIS range. The light sensor is arranged at a defined angle to the radiation direction of the light source, so that light normally does not hit the light sensor or only to a very small extent. Light is only detected by the light sensor when particles (e.g. smoke or dust) enter the optical path between the light source and the light sensor and scatter the light from the light source onto the light sensor. Depending on the increase in signal, a fire, for example, is detected and issued as an alarm. Additionally or alternatively, other physical variables, e.g. temperature and/or CO content, can be used as criteria for fire detection.

The publication DE 10 2010 041 693 A1 describes a method for testing the functionality of a photoelectric smoke detector with a transmitter element and a sensor element. To test functionality, the transmitting element is activated to emit a test beam and detected by a sensor element as a measurement signal. The recorded measurement signal is compared with a reference signal and, based on this, the functionality of the smoke detector, in particular of the transmitting element and sensor element, is determined.

The determination of fire using such fire detectors is based on the use of a relatively slow increase in the measured variable and the exceeding of a threshold value. As a result, fires are often only detected at a late stage.

The publication US 2018/350220 A1 relates to a method and a device for monitoring an area, wherein signals are received from a smoke detector and one or more minutiae are created based on the signals, one or more time windows are determined based on the minutiae and smoke is detected in the time windows based on the minutiae. or fire types can be determined.

It is therefore desirable to have a method for improved fire detection with a fire detector, in particular in order to be able to detect a fire at an early stage, in particular in the development process, and/or to be able to distinguish between types of fire and/or fire sources. Improved robustness against sources of interference would also be desirable.

Disclosure of the invention

According to the invention, a method for fire detection with a fire detector with the features of claim 1 is proposed. A fire detector, a computer program and a machine-readable storage medium are also proposed. Preferred and/or advantageous embodiments emerge from the subclaims and the description.

A method for detecting fire using a fire detector is proposed. The fire detector is designed to detect a fire, in particular smoke, flames, embers and/or a smoldering fire. The fire detector is preferably designed as an optical fire detector, in particular as a smoke detector with a scattered light detector (scattered light smoke detector). Scattered light smoke detectors use a light detector to measure the light scattered by smoke from a light source, the light detector being arranged in such a way that the light detector can only detect scattered light but not direct light from the light source. The light detector and light source are arranged in a measuring chamber, with the smoke penetrating into the measuring chamber. Alternatively and/or additionally, the fire detector is for detecting the fire based on a thermal quantity, for example the temperature, a weakening, for example of ionizing radiation, and/or a Conductivity developed. In particular, the fire detector can include a camera for image-based fire detection.

The fire detector has at least one sensor device for detecting a measured variable. The sensor device can comprise and/or form a light detection device, a scattered light detection device, a signal attenuation detection device, a temperature detection device, a carbon monoxide detection device and/or other sensor device for detecting a physical and/or chemical quantity. The sensor device is designed to record a measured variable. The measurement variable is, for example, an amount of light, a temperature or an attenuation. The sensor device is designed to output the measured variable as a measurement signal, whereby the measurement signal can include the measured variable in a converted form, for example measuring a quantity of light as a measured variable and outputting it as a voltage and/or current signal.

The sensor devices preferably include a light source and a light sensor. The light source and light sensor are arranged in such a way that the light from the light source is not detected by the light sensor without scattering from particles and/or smoke, and is only detected by the light sensor through scattering of the light emitted by the light source from particles, dirt, moisture and/or smoke becomes. The light sensor is designed in particular to measure a quantity of light and record it as a measurement variable. The recorded measurement variable is output, for example, as voltage and/or current. The output measured variable forms, in particular, the measurement signal. The light sensor is preferably designed as a photodiode. The measurement signal preferably represents the amount of light detected in millivolts. The measurement variable is recorded by the sensor devices in particular continuously and/or cyclically, for example every second or faster. The measurement signal includes in particular the recorded measurement variable and forms, for example, a time course of the measurement variable.

