Integrated system for gas leak detection and analysis, and method of calibration of an ai trained to detect gas leaks

Description

1 / 23 Integrated system for gas leak detection and analysis, and calibration method for an AI trained to detect leaks. GAS FIELD OF THE INVENTION

[001] The present invention falls within the field of gas leak detection. More specifically, the present invention relates to the detection of leaks of toxic and flammable gases with real-time image generation of gas clouds, supporting decision-making during emergency scenarios. FUNDAMENTALS OF THE INVENTION

[002] The effective detection and analysis of toxic and flammable gas leaks in industrial environments is of vital importance. Such leaks pose significant risks to human and environmental safety, and can lead to disasters such as explosions and severe contamination. Accurate and rapid detection of these leaks is crucial to avoid such consequences; however, existing methods face significant difficulties due to the complexity of industrial facilities where gases are processed under extreme pressure and temperature conditions.

[003] Traditional leak detection systems often rely on technologies that detect only specific types of gases or operate in isolation without effective data integration, resulting in a limited ability to identify and quantify leaks in real time. These systems may include point detectors that need to be close to the leak location, or air sampling technologies, which are slow and ineffective in dynamic and ventilated environments. Furthermore, many of these systems do not provide comprehensive views of the facility's status, limiting their usefulness in emergencies.

[004] Another limitation of conventional approaches is the inability to handle multiple types of gases simultaneously, each requiring different detection methods due to their chemical and physical properties and necessitating the installation of multiple devices specific to each type of gas of interest. This not only increases the complexity and installation costs, but also makes maintenance and calibration more difficult. Petition 870250016671, dated 28 / 02 / 2025, page 43 / 73 2 / 23 of the system is more complex. Point gas detectors, for example, may fail to identify small leaks or leaks occurring outside the direct range of the sensors. Interference from dust, humidity, or other chemicals in the air can further reduce the accuracy of traditional sensors. Furthermore, areas with restricted or hazardous access complicate the installation and maintenance of gas detectors, challenging continuous monitoring.

[005] Air sampling systems, while useful in some contexts, are limited by the speed at which air samples can be analyzed and by the fact that detection is often late, after the gas has already dispersed substantially. This reduces the usefulness of these systems in preventing incidents or minimizing the consequences of a leak.

[006] In summary, the state-of-the-art approach lacks an integrated solution that can offer rapid, accurate, and comprehensive gas leak detection. The need for a technology capable of integrating data from different sensory sources in real time is clear, aiming to overcome the limitations of existing systems and provide a more effective response to potentially dangerous situations. STATE OF THE ART

[007] Document CN220817471U, entitled “A multi-dimensional visual inspection system for pipes”, discloses a multi-dimensional visual inspection system for a pipeline and aims to solve the problem of low inspection efficiency caused by the fact that a worker on an existing inspection system cannot visually observe the position of a leak point.Comprising a switching unit, a wireless communication unit, a data transmission unit, and a data acquisition unit, a power supply module and a plurality of interfaces are integrated into the switching unit; the input end of the wireless communication unit is connected to the switching unit via a network cable, the output end of the wireless communication unit is connected to a base station, the input end of the data transmission unit is connected to the switching unit via a network cable, and the output end of the unit... Petition 870250016671, dated 28 / 02 / 2025, page 44 / 73 3 / 23 Data transmission is communicated with a mobile terminal. According to the inspection system provided by the utility model, real-time monitoring and positioning of pipeline leakage conditions are performed through the acquisition and processing of optical images and sound information. Leakage conditions and the positions of leakage points can be clearly observed through visualization of sound information and superimposition of sound information and optical images, thus improving detection efficiency. The technology also includes an IR sensor with an acoustic sensor for precise leak detection.

[008] Document US2023071544A1, entitled “ACOUSTIC DETECTION DEVICE AND SYSTEM WITH REGIONS OF INTEREST”, discloses a system that includes: an acoustic detection device comprising a plurality of microphones; an output or means of communication for sending a signal or message; a memory for storing a location of at least one region of interest and an associated threshold value; a processing circuit for detecting a location and a sound pressure level of a potential sound source; for testing whether the detected sound source is located in a region of interest, and for testing whether the detected sound pressure level is above the threshold, and if a result of these tests is true, sending a signal or message to operate a component, such as an audiovisual component or a valve or similar.

[009] The technology also provides an acoustic detection device comprising: a plurality of microphones spaced in at least two directions and configured to convert acoustic waves originating from a scene into a plurality of analog or digital sound signals; an output for sending a signal to an external component and / or a communication medium for sending a message to an external processor; a memory for storing information from at least one region of interest, including such location information and at least one threshold level; a processing circuit connected to said microphones, said output and / or said communication medium, and said memory, and configured to: a) receive the plurality of sound signals and detect a location and a sound pressure level or Petition 870250016671, dated 28 / 02 / 2025, p. 45 / 73 4 / 23 a value derived therefrom from one or more potential sound sources in the said scene; and b) for at least one detected sound source, test whether the detected sound source is located in a Region of Interest; and test whether the detected Sound Pressure Level or the value derived therefrom is greater than at least one threshold associated with that region; and if both conditions are met, send a first signal and / or a first message indicating an alarm.

