Methane leakage laser detection method and system based on industrial vision

Abstract

The present application relates to the technical field of methane leakage laser detection based on industrial vision, and particularly relates to a methane leakage laser detection method and system based on industrial vision, which comprises the following steps: monitoring laser signal intensity, and capturing a signal intensity drop event when the laser signal intensity drops by more than a preset intensity threshold; for the signal intensity drop event, extracting the speed, amplitude and fluctuation frequency of the signal intensity drop to obtain laser attenuation mode characteristics; analyzing the color, transparency, shape, internal texture, edge and moving track of objects in the laser path area to obtain visual characteristics; correlating the laser attenuation mode characteristics and the visual characteristics, and according to the correlation result, determining whether the signal intensity drop is caused by target gas absorption or non-target aerosol scattering to obtain a determination result; and outputting corresponding warning information according to the determination result. The above can improve the accuracy and reliability of detection.

Description

Technical Field

[0001] This invention relates to the technical field of laser detection of methane leaks based on industrial vision, and specifically to a method and system for laser detection of methane leaks based on industrial vision. Background Technology

[0002] In industrial production and safety monitoring, early and accurate detection of leaks of flammable gases such as methane is crucial. Laser absorption spectroscopy, due to its high sensitivity and rapid response, is widely used for methane leak detection. These systems typically determine the presence and concentration of methane by emitting a laser beam of a specific wavelength through the monitored area and measuring the light intensity attenuation at the receiver. However, in complex industrial environments, in addition to the absorption of the target gas methane, there are often non-target aerosols such as water vapor, dense fog, or dust. These aerosols cause significant attenuation of the laser signal through scattering, leading to misjudgments of methane leaks. While traditional industrial vision systems can capture images of the scene, their image processing logic often struggles to effectively identify these variable-form and-concentration interfering substances and distinguish them from actual leaks, thus affecting the accuracy and reliability of detection. Summary of the Invention

[0003] The purpose of this invention is to address the aforementioned shortcomings by proposing a laser detection method and system for methane leakage based on industrial vision.

[0004] The present invention adopts the following technical solution:

[0005] A laser detection method for methane leaks based on industrial vision, comprising the following steps:

[0006] Monitor the laser signal intensity and capture the signal intensity drop event when the laser signal intensity drops below a preset intensity threshold;

[0007] For signal strength degradation events, the speed, amplitude, and fluctuation frequency of signal strength degradation are extracted to obtain laser attenuation mode characteristics;

[0008] Simultaneously, the color, transparency, shape, internal texture, edges, and movement trajectory of objects within the laser path area are analyzed to obtain visual features;

[0009] The laser attenuation mode features are correlated with visual features, and based on the correlation results, it is determined whether the signal intensity decrease is caused by absorption by the target gas or by scattering by non-target aerosols, thus obtaining the discrimination result;

[0010] Based on the judgment results, the corresponding warning information is output.

[0011] This technical solution effectively combines laser attenuation mode characteristics and industrial vision features to accurately distinguish between methane leaks and non-target aerosol interference, significantly reducing false alarm rates and improving detection accuracy and reliability.

[0012] Furthermore, the method also includes the following steps:

[0013] The laser attenuation mode characteristics and visual characteristics were standardized to obtain the original evidence.

[0014] Configure an evidence weight set for complex aerosol scenarios. The weight set is set based on the discriminative power of different features in distinguishing between methane leakage and complex aerosol interference.

[0015] For each characteristic piece of evidence in the original evidence, calculate its strength of evidence for the two possibilities of methane leakage and complex aerosol interference. The strength of evidence is calculated using a preset range of characteristic values ​​and the corresponding strength function.

[0016] By combining the evidence weight set for complex aerosol scenarios, the evidence strength of all features and their corresponding weights are comprehensively calculated to obtain two overall confidence scores for methane leakage and complex aerosol interference.

[0017] When both overall confidence scores are lower than the preset high confidence threshold, or when the difference between the two overall confidence scores is less than the preset discrimination margin, it is determined to be a fuzzy discrimination and the enhanced analysis mode is triggered. The enhanced analysis mode includes increasing the laser signal sampling frequency, extending the sampling time, increasing the video capture frame rate, and performing higher resolution local image acquisition.

[0018] At the same time, the system outputs preliminary judgment results with low confidence indicators to the operator and displays the main features that lead to fuzzy judgments.

[0019] Furthermore, the steps to trigger the enhanced analytics mode include:

[0020] Real-time monitoring of processor load, memory usage, and data transfer bandwidth provides information on system resource utilization.

[0021] Based on the difference between the overall confidence scores of the system resource usage and the fuzzy discrimination, the data acquisition parameters of the augmentation analysis are dynamically adjusted to obtain the augmented laser signal and visual image data.

[0022] The acquired laser signals and visual image data are compressed and feature dimensionality reduced in real time to obtain preprocessed and dimensionality-reduced enhanced feature data.

[0023] Using a multi-threaded or parallel processing architecture, data acquisition, preprocessing, and feature extraction tasks are assigned to different processing cores, which helps to improve processing speed.

[0024] An event-triggered transmission mechanism is used to transmit preprocessed and dimensionality-reduced enhanced feature data.

[0025] Using the preprocessed and dimensionality-reduced enhanced feature data, cross-modal information association and discrimination are re-performed.

[0026] Furthermore, the steps for standardizing the laser attenuation mode characteristics and visual features to obtain the original evidence include:

[0027] Continuously monitor the ambient light intensity, temperature, and air humidity in the laser path area to obtain environmental parameters;

[0028] Real-time analysis of the historical fluctuation range and current trend of environmental parameters yields dynamic information on environmental parameters;

[0029] Based on the identified types of interference, the typical dynamic range of the laser attenuation mode characteristics and visual characteristics of that type of interference under different environmental conditions is obtained.

[0030] Based on environmental parameters, dynamic information of environmental parameters, and typical dynamic range, the parameters of the standardized process are adaptively adjusted.

[0031] Based on the adaptively adjusted standardized processing parameters, the laser attenuation mode features and visual features are standardized to obtain the original evidence.

[0032] Furthermore, the steps for configuring the evidence weight set for complex aerosol scenarios include:

[0033] Continuously monitor the ambient light intensity, temperature, air humidity, and air pressure in the laser path area to obtain environmental monitoring parameters;

[0034] The composition, morphology, and physical properties of composite aerosols are analyzed in real time using an industrial vision system to obtain aerosol characteristics.

[0035] Based on environmental monitoring parameters and aerosol characteristics, the discriminative power of each feature in distinguishing between methane leakage and complex aerosol interference was evaluated.