The measurement signal includes in particular fluctuations, noise and/or scatter. The measurement of the measurand is an error-prone measurement, so that, for example, the measurement of a constant quantity also leads to deviations in the recorded measurand, these deviations being, for example, a Have a spread, a standard deviation and/or a variance. The deviations are understood in particular as noise and/or scatter, in particular around the real measured value and/or an average value. The noise and/or scattering is based in particular on electronic noise, measurement accuracies and/or sensitivity of the sensor devices, fluctuations in environmental parameters and/or other sources of error. The noise and/or scattering is in particular a deviation on small time scales, in particular less than 1 second. In particular, the measurement variable is recorded with a sampling rate of less than 1 second, in particular less or in the range of statistical noise and/or scattering. Furthermore, the measurement variable is detected in particular with a resolution of less than 1 V, in particular less than 1 mV, preferably with a resolution less than the statistical and/or electronic noise and/or scattering. In particular, the measurement signal can be a plurality and/or a superposition of different noise and/or scattering, for example the measurement signal is a measured mean value and/or real value plus a first noise, for example electronic noise, plus a second noise, for example a change in the Environmental parameters. The measurement signal is preferably designed as an analog signal, in particular a current or voltage signal. The fluctuations are preferably also based on changes in the measured variable, for example when a fire breaks out.

The method includes detecting the measurement signal from the sensor device and/or detecting multiple measurement signals from multiple sensor devices for and/or over at least one evaluation time interval. The evaluation time interval is in particular designed to include at least 1000, preferably at least 10,000 and in particular at least 100,000 measuring points, with a measuring point describing a measured variable recorded at a time. For example, the measurement signal is recorded for at least 5 minutes, in particular at least 1 hour and in particular at least one day, with the temporal resolution of the measured variable acquisition being, for example, less than or equal to 1 second and in particular less than or equal to 500 milliseconds. For example, the light sensor's photodiode is designed to output a voltage signal, with the resolution and/or scale division of the measurement signal is less than or equal to 1 mV. Preferably, detecting the measurement signal includes storing the measurement signal, in particular as a measurement signal curve.

The method provides a time series analysis for the captured measurement signal and/or for the captured measurement signals. The time series analysis is carried out for the evaluation time interval. For time series analysis, for example, the evaluation time interval is divided into sub-intervals, with the sub-intervals in particular being of the same size. The subintervals in particular have a subinterval length s. For the time series analysis, the time interval is preferably subdivided in different ways, for example for different sub-interval lengths. For example, the time series analysis includes a number of sub-evaluations for different sub-intervals. The time series analysis can in particular be based on known statistical, mathematical and/or stochastic methods and/or models. The time series analysis is preferably carried out using a computer and/or implemented in software. In particular, analysis parameters are determined, calculated and/or estimated through time series analysis. The analysis parameters obtained are, for example, analysis results of time series analysis.

Based on the time series analysis, a fire event is detected and/or recognized. The fire event is preferably detected based on the and/or the analysis parameters obtained. For example, a determination of a deviation of the analysis parameter(s) from a target value or target range is carried out, whereby, for example, if the upward and/or downward deviation is too large, contamination is considered to be detected and/or present.

The invention is based on the idea that by evaluating a measurement signal in an evaluation time interval, detection, in particular early detection, of a fire event, for example in the development phase, is possible. While a current or current assessment of the magnitude of the measured value was previously used to evaluate and/or determine a fire, time series analysis of a measurement signal in a longer time interval can enable more precise, better and more error-free detection of fires.