[0010] Document CN112116566, entitled “Land oil and gas pipeline defect diagnosis method based on hyperspectral remote sensing technology”, discloses a method for detecting gas leaks along a pipeline that uses hyperspectral images and machine learning and neural network techniques to determine whether or not a leak is occurring in real time. More specifically, the method in this reference comprises a method for diagnosing defects in onshore oil and gas pipelines based on hyperspectral remote sensing technology, comprising the following steps: Step 1: Hyperspectral image acquisition; Step 2: Surface data extraction; Step 3: Data classification; Step 4: Classification of hyperspectral pipeline data by the BP neural network and defect area detection; Step 5: Verification of the classification results from step 4 with a KNN nearest neighbor node machine learning algorithm.

[0011] Document WO2021034910A1, entitled “System and method for cyberphysical inspection and monitoring of nonmetallic structures”, discloses a technological solution for analyzing a sequence of image frames of the electromagnetic spectrum of a nonmetallic asset and detecting or predicting an aberration in the asset, including a detected or predicted location of the aberration.The technological solution involves receiving electromagnetic spectrum image frames from a pair of different types of machine learning systems, applying a machine learning algorithm to the electromagnetic spectrum image frames to stratify the electromagnetic spectrum images into levels of abstraction according to an image topology and generate the first aberration determination information, applying a second machine learning algorithm to the electromagnetic spectrum image frames to detect patterns in electromagnetic spectrum images over time and produce... Petition 870250016671, dated 28 / 02 / 2025, page 46 / 73 5 / 23 second aberration determination information, generate an aberration assessment based on the first and second aberration determination information, and transmit the aberration assessment to a communication device, including prediction of an aberration and a location of the aberration on the non-metallic asset.

[0012] The disclosed technology comprises a field transducer (FT) device 20 that may include a gas sensor that can detect, measure, or monitor one or more types of gases. The radiant energy sensor may function as a gas sensor in applications where a sequence of images of a gas can be captured by the radiant energy sensor and the images analyzed to detect or predict the gas. The gas sensor may include, for example, an electrochemical sensor, a catalytic sphere sensor, an IR camera, a FLIR camera, a hyperspectral camera, or any other sensor device that can detect a variety of different gases that may be contained in the asset 10. The gas sensor may include one or more spectral or hyperspectral sensors, each configured to collect image data in a narrow spectral band, including image data relating to the transmittance, absorption, or reflectance of electromagnetic energy by gas molecules.An additional machine learning system (e.g., CNN or RNN or ENN) or an additional layer on the CNN or RNN can be applied to distinguish between different gases and classify gases according to gas type, concentration, flow vector (including, for example, flow direction, velocity, magnitude, and changes in flow direction, velocity, or magnitude as a function of time). An FT 20 device equipped with a hyperspectral camera can use the camera both as a radiant energy sensor and as a gas sensor, capturing images in different regions of the electromagnetic spectrum.

[0013] Document CN114509214A, entitled “Monitoring system and method integrating multiple sensors”, discloses a monitoring system and method that integrates multiple sensors, and the system comprises an intelligent patrol terminal that is provided with a robotic arm, a first communication module and a first data processing module, and a pressure sensing device that is arranged in a pipe connection part and is provided with a second communication module. The data processing server is used to process data. Petition 870250016671, dated 28 / 02 / 2025, page 47 / 73 6 / 23 acquired by at least one intelligent patrol terminal and / or at least one pressure detection device; and the central server is connected to the data processing server and is used to judge whether or not there is a gas leak according to the data transmitted by the data processing server and send a processing instruction according to a judgment result. The intelligent patrol terminal and the pressure detection device are used to collect data, the data processing server is used to process and analyze the collected data, then the central server comprehensively assesses whether or not there is a gas leak, the pipeline leak condition can be accurately assessed, and the leak location can be precisely located.

[0014] The multisensor fusion monitoring system includes: an intelligent patrol terminal, which is equipped with a robotic arm, a first communication module and a first data processing module; a pressure detection device located at the pipeline connection, the pressure detection device is equipped with a second communication module; a data processing server for processing data collected by at least one of the intelligent patrol terminals and / or at least one of the pressure detection devices; a central server connected to the data processing server to judge whether there is a gas leak according to the data transmitted by the data processing server and send processing instructions according to the result of the judgment.The central server is also configured to send a data reprocessing instruction to the data processing server or send a data acquisition instruction to the smart patrol terminal and / or the pressure detection device when it cannot be determined whether there is a gas leak according to the existing data. SUMMARY OF THE INVENTION

