Method and device for the detection of forest fires

Abstract

The invention relates to a method for determining ML data (machine learning data) for a system for forest fire early detection comprising the following method steps: recording first measurement data from measurements of first parameters of a forest; recording second measurement data from measurements of second parameters of a forest; generating first ML data from the first measurement data; and generating second ML data from the second measurement data, wherein the second ML data is different from the first ML data. The invention also relates to a forest fire early detection system comprising a network having a first terminal, wherein the first terminal has a sensor device, an evaluation device for evaluating measurement signals provided by the sensor device, a device for power supply, a first control device, wherein the first control device has a first memory unit, wherein a first ML data set is stored on the first memory unit, a second terminal, wherein the second terminal has a sensor device, an evaluation device for evaluating measurement signals provided by the sensor device, a device for power supply, a second control device, wherein the second control device has a second memory unit, wherein a second ML data set is stored on the second memory unit, wherein the second ML data set is different from the first ML data set, as well as a network server.

Description

[0001] V E R F A H R E N U N D VO R R I C H T U N G Z U R D E T E K T I O N VO N WA L D B RÄ N D E N

[0002] The invention relates to a method for determining ML data (machine learning data) for a system for early forest fire detection, comprising the method steps of acquiring first measurement data from measurements of first parameters of a forest, acquiring second measurement data from measurements of second parameters of a forest, creating first ML data from the first measurement data, and creating second ML data from the second measurement data, wherein the second ML data is different from the first ML data.The invention further relates to a forest fire early detection system with a network having a first terminal, wherein the first terminal has a sensor device, an evaluation device for evaluating measurement signals supplied by the sensor device, a device for supplying energy, a first control device, wherein the first control device has a first memory unit, wherein a first ML data set is stored on the first memory unit, a second terminal, wherein the second terminal has a sensor device, an evaluation device for evaluating measurement signals supplied by the sensor device, a device for supplying energy, a second control device, wherein the second control device has a second memory unit, wherein a second ML data set is stored on the second memory unit, wherein the second ML data set is different from the first ML data set, and a network server.

[0003] State of the art

[0004] The larger a forest fire, the more difficult it is to determine its direction and speed of spread. Weather, wind, soil conditions, and vegetation determine its path and speed of spread, which can change quickly. It is therefore very important to detect a forest fire very early in order to minimize damage and keep the forest fire controllable, or to give the fire service a decisive time advantage. During a forest fire, the complex thermal decomposition processes (distillation, pyrolysis, charring, and the oxidation of the gas products produced during flame combustion) occur simultaneously and often in close proximity to one another. The thermal decomposition of fuels occurs in front of and along the fire line, while enclaves with intermittent open flames often persist far behind the flame front.

[0005] Flame combustion generally occurs between 800°C - 1200°C. Smoldering ground fires occur between 300°C - 600°C. Combustible gases, particularly volatile organic compounds (VOCs), are formed more quickly at temperatures above 200°C and reach their peak at 320°C. VOCs are the collective term for organic, carbon-containing substances that evaporate into the gas phase at room temperature or higher temperatures, particularly terpenes. Various organic compounds are also formed, such as methanol, carbon dioxide, carbon monoxide, and molecular hydrogen. Flammable combustion only begins at 425°C to 480°C. Flame temperatures of 700°C to 1300°C are the most common. In this temperature range, carbon dioxide, nitrogen oxides, and volatile sulfur-containing compounds (VSCs), particularly sulfur dioxide, are mainly formed. Smoldering fires spread slowly, approx.3 cm / h, they can generate ground temperatures above 300°C for several hours with peak temperatures of 600°C.

[0006] The following table shows the gases formed during a forest fire, graded by temperature:

[0007] Earth observation data, particularly in the form of aerial and satellite imagery, can potentially help in detecting wildfires. The sharp increase in available earth observation data, particularly aerial and satellite imagery, enables widespread detection of wildfires. While satellite data are useful for detecting and fighting fires, they have one drawback: They usually reach emergency responders only with a delay. Geostationary satellites, due to their great distance, provide only low image resolutions, and non-geostationary satellites must orbit the Earth before they can provide new images.

[0008] Another option for detecting forest fires is to install a network of gas sensors directly in the forest. These sensors detect gases released during the development of forest fires, thus enabling early detection of forest fires before they can be detected by optical systems from a distance. However, due to the varying vegetation types and soil compositions in forests, different gases and gas concentrations are produced, making accurate detection very difficult. Furthermore, the increasing temperatures during the different phases of forest fire development alone result in different gases and gas concentrations. However, since deploying emergency personnel is very expensive, the detection accuracy needs to be improved.

[0009] It is therefore an object of the present invention to provide a method for implementing ML data in a system for early forest fire detection and / or early forest fire detection, which works reliably, is expandable as required and is cost-effective in installation and maintenance.

[0010] It is also an object of the invention to provide a forest fire early detection system that operates reliably, can be expanded as required and is cost-effective to install and maintain.

[0011] The problem is solved by means of the method for determining ML data for a forest fire early detection system. Advantageous embodiments of the invention are set forth in the following subclaims.

[0012] The inventive method for determining ML data for a system for early forest fire detection has four method steps: In the first method step, first measurement data are determined from measurements of first parameters of a forest. Such parameters include, for example, temperature, humidity, wind direction and strength and / or the composition of gases. First measurement data are created from these parameters. In the second method step, second measurement data are determined from measurements of second parameters of a forest. Second measurement data are created from the measured second parameters, such as, for example, temperature, humidity, wind direction and strength and / or the composition of gases. The values ​​of the first parameters are different from the values ​​of the second parameters, and consequently the first measurement data are different from the second parameters.

[0013] In the third method step, first ML data is created from the first measurement data. ML data is data created from the measurement data using the algorithm of a machine learning model. The first ML data is created using the first measurement data. In the fourth method step, second ML data is created from the second measurement data. Due to the use of different first and second measurement data, the second ML data is different from the first ML data according to the invention.

[0014] ML data within the meaning of this invention is data to which a machine learning algorithm has been applied. The ML data contains measurement data on the type and concentrations of various gases, as well as their type and concentrations at different temperatures during the various phases of forest fire development. In particular, it contains data on, for example, volatile organic compounds (VOCs), volatile sulfur-containing compounds (VSCs), as well as carbon dioxide, carbon monoxide, and molecular hydrogen, preferably in the ppb range. Furthermore, optionally included are data on air humidity, ambient temperature, and air pressure, also recorded at different temperatures during the various phases of forest fire development and created by the machine learning algorithm.

[0015] In a further development of the invention, the first ML data and the second ML data can be applied to measurement data of the same type and / or measurement data of the same sensor type. Thus, the first ML data and the second ML data can both be applied to the measurement data of a temperature sensor and / or both to the measurement data of a gas sensor and / or both to the measurement data of a humidity sensor. Measurement data of the same type within the meaning of this document is measurement data acquired by a sensor of the same design. With the help of the machine learning algorithm, the forest fire early warning system is enabled to recognize patterns and regularities based on existing data sets and algorithms. The findings can be generalized and used for new problem solutions or for the analysis of previously unknown data. Reinforcement learning is preferably used for this purpose.Through reward and punishment, the algorithm learns a tactic for how to act in potentially occurring situations in order to maximize the system's utility. The machine learning algorithm interacts with the environment and is evaluated by a cost function or reward system. In reinforcement learning, the machine learning algorithm is not shown which action is the right one in which situation; instead, it receives positive or negative feedback through the cost function. Based on the cost function, it then assesses which action is the right one at a given time. Thus, the system learns to maximize the reward function through "reinforcement" through praise or punishment.

[0016] A neural network can also be used. These networks are capable of automatically learning key features relevant to solving a problem from the data. Unlike traditional machine learning methods, in which the features must be laboriously designed and defined by humans, neural networks independently learn the correlations and features from the data.

[0017] The model's algorithm enables improved application-specific detection of the gases to be detected. In particular, the sensitivity of a forest fire early detection system is increased. The algorithm also corrects the recorded measured values. In particular, the accuracy is increased, i.e. false positives are reduced or eliminated. Forest fire detection based on threshold values ​​is error-prone and generates many false positives. With machine learning models, especially models created based on locally acquired data, the accuracy can be increased enormously, as the model adapts to reality. In addition, the baseline and long-term deviations of the measured values ​​are compensated. By evaluating the measured values ​​recorded during a forest fire, statements can be made about the current situation, e.g., spread speed and direction, during forest fires.