According to the invention, the time series analysis for the fluctuations is carried out. The method provides for a time series analysis of the course and/or behavior of fluctuations over time. This is based on the idea that the average value or main portion of the measurement signal that was previously used to determine fire is already subject to characteristic fluctuations when a fire occurs. For example, the smoke concentration is not yet sufficient to overwrite a predetermined threshold value, but this small amount of smoke is sufficient for characteristic fluctuations in the measurement signal. By evaluating and time series analysis of these small fluctuations in the form of fluctuations, noise and/or scatter, even minor changes in the fire development process can be detected. For example, it is detected and/or analyzed how the fluctuations, the noise and/or the scatter change over time in the evaluation time interval. For example, a width of the fluctuations, a spread width and/or width of the noise can be determined, analyzed and/or used.

The measurement signal is composed in particular of a dominant mean component, a slowly changing trend component and/or a quasi-periodic trend component. The average proportion corresponds, for example, to a certain particle concentration, which, for example, reacts with a sharp increase in the event of a fire. The slow trend component is based, for example, on contamination, in particular in the form of dust deposits, moisture or aging of the light source and/or the light sensor. A quasi-periodic trend component is understood to mean, for example, a swirling formation of dust, for example due to ventilation, blowing and/or thermals. In particular, the mean share as a trend can have a characteristic functional behavior in the event of a fire, for example an exponential increase. The method provides, for example, that the measurement signal is detrended before the time series analysis. In particular, the detrending of the measurement signal for the acquired measurement signal can take place in the evaluation time interval after the acquisition of the measurement signal but before the time series analysis. The trend can, for example, be linear, exponential, be a quadratic or any polynomial trend. The time series analysis takes place for the trend-corrected measurement signal. Not according to the invention, the trend adjustment can also take place during the actual time series analysis, for this purpose the measurement signal of the evaluation time interval is divided into the sub-intervals and the trend adjustment is carried out for the respective sub-intervals, with the further actual time series analysis, in particular of the fluctuations, the noise and / or the scatter, for and/or in the detrended subintervals. This design is based on the idea that some time series analyses, in particular statistical, stochastic and/or mathematical methods, are not possible for trending variables, signals and/or courses.

The detrending is preferably based on a parameter-free method, for example an empirical mode decomposition (EMD), a Hilbert-Huang transformation and/or a spline approximation. Alternatively and/or additionally, the trend adjustment can be based on a numerical and/or analytical method and/or fit. For example, the analytical context of the trend, e.g. based on a physical law, can be known and used to cleanse the measurement signal of the trend. In particular, this creates a detrended measurement signal that essentially fluctuates around a constant mean, e.g. zero, with the time series analysis evaluating these fluctuations.

According to the invention, the time series analysis is designed as a fluctuation analysis. In particular, the fluctuation analysis is designed to mathematically analyze a time series and/or measurement series, here the measurement signal, and according to the invention to determine and/or quantify a long-term correlation. The fluctuation analysis according to the invention can be implemented analytically and/or numerically. Fluctuation analysis can be designed for autocorrelations and/or cross-correlations. This configuration is based on the idea that changes in the measurement signal when a fire occurs do not necessarily have to be immediately accompanied by a change in the mean value, but can also cause a slow change that remains unnoticed for a long time. For this reason, for example, a correlation of the measurement signal, in particular the Fluctuations, noise and/or scatter are evaluated and used for early fire detection. Smallest changes in Measurement signals that do not yet have an effect on the mean value can already manifest themselves in the fluctuations, in the scatter and/or in the noise. In addition to the noise width, the correlation in particular is a sensitive measure of the changes and can therefore detect changes on the smallest scale at an early stage can.

It is therefore particularly preferred that the time series analysis includes a correlation analysis and/or an autocorrelation analysis. By determining correlations in the measurement signal, physically relevant information about processes, fires, smoke, dust, physical and/or chemical processes in the area surrounding the fire detector can be detected at an early stage without waiting for the slow trend and/or a time-delayed reaction in the average value of the measurement signal must. A method is thus provided that can detect and determine fire events at an early stage. In particular, the time series analysis also excludes processes that have different correlation behavior than fire events as fire events. A method for detecting a fire event that is resistant to false detections and/or false alarms is therefore also provided.