[0015] The proposed invention addresses these challenges through an integrated system that employs sensing technologies across multiple ranges of the electromagnetic spectrum and in the acoustic domain to detect, monitor, and quantify leaks in real time. This system overcomes the limitations of existing technologies through Petition 870250016671, dated 28 / 02 / 2025, page 48 / 73 7 / 23 Data fusion from different sensors, increasing the accuracy, speed, and reliability of detection. The present invention describes a system composed of two or more optical image-forming sensors, operating in different spectral ranges, such as MWIR (Medium Wave Infrared), LWIR (Long Wave Infrared), and the visible spectrum, capable of generating two-dimensional images and, together, measuring the absorbance of gases sensitive to the sensor ranges. The system also includes a microphone array, allowing for the precise location of sound sources, and combining this data with the optical sensors to locate, characterize, and quantify gas leaks. The data collected by the optical and acoustic sensors are processed by artificial intelligence (AI) algorithms, which interpret information related to the detection, location, classification, and quantification of gases.These algorithms utilize supervised learning techniques, which are refined and calibrated in the operating environment to increase the accuracy of measurements. The calibration process uses computer graphics-based simulations to generate simulated leaks on images collected from the final operating environment, allowing for the training and refinement of AI models. The calibration process is completed within minutes of installation using a high-performance embedded computing platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The present invention will now be described with reference to typical embodiments thereof and also with reference to the accompanying drawings, in which:

[0017] Figure 1A is a block representation of the integrated gas leak detection system according to the present invention;

[0018] Figure 1B is a block representation of the central plate according to the present invention;

[0019] Figure 1C is an illustrative representation of an embodiment of the present invention;

[0020] Figure 1D is an alternative view of figure 1C;

[0021] Figure 2 is a representation of a training modality of Petition 870250016671, dated 28 / 02 / 2025, page 49 / 73 8 / 23 Embedded AI according to the present invention;

[0022] Figure 3 is a representation of the image mask generated by the AR system according to the present invention;

[0023] Figure 4 is a representation of an interface containing data captured by the sensors, according to the present invention.

[0024] Figure 5 shows examples of absorbance responses of some gases in the electromagnetic spectrum, more specifically within a long wavelength infrared range (Figure 5A) and a narrow medium wavelength range (Figure 5B). DETAILED DESCRIPTION OF THE INVENTION

[0025] Specific embodiments of this disclosure are described below. In an effort to provide a concise description of these embodiments, all features of an actual implementation may not be described in the descriptive report. It should be appreciated that in the development of any actual implementation, as in any engineering or design project, numerous specific implementation decisions must be made to achieve the specific objectives of the developers, such as compliance with system and business-related constraints, which may vary from one implementation to another. Furthermore, it should be appreciated that such a development effort may be complex and time-consuming, but would nevertheless be a routine design and manufacturing undertaking for those of ordinary skill having the benefit of this disclosure.

[0026] The present invention is described generally and without limitation with reference to Figures 1A and 1B. The technology disclosed herein consists of an integrated system (100) of multiple sensor devices communicating with a central board (110) comprising a processing unit (106), a previously trained embedded Artificial Intelligence (107) that uses the data detected by the multiple sensor devices as input to determine if a gas leak is occurring, and a communication module (108). As seen in Figure 1A, the integrated system (100) of the present invention comprises at least one thermal sensor (101), a set of spectral range sensors (102) comprising at least one range sensor. Petition 870250016671, dated 28 / 02 / 2025, pp. 50 / 73 9 / 23 spectral, but preferably multiple spectral band sensors (band 1... band N), and an array of acoustic sensors (103).

[0027] The integrated system (100) may also comprise a visible spectrum camera (104) and an Augmented Reality (AR) system (105) also communicating with the central board (110) to generate an AR image relating to the detected leak, as will be explained in more detail later.

[0028] The visible spectrum camera (104) is configured to capture light reflected by objects, rather than thermal emissions, operating in the wavelength range of 400 to 700 nanometers (nm). Preferably it uses CCD (charge coupled device) or CMOS (complementary metal oxide semiconductor) type sensors to convert visible light into electrical signals, which are processed to form digital images. CCD offers high sensitivity and image quality, while CMOS is more energy efficient.

[0029] In a preferred embodiment, the multiple spectral band sensors comprise at least one MWIR (Medium Wave Infrared) and LWIR (Long Wave Infrared) sensor.

[0030] The MWIR sensor is a cooled infrared thermal module designed to detect gas leaks. It uses materials such as InSb (indium antimonide) or HgCdTe (mercury-cadmium telluride) in its high-sensitivity infrared detector with a NETD (noise equivalent temperature difference) of at least 15 mKF and typically operates in the wavelength range between 3.2 and 3.5 pm. Preferably, this sensor has a resolution of at least 320x256 pixels and a pixel pitch (distance between pixel centers) of 30 pm, with resolution dependent on the distance and area to be imaged, and can be optimized for the operating range of interest. The use of cooling is necessary to reduce thermal noise, improve measurement accuracy, and detect small temperature variations, especially in gas leaks.

[0031] The LWIR sensor typically operates in the wavelength range between 7.5 and 13.5 pm and uses an uncooled microbolometer, usually made of VOx (oxide). Petition 870250016671, dated 28 / 02 / 2025, pp. 51 / 73 (10 / 23 vanadium). This sensor detects thermal radiation and converts it into thermal images, being widely used due to its ability to operate without cooling, while maintaining good thermal sensitivity and efficiency for real-time temperature measurements.