[0036] Based on the discriminative power of the assessment, the set of evidence weights for complex aerosol scenarios is dynamically adjusted.

[0037] Furthermore, for each characteristic piece of evidence in the original evidence, the strength of its evidence for the two possibilities of methane leakage and complex aerosol interference is calculated. The steps for calculating the strength of evidence using a preset characteristic value range and the corresponding strength function include:

[0038] Continuously monitor the ambient light intensity, temperature, air humidity, and air pressure in the laser path area to obtain environmental monitoring parameters;

[0039] The composition, morphology, and physical properties of composite aerosols are analyzed in real time using an industrial vision system to obtain aerosol characteristics.

[0040] Based on environmental monitoring parameters and aerosol characteristics, the range of characteristic values ​​used to calculate the strength of evidence is dynamically updated.

[0041] At the same time, the shape and slope of the intensity function curve are dynamically adjusted based on environmental monitoring parameters and aerosol characteristics;

[0042] Based on the dynamically updated eigenvalue range and the dynamically adjusted intensity function, the strength of evidence for each characteristic in the original evidence against the two possibilities of methane leakage and complex aerosol interference is calculated.

[0043] Furthermore, the steps for comprehensively calculating the evidence strength of all features and their corresponding weights to obtain two overall confidence scores for methane leakage and complex aerosol interference include:

[0044] The evidentiary strength of laser attenuation mode features, the evidentiary strength of visual features, and their corresponding weights;

[0045] The evidentiary strength of laser attenuation pattern features and visual features were normalized.

[0046] The evidence strength of the normalized laser attenuation pattern features and the evidence strength of the visual features are initially weighted according to the weights to obtain the preprocessed data.

[0047] The preprocessed data is input into the first layer of the multilayer sensing network;

[0048] The first layer input is transformed by a non-linear activation function, and then non-linear transformation and feature combination are performed through the hidden layer to obtain the deep fused features.

[0049] Based on the deeply fused features, the output layer generates two overall confidence scores for methane leakage and complex aerosol interference.

[0050] Furthermore, the steps of transforming the first-layer input using a non-linear activation function include:

[0051] The laser attenuation mode features and visual features in the first layer input are preprocessed to obtain the preprocessed features.

[0052] The degree of deviation between the preprocessed features and the historical average and standard deviation is calculated to obtain information on the degree of deviation.

[0053] Based on the deviation information, the input to the activation function is dynamically adjusted to obtain the adjusted activation function input;

[0054] Based on the dynamic range and rate of change of the preprocessed features, the slope parameter of the activation function is adaptively adjusted to obtain the adjusted slope parameter.

[0055] By combining the adjusted activation function input and the adjusted slope parameter, the first layer input is transformed using a nonlinear activation function.

[0056] Furthermore, the steps for dynamically adjusting the input to the activation function include:

[0057] Determine whether the activation function input before adjustment is an extremely high or extremely low value;

[0058] When the value is determined to be extremely high, the activation function input before adjustment is non-linearly compressed using a preset logarithmic compression function to obtain the adjusted activation function input. The adjusted activation function input is mapped to the effective processing limit of the subsequent processing module.

[0059] When the value is determined to be extremely low, the activation function input before adjustment is nonlinearly extended through a preset exponential expansion function to obtain the adjusted activation function input. The adjusted activation function input is mapped to the effective processing lower limit above the subsequent processing module.

[0060] When the value is determined to be non-extreme, the activation function input before adjustment is adjusted by a linear scaling function to obtain the adjusted activation function input. The adjusted activation function input is kept within the effective processing range of the subsequent processing modules.

[0061] This application also discloses a methane leak laser detection system based on industrial vision, applied to a methane leak laser detection method based on industrial vision. The system includes:

[0062] The monitoring module is used to monitor the laser signal intensity and capture the signal intensity drop event when the laser signal intensity drops below a preset intensity threshold.

[0063] The extraction module extracts the speed, amplitude, and fluctuation frequency of the signal intensity decrease for signal intensity decrease events, thereby obtaining the laser attenuation mode characteristics;

[0064] The analysis module is used to analyze the color, transparency, shape, internal texture, edges, and movement trajectory of objects within the laser path area to obtain visual features;

[0065] The discrimination module is used to correlate laser attenuation mode features with visual features, and based on the correlation results, to determine whether the signal intensity decrease is caused by absorption by the target gas or by scattering by non-target aerosols, and obtain the discrimination result.

[0066] The early warning output module outputs corresponding early warning information based on the judgment results.

[0067] This application improves the accuracy and reliability of methane leak detection, reduces unnecessary production interruptions and economic losses, and enhances staff's trust in automated monitoring systems.

[0068] To further understand the features and technical content of the present invention, please refer to the following detailed description and drawings of the present invention. However, the drawings provided are for reference and illustration only and are not intended to limit the present invention. Attached Figure Description

[0069] Figure 1 This is a flowchart of a laser detection method for methane leakage based on industrial vision according to the present invention.

[0070] Figure 2 This is a schematic diagram of the structure of a methane leak laser detection system based on industrial vision according to the present invention. Detailed Implementation

[0071] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can understand the advantages and effects of the present invention from the content disclosed in this specification. The present invention can be implemented or applied through other different specific embodiments, and various details in this specification can also be modified and changed based on different viewpoints and applications without departing from the spirit of the present invention. Furthermore, the accompanying drawings of the present invention are for simple illustrative purposes only and are not depictions of actual dimensions; this is stated in advance. The following embodiments will further describe the relevant technical content of the present invention in detail, but the disclosed content is not intended to limit the scope of protection of the present invention.

[0072] This embodiment provides a laser detection method and system for methane leaks based on industrial vision, combined with... Figure 1 and Figure 2 As shown.

[0073] refer to Figure 1 A laser detection method for methane leaks based on industrial vision, comprising the following steps:

[0074] Monitor the laser signal intensity and capture the signal intensity drop event when the laser signal intensity drops below a preset intensity threshold;

[0075] For signal strength degradation events, the speed, amplitude, and fluctuation frequency of signal strength degradation are extracted to obtain laser attenuation mode characteristics;

[0076] Simultaneously, the color, transparency, shape, internal texture, edges, and movement trajectory of objects within the laser path area are analyzed to obtain visual features;

[0077] The laser attenuation mode features are correlated with visual features, and based on the correlation results, it is determined whether the signal intensity decrease is caused by absorption by the target gas or by scattering by non-target aerosols, thus obtaining the discrimination result;

[0078] Based on the judgment results, the corresponding warning information is output.