As one of the alternatives, time series analysis includes Hurst analysis. Using the Hurst analysis, an exponent H, the so-called Hurst exponent, is determined as an analysis parameter, for example. For example, a Hurst R/S analysis is carried out as a time series analysis. For this purpose, for example, for the evaluation time interval divided into partial intervals with a partial interval length (s), the range (R) of the cumulative and, if necessary, mean-adjusted time series, here the measurement signal, is determined and set in relation to the standard deviation (S) of the non-cumulative time series, so that R/S is determined. By choosing different observation lengths (subinterval length s) for the subintervals, a functional relationship of R/S can be determined depending on the observation length s. The functional connection is evaluated in particular as a power law R/S - s ^H and H is determined as the Hurst exponent. A deviation of the Hurst exponent from one or more reference values is preferably used to determine the fire event. For example, a first, a second and third reference value is defined, with the first Reference value for example C ₁ =1.5, the second reference value C ₂ =1.0 and the third reference value C ₃ =0.5. In particular, the following can be used for detection: H> C ₁ is smoldering fire, C ₂ <H< C ₁ is open fire, C ₃ <H< C ₂ is frying/deep-frying and H<C ₃ is water vapor.

The Hurst exponent is particularly limited to the range 0<H<1! In particular, the Hurst R/S analysis can only be used for strictly stationary series. The smallest instationarities/trends (e.g. especially when the signal increases in the event of a fire) distort the Hurst exponent. In the case of non-stationary series, the exponent essentially remains at H=1. The DFA, on the other hand, is also suitable for trendy and non-stationary series. Here the correlation exponent α (which in the stationary case is equivalent to the Hurt exponent H) can also display values greater than 1, for example if instead of a trendy noise (e.g. a so-called fractional Gaussian noise) there is actually a movement (e.g. a so-called fractional Brownian motion). ) takes place. In this case the series is said to be unbounded. Depending on the type of event (e.g. a fire), such a case can also occur under certain circumstances, so that it is advisable here, for example, to carry out a DFA and use the correlation exponent α. It may be useful to also carry out a multifractal DFA (MF-DFA) with a variable statistical moment q and to take into account any multifractality that may be present as a further feature, which occurs, for example, if α depends significantly on the statistical moment q. In particular, a multifractality is examined via an MF-DFA rather than via a Hurst R/S analysis, even though α(q) is sometimes referred to in the literature as the "generalized" Hurst exponent H(q). The multifractal correlation exponent α(q) can, but does not necessarily have to be between 0<α(q)<1, but can also be above 1 compared to the exponent from the Hurst analysis. In the case of a non-stationary series, such as a fractional Brownian motion (fBm), a Hurst R/S analysis (after prior detrending if necessary) can be applied to the incremental series (the incremental series simply corresponds to the derivative of the original series, i.e. x _i ←x _i -x _i-1 ). For example, if the case is α=1.7, i.e. a positively correlated fractional Brownian motion, for example in the case of superdiffusion, then the corresponding correlation exponent of the incremental series is just α'=1-α=0.7. This can be used, for example, to check whether there is really a factional Brownian motion is present. If the series has been completely cleaned of trends before the incrementation and if, for example, a Hurst R/S analysis is carried out on this cleaned incremental series, then in this case α'=H also applies to the incremental series. This can be used as a second validation option. In the time range of fire detection, i.e. when the signal has not yet moved significantly above the resting average, a Hurst R/S analysis can be sufficient to detect an event early, i.e. to detect changes in the correlation behavior of the supposed background noise. However, as soon as the signal increases sharply (in the fire classification phase), in addition to the fluctuation superimposed on the increase/trend, there can be either a fractional Gaussian noise with 0<α ≤1 or a fractional Brownian motion α>1. One embodiment of the invention provides that the time series analysis includes a plurality of individual time series analyses. The individual time series analyzes are in particular time series analyzes based on different statistical moments. For example, individual fluctuation, correlation, autocorrelation and/or Hurst analyzes are carried out for different statistical moments q. As explained above, the dependence of the correlation exponent on the statistical moment relates particularly to multifractal DFA. For example, for the commonly used variance and standard deviation, the statistical moment is q=2. Integer values between -10 and 10 are preferably used as statistical moments. For example, the Hurst exponents H are determined as H(q) for the different statistical moments. Such an analysis is particularly referred to as a multifractal spectrum. In particular, the contamination determination and/or operational readiness determination is based on an evaluation of the functional relationship of H(q).