[0032] In a preferred embodiment, the thermal sensor (101) is a thermal camera, the MWIR and LWIR sensors are infrared cameras and the acoustic sensors (103) are a microphone array.

[0033] In a preferred embodiment, the acoustic sensor array is a microphone array, configured to create an acoustic image by capturing sounds from different sources in the environment. The microphones detect the arrival time and intensity of the sound waves, processing this data to generate a visual map that indicates the location of the sound sources. This process is essential for identifying gas leaks, machine noises, or other sound sources in complex environments.

[0034] The integrated system (100) also comprises a high-performance embedded platform for processing data efficiently. Preferably, a GPU (graphics processing unit) with a computational capacity of at least 275 TOPS (trillions of operations per second) and a CPU (central processing unit) with at least 12 cores are used. Preferably, the platform should have multiple input and output interfaces and be energy efficient, making the system 100 suitable for field applications such as gas leak monitoring and detection.

[0035] A preferred embodiment of the integrated system (100) of the present invention is shown in Figures 1C and 1D, where all the sensor devices mentioned above are mounted in a single encapsulation (120) to be installed in the area to be monitored, it being understood that the preferred encapsulation (120) is not limiting.

[0036] The central plate present in the encapsulation (120) may also contain a Petition 870250016671, dated 28 / 02 / 2025, pp. 52 / 73 11 / 23 Communication module (not shown) wired or wireless for communication with a remote station to send and receive parameter data from local servers or a cloud. The communication module may include, without limitation, fiber optic, LAN, WAN, Wi-Fi, Bluetooth, radio frequency, or any other communication technology deemed suitable.

[0037] The deployment of the technology disclosed herein consists of installing one or more of the encapsulations (120) exemplified in figures 1C and 1D in an industrial plant and commissioning the communication and data processing subsystems. Part of this commissioning is performed using multiple images obtained from the multiple integrated optical and acoustic sensors of the invention, used as input for the embedded AI (107) which uses an advanced segmentation model that combines simultaneous detection and classification of regions of interest in images. This model is designed to identify specific objects and events, such as leaks, with high precision and speed, even in complex scenarios. It is based on architectures that perform real-time processing, optimized to locate and categorize patterns in images, leveraging heat maps to highlight critical regions.Initial training is conducted using real data from controlled environments, where variables such as quantity, type of gas, and leakage conditions are adjusted. Subsequently, the model is refined with synthetic data integrated with real images captured in the monitoring environment, increasing its robustness and reliability.

[0038] One or more encapsulations (120) perform the monitoring and detection of gas leaks in pipelines, valves and equipment. The sensor devices provide spectroscopy, acoustic and thermal maps, evidencing the detection, location and quantification of gas leaks through sensing by thermal (101), spectral (102) cameras and microphone array (103).

[0039] With measurements throughout the pipelines and equipment, it is possible to detect leaks of toxic and flammable gases by generating images in different ranges of the electromagnetic spectrum and in the acoustic domain in real time of the gas clouds. Analysis techniques based on hyperspectral imaging allow the estimation of the gas type by spectroscopy, the concentration by signal intensity and Petition 870250016671, dated 28 / 02 / 2025, pp. 53 / 73 12 / 23 the gas volume with the reconstruction of the gas cloud. The data obtained are processed in the processing unit (106), employing pattern identification algorithms, such as machine learning and image processing to generate real-time information and historical trend analysis, enabling decisions for maintenance and risk mitigation. Based on the images obtained by the sensor devices, the embedded AI (107) determines whether or not a gas leak has been detected.

[0040] AI training is provided by acquiring images of leak tests in diverse environments, forming an image dataset. The dataset is divided into training data and validation data, which are subjected to a neural network architecture. The image dataset must contain images of leaks under various conditions and at least an equal number of images of masks, i.e., templates showing where gas leaks exist. This set of image pairs is divided into 70% for training and 30% for validation of the neural network. The neural network used is based on an advanced convolutional segmentation architecture, manually designed, composed of 11 main layers. This architecture is optimized to capture details at both global and local levels, allowing for the precise segmentation of complex patterns in images.It utilizes dense connections between layers to improve feature extraction and information reconciliation at different scales, resulting in greater accuracy and efficiency in visual data processing. The artificial neural network is designed for image segmentation with an input of at least 224x224 pixels and 3 channels (RGB). It has a contraction path (encoder) composed of 5 convolutional blocks with 16, 32, 64, 128, and 256 filters, followed by MaxPooling2D and Dropout layers for regularization. The expansion path (decoder) uses 5 convolutional transposition layers (Conv2DTranspose) to increase the spatial dimension, each concatenated with the respective encoder layers. The output layer is a 1x1 convolution with sigmoid activation, designed to produce a binary segmentation mask.The model uses the RMSprop optimizer and the binary_crossentropy loss function, measuring performance with the IoU (Intersection over Union) metric. The network is trained with 128 epochs and batches of 4 images.