[0018] In one development of the invention, the first ML data are implemented in a first terminal device of the forest fire early detection system. The first ML data are implemented in a first terminal device instead of in a central unit, e.g., a network server. A forest fire early detection system typically has a plurality of terminal devices that are widely distributed and have self-sufficient energy supply systems. The implementation of the ML data, i.e., the storage of the ML data on a first terminal device and the application of the ML data to measurement data acquired by the first terminal device in the first terminal device, enables the machine learning model to be adapted and applied to the local conditions of the first terminal device. At the same time, the energy consumption of an individual terminal device is reduced because only a reduced amount of data needs to be transmitted.

[0019] In a further embodiment of the invention, the second ML data are implemented in a second terminal device of the forest fire early warning system. The second ML data are implemented in a second terminal device instead of in a central unit, e.g., a network server, wherein the second ML data are different from the first ML data, which are implemented in a first terminal device. A forest fire early warning system typically has a plurality of terminal devices that are widely distributed. The implementation of the second ML data, i.e., the storage of the second ML data on a second terminal device and the application of the second ML data to measurement data acquired by the second terminal device in the second terminal device, enables the machine learning model to be adapted and applied to the local conditions of the second terminal device.

[0020] In a further embodiment of the invention, first terminals acquire measurement data in a first region, and second terminals acquire measurement data in a second region, wherein the first region differs from the second region in terms of vegetation or similar characteristics. The terminals acquire measurement data for detecting a forest fire, wherein the terminals are located in different regions of the forest to be monitored.

[0021] In a further embodiment of the invention, the first parameters of the forest are different from the second parameters of the forest. In particular, the values ​​of the first parameters are different from the second parameters. The first measurement data created from the first parameters are therefore different from the second measurement data created using the second parameters.

[0022] In an advantageous embodiment of the invention, the first parameters of the forest are characteristic of a first region of the forest, and the second parameters of the forest are characteristic of a second region of the forest. The first region and the second region differ from one another. The first region and the second region differ in particular with regard to, for example, their vegetation, their climatic conditions such as prevailing wind speeds and directions, soil moisture of their terrain, and the nature and density of the combustible material. The differences between the regions result in different behavior of a forest fire in the different regions. In particular, the speed and direction of spread are different.Recording the different characteristic parameters of the first and second regions enables firefighting units to localize a forest fire more precisely and fight it more effectively in the first and second regions.

[0023] In a further advantageous embodiment of the invention, the first and / or second measurement data are determined from measurements of tn parameters of a forest under tn conditions. For the purposes of this document, tn parameters (true negative parameters) are parameters recorded from a forest that is not experiencing a forest fire. Such parameters include, for example, temperature, humidity, wind direction and strength, and / or the composition of gases. The tn measurement data recorded with these tn parameters (true negative measurement data) are accordingly measurement data measured from a forest that is also not experiencing a forest fire. The recorded tn measurement data therefore represent measurement data from a forest that is not experiencing a forest fire, wherein the first measurement data differs from the second measurement data in that they have different tn parameters. This generates different tn parameters and likewise different first and second measurement data.By collecting tn measurement data, the baseline and long-term deviations of the measured values ​​are compensated. Furthermore, design-related deviations of the installed sensors are recorded, averaged, and compensated. The tn measurement data are obtained from a laboratory test and / or in a natural environment. Under laboratory conditions, the tn measurement data are recorded under idealized conditions and are reproducible. Alternatively, the tn measurement data are obtained from the natural environment for which the forest fire early detection system is intended. The collected tn measurement data are therefore specific to the respective forest and, in particular, specific to the region in which the tn measurement data are determined.

[0024] In a further advantageous embodiment of the invention, the first and / or second measurement data are determined from measurements of tp parameters of a forest under tp conditions. Acquiring tp (true-positive) measurement data—that is, for events whose measurement data represent a forest fire—and subsequently generating an ML data set containing tp parameters increases the reliability of the measured values ​​acquired by a sensor, to which both the tn-ML data set and the tp-ML data set are applied. The rate of false alarms is thereby significantly reduced.

[0025] When using only tn measurement data and generating tn-ML data, baseline and long-term deviations are compensated. However, the forest fire early detection system can trigger a false alarm if the recorded measurement data and the resulting result data from a terminal device deviate from these tn-ML data, even if no forest fire is present in the area being monitored.

[0026] In a further embodiment of the invention, the first and / or second ML data are created from the tn measurement data. ML data are data created from the tn measurement data using the algorithm of a machine learning model. The first and / or second ML data have tn parameters (true negative parameters) that are recorded from a forest that does not have a forest fire. This generates different tn parameters and also different first and second measurement data. By collecting tn measurement data, the baseline and long-term deviations of the measured values ​​are compensated.

[0027] In a further embodiment of the invention, the second ML data is created from the tp measurement data. ML data is data created from the tp measurement data using the algorithm of a machine learning model. Creating tp (true-positive) ML data—that is, for events whose measurement data represent a forest fire—increases the reliability of the measured values ​​recorded by a sensor, to which both the tn ML data set and the tp ML data set are applied. This significantly reduces the rate of false alarms.

[0028] In an advantageous embodiment of the invention, tn-ML data and / or tp-ML data are implemented into a forest fire early warning system for detecting forest fires. The ML data thus comprises tp-ML data and tn-ML data, which are stored in the forest fire early warning system and applied to measurement data from the terminal device. This significantly reduces the rate of false alarms.

[0029] In a further embodiment of the invention, the first and / or second measurement data are acquired by a permanently installed ML module. Using an ML module, parameters are measured and measurement data is generated. The measurement data is sent via a network to a control device equipped with a machine learning algorithm. The permanently installed ML module is fixed within a forest fire early detection system, e.g., on a tree. This measurement data is thus assigned to a specific region of the forest to be monitored.

[0030] In an advantageous development of the invention, the first and / or second measurement data are recorded by several permanently installed ML modules, wherein the individual permanently installed ML modules are installed in different regions of the forest to be monitored. The individual ML modules therefore record parameters that are characteristic of each region and, in particular, different from one another. In a further advantageous embodiment of the invention, the permanently installed ML module takes on the function of a terminal device, provided that the permanently installed ML module does not record first or second measurement data. ML modules then function like a conventional terminal device for early forest fire detection. The terminal device and ML module can therefore be constructed identically to one another, thereby reducing the costs and operation of the early forest fire detection system.

[0031] In a further embodiment of the invention, the first and / or second measurement data are acquired by a mobile ML module. The mobile ML module is arranged in a vehicle that is remotely controlled, automatically, and / or autonomously moved. Advantageously, the vehicle is a flight-capable drone that can fly over the forest to be monitored. The vehicle has suitable sensors and a control device for recording first and / or second parameters.

[0032] In a further embodiment of the invention, the mobile ML module changes location between the acquisition of the first and second measurement data. The mobile ML module acquires first measurement data in the first region, then changes location and acquires second measurement data in a second region. Using the mobile ML module, different measurement data from different regions can therefore be acquired in a short period of time. Furthermore, the mobile ML module, like a permanently installed ML module, can be configured to operate as a terminal device for early forest fire detection. The mobile ML module can therefore also perform the function of a terminal device, namely the acquisition or detection of a forest fire.

[0033] The problem is also solved by means of a forest fire early detection system with a network. Advantageous embodiments are set forth in the following subclaims.

[0034] The forest fire early detection system according to the invention with a network has a first terminal. The first terminal has a sensor device that has one or a plurality of sensors, e.g., for gas analysis. The first terminal also has a first control device that has a first memory unit. A first ML data set is stored on the first memory unit. The first terminal also has an evaluation device for evaluating the measurement signals supplied by the sensor device and a power supply device. The power supply device enables autonomous operation of the terminal, for example by charging a battery via solar cells.

[0035] The forest fire early detection system according to the invention with a network also has a second terminal. The second terminal also has a sensor device that has one or a plurality of sensors. The second terminal also has a second control device that has a second memory unit. A second ML data set is stored on the second memory unit. The second terminal also has an evaluation device for evaluating the measurement signals supplied by the sensor device and a power supply device.

[0036] The forest fire early detection system according to the invention also includes a network server. The network server has interfaces to other applications that can be used, for example, to determine the direction and speed of a forest fire's spread.

[0037] According to the invention, the first ML data set and the second ML data set are different from each other. A first ML data set is created using first measurement data, while the second ML data set is created using second measurement data. The first measurement data are typically recorded in a first region of the forest to be monitored by the forest fire early warning system, and the second measurement data are recorded in a second region, with the first and second regions being different from each other.

[0038] In one development of the invention, the first control device is suitable and provided for accessing a first memory containing a first ML data set from the adaptation and application of a machine learning model, and / or the second control device is suitable and provided for accessing a second memory containing a second ML data set from the adaptation and application of a machine learning model. The first and second control devices are arranged in and / or connected to different terminal devices. The algorithm of the machine learning model enables improved application-specific detection of the gases to be detected. In addition, the baseline and long-term deviations of the measured values ​​are compensated.For this purpose, data on different gas compositions and their concentrations are made available to the sensor system, which is compared with the gas compositions and their concentrations determined by the sensor.