According to the invention, the time series analysis comprises and/or forms a Hurst analysis and/or a detrended fluctuation analysis (DFA) or a multifractal DFA (MF-DFA) with a multifractal exponent.

In particular, at least one analysis parameter is determined by means of the time series analysis, with in one embodiment at least one of the analysis parameters forming and/or describing a scale parameter. For example, based on the analysis parameter, in particular the scale parameter, a distinction is made between types of fire, such as open fire and smoldering fire, and/or between disturbance variables, such as water vapor, fat, dust and/or cigarette smoke, and/or between types of fire and disturbance variables.

One embodiment of the method provides that the measurement signal is recorded for a plurality of evaluation time intervals, for example at least twice, preferably at least ten times and in particular at least 100 times. The evaluation time intervals are preferably connected flush with one another, for example, after the first evaluation time interval has ended, the next evaluation time interval follows directly; alternatively, the evaluation time intervals can have overlaps or can be designed without overlap, so that, for example, there is a pause in the detection between two evaluation time intervals. The individual evaluation time intervals are evaluated using time series analysis. The time series analysis, in particular the determination of the Hurst exponent and/or the analysis parameters, is carried out for the recorded measurement signals of the plurality of evaluation time intervals. For example, the analysis parameters and/or Hurst parameters are compared. Alternatively and/or additionally, a time course, a change, a correlation or a functional connection is determined for the Hurst exponents and/or analysis parameters, the determination of the fire event preferably being based on the comparison, determination and/or evaluation.

The evaluation time intervals are preferably based on a rolling window. Rolling windows are sometimes also referred to as sliding windows. The rolling window has, for example, a fixed interval length, in particular the evaluation time interval length, with the rolling window being shifted to determine the plurality of evaluation time intervals, the shifting representing a temporal shift of the detection point. The rolling window can be moved continuously or discretely, for example with a time offset equal to the evaluation time interval length.

Optionally, an additional environmental variable is determined based on the time series analysis of the measurement signal. The additional environmental variable is, for example, a size and/or evaluation of the environment of the fire detector, for example the air, the temperature and/or lighting conditions. For example, air quality and/or air flow is determined as an additional environmental variable. The air quality can, for example, describe a carbon monoxide content, a carbon dioxide content and/or a dust pollution. The air flow can, for example, describe a draft. This embodiment is based on the idea that, for example, dust particles in and/or around the fire detector lead to fluctuations in the measurement signal, whereby these fluctuations can be used as characteristics for evaluating the air quality using time series analysis. A further object of the invention is a fire detector for detecting a fire event, in particular a fire and/or smoke. The fire detector has a sensor device for detecting a measured variable and for outputting a measurement signal. The fire detector has an evaluation unit, which can be designed with software or hardware technology. According to the invention, the evaluation unit is set up and/or designed to carry out and/or carry out the previously described method. The evaluation unit is designed to detect the measurement signal for the evaluation time interval, in a special embodiment for a plurality of evaluation time intervals. The evaluation unit is designed to carry out the time series analysis for this and/or for the evaluation time intervals. Based on the results, for example the analysis parameter and in particular the Hurst exponent, a fire event is determined by the evaluation unit.