[0079] The method proposed in this application is mainly applied to gas leak detection in industrial environments, particularly for leaks of target gases such as methane. In this method, "laser signal intensity" refers to the optical power detected by the receiver after the laser beam passes through the monitored area. When the laser signal intensity drops below a "preset intensity threshold," it indicates possible gas absorption or scattering, and the system captures a "signal intensity decrease event." "Laser attenuation mode characteristics" refer to features obtained by analyzing the rate, amplitude, and fluctuation frequency of signal intensity decrease; these features reflect the dynamic process of laser attenuation. "Visual features" refer to features obtained by analyzing the color, transparency, shape, internal texture, edges, and movement trajectory of objects within the laser path area using an industrial vision system; these features are used to identify and distinguish different interfering substances or target gases. "Target gas absorption" refers to the absorption of specific wavelength laser light by target gas molecules such as methane, leading to laser intensity attenuation. "Non-target aerosol scattering" refers to the scattering of laser light by non-target aerosol particles such as water vapor, dense fog, and dust, also resulting in laser intensity attenuation.

[0080] Firstly, there are several ways to monitor laser signal intensity. For example, a continuous-wave laser can emit a laser beam of a specific wavelength, and a photodetector can receive the laser signal in real time. The photodetector converts the optical signal into an electrical signal, which is then digitized by an analog-to-digital converter. The digitized signal intensity data is continuously monitored. When the continuously monitored laser signal intensity value falls below a preset intensity threshold, the system records this event. Another approach is to use a pulsed laser to emit laser light and monitor the laser signal intensity by measuring the peak intensity of each pulse. When the peak intensity of multiple consecutive pulses falls below a preset threshold, the system detects a signal intensity decrease event.

[0081] Secondly, for the captured signal strength decline events, it is necessary to extract the laser attenuation mode features. For example, by analyzing the time series data before and after the signal strength decline event, the time required for the signal strength to decrease from the normal level to the lowest point can be calculated, which is the decline rate. The magnitude of the signal strength decline can be determined by the difference between the normal strength and the lowest strength. The fluctuation frequency can be obtained by performing Fourier transform or wavelet analysis on the signal fluctuations during the signal strength decline process. The extraction of these features helps to quantify the dynamic characteristics of laser attenuation.

[0082] Simultaneously, to acquire visual features, industrial vision systems analyze the visual information of objects within the laser path area. For example, a high-resolution industrial camera can continuously capture a video stream of the laser path area. Image processing algorithms can analyze video frames in real time, extracting object color information (such as RGB or HSV values), transparency (assessed through background transmittance), shape (such as circular or irregular shapes), internal texture (such as uniformity and graininess), edges (obtained through edge detection algorithms such as the Canny operator), and movement trajectory (achieved through target tracking algorithms such as Kalman filtering). These visual features provide detailed information about the presence and nature of interference objects in the laser path.

[0083] Subsequently, the extracted laser attenuation pattern features are correlated with visual features. For example, laser signal intensity data can be synchronized with industrial visual image data through timestamp alignment. When a signal intensity decrease event is captured, the system retrieves the corresponding visual image data within that time period and matches the laser attenuation pattern features with the visual features extracted from these images. This correlation can be achieved by constructing a multimodal feature vector, which includes information such as the speed, amplitude, and fluctuation frequency of laser attenuation, as well as the color, transparency, shape, texture, edges, and movement trajectory of the visual object.

[0084] Finally, based on the correlation results, it is determined whether the signal intensity decrease is caused by target gas absorption or non-target aerosol scattering, and corresponding warning information is output. For example, machine learning models (such as support vector machines, neural networks, or decision trees) can be used to classify the correlated multimodal feature vectors. During the training phase, this model is trained using a large amount of data on known methane leak events (target gas absorption) and aerosol interference events (non-target aerosol scattering). The trained model can output a discrimination result based on the input feature vector, indicating whether the current event is a methane leak or aerosol interference. If the discrimination result is a methane leak, the system will output a "methane leak warning" message; if the discrimination result is aerosol interference, the system will output an "aerosol interference warning" message, or will not output a methane leak warning.

[0085] This application further proposes a laser detection method for methane leaks based on industrial vision, which includes the following steps:

[0086] The laser attenuation mode characteristics and visual characteristics were standardized to obtain the original evidence.

[0087] Configure an evidence weight set for complex aerosol scenarios. The weight set is set based on the discriminative power of different features in distinguishing between methane leakage and complex aerosol interference.

[0088] For each characteristic piece of evidence in the original evidence, calculate its strength of evidence for the two possibilities of methane leakage and complex aerosol interference. The strength of evidence is calculated using a preset range of characteristic values ​​and the corresponding strength function.

[0089] By combining the evidence weight set for complex aerosol scenarios, the evidence strength of all features and their corresponding weights are comprehensively calculated to obtain two overall confidence scores for methane leakage and complex aerosol interference.

[0090] When both overall confidence scores are lower than the preset high confidence threshold, or when the difference between the two overall confidence scores is less than the preset discrimination margin, it is determined to be a fuzzy discrimination and the enhanced analysis mode is triggered. The enhanced analysis mode includes increasing the laser signal sampling frequency, extending the sampling time, increasing the video capture frame rate, and performing higher resolution local image acquisition.

[0091] At the same time, the system outputs preliminary judgment results with low confidence indicators to the operator and displays the main features that lead to fuzzy judgments.

[0092] Specifically, standardization of laser attenuation pattern features and visual features aims to eliminate differences in the dimensions and numerical ranges of different features, making them comparable and thus providing a unified basis for subsequent calculations of evidence strength. Standardization can employ methods such as Z-score standardization, Min-Max standardization, or Decimal Scaling standardization to ensure that all feature data are compared on a uniform scale.

[0093] The configuration of evidence weight sets for complex aerosol scenarios aims to assign different levels of importance to different features based on their actual discriminative power in distinguishing between methane leaks and complex aerosol interference. For example, certain visual features (such as specific shapes or internal textures) may have higher discriminative power in identifying specific aerosol types, while certain laser attenuation pattern features (such as attenuation rate or amplitude) may be more sensitive to methane absorption. The weight set can be optimized based on expert experience, historical data analysis, or machine learning algorithms to reflect the actual contribution of the features.

[0094] In practical applications, for each feature in the original evidence, its strength of evidence for the two possibilities of methane leakage and complex aerosol interference is calculated. The purpose is to quantify the degree to which each feature supports both possibilities. The strength of evidence can be calculated by mapping feature values ​​to a predefined strength function (e.g., a sigmoid function, a Gaussian function, or a piecewise linear function), which converts feature values ​​into confidence scores between 0 and 1. The predefined range of feature values ​​defines the correspondence between feature values ​​and strengths.