[0039] In an advantageous embodiment of the invention, the machine learning model comprises tn-ML data. tn-ML data is data created from tn measurement data using the algorithm of a machine learning model. The tn measurement data (true negative measurement data) acquired with these tn parameters are measurement data measured from a forest that also does not have a forest fire. The acquired tn measurement data therefore represents measurement data from a forest that does not have a forest fire. By acquiring tn measurement data, the baseline and long-term deviations of the measured values ​​are compensated. In addition, design-related deviations of the sensors installed in the terminal device are recorded, averaged, and compensated.

[0040] In a further development of the invention, the machine learning model comprises tp-ML data. Collecting tp measurement data—that is, for events whose measurement data represent a forest fire—and subsequently generating a tp-ML dataset increases the reliability of the measured values ​​acquired by a terminal device and the resulting result data, to which both the tn-ML dataset and the tp-ML dataset are applied. This significantly reduces the rate of false alarms.

[0041] In a further embodiment of the invention, the forest fire early detection system has an ML module that is suitable and intended to collect measurement data for creating ML data. Using an ML module, parameters are measured and measurement data is created. The measurement data is sent via a network to a control device that has a machine learning algorithm. In a further embodiment of the invention, the forest fire early detection system has a permanently installed ML module. Using a permanently installed ML module, parameters are measured and measurement data is created. The measurement data is sent via a network to a control device that has a machine learning algorithm. The permanently installed ML module is arranged in a fixed location within a forest fire early detection system, e.g., on a tree.

[0042] In a further development of the invention, the forest fire early detection system comprises several permanently installed ML modules, with at least two of the ML modules located in different regions of the forest to be monitored. The individual ML modules therefore record parameters that are characteristic of each region and, in particular, different from one another.

[0043] In a further embodiment of the invention, the forest fire early detection system comprises a mobile ML module. The mobile ML module is arranged in a vehicle that is remotely controlled, automatically, and / or autonomously moved. Advantageously, the vehicle is a flight-capable drone that can fly over the forest to be monitored. The vehicle has suitable sensors and a control device for recording first and / or second parameters. The mobile ML module can therefore also perform the function of a terminal device, namely the detection or recognition of a forest fire.

[0044] In a further embodiment of the invention, the mobile ML module is suitable for collecting measurement data from multiple regions of the forest to be monitored in order to create ML datasets. The mobile ML module collects first measurement data in the first region, then changes location and collects second measurement data in a second region. Using the mobile ML module, different measurement data from different regions can therefore be collected in a short time. Furthermore, the mobile ML module, like a permanently installed ML module, can be configured as a terminal for early forest fire detection. This task is further achieved using the inventive method for determining ML data for a system for early forest fire detection.

[0045] The inventive method for determining ML data for a system for early forest fire detection has four method steps: In the first method step, first measurement data are determined from measurements of first parameters of a forest. Such parameters include, for example, temperature, humidity, wind direction and strength and / or the composition of gases. First measurement data are created from these parameters. In the second method step, second measurement data are determined from measurements of second parameters of a forest. Second measurement data are created from the measured second parameters, such as, for example, temperature, humidity, wind direction and strength and / or the composition of gases. The values ​​of the first parameters are different from the values ​​of the second parameters, and consequently the first measurement data are different from the second parameters.

[0046] In the third method step, first ML data is created from the first measurement data. ML data is data created from the measurement data using the algorithm of a machine learning model. The first ML data is created using the first measurement data. In the fourth method step, second ML data is created from the second measurement data. Due to the use of different first and second measurement data, the second ML data is different from the first ML data according to the invention.

[0047] ML data, within the meaning of this invention, is data to which a machine learning algorithm has been applied. The ML data contains measurement data on the type and concentrations of various gases, as well as their type and concentrations at different temperatures during the various phases of forest fire development. In particular, it contains data on volatile organic compounds (VOCs), volatile sulfur-containing compounds (VSCs), as well as carbon dioxide, carbon monoxide, and molecular hydrogen, preferably in the ppb range. Furthermore, it contains data on air humidity, ambient temperature, and air pressure, also recorded at different temperatures during the various phases of forest fire development and created by the machine learning algorithm.

[0048] With the help of the machine learning algorithm, the forest fire early warning system is able to recognize patterns and regularities based on existing data sets and algorithms. The findings can be generalized and used to solve new problems or analyze previously unknown data. Reinforcement learning is preferred for this. Through rewards and punishments, the algorithm learns a strategy for how to act in potentially occurring situations in order to maximize the system's benefit. The machine learning algorithm interacts with the environment and is evaluated by a cost function or reward system. With reinforcement learning, the machine learning algorithm is not shown which action is the right one in which situation, but rather receives positive or negative feedback via the cost function.Based on the cost function, it then assesses which action is the right one at which point in time. Thus, the system learns to maximize the reward function through "reinforcement" through praise or punishment.

[0049] A neural network can also be used. These networks are capable of automatically learning key features relevant to solving a problem from the data. Unlike traditional machine learning methods, in which the features must be laboriously designed and defined by humans, neural networks independently learn the correlations and features from the data.

[0050] The model's algorithm enables improved application-specific recording of the gases to be detected. In particular, the sensitivity of a forest fire early detection system is increased. In addition, the accuracy is increased, i.e. false positives are reduced or eliminated. Forest fire detection based on threshold values ​​is error-prone and generates many false positives. With machine learning models, especially models created based on locally acquired data, the accuracy can be increased enormously, as the model adapts to reality. The algorithm also corrects the recorded measured values. In addition, the baseline and long-term deviations of the measured values ​​are compensated. By evaluating the measured values ​​recorded during a forest fire, statements can be made about the current situation, e.g., spread speed and direction, during forest fires.

[0051] In one development of the invention, the first and second ML data are implemented in the same terminal device of the forest fire early warning system. The first ML data is implemented in a first terminal device instead of in a central unit, e.g., a network server. A forest fire early warning system typically has a plurality of terminal devices that are widely distributed and have self-sufficient energy supply systems. The implementation of the ML data, i.e., the storage of the ML data on the terminal devices and the application of the ML data to measurement data acquired by the terminal devices in the terminal device, enables the machine learning model to be adapted and applied to the local conditions of the terminal devices. At the same time, the energy consumption of an individual terminal device is reduced because only a reduced amount of data needs to be transmitted.

[0052] In a further embodiment of the invention, the first measurement data and the second measurement data are recorded at different times of day. First and second measurement data recorded at different times of day are typically also different. Not only is the air temperature typically different at midday than at night, for example, but the gas composition of the ambient air and its humidity are also different due to the different insolation. These different measurement data are incorporated into the ML data created using a machine learning model, increasing the reliability and sensitivity of early detection of a forest fire.

[0053] In a further embodiment of the invention, the first measurement data and the second measurement data are recorded at different times of the year. First and second measurement data recorded at different times of the year are typically also different. Not only is the air temperature typically different in winter than in summer, for example, but the gas composition of the ambient air and its humidity are also different due to the different insolation. These different measurement data are incorporated into the ML data created using a machine learning model, increasing the reliability and sensitivity of early detection of a forest fire.

[0054] In a further advantageous embodiment of the invention, the first and / or second measurement data are determined from measurements of tn parameters of a forest under tn conditions. For the purposes of this document, tn parameters (true negative parameters) are parameters recorded from a forest that is not experiencing a forest fire. Such parameters include, for example, temperature, humidity, wind direction and strength, and / or the composition of gases. The tn measurement data recorded with these tn parameters (true negative measurement data) are accordingly measurement data measured from a forest that is also not experiencing a forest fire. The recorded tn measurement data therefore represent measurement data from a forest that is not experiencing a forest fire, wherein the first measurement data differs from the second measurement data in that they have different tn parameters. This generates different tn parameters and likewise different first and second measurement data.By collecting tn measurement data, the baseline and long-term deviations of the measured values ​​are compensated. Furthermore, design-related deviations of the installed sensors are recorded, averaged, and compensated. The tn measurement data are obtained from a laboratory test and / or in a natural environment. Under laboratory conditions, the tn measurement data are recorded under idealized conditions and are reproducible. Alternatively, the tn measurement data are obtained from the natural environment for which the forest fire early detection system is intended. The collected tn measurement data are therefore specific to the respective forest and, in particular, specific to the region in which the tn measurement data are determined.

[0055] In a further embodiment of the invention, the first and / or second ML data are created from the tn measurement data. ML data are data created from the tn measurement data using the algorithm of a machine learning model. The first and / or second ML data have tn parameters (true negative parameters) that are recorded from a forest that does not have a forest fire. This generates different tn parameters and also different first and second measurement data. By collecting tn measurement data, the baseline and long-term deviations of the measured values ​​are compensated.