A further subject of the invention is a computer program for execution on a computer and/or the fire detector as previously described. In particular, the computer program includes and/or the computer program is based on a program code with program code means. The computer program is designed to carry out the steps of the method as previously described when executed on the computer and/or the fire detector. In particular, the computer program is implemented in the fire detector, in particular in the evaluation unit, so that the evaluation unit The fire event is detected by carrying out the computer program and thus the procedure.

A further subject of the invention is a machine-readable storage medium, for example a DVD, CD, diskette or other residual storage medium. The computer program, in particular the program code and/or the program code means, is stored on the storage medium.

• Figuren 1a, b einen Brandmelder in unterschiedlichen Zuständen;
• Figur 2a, b, c, d Messsignal und Beispiel einer Zeitreihenanalyse;
• Figuren 3a, b Beispiel eines realen Messsignals und Analyseparameter;
• Figuren 4a, b, c beispielhafte Branderkennung mittels Zeitreihenanalyse;
• Figuren 5a, b, c, d Beispiel einer Branderkennung mittels Rolling Window.

Further advantages, refinements and effects result from the attached figures and their description. Show:

• Figures 1a, b a fire alarm in different states;
• Figure 2a, b , c, d Measurement signal and example of a time series analysis;
• Figures 3a, b Example of a real measurement signal and analysis parameters;
• Figures 4a, b, c exemplary fire detection using time series analysis;
• Figures 5a, b , c, d Example of fire detection using rolling windows.

The Figures 1a and b show an exemplary embodiment of a photoelectric fire detector 1. The fire detector 1 is designed for mounting on a ceiling 2. The fire detector 1 comprises a sensor device, the sensor device comprising a light source 3 and a light detector 4. Furthermore, the fire detector 1 has a measuring channel 5, the measuring channel 5 also being referred to as a scattered light path and/or a smoke chamber. The light detector 4 is designed to detect incident light as a measurement variable and to output a voltage signal as a measurement signal.

The fire alarm 1 in Figure 1a shows the fire detector 1 in a normal state, also called idle state. The normal state is defined as the condition of the fire detector in a fully functional condition without smoke and/or fire, pollution, dust and/or moisture. The light source 3 and light detector 4 are in a common plane, for example (note: the The plane can also run differently) arranged in the same direction as the ceiling 2. The light source 3, for example an LED, emits light, in particular IR radiation and/or visible light. The radiation can in particular be directed or spherical. The light emitted by the light source 3 can be described in the direction of the light detector 4 by a beam path 6. In the present state, no light emitted by the light source 3 is detected by the light detector 3, since the normal beam path 6 does not extend to the detector. Detection could only take place if the emitted light is scattered to the light detector 4. Since in this state there are no particles, such as smoke or dirt, that could scatter the light into the light detector 4, no light is detected by the light detector 4. The measurement signal here is a voltage around a mean zero value, for example 0 mV or 0 mV plus an offset, with the measurement signal having a scatter and/or noise around the mean zero value.

Figure 1b shows fire detector 1 off Figure 1a in the event of a fire. Smoke 7 has now entered measuring channel 5. The smoke 7 particularly includes particles and has reflective properties. The light emitted by the light source 3 is scattered by the smoke 7 in the measuring channel 5, with part of the scattered light following an extended beam path 8 to the light detector 4. The scattered light is detected by the light detector 4 and output as a measurement variable or measurement signal, the measurement signal deviating significantly from the mean zero value. The fire is finally detected based on the measurement signal.

The Figures 2a - d show an example of the principle of a trend-corrected fluctuation analysis of a measurement signal S. The measurement signal S has fluctuations and therefore fluctuates or noises around the value S = 0, the mean zero value of the existing measurement signal S. Although there is no actual trend or course to be seen in the noise itself, Early fire detection can be achieved through time series analysis of the signal, especially the fluctuations and noise. While a minimum amount of smoke is required to scatter the light for a sharp increase in the mean value of the measurement signal S, even the smallest amounts of smoke can lead to noticeable fluctuations in the measurement signal when a fire starts and/or a smoldering fire occurs. By evaluating the fluctuations using time series analysis, fires can be detected early.