[0095] Furthermore, by combining the evidence weight set for complex aerosol scenarios, the evidence strength of all features and their corresponding weights are comprehensively calculated to obtain two overall confidence scores for methane leakage and complex aerosol interference. This comprehensive calculation can employ methods such as weighted average, fuzzy integral, or Dempster-Shafer evidence theory to integrate evidence from different features, thereby obtaining an overall confidence assessment of the two possibilities.

[0096] When both overall confidence scores are below a preset high confidence threshold, or the difference between the two overall confidence scores is less than a preset discrimination margin, the system is judged as fuzzy. This indicates that the current discrimination result lacks sufficient certainty or discriminative power. In this case, the system will trigger an enhanced analysis mode, which aims to acquire richer discrimination information by collecting more and higher-quality data. Specifically, the enhanced analysis mode includes increasing the laser signal sampling frequency to capture finer attenuation details, extending the sampling time to observe signal changes over a longer period, increasing the video capture frame rate to obtain more continuous visual information, and performing higher-resolution local image acquisition to reveal more microscopic visual features.

[0097] Simultaneously, the system outputs preliminary judgment results with low confidence indicators to the operator and displays the main features leading to fuzzy judgments. This allows the operator to promptly understand the uncertainty of the current judgment and, based on the main features prompted by the system, to conduct targeted manual review or take further measures.

[0098] In some preferred embodiments, it is assumed that a laser detection system detects a decrease in laser signal intensity in a chemical plant area. The system first standardizes the captured laser attenuation pattern features (e.g., attenuation rate of 0.5% / s, amplitude of 20%) and visual features (e.g., white haze within the path area, blurred edges, and uniform internal texture). Then, based on a preset weight set (e.g., in the current humidity environment, the weight of "white haze" in the visual features is higher than that of "blurred edges"), the system calculates the strength of evidence for each standardized feature for both methane leakage and water vapor (a complex aerosol). For example, the attenuation rate of 0.5% / s has a strength of evidence of 0.7 for methane leakage and 0.3 for water vapor; the white haze has a strength of evidence of 0.2 for methane leakage and 0.8 for water vapor. Next, the system combines the weight set to comprehensively calculate the strength of evidence for all features with their corresponding weights, resulting in an overall confidence score of 0.45 for methane leakage and 0.55 for water vapor. Assuming a preset high-confidence threshold of 0.6 and a discrimination margin of 0.1, the system determines a fuzzy discrimination because both overall confidence scores are below 0.6, and the difference of 0.1 is less than or equal to the discrimination margin of 0.1. At this point, the system immediately triggers an enhanced analysis mode, increasing the laser signal sampling frequency from 100Hz to 500Hz, the video capture frame rate from 30fps to 60fps, and acquiring higher-resolution local images of the foggy area. Simultaneously, the system outputs a preliminary discrimination result to the operator: "Possibly water vapor, low confidence," and displays that the main characteristic causing the fuzzy discrimination is "insufficient combined discriminative power of laser attenuation pattern features and visual features." After receiving this information, the operator can either review the data acquired in the enhanced analysis mode or wait for the system to re-discriminate based on the new data, thus avoiding erroneous decisions under uncertain circumstances.

[0099] This application further proposes the following steps for triggering the enhanced analysis mode:

[0100] Real-time monitoring of processor load, memory usage, and data transfer bandwidth provides information on system resource utilization.

[0101] Based on the difference between the overall confidence scores of the system resource usage and the fuzzy discrimination, the data acquisition parameters of the augmentation analysis are dynamically adjusted to obtain the augmented laser signal and visual image data.

[0102] The acquired laser signals and visual image data are compressed and feature dimensionality reduced in real time to obtain preprocessed and dimensionality-reduced enhanced feature data.

[0103] Using a multi-threaded or parallel processing architecture, data acquisition, preprocessing, and feature extraction tasks are assigned to different processing cores, which helps to improve processing speed.

[0104] An event-triggered transmission mechanism is used to transmit preprocessed and dimensionality-reduced enhanced feature data.

[0105] Using the preprocessed and dimensionality-reduced enhanced feature data, cross-modal information association and discrimination are re-performed.

[0106] Specifically, real-time monitoring of processor load, memory usage, and data transmission bandwidth refers to the system continuously acquiring indicators such as CPU utilization, RAM usage, and network interface throughput of the current computing unit to comprehensively understand the system's operating status and available resources. This system resource usage data can serve as a basis for subsequent dynamic adjustments. Specifically, dynamically adjusting the data acquisition parameters for augmented analysis based on the difference between the system resource usage and the overall confidence scores of the fuzzy discrimination can be understood as the system intelligently adjusting parameters such as the laser signal sampling frequency, sampling time, video capture frame rate, and local image acquisition resolution based on the currently available computing and transmission resources and the degree of fuzziness in the discrimination results (i.e., the closeness of the two overall confidence scores). For example, when system resources are sufficient and the discrimination fuzziness is high, the acquisition parameters can be appropriately increased; when resources are scarce, lower but still effective acquisition parameters may be selected while ensuring discrimination accuracy to avoid system overload. In practical applications, real-time compression and feature dimensionality reduction are performed on the enhanced laser signals and visual image data. For example, lossless or lossy compression algorithms (such as JPEG, MPEG, Run-Length Encoding, etc.) can be used to compress the image data, and principal component analysis, linear discriminant analysis, or deep learning encoders can be used to reduce the dimensionality of the laser signals and visual features. The aim is to significantly reduce the data volume while retaining key information, thereby reducing the burden on subsequent processing and transmission. Furthermore, a multi-threaded or parallel processing architecture is employed. For instance, data acquisition tasks can be assigned to one thread or core, data preprocessing to another, and feature extraction to a third. This aims to fully utilize the computing power of multi-core processors to achieve concurrent task execution, thereby significantly shortening the overall data processing time. An event-triggered transmission mechanism is used, which means that data transmission is only initiated when new enhanced feature data is ready or specific transmission conditions are met. This aims to avoid unnecessary continuous data flow, reduce network bandwidth consumption, and ensure the timeliness and effectiveness of data transmission. Finally, the enhanced feature data after preprocessing and dimensionality reduction is used to re-perform cross-modal information association and discrimination. This means that the optimized enhanced feature data is re-input into the discrimination module, and the association analysis between laser attenuation mode features and visual features is performed again. Discrimination is then performed based on richer and more refined data in order to obtain more accurate and higher confidence discrimination results.