[0056] In a further advantageous embodiment of the invention, the first and / or second measurement data are determined from measurements of tp parameters of a forest under tp conditions. Acquiring tp (true-positive) measurement data—that is, for events whose measurement data represent a forest fire—and subsequently generating an ML data set containing tp parameters increases the reliability of the measured values ​​acquired by a sensor, to which both the tn-ML data set and the tp-ML data set are applied. The rate of false alarms is thereby significantly reduced.

[0057] When using only tn measurement data and generating tn-ML data, baseline and long-term deviations are compensated. However, the forest fire early detection system can trigger a false alarm if the recorded measurement data and the resulting result data from a terminal device deviate from these tn-ML data, even if no forest fire is present in the area being monitored.

[0058] In an advantageous embodiment of the invention, tn-ML data and / or tp-ML data are implemented into a forest fire early warning system for detecting forest fires. The ML data thus comprises tp-ML data and tn-ML data, which are stored in the forest fire early warning system and applied to measurement data from the terminal device. This significantly reduces the rate of false alarms.

[0059] In a further embodiment of the invention, the first and / or second measurement data are acquired by a permanently installed ML module. Using an ML module, parameters are measured and measurement data is generated. The measurement data is sent via a network to a control device equipped with a machine learning algorithm. The permanently installed ML module is fixed within a forest fire early detection system, e.g., on a tree. This measurement data is thus assigned to a specific region of the forest to be monitored.

[0060] In a further advantageous embodiment of the invention, the permanently installed ML module assumes the function of a terminal device, provided the permanently installed ML module is not collecting first or second measurement data. ML modules then function like a conventional terminal device for early forest fire detection. The terminal device and ML module can therefore be constructed identically, thereby reducing the costs and operation of the early forest fire detection system.

[0061] In a further embodiment of the invention, the first and / or second measurement data are acquired by a mobile ML module. The mobile ML module is arranged in a vehicle that is remotely controlled, automatically, and / or autonomously moved. Advantageously, the vehicle is a flight-capable drone that can fly over the forest to be monitored. The vehicle has suitable sensors and a control device for recording first and / or second parameters.

[0062] In a further embodiment of the invention, there is a time interval of several hours, preferably several weeks, and particularly preferably several months, between the acquisition of the first and second measurement data. This allows measurement data to be acquired at different times of day and particularly preferably at different times of the year. For example, the parameters air temperature, gas composition of the ambient air, and its humidity are different from each other at these times. These different measurement data are incorporated into the ML data created using a machine learning model, increasing the reliability and sensitivity of early detection of a forest fire.

[0063] The task is further achieved by means of the inventive method for determining ML data for a system for early forest fire detection.

[0064] The inventive method for determining ML data for a forest fire early detection system comprises four method steps: In the first method step, first measurement data are determined from measurements of first parameters of a forest. Such parameters include, for example, temperature, humidity, wind direction and strength, and / or the composition of gases. First measurement data are created from these parameters. In the second method step, second measurement data are determined from measurements of second parameters of a forest. Second measurement data are created from the measured second parameters, such as, for example, temperature, humidity, wind direction and strength, and composition of gases. The values ​​of the first parameters are different from the values ​​of the second parameters, and consequently the first measurement data are different from the second parameters.

[0065] In the third process step, initial ML data is created from the initial measurement data. ML data is data that is created from the measurement data using the algorithm of a machine learning model. The initial ML data is created using the initial measurement data. In the fourth process step, second ML data is created from the initial measurement data.

[0066] According to the invention, the first ML data and the second ML data are acquired by one and the same ML module at different values ​​of the same environmental parameter. Due to the use of different first and second measurement data, the second ML data are different from the first ML data.

[0067] ML data within the meaning of this invention is data to which a machine learning algorithm has been applied. The ML data contains measurement data on the type and concentrations of different gases, as well as their type and concentrations at different temperatures during the different phases of forest fire development. In particular, it contains data on, for example, volatile organic compounds (VOCs), volatile sulfur-containing compounds (VSCs), as well as carbon dioxide, carbon monoxide, and molecular hydrogen, preferably in the ppb range. It also contains data on air humidity, ambient temperature, and air pressure, also recorded at different temperatures during the different phases of forest fire development and created by the machine learning algorithm. With the help of the machine learning algorithm, the forest fire early detection system is able to recognize patterns and regularities based on existing data sets and algorithms.The findings can be generalized and used to solve new problems or analyze previously unknown data. Reinforcement learning is preferred for this. Through reward and punishment, the algorithm learns a strategy for how to act in potentially occurring situations in order to maximize the system's benefit. The machine learning algorithm interacts with the environment and is evaluated by a cost function or reward system. With reinforcement learning, the machine learning algorithm is not shown which action is the right one in which situation, but rather receives positive or negative feedback from the cost function. Based on the cost function, it is then assessed which action is the right one at a given time. In this way, the system learns to maximize the reward function through "reinforcement" through praise or punishment.

[0068] A neural network can also be used. These networks are capable of automatically learning key features relevant to solving a problem from the data. Unlike traditional machine learning methods, in which the features must be laboriously designed and defined by humans, neural networks independently learn the correlations and features from the data.

[0069] The model's algorithm enables improved application-specific recording of the gases to be detected. In particular, the sensitivity of a forest fire early warning system is increased. In addition, the accuracy is increased, i.e. false positives are reduced or eliminated. Forest fire detection based on threshold values ​​is error-prone and generates many false positives. With machine learning models, especially models created based on locally acquired data, the accuracy can be increased enormously, as the model adapts to reality. Furthermore, the algorithm corrects the recorded measured values. In addition, the baseline and long-term deviations of the measured values ​​are compensated. By evaluating the measured values ​​recorded during a forest fire, statements can be made about the current situation, e.g.

[0070] Spreading speed and direction during wildfires.

[0071] In a further development of the invention, the environmental parameters comprise one or more parameters from the group of temperature, temperature profile, air humidity, air humidity profile, soil moisture, soil moisture profile or similar environmental parameters. The temperature of the ambient air and its particularly short-term temperature profile are a key indicator for the occurrence and presence of a forest fire. The relative humidity and its profile over time as well as the soil moisture and its profile over time are indicators, in particular, of the risk of forest fires. Other parameters include, for example, the composition of the gases in the ambient air that characterize a forest fire, e.g., volatile organic compounds (VOCs). The forest fire early detection system has sensors for determining the air humidity, air pressure and a temperature sensor.Optionally or additionally, a terminal device has sensors for gas analysis and for detecting the prevailing wind direction, with which the composition and concentration of gases as well as their direction of propagation are determined.

[0072] In a further embodiment of the invention, the first measurement data and the second measurement data are recorded at different times of day. First and second measurement data recorded at different times of day are typically also different. Not only is the air temperature typically different at midday than at night, for example, but the gas composition of the ambient air and its humidity are also different due to the different insolation. These different measurement data are incorporated into the ML data created using a machine learning model, increasing the reliability and sensitivity of early detection of a forest fire.

[0073] In a further embodiment of the invention, the first measurement data and the second measurement data are recorded at different times of the year. First and second measurement data recorded at different times of the year are typically also different. Not only is the air temperature typically different in winter than in summer, for example, but the gas composition of the ambient air and its humidity are also different due to the different insolation. These different measurement data are incorporated into the ML data created with a machine learning model, increasing the reliability and sensitivity of early detection of a forest fire.

[0074] In a further advantageous embodiment of the invention, the first and / or second measurement data are determined from measurements of tn parameters of a forest under tn conditions. For the purposes of this document, tn parameters (true negative parameters) are parameters recorded from a forest that is not experiencing a forest fire. Such parameters include, for example, temperature, humidity, wind direction and strength, and / or the composition of gases. The tn measurement data recorded with these tn parameters (true negative measurement data) are accordingly measurement data measured from a forest that is also not experiencing a forest fire. The recorded tn measurement data therefore represent measurement data from a forest that is not experiencing a forest fire, wherein the first measurement data differs from the second measurement data in that they have different tn parameters. This generates different tn parameters and likewise different first and second measurement data.By collecting tn measurement data, the baseline and long-term deviations of the measured values ​​are compensated. Furthermore, design-related deviations of the installed sensors are recorded, averaged, and compensated. The tn measurement data are obtained from a laboratory test and / or in a natural environment. Under laboratory conditions, the tn measurement data are recorded under idealized conditions and are reproducible. Alternatively, the tn measurement data are obtained from the natural environment for which the forest fire early detection system is intended. The collected tn measurement data are therefore specific to the respective forest and, in particular, specific to the region in which the tn measurement data are determined.