In addition to the actual measurement signal S, a cumulative series 9 is also shown in FIG. 6a. The cumulative series 9 is obtained, as in a classic Hurst analysis, by successively adding up the time series, with the time series forming the measurement signal S.

The Figures 2b and 2c They also show the measurement signal S and the cumulative series 9. Now the analysis time interval is divided into equal segments (subintervals) of length s. For the time series analysis, the analysis time interval is divided in different ways, with the types differing in the chosen length s. Figure 2b shows a division into smaller subintervals, i.e. smaller length s, than that in Figure 2c subdivision shown. In each segment, a polynomial of the nth degree is adapted to the time series and subtracted from the cumulative series. The variance in each segment is now determined from the residual obtained in this way and averaged over the number of segments. The root is now taken from this averaged variance (forming the standard deviation). The result is referred to, for example, as the fluctuation parameter F(s).

This is done for the different lengths s and then the functional connection of the form F(s)=Const. ₁ *s ^α +const. ₂ determined, where α essentially corresponds to the Hurst exponent. A logarithmic plot is in Figure 2d shown, where α describes the slope of the straight line. The different lines correspond to different orders n of the fit polynomial.

Figure 3a shows an exemplary time series analysis of a real measurement signal S issued by a fire detector 1. The measurement signal shows an increase as a trend, which is caused, for example, by smoke from an emerging fire. While conventional fire detectors, for example, wait until the measurement signal has exceeded a threshold value, a fire can be detected beforehand according to the method. For this purpose, a time series analysis, in particular a trend-adjusted fluctuation analysis, is carried out for the measurement signal.

The result for some fluctuation analyzes is in Figure 3b shown. The Fluctuation analyzes are based on trend adjustment with different polynomial degrees n of the fit polynomial. From this context, for example, the Hurst exponent H, understood here as the scale parameter α , will be determined as the analysis parameter. Based on the scale parameter α, a fire is concluded and/or discrimination is made between types of fire.

Figures 4a, b, c show schematically a time series analysis of the measurement signal S in a rolling window analysis. Figure 6a shows the course of a fire event Z as a time course. The fire event starts suddenly at a time to. Figure 4b shows the associated measurement signal S from fire detector 1 in the same period. Unlike the fire event Z, the measurement signal does not change suddenly at to, but responds with a time delay with τ ₁ . Only after t ₀ + τ ₁ does the measurement signal S, or its mean value, exceed the threshold value X. Figure 4c shows the analysis parameter α , for example the Hurst exponent. A change in the analysis parameter α can be noticed after a time delay τ ₂ , where τ ₂ « τ ₁ . Time series analysis enables earlier fire detection.

Figures 5a-d show an example of a time series analysis of the measurement signal S based on a rolling window. The measurement signal S in 5a is based, for example, on the formation of a smoldering fire at time to. As in to Figures 4 explains: The measurement signal S or its mean value only reacts with a time delay with an increase.

Figure 5b shows the detrended measurement signal S* for the measurement signal S Figure 5a . This essentially forms a fluctuation around a constant value, here S=125. The time series analysis is carried out for the detrended measurement signal S*. For this purpose, the rolling window method is used, whereby an analysis time interval is shifted 11 in time as a “time window”. A first analysis time interval A ₁ and a second analysis time interval A ₂ are shown. The analysis time intervals A _1, A ₂ are of the same length and only offset in time. For further analysis, the analysis time interval A ₂ is further shifted in time.