[0107] In some preferred embodiments, assuming a methane leak laser detection system at an industrial site, during initial discrimination, has two overall confidence scores of 0.52 and 0.48, with a difference of 0.04, lower than the preset discrimination margin of 0.05. Therefore, it is judged as a fuzzy discrimination, requiring the triggering of enhanced analysis mode. At this time, the system first monitors in real time that the processor load is 70%, the memory usage is 85%, and the data transmission bandwidth is 60% utilization. Based on these system resource usages, the system judges that current resources are relatively tight. Simultaneously, considering the small difference between the two overall confidence scores for fuzzy discrimination, it indicates a high discrimination difficulty. Based on this, the system dynamically adjusts the data acquisition parameters for enhanced analysis: increasing the laser signal sampling frequency from 100Hz to 200Hz, and adjusting the video capture frame rate from the planned 60fps to 45fps to balance data volume and resource consumption. Simultaneously, the resolution of local image acquisition is adjusted from 1920x1080 to 1280x720, focusing on key details in the laser path area. The acquired laser signal and visual image data are then compressed in real time. For example, visual image data is compressed using H.264 encoding, while laser signal data undergoes feature dimensionality reduction via wavelet transform to retain key frequency components. In terms of processing architecture, data acquisition is handled by a dedicated core, data compression and feature dimensionality reduction are processed in parallel by another core, and feature extraction is performed by a third core. When the preprocessed and dimensionality-reduced enhanced feature data is ready, the system uses an event-triggered transmission mechanism to send data packets to the discrimination module via a local area network. Upon receiving this optimized enhanced feature data, the discrimination module re-performs cross-modal information association and discrimination. For example, through recalculation, the new overall confidence score might become 0.85 and 0.10, a significantly increased difference, thus clearly identifying the methane leak event and outputting a high-confidence warning message.

[0108] The steps described above for standardizing laser attenuation mode features and visual features to obtain original evidence include:

[0109] Continuously monitor the ambient light intensity, temperature, and air humidity in the laser path area to obtain environmental parameters;

[0110] Real-time analysis of the historical fluctuation range and current trend of environmental parameters yields dynamic information on environmental parameters;

[0111] Based on the identified types of interference, the typical dynamic range of the laser attenuation mode characteristics and visual characteristics of that type of interference under different environmental conditions is obtained.

[0112] Based on environmental parameters, dynamic information of environmental parameters, and typical dynamic range, the parameters of the standardized process are adaptively adjusted.

[0113] Based on the adaptively adjusted standardized processing parameters, the laser attenuation mode features and visual features are standardized to obtain the original evidence.

[0114] Specifically, continuous monitoring of ambient light intensity, temperature, and air humidity in the laser path area refers to the real-time acquisition of environmental data within the laser path area using environmental sensors deployed in the detection system. This environmental data is considered environmental parameters used to reflect the current physical state of the detection environment. For example, ambient light intensity can be acquired using a photosensor, while temperature and air humidity can be measured using a temperature and humidity sensor.

[0115] Real-time analysis of the historical fluctuation range and current trend of environmental parameters refers to the system's continuous data accumulation and analysis of the acquired environmental parameters. Through statistical analysis of historical data, the normal fluctuation range of environmental parameters can be established, and the trend of current environmental parameters relative to historical data can be identified, such as whether it is rising, falling, or stable. This information constitutes the dynamic information of environmental parameters, providing a basis for subsequent adaptive adjustments.

[0116] In practical applications, based on the identified type of interference, obtaining the typical dynamic range of the laser attenuation mode characteristics and visual features corresponding to that type of interference under different environmental conditions refers to the system pre-storing or real-time identification of the typical numerical range or distribution characteristics of the laser attenuation mode characteristics (such as attenuation rate, amplitude, and fluctuation frequency) and visual features (such as color, transparency, and shape) caused by different types of non-target aerosols (such as water mist, smoke, dust, etc.) under different environmental conditions (such as high humidity, low temperature, strong light, etc.). These typical dynamic ranges are key references for accurate standardization processing.

[0117] Furthermore, based on environmental parameters, dynamic information about these parameters, and typical dynamic ranges, the parameters of the standardization process are adaptively adjusted. This means that the system dynamically adjusts the parameters in the standardization algorithm (such as Z-score standardization, Min-Max standardization, etc.) according to the currently monitored environmental parameters, the analyzed dynamic information about these parameters, and the typical dynamic range for the currently identified interference type. These parameters include the mean, standard deviation, maximum value, minimum value, or scaling factor. This adaptive adjustment ensures that the standardization process can better adapt to the current environment and the characteristics of the interference.

[0118] Therefore, based on the adaptively adjusted standardized processing parameters, the laser attenuation pattern features and visual features are standardized to obtain the original evidence. This means applying the dynamically adjusted standardized parameters to the original data of laser attenuation pattern features and visual features, thereby transforming these features to a uniform scale or distribution. The standardized feature data constitutes the original evidence. It eliminates biases caused by the environment and the type of interference, making different features comparable and features collected at different time points comparable, thus providing more reliable input for subsequent discrimination.

[0119] In some preferred embodiments, it is assumed that a laser detection system is in operation in an industrial park.

[0120] First, the system continuously monitored the ambient light intensity as it gradually transitioned from strong daylight to weak light in the evening, while air humidity increased significantly due to rainfall and temperature decreased slightly. These environmental parameters were collected in real time.

[0121] Next, the system analyzed historical data on these environmental parameters and found that the current changes in light intensity and humidity exceeded the average daily fluctuation range, showing a clear downward and upward trend. This dynamic information was recorded.

[0122] Meanwhile, the industrial vision system identifies a large amount of water mist (a type of composite aerosol) within the laser path area through image analysis. The system then retrieves the typical dynamic range of the water mist's laser attenuation mode characteristics (e.g., typically large attenuation amplitude and low fluctuation frequency) and visual characteristics (e.g., reduced transparency and diffused morphology) under "weak light and high humidity" environmental conditions from a pre-set database.

[0123] Based on these real-time environmental parameters, dynamic information on environmental parameters, and the typical dynamic range of water mist, the system adaptively adjusts the parameters of the normalization process. For example, for the laser attenuation amplitude, the mean and standard deviation of the normalization process are adjusted to accommodate the large attenuation caused by water mist; for the transparency of visual features, the scaling factor is adjusted to better reflect the transparency changes caused by water mist.