[0075] In a further embodiment of the invention, the first and / or second ML data are created from the tn measurement data. ML data are data created from the tn measurement data using the algorithm of a machine learning model. The first and / or second ML data have tn parameters (true negative parameters) that are recorded from a forest that does not have a forest fire. This generates different tn parameters and also different first and second measurement data. By collecting tn measurement data, the baseline and long-term deviations of the measured values ​​are compensated.

[0076] When using only tn measurement data and generating tn-ML data, baseline and long-term deviations are compensated. However, the forest fire early detection system can trigger a false alarm if the recorded measurement data and the resulting result data from a terminal device deviate from these tn-ML data, even if no forest fire is present in the area being monitored.

[0077] In an advantageous embodiment of the invention, tn-ML data and / or tp-ML data are implemented into a forest fire early warning system for detecting forest fires. The ML data thus comprises tp-ML data and tn-ML data, which are stored in the forest fire early warning system and applied to measurement data from the terminal device. This significantly reduces the rate of false alarms.

[0078] In a further embodiment of the invention, the first and / or second measurement data are acquired by a permanently installed ML module. Using an ML module, parameters are measured and measurement data is generated. The measurement data is sent via a network to a control device equipped with a machine learning algorithm. The permanently installed ML module is fixed within a forest fire early detection system, e.g., on a tree. This measurement data is thus assigned to a specific region of the forest to be monitored.

[0079] In a further advantageous embodiment of the invention, the permanently installed ML module assumes the function of a terminal device, provided the permanently installed ML module does not record first or second measurement data. ML modules then operate like a conventional terminal device for early forest fire detection. The terminal device and ML module can therefore be constructed identically to one another, thereby reducing the costs and operation of the early forest fire detection system. In a further embodiment of the invention, the first and / or second measurement data are recorded by a mobile ML module. The mobile ML module is arranged in a vehicle that is remotely controllable, automatically and / or autonomously moved. Advantageously, the vehicle is a flight-capable drone that can fly over the forest to be monitored. The vehicle has suitable sensors and a control device for recording first and / or second parameters.

[0080] In a further embodiment of the invention, there is a time interval of several hours, preferably several weeks, and particularly preferably several months, between the acquisition of the first and second measurement data. This allows measurement data to be acquired at different times of day and particularly preferably at different times of the year. For example, the parameters air temperature, gas composition of the ambient air, and its humidity are different from each other at these times. These different measurement data are incorporated into the ML data created using a machine learning model, increasing the reliability and sensitivity of early detection of a forest fire.

[0081] The task is further achieved by means of the inventive method for updating ML data on end devices of a forest fire early warning system.

[0082] The inventive method for updating ML data on end devices of a forest fire early warning system comprises three steps: In the first step, measurement data from measurements of forest parameters are collected. Such parameters include, for example, temperature, humidity, wind direction and force, and / or the composition of gases. Initial measurement data is created from these parameters.

[0083] In the second process step, ML data is created from the measurement data. ML data is data created from the measurement data using the algorithm of a machine learning model. The ML data is created using the measurement data. ML data within the meaning of this invention is data to which a machine learning algorithm has been applied. The ML data contains measurement data on the type and concentrations of various gases, as well as their type and concentrations at different temperatures during the different phases of forest fire development. In particular, data on volatile organic compounds (VOCs), volatile sulfur-containing compounds (VSCs), as well as carbon dioxide, carbon monoxide, and molecular hydrogen, preferably in the ppb range, are included.Furthermore, data on humidity, ambient temperature and air pressure are included, also recorded at different temperatures in the different phases of forest fire development and created by the machine learning algorithm.

[0084] With the help of the machine learning algorithm, the forest fire early warning system is able to recognize patterns and regularities based on existing data sets and algorithms. The findings can be generalized and used to solve new problems or analyze previously unknown data. Reinforcement learning is preferred for this. Through rewards and punishments, the algorithm learns a strategy for how to act in potentially occurring situations in order to maximize the system's benefit. The machine learning algorithm interacts with the environment and is evaluated by a cost function or reward system. With reinforcement learning, the machine learning algorithm is not shown which action is the right one in which situation, but rather receives positive or negative feedback via the cost function.Based on the cost function, it then assesses which action is the right one at which point in time. Thus, the system learns to maximize the reward function through "reinforcement" through praise or punishment.

[0085] A neural network can also be used. These networks are able to automatically learn crucial features relevant to solving a problem from the data. In contrast to classic machine learning methods, in which the features must be laboriously designed and defined by humans, neural networks independently learn the correlations and features from the data. The model's algorithm enables improved application-specific recording of the gases to be detected. In particular, the sensitivity of a forest fire early detection system is increased. The algorithm also corrects the recorded measured values. Furthermore, the accuracy is increased, i.e. false positives are reduced or eliminated. Forest fire detection based on threshold values ​​is error-prone and generates many false positives.Machine learning models, especially those based on locally acquired data, can significantly increase accuracy because the model adapts to reality. In addition, baseline and long-term deviations in the measured values ​​are compensated for. By evaluating the measured values ​​recorded during a forest fire, conclusions can be drawn about the current situation, such as the speed and direction of spread, during forest fires.

[0086] In the third step, the ML data is transmitted to a terminal device of the forest fire early warning system. A forest fire early warning system typically comprises a large number of terminal devices that are widely distributed and have self-sufficient power supply systems. The implementation of the ML data—that is, the storage of the ML data on the terminal devices and the application of the ML data to the measurement data acquired by the terminal devices—enables the adaptation and application of the machine learning model to the local conditions of the terminal devices. At the same time, the energy consumption of an individual terminal device is reduced, since only a smaller amount of data needs to be transmitted.

[0087] According to the invention, the ML data is transmitted wirelessly. For this purpose, the forest fire early detection system has a wireless network or is integrated into such a wireless network. A LoRa or LoRaWAN mesh gateway network, for example, can be used. ML data sets are sent to the end devices at preferred intervals. The intervals can be time-based and / or data volume-based. The ML data sets are stored in a memory of the end device and applied to the acquired measurement data by the end device's control unit. In an optional embodiment of the invention, the ML algorithm is also installed on the end device. Here, "edge computing" is used to evaluate the data locally in the end device. This is important and advantageous because it eliminates the need to transmit the data to a central cloud application, thus reducing the load on the (very narrowband) LoRaWAN network, for example.

[0088] In a further embodiment of the invention, the ML data is created from the acquired measurement data on an application server that is suitable and intended to execute a machine learning program. The application server has a system that has a machine learning algorithm. The machine learning algorithm uses training data to improve the machine learning model.

[0089] In a further embodiment of the invention, the acquired measurement data is transmitted to the network server. On the network server, the acquired measurement data is available to other applications that detect and record a forest fire. By integrating the measurement data into a machine learning model, models for detecting forest fires are also created using the network server. The acquired measurement data can include the result of comparing the measurement data with the machine learning data, simply an evaluation of the comparison of the measurement data with the machine learning data, and / or a simple warning signal. In any case, the acquired measurement data indicates whether a forest fire was detected by the sensor of the end device.

[0090] In a further embodiment of the invention, the network server is coupled to an application server, to which the network server sends the acquired measurement data. The application server is suitable and intended to execute a machine learning program. The application server has a system that has a machine learning algorithm. The machine learning algorithm uses training data to improve the machine learning model. The application server has access to the measurement data acquired by the terminal device. The measurement data acquired by the terminal device is training data with which a machine learning algorithm of the application server is trained. In a further embodiment of the invention, the measurement data is acquired by a permanently installed ML module. Using an ML module, parameters are measured and measurement data is created.The measurement data is sent via a network to a control device equipped with a machine learning algorithm. The permanently installed ML module is located at a fixed location within a forest fire early detection system, for example, on a tree. This measurement data is thus assigned to a specific region of the forest being monitored.

[0091] In a further embodiment of the invention, the measurement data is acquired by a mobile ML module. The mobile ML module is mounted in a vehicle that can be remotely controlled, automatically, and / or autonomously moved. Advantageously, the vehicle is a flight-capable drone that can fly over the forest to be monitored. The vehicle has suitable sensors and a control device for recording parameters.

[0092] Embodiments of the method according to the invention for determining ML data for a system for early forest fire detection and of the forest fire early detection system according to the invention are shown in a simplified schematic form in the drawings and are explained in more detail in the following description.