[0041] A convolutional segmentation network like the one used, works Petition 870250016671, dated 28 / 02 / 2025, pp. 54 / 73 13 / 23 identifying and separating specific regions of interest within an image, in this case, the areas where there is a gas leak. The process involves feeding the neural network with pairs of images and their corresponding segmentation masks, i.e., the image pairs, which act as templates to indicate the exact areas where the gas is present. The training set is used to adjust the network parameters, while the validation set evaluates the model's performance on unseen data, monitoring for possible signs of training bias. The IoU performance metric measures the accuracy of the overlap between the predicted and actual masks.

[0042] Diversity ensures greater generalization of the application, allowing the same neural network to be used in different environments. However, this condition impairs the system's efficiency in the final application in noisy and dynamic environments. To overcome this, a calibration system is used. When the camera with the trained AI is positioned, a set of images of the environment to be observed is acquired. This set is passed through a simulator or artificial plume generator (onboard or on an external server) that generates gas events which, in turn, are allocated to the database for calibration. Calibration is performed by loading the previously trained base network and adjusting its parameters with new training that uses real images of the monitored environment, combined with leak simulations generated from synthetic data.After the adjustment, network validation is conducted with new synthetic images containing simulated leaks and other artifacts designed to confuse the model. Validation is only accepted when the network reaches a pre-established confidence level, demonstrating accurate segmentation capability without generating significant false positives or false negatives, as assessed by specific performance metrics such as accuracy, recall, and F1-score.

[0043] Therefore, the idea of ​​training is to have enough data so that the network can function properly in different environments. For example, an image dataset comprising at least 350 images, plus 350 corresponding mask images, is usually considered sufficient for pre-training. Finally, calibration adapts the AI ​​to the new environment, allowing it to better adapt to the location, reducing detection errors. This calibration is performed in Petition 870250016671, dated 28 / 02 / 2025, pp. 55 / 73 14 / 23 minutes, allowing for almost immediate field installation without the need for additional neural network assessments and training.

[0044] The present invention is configured to combine data obtained from sensor devices, i.e., thermal sensors (101), spectral sensors (102), acoustic sensors (103), visible spectrum camera (104) and AR system (105), to create representations based on the fusion of imaging data in different ranges of the electromagnetic spectrum and in the acoustic domain, enabling an augmented reality system, map generation, activation of inhibitor mechanisms and alarms. Preferably, said representations are created by the embedded platform. This approach facilitates punctual campaigns as it does not depend on data infrastructure or electrical power. In another embodiment, said representations can be created on the external server, which is more suitable for permanent installations where the use of centralized server processing is a more optimized solution.

[0045] Optionally, the processing unit (106) may comprise multiple processing units, each specialized to process the imaging data received from each of the multiple sensor devices (101), (102), (103) and (104) to obtain more refined processing.

[0046] Sensors (101), (102), and (104) are chosen to have the ability to capture images in various ranges of the electromagnetic spectrum, including visible, ultraviolet, and infrared spectra, providing a comprehensive view of the monitored environment. This allows visualization of a wide range of phenomena that would be invisible to the human eye, from subtle temperature variations to the presence of hazardous gases in the air.

[0047] The use of the visible spectrum camera (104) together with the thermal camera (101) and cameras sensitive to specific gas spectra (102), such as narrow wavelength bands of ultraviolet and infrared, offers a comprehensive view of the monitored environment. Each gas has a unique absorbance spectrum, which is said to be a signature that defines how that gas absorbs different wavelengths of light. Absorbance refers to the ability of a substance to absorb light at a given wavelength. When light passes through a gas, some of it is absorbed. Petition 870250016671, dated 28 / 02 / 2025, pp. 56 / 73 15 / 23 by the gas at specific wavelengths, depending on the chemical composition of the gas. This absorption reduces the intensity of light passing through the gas at certain wavelengths, creating a unique pattern that can be detected and analyzed. By equipping the integrated system (100) with cameras (102) capable of detecting these specific absorption patterns, it is possible to identify the presence of different gases in the monitored environment. This is particularly useful for the detection of toxic and flammable gases, which can pose significant risks to safety and health. These cameras, by focusing on the absorbance spectra that correspond to the gases of interest, can detect the presence of these gases even at low concentrations, before they become dangerous.

[0048] In particular, the spectral band camera array (102) shall include at least one spectral band camera configured to be sensitive to the absorbance of a specific gas to be detected. Optionally, the spectral band camera array 102 may include multiple spectral band cameras, each configured to be sensitive to a specific gas. Thus, the present invention can be adjusted to detect the leakage of more than one type of gas using only one encapsulation (120), enabling the identification of toxic and / or flammable gases based on their unique spectral signatures. This approach allows detection even at low concentrations, anticipating safety risks.

[0049] The microphone array (103) acts as an acoustic imaging system, playing a crucial role in the detection and location of gas leaks. These sensors are designed to capture and locate sounds and vibrations, including those imperceptible to the human ear, which may be indicative signs of the presence or onset of a gas leak. The microphone array (103) is configured to use beamforming to isolate and locate the source of a leak sound with a wide acoustic spectral range, even in noisy environments. Beamforming is a process in which signals captured by multiple microphones are combined in such a way as to focus on a specific direction while attenuating sounds from other directions. This is achieved by adjusting the phases of the captured signals, allowing the system to direct its hearing to the sound source. In the case of detection Petition 870250016671, dated 28 / 02 / 2025, pp. 57 / 73 16 / 23 of gas leaks, this technique is particularly valuable because microphones can be oriented to isolate and pinpoint the sound source of a leak, even in a noisy environment where multiple sounds may be present.