[0124] Finally, based on these adaptively adjusted standardized processing parameters, the currently acquired laser attenuation mode features and visual features are standardized to obtain the original evidence. In this way, even under complex environmental conditions and the presence of specific interfering substances, the standardized feature data can still accurately and effectively reflect the possibility of methane leakage, avoiding misjudgments caused by environmental changes or differences in the type of interfering substances.

[0125] This application further proposes steps for configuring a set of evidence weights for complex aerosol scenarios, including:

[0126] Continuously monitor the ambient light intensity, temperature, air humidity, and air pressure in the laser path area to obtain environmental monitoring parameters;

[0127] The composition, morphology, and physical properties of composite aerosols are analyzed in real time using an industrial vision system to obtain aerosol characteristics.

[0128] Based on environmental monitoring parameters and aerosol characteristics, the discriminative power of each feature in distinguishing between methane leakage and complex aerosol interference was evaluated.

[0129] Based on the discriminative power of the assessment, the set of evidence weights for complex aerosol scenarios is dynamically adjusted.

[0130] Specifically, continuous monitoring of ambient light intensity, temperature, air humidity, and air pressure within the laser path area aims to identify key environmental factors affecting laser signal attenuation and visual feature recognition. Ambient light intensity can influence the quality of visual feature capture, temperature and air pressure affect gas density and diffusion, while air humidity can affect aerosol formation and stability. These environmental monitoring parameters can be acquired in real time by multiple sensors deployed within the laser path area, such as light sensors, temperature sensors, humidity sensors, and air pressure sensors. The acquired data is continuously transmitted to the processing unit to generate real-time environmental monitoring parameters.

[0131] This involves using industrial vision systems to analyze the composition, morphology, and physical properties of complex aerosols in real time, aiming to gain a deeper understanding of the nature of these interfering substances. Industrial vision systems utilize high-resolution cameras to capture images or video streams within a laser path area and, through image processing and pattern recognition algorithms, identify and analyze the aerosol's components (e.g., through spectral analysis or color features), morphology (e.g., particle size, shape distribution), and physical properties (e.g., density, motion trajectory). For example, deep learning models can be used to analyze visual images to identify different types of aerosol particles and quantify their key characteristics, thereby obtaining the aerosol properties.

[0132] In practical applications, evaluating the discriminative power of various features in distinguishing between methane leaks and complex aerosol interference, based on environmental monitoring parameters and aerosol characteristics, involves assessing the effectiveness of laser attenuation pattern features and visual features (such as the rate, amplitude, and frequency of signal intensity decrease, object color, transparency, shape, internal texture, edges, and movement trajectory) under different environmental conditions and aerosol types. For example, at high humidity levels, water mist may cause laser scattering, potentially enhancing the discriminative power of visual features (such as the morphology of water mist), while some aspects of laser attenuation pattern features may be affected. This evaluation process can be based on pre-trained models, expert knowledge bases, or real-time data analysis.

[0133] Therefore, the set of evidence weights for complex aerosol scenarios is dynamically adjusted based on the assessed discriminative power. This means that the system can adjust the weights assigned to different features in real time when environmental or aerosol characteristics change. For example, under a specific aerosol type, if a visual feature is assessed as having higher discriminative power, its corresponding weight will be increased, while the weight of a laser attenuation mode feature may be decreased accordingly, and vice versa. This dynamic adjustment ensures that at any given moment, the system can make decisions using the most discriminative feature combination, thereby optimizing discrimination accuracy.

[0134] In some preferred embodiments, it is assumed that a laser detection system is operating in a chemical plant area. At some point, the ambient humidity suddenly increases, and simultaneously, the industrial vision system detects a large number of fine water droplets appearing within the laser path area. The morphology and transparency of these water droplets are significantly different from the gas clouds produced by a methane leak.

[0135] At this point, the system first continuously monitors changes in environmental parameters such as humidity and air pressure. Simultaneously, the industrial vision system performs real-time analysis on the composition (water), morphology (uniformly distributed fine particles), and physical properties (low transparency, scattering laser) of the water mist to obtain aerosol characteristics.

[0136] Based on these environmental monitoring parameters and aerosol characteristics, the system evaluation found that under the current conditions of high humidity and water mist, the discriminative power of specific wavelength attenuation related to methane absorption in the laser attenuation mode characteristics may be reduced by interference from water mist scattering, while the discriminative power of particle morphology, transparency, and movement trajectory in the visual characteristics is relatively enhanced, because the visual characteristics of water mist are significantly different from those of methane clouds.

[0137] Based on this evaluation, the system dynamically adjusts the set of evidence weights for complex aerosol scenarios. Specifically, the weights assigned to visual features (e.g., particle morphology, transparency) are increased, while the weights assigned to certain laser attenuation pattern features that are significantly affected by water mist are correspondingly decreased.

[0138] Subsequently, when a new signal strength decrease event occurs, the system combines the adjusted weight set to comprehensively calculate the evidentiary strength of laser attenuation mode features and visual features, thereby more accurately determining whether the signal strength decrease is caused by methane leakage or water mist scattering. Through this dynamic adjustment, even in complex environments with water mist interference, the system can effectively distinguish between methane leakage and non-target aerosol scattering, avoiding misjudgments caused by environmental changes and improving the accuracy and reliability of detection.

[0139] In some embodiments described above in this application, the strength of evidence for each characteristic piece of evidence in the original evidence against the two possibilities of methane leakage and interference from complex aerosols is calculated using a preset characteristic value range and a corresponding intensity function. However, in actual industrial applications, environmental conditions (such as ambient light intensity, temperature, air humidity, and air pressure) and the composition, morphology, and physical properties of complex aerosols are dynamically changing. Relying solely on statically preset characteristic value ranges and intensity functions may not accurately reflect the actual situation under the current environment, thus affecting the accuracy of the evidence strength calculation and the reliability of the judgment results.

[0140] For each characteristic piece of evidence in the original evidence, the strength of evidence for the two possibilities of methane leakage and complex aerosol interference is calculated. The steps for calculating the strength of evidence using a preset characteristic value range and the corresponding strength function include:

[0141] Continuously monitor the ambient light intensity, temperature, air humidity, and air pressure in the laser path area to obtain environmental monitoring parameters;

[0142] The composition, morphology, and physical properties of composite aerosols are analyzed in real time using an industrial vision system to obtain aerosol characteristics.

[0143] Based on environmental monitoring parameters and aerosol characteristics, the range of characteristic values ​​used to calculate the strength of evidence is dynamically updated.