[0093] They show:

[0094] Fig. 1 : The forest to be monitored with four different regions

[0095] Fig. 2: Structure of a forest fire early detection system comprising a LoRa radio network, ML units fixed

[0096] Fig. 3 a: Detailed view of a forest fire early warning system

[0097] Fig. 3 b: Structure of a terminal arranged in the second region

[0098] Fig. 3 c: Structure of a terminal arranged in the third region

[0099] Fig. 3 d: Structure of a terminal arranged in the fourth region

[0100] Fig. 4 a: Structure of a permanently installed ML unit with a double

[0101] Forest fire detection sensor

[0102] Fig. 4 b: Structure of a terminal device with a forest fire detection sensor Fig. 5 a: Structure of a permanently installed ML unit with two dual forest fire detection sensors

[0103] Fig. 5 b: Structure of a terminal with two forest fire detection sensors

[0104] Fig. 6 a: Structure of a permanently installed ML unit with two triple forest fire detection sensors

[0105] Fig. 6 b: Structure of a terminal with two forest fire detection sensors

[0106] Fig. 7 a: Structure of a mobile forest fire detection unit

[0107] Fig. 7 b: Structure of a terminal with three forest fire detection sensors

[0108] Fig. 1 shows an embodiment of a forest W to be monitored, in which four different regions R1, R2, R3, R4 are arranged, highlighted by different hatching. The regions R1, R2, R3, R4 are clearly located, clearly demarcated from one another, adjoin one another seamlessly and do not overlap. In this and all following embodiments, the regions R1, R2, R3, R4 differ in their altitude, the amount of precipitation and soil moisture occurring in the region R1, R2, R3, R4 and, as a result, in the type of vegetation: the highest, rocky region R1 has a coniferous forest, the lowest region R2 with a flowing water body F has swampy areas with riparian vegetation, the region R3 has a mixed forest and the region R4 has a deciduous forest with moors.

[0109] Fig. 2 shows an embodiment of the forest fire early detection system 1 according to the invention, arranged in the forest W to be monitored with the different regions R1, R2, R3, R4, as described in the previous embodiment (see Fig. 1). The forest fire early detection system 1 has a network 10, in this embodiment a LoRaWAN mesh gateway network 10 with a plurality of terminal devices EDn. The first terminal devices ED1 are arranged in the first region R1, the second terminal devices ED2 in the second region R2, the third terminal devices ED3 in the third region R3 and the fourth terminal devices ED4 in the fourth region R4. The terminal devices EDn are each structurally identical to one another. To detect a forest fire, an individual terminal device EDn has a sensor device S which has sensors for determining the air humidity, the air pressure and a temperature sensor.Optionally or additionally, a terminal device EDn has sensors for gas analysis and for detecting the prevailing wind direction, which are used to determine the composition and concentration of gases as well as their direction of propagation. Furthermore, a terminal device EDn has a communication unit K, a control device C, and a storage unit Mn connected to the control device C. An ML data set MLn is stored in each storage unit Mn of each terminal device EDn.

[0110] The forest fire early detection system 1 also includes a plurality of permanently installed ML units ML, which are also arranged in different regions R1, R2, R3, R4. These ML units ML are connected to gateways G via a single-hop connection FSK like a terminal device EDn. Alternatively or additionally, the forest fire early detection system 1 can have one or a plurality of mobile ML modules 310 (see Fig. 7). These mobile ML modules 310 are suitable for collecting measurement data from several regions R1, R2, R3, R4 of the forest W to be monitored in order to create ML data sets ML1, ML2, ML3, ML4.

[0111] A gateway G has a communication interface to both a terminal device EDn for data exchange and a border gateway BGD. The connection to the border gateway BGD can be established via a meshed multi-hop network (MHF), while the connection to the terminal device ED is a single-hop connection (FSK). The two communication interfaces of the gateway G use different communication channels, so the sender can be identified via the communication channel used.

[0112] A border gateway (BGD) has a communication interface to a gateway (G) and to the network server (NS). The border gateway (BGD) then sends the data from an end device (ED), which was sent to the border gateway (BGD) via a single-hop and multi-hop connection, directly to the network server (NS) using the Internet Protocol (IP). Communication between the border gateway (BGD) and the network server (NS) can be via wired (WN) or wireless (IP). Each communication interface of the border gateway (BGD) uses its own communication channel, which is different from the other communication interfaces.

[0113] The network server NS is connected to a second control unit, which is suitable and intended to execute a machine learning program. In particular, the ML data set MLn of a terminal device EDn is updated via the second control unit, preferably at regular intervals.

[0114] Fig. 3 shows a section of the forest fire early detection system 1 according to the invention with terminals ED2, ED3, ED4 arranged therein. The forest fire early detection system 1 is the one described in the previous exemplary embodiment (see Fig. 2). The section (Fig. 3 a) shows three different regions R2, R3, R4 of the forest W to be monitored, with several second terminals ED2 being arranged in region R2, several third terminals ED3 being arranged in region R3, and several fourth terminals ED4 being arranged in the fourth region. The gateway G is connected to other gateways G, BGD as well as to the terminals EDn. A permanently installed ML module ML records measurement data to create ML data.

[0115] The terminal devices ED2, ED3, ED4 arranged in the different regions R2, R3, R4 are identical in construction (see Figures 4, 5, 6) and have a sensor device S that has one or more sensors for detecting a forest fire. In addition, each terminal device ED2, ED3, ED4 has a communication unit K for communicating with a gateway G and a power supply device E. The power supply device E enables autonomous operation of the terminal device and is a rechargeable battery that can be charged via solar cells. All of these components of a terminal device ED2, ED3, ED4 are connected to the control device C and are controlled by it. The control device C is simultaneously an evaluation device for evaluating the measurement signals supplied by the sensor device S and has a suitable software program for this purpose. A memory unit Mn is also connected to the control device C.The second control device C of the second terminal ED2 (Fig. 3 b), which is arranged in the second region R2, is connected to the second storage unit M2. The second storage unit M2 has a second ML data set ML2. Analogously, the third control device C of the third terminal ED3 (Fig. 3 c), which is arranged in the third region R3, is connected to the third storage unit M3. The third storage unit M3 has a third ML data set ML3. The fourth control device C of the fourth terminal ED4 (Fig. 3 d), which is arranged in the fourth region R4, is likewise connected to the fourth storage unit M4, which has a fourth ML data set ML4.

[0116] Not shown in this figure is the first region R1, in which the first terminal devices ED1 are arranged. These first terminal devices ED1 are also structurally identical to the terminal devices ED2, ED3, and ED4 shown here and have the same components. The first control device C of the first terminal device ED1 is connected to the first storage unit M1, which has a first ML data set ML1.

[0117] To determine the different ML data sets ML1, ML2, ML3, ML4, the permanently installed ML units measure ML parameters depending on the sensors S1.1, S1.2, S1.3, S1.4, S2.1, S2.2, S2.3, S2.4, S3.1, S3.2, S3.3, S3.4 arranged in the ML units. In this embodiment, the parameters temperature of the ambient air are measured by the sensors S1.1, S1.2, S1.3, S1.4, the humidity of the ambient air is measured by the sensors S2.1, S2.2, S2.3, S2.4 and the air pressure is measured by the sensors S3.1, S3.2, S3.3, S3.4.

[0118] The first measurement data and the second measurement data are recorded at different times of day. First and second measurement data recorded at different times of day are usually also different. Not only is the air temperature, for example, usually different at midday than at night, but the gas compositions of the ambient air and its humidity are also different due to the different insolation. In a further embodiment of the invention, the first measurement data and the second measurement data are recorded at different times of the year. First and second measurement data recorded at different times of the year are usually also different. Not only is the air temperature, for example, usually different in winter than in summer, but the gas compositions of the ambient air and its humidity are also different due to the different insolation.These different measurement data are incorporated into the ML data created using a machine learning model, increasing the reliability and sensitivity of early detection of a forest fire. In this example, the first and second measurement data are collected every 12 hours by the ML units. Furthermore, the first and second measurement data are collected by the same ML unit.

[0119] The control device C of an ML unit ML collects the parameters of the sensors S1.1, S1.2, S1.3, S1.4, S2.1, S2.2, S2.3, S2.4, S3.1, S3.2, S3.3, S3.4 of the ML unit ML, stores them in the memory unit of the ML unit ML and creates measurement data from them.

[0120] The values ​​of the measured parameters and consequently the measurement data created differ depending on the region R1, R2, R3, R4 in which the permanently installed ML units ML are arranged and are therefore characteristic for the individual regions R1, R2, R3, R4: The ML units ML arranged in the first region R1 create a first measurement data set, the ML units ML arranged in the second region R2 create a second measurement data set, the ML units ML arranged in the third region R3 create a third measurement data set, the ML units ML arranged in the fourth region R2 create a fourth measurement data set.