[0050] The integration of data obtained by sensors (101), (102), and (104) and by acoustic sensors (103) is managed by the processing unit (106) of the central board (110), which may use hardware accelerators with artificial intelligence, responsible for real-time coordination, self-calibration and automatic adjustment of system parameters to optimize detection in different environmental conditions. The self-calibration or automatic adjustment is performed by AI previously trained and retrained in the monitoring environment during the decommissioning period. The self-calibration capability is crucial to maintain the accuracy of the system in dynamic environments, where factors such as ambient light and background noise can vary significantly.

[0051] As mentioned earlier, calibration is achieved by obtaining real data from the final installation environment and simulating events on it so that the AI ​​can undergo a further training process for the specific environment. To perform self-calibration, the embedded platform is additionally configured to take a set of images of the final environment through the visible spectrum camera (104). For example, the set of images of the final environment may have 30 images. The images are added to the original database, making it more specific to the environment where the system (100) of the present invention is installed. The image pairs (original and with simulated defects) are split in a 70% to 30% ratio again. The previously generalized network is loaded into the system and retrained with only 12 epochs, using batches of 4 images, performing fine-tuning to improve accuracy in the specific environment.Self-calibration under these conditions takes about 5 minutes, which is very fast.

[0052] The embedded AI (107) is based on artificial neural networks, specifically segmentation networks, playing a fundamental role in the integrated system (100) of the present invention. Segmentation networks are a type of neural network designed to perform image segmentation tasks, i.e., dividing a Petition 870250016671, dated 28 / 02 / 2025, pp. 58 / 73 17 / 23 image in parts or segments, identifying and isolating specific objects within it and painting the area with a surface that is called a mask. An example is provided in Figures 3 and 4.

[0053] The embedded AI (107) is pre-trained to recognize patterns associated with different types of gases, enabling a detailed analysis of detected leaks. The training is carried out in such a way as to enable the embedded AI (107) to recognize and visually isolate specific gas events in the environment, based on their spectral signatures, sound origin and visual patterns. This allows the integrated system (100) to identify the presence and location of gas leaks accurately and in real time. For each acquired domain, whether electromagnetic or acoustic, the embedded AI (107) performs a segmentation of the respective image to highlight the pixels where an anomaly is identified. The segmentation process can be performed in both 2D (each isolated image) and 3D (the temporal evolution of the image).This exponentially expands the system's capacity, allowing for a deeper understanding of the extent and dynamics of leaks over time. Since 2D and 3D segmentation networks are widely known and consolidated in the technique, their detailed description is not the focus here. However, it is important to emphasize that, in the context of this application, these networks are optimized to efficiently identify anomalies associated with gas leaks, considering the unique spectral, acoustic, and visual characteristics of each gas type. Furthermore, their implementation is configured to ensure high precision in real-time segmentation, essential for the immediate monitoring and control of detected events, maximizing the effectiveness of the integrated system (100).

[0054] Optionally, the embedded AI (107) may comprise multiple AIs, each specifically chosen and trained to analyze the image generated by a specific sensor device. For example, the embedded AI (107) may comprise an AI specially adapted to analyze images obtained by the thermal camera (101), an embedded AI specially adapted to analyze images obtained by the microphone array (103), and an AI specially adapted for each of at least one or more spectral band cameras in the spectral band camera array 102. Petition 870250016671, dated 28 / 02 / 2025, pp. 59 / 73 18 / 23 In this way, the detection of potential leaks can be carried out even more accurately.

[0055] In the case of 2D segmentation, training focuses on static images, identifying and marking the presence of gases with pixel-by-pixel precision, which is crucial for determining the exact location and extent of the leak at a given time. This process is enriched by the use of masks, which visually highlight the affected areas, making identification intuitive and immediate for system operators. On the other hand, the inclusion of 3D segmentation introduces a temporal dimension to the analysis process, allowing the system to track and visualize the dynamics of gas leaks over time. Using video sequences, the 3D segmentation network analyzes how gases move, disperse, or increase in volume, offering valuable insight into leak behavior, as well as greater robustness.

[0056] To train segmentation networks, a database is needed, consisting of images or videos that have already been labeled or categorized with relevant information, such as the location and type of gases present. The database serves as a guide for the neural network, allowing it to learn to recognize specific patterns associated with different types of gases. Training the network involves adjusting the weights of the connections between neurons to minimize the error between the network's predictions and the actual labels provided by the database.

[0057] To train segmentation networks, a database is needed, consisting of images or videos that have already been labeled or categorized with relevant information, such as the location and type of gases present. The database serves as a guide for the neural network, allowing it to learn to recognize specific patterns associated with different types of gases. Training the network involves adjusting the weights of the connections between neurons to minimize the error between the network's predictions and the actual labels provided by the database.