[0144] At the same time, the shape and slope of the intensity function curve are dynamically adjusted based on environmental monitoring parameters and aerosol characteristics;

[0145] Based on the dynamically updated eigenvalue range and the dynamically adjusted intensity function, the strength of evidence for each characteristic in the original evidence against the two possibilities of methane leakage and complex aerosol interference is calculated.

[0146] Specifically, continuous monitoring of ambient light intensity, temperature, air humidity, and air pressure in the laser path area aims to identify key environmental factors affecting laser propagation and visual perception. These environmental monitoring parameters can be collected in real time by various sensors deployed within the detection area, such as light sensors, temperature sensors, humidity sensors, and air pressure sensors, with the purpose of providing basic data support for subsequent dynamic adjustments.

[0147] This study utilizes an industrial vision system to analyze the composition, morphology, and physical properties of complex aerosols in real time. The aim is to identify and quantify the interference characteristics of non-target aerosols that may exist within the laser path area. The industrial vision system uses a high-resolution camera to capture images of aerosols and, through image processing and pattern recognition algorithms, analyzes their visual features such as color, transparency, morphology, internal texture, and edges. It can even combine spectral analysis techniques to infer their composition, thereby obtaining the aerosol's properties. The goal is to accurately identify the nature of the interference source, providing crucial information for distinguishing between methane leaks and aerosol scattering.

[0148] In practical applications, the range of characteristic values ​​used to calculate the strength of evidence is dynamically updated based on environmental monitoring parameters and aerosol characteristics. The characteristic value range refers to the numerical interval within which a specific characteristic (such as laser attenuation rate, amplitude, fluctuation frequency, or visual characteristics like color and transparency) is considered valid evidence when determining whether a methane leak or aerosol interference is present. For example, under specific humidity levels, a certain aerosol may cause changes in the range of laser attenuation amplitude. By acquiring environmental monitoring parameters and aerosol characteristics in real time, the system can adaptively adjust the effective range of these characteristics, ensuring the accuracy of the evidence strength calculation.

[0149] Simultaneously, the shape and slope of the intensity function curve are dynamically adjusted based on environmental monitoring parameters and aerosol characteristics. The intensity function is a mathematical model that maps eigenvalues ​​to the strength of evidence (e.g., a confidence level between 0 and 1). Its curve shape and slope determine the sensitivity of eigenvalue changes to the strength of evidence. For example, in some environments, a small change in laser attenuation may have higher discriminative power, in which case the slope of the intensity function will be adjusted to be steeper; while in other environments, a more lenient discrimination range may be required. This dynamic adjustment ensures that the evidence strength calculation model maintains optimal discriminative performance under different operating conditions.

[0150] Therefore, based on the dynamically updated eigenvalue range and dynamically adjusted intensity function, the evidentiary strength of each feature in the original evidence is calculated for both the methane leak and complex aerosol interference possibilities. This step ensures that, under constantly changing environmental and aerosol conditions, the system can evaluate the discriminative power of each feature based on a model that most closely approximates reality, thus providing more reliable input for subsequent comprehensive discrimination.

[0151] This application further proposes a step to comprehensively calculate the evidentiary strength of all features and their corresponding weights to obtain two overall confidence scores for methane leakage and complex aerosol interference, including:

[0152] The evidentiary strength of laser attenuation mode features, the evidentiary strength of visual features, and their corresponding weights;

[0153] The evidentiary strength of laser attenuation pattern features and visual features were normalized.

[0154] The evidence strength of the normalized laser attenuation pattern features and the evidence strength of the visual features are initially weighted according to the weights to obtain the preprocessed data.

[0155] The preprocessed data is input into the first layer of the multilayer sensing network;

[0156] The first layer input is transformed by a non-linear activation function, and then non-linear transformation and feature combination are performed through the hidden layer to obtain the deep fused features.

[0157] Based on the deeply fused features, the output layer generates two overall confidence scores for methane leakage and complex aerosol interference.

[0158] Specifically, after receiving the evidence strength of laser attenuation pattern features, the evidence strength of visual features, and their corresponding weights, the first step is to normalize these evidence strengths. Normalization aims to eliminate differences in the units and numerical ranges of different feature evidence strengths, bringing them to a uniform scale, such as scaling them to between 0 and 1, to facilitate subsequent weighting and network processing. Normalization methods can employ various approaches, such as min-max normalization and Z-score standardization, with the goal of ensuring that the impact of all features on subsequent network inputs is fair and controllable.

[0159] Subsequently, the evidentiary strength of the normalized laser attenuation mode features and the evidentiary strength of the visual features are initially weighted according to preset weights. This initial weighting can be understood as a preliminary adjustment of the initial importance of different modalities or features in the discrimination process, to reflect their prior discriminative power in distinguishing between methane leakage and complex aerosol interference. This yields preprocessed data containing feature information that has undergone preliminary scaling and importance allocation.

[0160] Furthermore, the preprocessed data is input into the first layer of the multilayer perceptron. A multilayer perceptron is a feedforward artificial neural network consisting of at least one input layer, one or more hidden layers, and one output layer. The first layer, as the network's input layer, is responsible for receiving the pre-processed feature data.

[0161] Within a multilayer perceptron, the input to the first layer is transformed using nonlinear activation functions. These nonlinear activation functions, such as ReLU (Rectified Linear Unit), Sigmoid, or Tanh, introduce nonlinear properties, enabling the network to learn and represent complex nonlinear relationships within the input data. This transformed data then undergoes nonlinear transformation and feature combination through one or more hidden layers. Neurons in the hidden layers perform a weighted summation of the input using weights and biases, and then again undergo nonlinear transformation using activation functions, thereby achieving deep abstraction of the original features and extraction of high-level features. Through multilayer nonlinear transformations, the network can learn more discriminative, deeply fused features from the original evidence strength.

[0162] Finally, based on these deeply fused features, the output layer generates two overall confidence scores for methane leakage and complex aerosol interference. The output layer typically employs a softmax activation function to ensure that the sum of the two output scores is 1, representing the probability or confidence level of methane leakage and complex aerosol interference, respectively. These two scores together reflect the system's final judgment on the cause of the current signal strength decline event.

[0163] In some preferred embodiments, assuming a laser detection system detects a signal intensity decrease event in an industrial setting, it extracts the evidence strength of laser attenuation pattern features (e.g., velocity intensity 0.8, amplitude intensity 0.7, fluctuation frequency intensity 0.6) and visual feature evidence strengths (e.g., color intensity 0.9, transparency intensity 0.75, shape intensity 0.85). These evidence strengths are first normalized to a range of 0-1. For example, if the original intensity range is 0-100, the normalized values ​​are 0.8, 0.7, 0.6, 0.9, 0.75, and 0.85, respectively.