[0121] These four different measurement data sets are sent by the ML units ML via a single-hop connection to one or more gateways G, from a gateway G via a multi-hop connection to a border gateway BGD and via an IP connection to the network server NS. The network server NS finally sends the four different measurement data sets to a second control unit, which is coupled to the network server NS. On the second control unit, a machine learning algorithm is applied to the four different measurement data sets, thus generating four different ML data sets ML1, ML2, ML3, ML4, whereby the first ML data set ML1 is generated from the first measurement data set, the second ML data set ML2 from the second measurement data set, the third ML data set ML3 from the third measurement data set, and the fourth ML data set ML4 from the fourth measurement data set.

[0122] In addition, ML data sets ML1, ML2, ML3, ML4 for the detection of forest fires are implemented in the forest fire early warning system 1. The ML unit ML records tn parameters from the forest W to be monitored, which is not experiencing a forest fire. Such parameters are the temperature of the ambient air via the sensors S1.1, S1.2, S1.3, S1.4, the humidity of the ambient air via the sensors S2.1, S2.2, S2.3, S2.4 and the air pressure via the sensors S3.1, S3.2, S3.3, S3.4 of a forest W that is not experiencing a forest fire. From these tn parameters, tn measurement data of the forest W to be monitored are generated. The recorded tn measurement data are therefore generated from a forest W that is also not experiencing a forest fire. The tn measurement data are preferably collected and generated from the natural forest W, for which the forest fire early warning system 1 is intended. The tn measurement data are also different for each region R1, R2, R3, R4.From these tn measurement data collected in this way, regionally different tn-ML data are then created, with the tn-ML data differing from region to region. The different tn-ML data are therefore assigned to different regions R1, R2, R3, and R4 of the forest W to be monitored and are compared with the measurement data collected by the terminal device EDn.

[0123] Advantageously, as explained, not only tn parameters and tn measurement data are recorded and generated from true-negative events, but tp measurement data are also recorded from true-positive events. In the simplest case, a true-positive event is a forest fire. For this purpose, for example, forest components such as the fauna found in forests, forest soil components, and / or loose material on the forest floor are heated and / or burned at various temperatures in a laboratory, and the resulting gases are detected. This creation of tp measurement data is also carried out for a forest W to be equipped with a forest fire early detection system 1, specifically according to region R1, R2, R3, R4.

[0124] The tp-ML dataset is determined from these experimentally obtained measurement data in the laboratory. Therefore, a region-specific ML dataset MLn is generated with data on true-positive events—that is, events whose measurement data represent a forest fire. To determine the result data, the measured values ​​recorded from a specific region of the forest W to be monitored are compared with the tn-ML data assigned to this region. This enables the terminal device EDn, with the aid of its control device C, to compare the recorded measurement data with the tn-ML data and tp-ML data. If the recorded result data match the stored tp-ML data, it sends a corresponding message to the network server NS via the communication interface K of the terminal device ED.

[0125] The ML model's algorithm enables improved application-specific detection of the gases to be measured. Furthermore, the algorithm corrects the measured gas concentration for the measured humidity. In addition, the baseline and long-term deviations of the measured values ​​are compensated. For this purpose, data on different gas compositions and their concentrations are provided to the end device, which is compared with the gas compositions and their concentrations determined by the sensor device.

[0126] The implementation of the ML data sets ML1, ML2, ML3, ML4 in the terminal device EDn enables adaptation and application of the machine learning model to the local conditions and specifically for the region R1, R2, R3, R4 of the respective terminal device EDn, at the same time the energy consumption of a single terminal device EDn is reduced.

[0127] A further embodiment of the invention utilizes reinforcement learning. The algorithm learns tactics through reward and punishment for how to act in potentially occurring situations in order to maximize the utility of the agent (i.e., the system to which the learning component belongs). The algorithm learns a function from given pairs of inputs and outputs. During the learning process, a "teacher" provides the correct function value for an input. The goal of supervised learning is that, after several computational runs with different inputs and outputs, the network is trained to make associations.These four generated region-specific ML data sets ML1, ML2, ML3, ML4 are sent via a multi-hop connection and a single-hop connection by radio to each individual terminal device EDn arranged in the forest fire early detection system 1, whereby the first ML data set ML1 is sent to the terminal devices ED1 arranged in the first region R1, the second ML data set ML2 is sent to the terminal devices ED2 arranged in the second region R2, the third ML data set ML3 is sent to the terminal devices ED3 arranged in the third region R3, and the fourth ML data set ML4 is sent to the terminal devices ED4 arranged in the fourth region R4. Within a region R1, R2, R3, R4, all EDn arranged therein therefore have the same ML data set MLn.

[0128] Fig. 4, Fig. 5, and Fig. 6 each show variants of exemplary embodiments of a permanently installed ML unit ML and an analog terminal EDn, which are arranged within a system 1 for early forest fire detection (see Fig. 2). The ML unit ML, like a terminal EDn, is also a sensor for detecting a forest fire.

[0129] The ML units ML and end devices EDn shown here each have a self-sufficient power supply E. In the simplest case, the power supply E is a battery, which can also be designed to be rechargeable. However, the use of capacitors, particularly supercapacitors, is also possible. A somewhat more complex and cost-intensive energy supply E, but one that offers a very long service life, is the use of solar cells.

[0130] By means of the communication interface K, ML data packets from the ML unit ML are sent wirelessly as a data packet to a gateway G using a single-hop FSK connection via LoRa (chirp frequency spread modulation) or frequency modulation. The control device C is connected to and controls the sensors S1, S2, S3, the power supply E and the communication interface K. The control device C also has the memory unit Mn, which has an ML data set Mn. In the simplest case (Fig. 4), the ML unit ML and the terminal device EDn have a temperature sensor S1 (Fig. 4 b), whereby the ML unit ML has two identically constructed temperature sensors S1.1, S1.2 (Fig. 4 a). A terminal EDn (Fig. 5 b) and an ML unit ML can also each have a temperature sensor S1 and a humidity sensor S2, wherein the ML unit ML in turn has two temperature sensors S1.1, S1 that are identical to one another.2 and two identically constructed humidity sensors S2.1, S2.2 (Fig. 5 a). A terminal device EDn (Fig. 6 b) and an ML unit ML can also each have a temperature sensor S1 and a humidity sensor S2, wherein the ML unit ML similarly has three identically constructed temperature sensors S1.1, S1.2, S1,3 and three identically constructed humidity sensors S2.1, S2.2, S2.3 (Fig. 6 a).

[0131] By arranging two (or more) identical sensors S1.1, S1.2, S2.1, S2.2, S3.1, S3.2, design-related deviations and inaccuracies of the installed sensors S1.1, S1.2, S2.1, S2.2, S3.1, S3.2 are recorded, averaged, and compensated for when measuring data. All of these components are housed in a housing to protect them from the elements. The permanently installed ML module ML assumes the function of an end device EDn unless the ML module ML is collecting measurement data for creating ML data sets MLn.

[0132] Fig. 7 shows an embodiment of a forest fire detection unit 300 and a terminal device EDn, each having three different sensors S1, S2, S3. The forest fire detection unit 300 (Fig. 7a) has the ML unit 310. The forest fire detection unit 300 is designed as a flight-capable drone that is autonomous, automatic, and / or remotely controllable. The forest fire detection unit 300 has a drive unit 320 with a plurality of rotors 322 driven by motors 321. The motors 321 are typically electric motors and are supplied with energy by a rechargeable energy storage device (battery). The forest fire detection unit 300 can be controlled by pivoting the rotors 322 and varying the speed of the individual motors 321. The forest fire detection unit 300 according to the invention also has a navigation sensor 350 that detects objects in the vicinity of the forest fire detection unit 300.The navigation sensor 350 has one or a plurality of cameras and / or time-of-flight-based sensors (e.g., radar, ultrasound, lidar) that detect obstacles during the flight of the forest fire detection unit 300. The obstacles are detected, recognized, and analyzed by the control unit arranged in the forest fire detection unit 300 such that the forest fire detection unit 300 automatically avoids the obstacles during its flight.

[0133] In addition, the forest fire detection unit 300 comprises the actual sensor unit S, which comprises a plurality of sensors S1.1, S1.2, S1.3, S1.4, S2.1, S2.2, S2.3, S2.4, S3.1, S3.2, S3.3, S3.4. The sensors S1.1, S1.2, S1.3, S1.4, S2.1, S2.2, S2.3, S2.4, S3.1, S3.2,

[0134] 53.3, S3.4 are connected to the computing unit C. The forest fire detection unit 300 comprises four identically constructed temperature sensors S1.1, S1.2, S1.3, S1.4, four identically constructed humidity sensors S2.1, S2.2, S2.3, S2.4 and four identically constructed pressure sensors S3.1, S3.2, S3.3, S3.4. Furthermore, sensors for gas analysis can be arranged in the forest fire detection unit 300. By arranging two (or more) identically constructed sensors S1.1, S1.2, S1.3, S1.4, S2.1, S2.2,

[0135] 52.3, S2.4, S3.1 , S3.2, S3.3, S3.4, design-related deviations and inaccuracies of the installed sensors S1.1 , S1.2,

[0136] 51.3, S1.4, S2.1, S2.2, S2.3, S2.4, S3.1, S3.2, S3.3, S3.4 are recorded, averaged, and compensated. To record tn-ML measurement data, the forest fire detection unit 300 has the ML unit 310, which is detachably connected to the forest fire detection unit 300 via the connection 312 in the receptacle 311. This movable forest fire detection unit 300 also assumes the function of an end device EDn, provided the forest fire detection unit 300 is not recording measurement data for creating ML data sets MLn.