[0058] With reference to Figure 2, the self-calibration or automatic calibration of the segmentation network is performed by retraining the neural network to adapt to new environments or conditions soon after it is installed in the location to be monitored. Petition 870250016671, dated 28 / 02 / 2025, pages 60 / 73 19 / 23 This is done by combining simulations with information collected in real time from the environment.

[0059] Initially, the segmentation network (AI) is pre-trained using a generalized database containing a wide range of gas leak scenes. This initial training provides the segmentation network with a solid knowledge base, allowing it to recognize common leak patterns and characteristics in a variety of scenarios. The goal of this phase is to equip the segmentation network with the ability to efficiently assess gas leaks, even before it is specifically calibrated for a new environment. Preferably, the training set comprises at least 350 images without events and their respective counterparts where an event occurs, real or simulated.

[0060] After the system is installed at its final location, the autocalibration step is performed as previously mentioned. The embedded platform is additionally configured to take images of the final environment using the visible spectrum camera (104). Preferably, this set of images of the final environment should have at least 30 images. The set of images of the final environment is submitted to an embedded artificial plume simulator or generator or on an external server to simulate gas clouds in the final environment using computer graphics techniques. The clouds are generated at random positions within the images and with variable, but realistic, characteristics. For each generated cloud, a segmentation mask is automatically created, providing the gas detection template simultaneously.These images are then added to the original image dataset, making it more specific to the environment where the system (100) of the present invention is installed.

[0061] In addition to plume simulations, thermal simulations are added to the virtual environment creation platform to simulate temperature events, essential for calibrating the thermal anomaly detection network. Similarly, images formed from data collected by acoustic sensors (103) are used to generate sound events. These simulated events help calibrate the network for sound anomaly detection, ensuring that the system can identify with Petition 870250016671, dated 28 / 02 / 2025, pp. 61 / 73 20 / 23 accurately pinpoints the origin of leaks through acoustic signals.

[0062] With the new dataset, including those generated specifically for the new environment, the AI ​​is retrained. This retraining phase adjusts the AI ​​to recognize and analyze gas leaks under the unique conditions of the installation site. The process considers both two-dimensional (2D) data, which includes hyperspectral image content fused with acoustic imaging of sound sources, and three-dimensional (3D) data, which includes temporal content, significantly enriching the system's detection and analysis capabilities. The inclusion of temporal data allows for the training of three-dimensional neural networks, similar to networks used in magnetic resonance imaging, which take into account not only 2D images but the entire set across layers. For 3D detection applied to the camera, the temporal axis is considered for image integration.

[0063] The image pairs (original and with simulated defects) from the original image dataset plus the final environment image dataset are split in a 70% to 30% ratio again. The previously generalized AI is loaded into the system and retrained with only 12 epochs, using batches of 4 images, performing a calibration to improve accuracy in the specific environment.

[0064] The AR system (105) gives the integrated system (100) the ability to integrate and present the information processed by the neural networks directly onto the image captured by the visible spectrum camera (104), creating a data-rich and interactive augmented reality interface. This approach allows users to see the real world through the visible spectrum (104) with additional layers of digitally superimposed data that provide detailed information about the detected gas leaks.

[0065] With reference to figures 3 and 4, the information presented in augmented reality is derived from mask processing via embedded AI (107), and includes data such as the mass of gas leaked, the time elapsed since the start of the leak, the instantaneous gas flow rate, the direction of the leak, the average and minimum ambient temperature, the location of the leak sound source and the mask itself. Petition 870250016671, dated 28 / 02 / 2025, pp. 62 / 73 21 / 23 gas shape. This data is derived from statistical analyses and profiling of the image set characteristics. The visual information from the masks throughout the images provides physical attributes of interest of the gas, such as the direction and velocity of the gas during the propagation of the image masks over the acquisition time of the system, as well as the number of pixels in the mask and opacity relative to the background image to extract the instantaneous gas flow rate and mass. In addition, factors such as temperature can be extracted directly from thermal spectrum camera information, while the sound source comes from acoustic imaging.

[0066] This data is visualized in real time, allowing the user a comprehensive understanding of the leak event. For example, visualization of the leaked mass and instantaneous flow rate helps assess the severity of the leak, while the direction of the leak can indicate where the gas is moving, vital information for decision-making during an emergency response. Ambient temperature and the location of the sound source complement the picture, offering a complete view of the situation, which is fundamental for assessing the associated risks and planning mitigation actions. In addition, additional object detection networks of interest, such as fire detection, can be implemented in the visible spectrum camera (14) as extra information modules.

[0067] This data is visualized in real time, allowing the user a comprehensive understanding of the leak event. For example, visualization of the leaked mass and instantaneous flow rate helps assess the severity of the leak, while the direction of the leak can indicate where the gas is moving, vital information for decision-making during an emergency response. Ambient temperature and the location of the sound source complement the picture, offering a complete view of the situation, which is fundamental for assessing associated risks and planning mitigation actions. In addition, additional object detection networks, such as fire detection, can be implemented in the visible spectrum camera 14 as extra information modules.