The Figures 5c and 5d show the analysis parameters for the time series analyses, where Figure 5c shows the analysis parameters α for the analysis time interval A ₁ and Figure 5d shows the analysis parameters α for the analysis time interval A ₂ . The analysis parameter a differs significantly for the two analysis time intervals A ₁ , A ₂ , where, for example, as shown in the schematic representation, α = 0.56 for A ₁ and α = 0.68 for A ₂ . Based on the analysis parameters α or the time series analysis, fire determination is already possible before the mean value of the measurement signal as in Figure 5a exceeds a threshold value X.

Claims (12)

1. Method for fire detection using a fire detector (1),

   wherein the fire detector (1) comprises a sensor device for recording a measurement variable and for outputting a measurement signal (S), wherein the measurement signal (S) has fluctuations,

   wherein the method comprises the following steps:

   - recording the measurement signal (S) of the sensor device for an evaluation time interval,

   - carrying out a time series analysis for the measurement signal (S) in the evaluation time interval, wherein the measurement signal (S) is detrended before the time series analysis, wherein the time series analysis is carried out for the detrended measurement signal,

   - detecting a fire event on the basis of the time series analysis, wherein the time series analysis is designed as a fluctuation analysis, wherein the fluctuation analysis is designed to mathematically analyse the measurement signal and to determine and/or to quantify a long-term correlation,

   wherein the time series analysis comprises and/or forms a Hurst analysis and/or a detrended fluctuation analysis DFA or a multifractal detrended fluctuation analysis MF-DFA with a multifractal exponent α(q).
2. Method according to Claim 1, characterized in that the measurement signal (S) has a dominant mean value component, a slow trend component and/or a quasi-periodic trend component.
3. Method according to Claim 2, characterized in that the detrending is based on an empirical mode decomposition, a Hilbert-Huang transformation and/or a spline approximation.
4. Method according to any of the preceding claims, characterized in that the time series analysis comprises a plurality of individual time series analyses, wherein the individual time series analyses are based on different statistical moments and/or are based on different degrees of a polynomial fit.
5. Method according to any of the preceding claims, characterized in that at least one analysis parameter is determined by means of the time series analysis, wherein at least one of the analysis parameters describes a scale parameter •, wherein a distinction is drawn between fire and disturbance and/or between types of fire and/or types of disturbance on the basis of the scale parameter •.
6. Method according to any of the preceding claims, characterized in that the measurement signal (S) of the sensor device is recorded for a plurality of evaluation time intervals (A₁, A₂), wherein the time series analysis is carried out for each of the measurement signals (S) of the evaluation time intervals (A₁, A₂), wherein the fire event is determined on the basis of a comparison of the plurality of time series analyses.
7. Method according to Claim 6, characterized in that the plurality of evaluation time intervals (A₁, A₂) are based on a rolling window.
8. Method according to any of the preceding claims, characterized in that a distinction between the types of fire: smouldering fire, open fire and/or deflagration is drawn on the basis of the time series analysis.
9. Method according to any of the preceding claims, characterized in that an additional environment variable, in particular an air quality, is determined on the basis of the time series analysis.
10. Fire detector (1) for detecting a fire event, a fire and/or smoke on the basis of a measurement variable and/or a measurement signal (S), wherein the fire detector (1) comprises a sensor device for recording the measurement variable and for outputting the measurement signal (S), wherein the measurement signal (S) has fluctuations, noise and/or dispersion, with an evaluation unit, wherein the evaluation unit is designed to carry out the method according to any of the preceding claims, wherein the evaluation unit is designed to record the measurement signal (S) for an evaluation time interval, to carry out a time series analysis of the measurement signal (S) in the evaluation time interval and to determine a fire event on the basis of the time series analysis.
11. Computer program for execution on a computer and/or the fire detector (1) according to Claim 10, characterized in that the computer program is designed to carry out the steps of the method according to any of Claims 1 to 9 upon execution on the computer and/or the fire detector (1).
12. Machine-readable storage medium, in particular nonvolatile machine-readable storage medium, on which the computer program according to Claim 11 is stored.

Images