[0137] The terminal EDn has an analogue structure with three different sensors S1, S2, S3 (Fig. 7 b). To detect a forest fire, the terminal EDn has a temperature sensor S1, a sensor S2 for measuring the humidity, and a sensor S3 for measuring the air pressure. In order to be able to install and operate the terminal EDn in inhospitable and particularly rural areas far from any power supply, the terminal EDn is equipped with a self-sufficient power supply E. The terminal EDn also has a communication interface K and a control device C, which is connected to the storage unit Mn.

[0138] The communication interface K of the terminal device EDn is wirelessly connected to the communication interfaces of the gateways G. The control device C is also connected to the communication interface K and the sensors S1, S2, S3 and controls them.

[0139] LIST OF REFERENCE SYMBOLS

[0140] 1 forest fire detection system

[0141] 10 Network

[0142] EDn, ED1, ED2, ED3, terminal / forest fire detection sensor ED4

[0143] G Gateway

[0144] NS Internet Network Server

[0145] IP Internet Protocol

[0146] MHF multi-hop radio network

[0147] ML ML unit

[0148] BGD Border Gateway

[0149] FSK FSK modulation

[0150] WN Wired connection

[0151] W Forest

[0152] R1, R2, R3, R4 regions

[0153] F Flowing water

[0154] 310 ml unit

[0155] 311 recording

[0156] 312 Detachable connection

[0157] 320 Flight propulsion / propulsion unit

[0158] 321 engine

[0159] 322 Rotor

[0160] 350 navigation sensor

[0161] C Control device

[0162] E Device for supplying energy

[0163] K Communication interface

[0164] Mn, M1, M2, M3, M4 storage unit

[0165] MLn, ML1, ML2, ML3,ML data set ML4 S sensor device

[0166] 51, S1.1, S1.2, S1.3, temperature sensor

[0167] S1.4

[0168] 52, S2.1, S2.2, S2.3, humidity sensor

[0169] S2.4

[0170] 53, S3.1, S3.2, S3.3, pressure sensor

[0171] S3.4

Claims

PA TE N CLAIMS Method for determining ML data for a system (1) for Early forest fire detection with the procedural steps • Determination of initial measurement data from measurements of initial parameters of a forest (W) • Determining second measurement data from measurements of second parameters of a forest (W) • Creating first ML data (ML1) from the first measurement data • Creating second ML data (ML2) from the second measurement data, wherein the second ML data (ML2) are different from the first ML data (ML1). Method for determining ML data for a system (1) for early forest fire detection according to claim 1, characterized in that the first ML data (ML1) are stored on a first terminal (ED1) of the Forest fire early detection system (1). Method for determining ML data for a system (1) for forest fire early detection according to claim 1 or 2, characterized in that the second ML data (ML2) are stored on a second terminal (ED2) of the forest fire early warning system (1) should be implemented.

4. Method for determining ML data for a system (1) for Early detection of forest fires after one or more of the preceding Claims, characterized in that first terminals (ED1) acquire measurement data in a first region (R1) and second End devices (ED2) measure data in a second region (R2), wherein the first region (R1) differs from the second region (R2) with regard to vegetation or the like.

5. Method for determining ML data for a system (1) for Early detection of forest fires after one or more of the preceding Claims, characterized in that the first parameters of the forest (W) are different from the second parameters of the forest (W).

6. Method for determining ML data for a system (1) for Forest fire early detection according to claim 5, characterized in that the first parameters of the forest (W) are characteristic of a first region (R1) of the forest (W) and the second parameters of the forest (W) are characteristic of a second region (R2) of the forest (W), wherein the first region (R1) of the forest (W) is different from the second region (R2) of the forest (W).

7. Method for determining ML data for a system (1) for Early detection of forest fires after one or more of the preceding Claims, characterized in that the first and / or second measurement data are determined from measurements of tn parameters of a forest (W) under tn conditions. Method for determining ML data for a system (1) for early forest fire detection according to one or more of the preceding claims, characterized in that the first and / or second measurement data are determined from measurements of tp parameters of a forest (W) under tp conditions. Method for determining ML data for a system (1) for early forest fire detection according to one or more of the preceding claims, characterized in that the first (ML1) and / or second ML data (ML2) are created from the tn measurement data. Method for determining ML data for a system (1) for early forest fire detection according to one or more of the preceding claims, characterized in that the first (ML1) and / or second ML data (ML2) are created from tp measurement data.Method for determining ML data for a system (1) for early forest fire detection according to one or more of the preceding. Claims, characterized in that tn-ML data and / or tp-ML data for the detection of forest fires are implemented in a system (1) for early forest fire detection (1). Method for determining ML data for a system (1) for Forest fire early detection according to one or more of the preceding claims, characterized in that the first and / or second measurement data are acquired by a permanently installed ML module (ML). Method for determining ML data for a system (1) for forest fire early detection according to claim 12, characterized in that the first and / or second measurement data are acquired by several permanently installed ML modules. (ML), wherein the individual permanently installed ML modules (ML) are installed in different regions (R1, R2, R3, R4) of the forest (W) to be monitored. Method for determining ML data for a system (1) for early forest fire detection according to claim 12 or 13, characterized in that the permanently installed ML module (ML) takes over the function of a terminal device (EDn), provided that the permanently installed ML module (ML) does not record first and / or second measurement data. Method for determining ML data for a system (1) for Early detection of forest fires after one or more of the preceding Claims, characterized in that the first and / or second measurement data are acquired by a mobile ML module (300). A method for determining ML data for a system (1) for early forest fire detection according to claim 15, characterized in that a change of location of the mobile ML module (300) occurs between the acquisition of the first and second measurement data. A forest fire detection system (1) with a network (10) comprising • a first terminal (ED1), wherein the first terminal (ED1) comprises a sensor device (S), a first control device (C) with a first memory unit (M1), an evaluation device for evaluating measurement signals supplied by the sensor device and a power supply device (E), wherein a first ML data set (ML1) is stored on the first memory unit (M1), • a second terminal (ED2), wherein the second terminal (ED2) has a sensor device, a second control device (C) with a second memory unit (M2), an evaluation device for evaluating measurement signals supplied by the sensor device and a power supply device (E), wherein a second ML data set (ML2) is stored on the second memory unit (M2), wherein the second terminal (ED2) is identical in construction to the first terminal (ED1), • a network server (NS), characterized in that the first ML data set (ML1) is different from the second ML data set (ML2). Forest fire early detection system (1) with a network (10) according to claim 17, characterized in that the first control device (C) is suitable and intended to access a first memory (M1) containing a first ML data set (ML1) from the adaptation and application of a machine learning model, and / or the second control device (C) is suitable and intended to access a second memory (M2) containing a second ML data set (ML2) from the adaptation and application of a machine learning model. Forest fire early detection system (1) with a network (10) according to claim 17 or 18, characterized in that the machine learning model comprises tn-ML data.A forest fire early detection system (1) with a network (10) according to one or more of claims 17 to 19, characterized in that the forest fire early detection system (1) has an ML module (ML, 300) that is suitable and intended to acquire measurement data for creating ML data (ML1, ML2). A forest fire early detection system (1) with a network (10) according to claim 20, characterized in that the forest fire early detection system (1) has a permanently installed ML module (ML). A forest fire early detection system (1) with a network (10) according to claim 21, characterized in that. The forest fire early detection system (1) comprises a plurality of permanently installed ML modules (ML), wherein at least two of the ML modules (ML) are arranged in different regions (R1, R2, R3, R4) of the forest (W) to be monitored.

23. A forest fire early detection system (1) with a network (10) according to one or more of claims 17 to 22, characterized in that the forest fire early detection system (1) comprises a mobile ML module (ML).

24. Forest fire early detection system (1) with a network (10) according to claim 23, characterized in that the mobile ML module is suitable for collecting measurement data from several regions (R1, R2, R3, R4) of the forest (W) to be monitored in order to create ML data sets (ML1, ML2, ML3, ML4).

Images