[0068] The system was designed to facilitate access to and integration of data with other platforms. All information is processed and displayed in augmented reality. Petition 870250016671, dated 28 / 02 / 2025, pp. 63 / 73 22 / 23 are simultaneously redirected to a cloud-based platform. This not only ensures that the data is accessible from anywhere, but also allows for seamless integration with other security and alert systems. The availability of data in the cloud opens up several possibilities for advanced analytics, long-term storage, and sharing critical information with emergency response teams, facility management, and regulatory authorities.

[0069] The present invention can also be configured so that the communication module (108) communicates with a cloud-based platform (not shown). This enables advanced analysis, long-term storage, and sharing of critical data with emergency response teams and regulatory authorities. Furthermore, the communication module can communicate with a remote station linked to the monitored environment, in order to report a detected leak so that preventive or emergency actions can be taken.

[0070] The main differentiators of the proposed solution are: speed and ability to identify, locate, and quantify leaks of toxic and flammable gases; coverage area encompassing the entire piping network and equipment; generation of historical data; and failure prediction. These capabilities are possible with the integration of multiple optical sensors operating in different spectral ranges capable of generating images of different gases, given their ability to detect various absorbances of these gases. Additionally, acoustic images generated from arrays of acoustic sensors combined with beamforming techniques integrated with the optical sensors are capable of feeding artificial neural networks that segment gas clouds, providing identification, location, and quantification of the gases of interest.

[0071] The proposed solution has the unique potential to establish a high spatial resolution, real-time sensing network for large industrial plants such as refineries, platforms, and FPSOs. The ability to quickly detect small leaks and / or leaks of explosive gases is a major safety advantage, mitigating risks to life and property. The continuous and omnipresent monitoring enabled by the proposed system impacts the risk analysis of processes and consequently the costs of maintenance processes and Petition 870250016671, dated 28 / 02 / 2025, pp. 64 / 73 23 / 23 security, including the value of plant insurance, benefits strongly advocated by digital transformation.

[0072] Comprehensive monitoring of the plant and its pipelines increases operational safety to unprecedented levels given the potential to generate real-time alerts informing of risk zones and identifying the level of risks.

[0073] It is important to highlight that the proposed system can, in addition to detecting, locating and quantifying gas leaks, generate numerous information for monitoring energy efficiency, structural integrity and various operational processes.

[0074] The impacts resulting from the development and application of the proposed system are: elimination and / or reduction and / or complementation of fixed detection systems with a significant reduction in the calibration costs of currently installed detectors; elimination of human exposure during toxic and / or flammable gas leak scenarios; and the ability to determine the direction and orientation of gas clouds as well as the overpressure generated in case of ignition, thus guiding various tactical emergency response actions.

[0075] Although aspects of the present disclosure may be susceptible to various modifications and alternative forms, specific embodiments have been shown as examples in the drawings and have been described in detail in this document. But it should be understood that the invention is not intended to be limited to the particular forms disclosed. Instead, the invention should cover all modifications, equivalents and alternatives that fall within the scope of the invention, as defined by the following appended claims. Petition 870250016671, dated 28 / 02 / 2025, pp. 65 / 73

Claims

1 / 2 CLAIMS 1. Integrated system (100) for gas leak detection and analysis, characterized in that it comprises: a thermal sensor (100); a spectral band sensor array (102) comprising at least one spectral band sensor; an acoustic sensor (103); and a central board (110) including a processing unit (106), an embedded AI (107), and a communication module (108), wherein the embedded AI (107) is pre-trained to detect gas leaks, wherein the processing unit (106) is configured to receive imaging data obtained from the thermal sensor (100), the spectral band sensor array (102) and the acoustic sensor (103) and use them as input data for the embedded AI (107).

2. Integrated system (100), according to claim 1, characterized in that it further comprises: a visible spectrum camera (104) and an AR system (105), wherein the processing unit (106) is additionally configured to receive imaging data obtained from the visible spectrum camera (104) and provide the data obtained from the thermal camera (100), the spectral band camera array (102), the microphone array (103) and the visible spectrum camera (104) to the AR system (105), wherein the AR system (105) is configured to generate an augmented reality image based on the data obtained from the processing unit (106).

3. Integrated system (100), according to claim 2, characterized in that the augmented reality image contains data relating to the mass of gas leaked, the time elapsed since the start of the leak, the instantaneous gas flow rate, the direction of the leak, the average and minimum ambient temperature, the location of the leak sound source and a mask of the shape of the leaked gas.

4. Integrated system (100), according to any of the preceding claims, characterized in that the processing unit (106) is additionally configured to calibrate the embedded AI (107).

5. Integrated system (100), according to claim 4, characterized in that the processing unit (106) is additionally configured to periodically recalibrate the embedded AI (107).

6. A calibration method for an AI trained to detect gas leaks, wherein the AI ​​was trained using a training image set comprising images of an environment without leaks and counterpart images of an environment with real or simulated leaks, the method characterized in that it comprises the steps of: obtaining visible spectrum images of a new environment; performing gas leak simulations on the visible spectrum images of the new environment; and retraining the AI ​​using the training image set, the visible spectrum images of the new environment, and the visible spectrum images of the new environment with simulated gas leaks. Petition 870250016671, dated 28 / 02 / 2025, pp. 67 / 73

Images