[0164] Subsequently, the normalized evidence strength is initially weighted according to preset weights (e.g., the weight set for laser attenuation pattern features is [0.3, 0.3, 0.4], and the weight set for visual features is [0.4, 0.3, 0.3]). For example, the initial weighted result for laser attenuation pattern features is (0.8*0.3+0.7*0.3+0.6*0.4)=0.24+0.21+0.24=0.69, and the initial weighted result for visual features is (0.9*0.4+0.75*0.3+0.85*0.3)=0.36+0.225+0.255=0.84. These initially weighted data (0.69 and 0.84) and the original normalized intensity values ​​together constitute the preprocessed data input to the first layer of the multilayer perceptron network.

[0165] This multilayer perceptron can be configured to include one input layer, two hidden layers, and one output layer. The input layer receives the preprocessed data described above. The first hidden layer may contain 16 neurons, and the second hidden layer may contain 8 neurons, both using ReLU as the non-linear activation function. For example, after the input data is weighted and activated by ReLU in the first layer, a new set of feature representations is generated. These feature representations are then used as input to the second hidden layer, and subjected to weighted summation and ReLU activation again to further extract and combine features. Finally, the output layer contains two neurons using the Softmax activation function, outputting two overall confidence scores for methane leakage and complex aerosol interference, respectively. For example, the output might be [0.95, 0.05], indicating a 95% confidence level for methane leakage and a 5% confidence level for complex aerosol interference. In this way, the network can learn how to more accurately determine the cause of signal degradation under different feature combinations, thus providing more refined and reliable discrimination results.

[0166] The steps of transforming the first-layer input using a non-linear activation function include:

[0167] The laser attenuation mode features and visual features in the first layer input are preprocessed to obtain the preprocessed features.

[0168] The degree of deviation between the preprocessed features and the historical average and standard deviation is calculated to obtain information on the degree of deviation.

[0169] Based on the deviation information, the input to the activation function is dynamically adjusted to obtain the adjusted activation function input;

[0170] Based on the dynamic range and rate of change of the preprocessed features, the slope parameter of the activation function is adaptively adjusted to obtain the adjusted slope parameter.

[0171] By combining the adjusted activation function input and the adjusted slope parameter, the first layer input is transformed using a nonlinear activation function.

[0172] Specifically, the laser attenuation pattern features and visual features in the first layer input are preprocessed to obtain preprocessed features. This step aims to perform preliminary cleaning and normalization of the original input data, such as zero-mean normalization, unit variance normalization, or range scaling, to eliminate dimensional differences between different features and ensure that the data is within the appropriate processing range of the activation function. The preprocessed features form the basis for subsequent dynamic adjustments.

[0173] This process involves calculating the degree of deviation between the preprocessed features and their historical mean and standard deviation to obtain deviation information. Deviation information can be understood as the degree of deviation of the current feature value from its historical statistical distribution. For example, the difference between the current feature value and the historical mean can be calculated and divided by the historical standard deviation to obtain a standardized deviation value. The purpose is to quantify the degree of anomaly of the current data point or its position within the overall distribution, providing a basis for subsequent dynamic adjustments.

[0174] In practical applications, the input to the activation function is dynamically adjusted based on the bias information, resulting in a modified activation function input. This dynamic adjustment aims to correct the original input to the activation function based on the characteristics of the input data. For example, when the bias information indicates that the input is an extreme value, the input can be compressed or expanded to prevent the activation function from saturating prematurely or the gradient from vanishing. Specific adjustment methods can include linear scaling, logarithmic transformation, or exponential transformation to ensure that the activation function processes data within its effective working range.

[0175] Furthermore, based on the dynamic range and rate of change of the preprocessed features, the slope parameter of the activation function is adaptively adjusted to obtain the adjusted slope parameter. The dynamic range refers to the difference between the maximum and minimum values ​​of a feature over a period of time, and the rate of change refers to how quickly the feature value changes over time. The purpose of adaptively adjusting the slope parameter is to enable the activation function to better adapt to the distribution of the input data. For example, when the feature dynamic range is large, the slope can be appropriately reduced to avoid saturation; when the rate of change is high, the slope can be increased to enhance sensitivity to changes. The adjustment of the slope parameter can be based on a pre-defined lookup table, empirical rules, or through an online learning algorithm.

[0176] Therefore, by combining the adjusted activation function input and the adjusted slope parameter, the first-layer input is transformed using a non-linear activation function. This step is the final feature transformation process. By applying the dynamically adjusted input and the adaptively adjusted slope parameter to a non-linear activation function (such as ReLU, Sigmoid, Tanh, etc.), more representative and robust deep fusion features can be generated, thus providing high-quality input for subsequent discrimination tasks.

[0177] This application further proposes a step for dynamically adjusting the input to the activation function, which optimizes the response characteristics of the activation function by employing differentiated processing strategies for input values ​​within different ranges. Specifically, the step for dynamically adjusting the input to the activation function includes:

[0178] Determine whether the activation function input before adjustment is an extremely high or extremely low value;

[0179] When the value is determined to be extremely high, the activation function input before adjustment is non-linearly compressed using a preset logarithmic compression function to obtain the adjusted activation function input. The adjusted activation function input is mapped to the effective processing limit of the subsequent processing module.

[0180] When the value is determined to be extremely low, the activation function input before adjustment is nonlinearly extended through a preset exponential expansion function to obtain the adjusted activation function input. The adjusted activation function input is mapped to the effective processing lower limit above the subsequent processing module.

[0181] When the value is determined to be non-extreme, the activation function input before adjustment is adjusted by a linear scaling function to obtain the adjusted activation function input. The adjusted activation function input is kept within the effective processing range of the subsequent processing modules.

[0182] The “activation function input before adjustment” refers to the data to be input into the activation function, determined before this step, based on the degree of deviation between the preprocessed features and the historical average and standard deviation.

[0183] Determining whether the input to the activation function before adjustment is an extreme high or low value can be achieved by setting a series of preset thresholds. For example, an upper threshold and a lower threshold can be defined. When the input value is higher than the upper threshold, it is considered an extreme high value, and when the input value is lower than the lower threshold, it is considered an extreme low value. These thresholds can be dynamically or statically set based on historical data distribution, model training experience, or the needs of specific application scenarios to ensure accurate identification of extreme values.

[0184] When an input is determined to be "extremely high," it is subjected to "non-linear compression" using a "preset logarithmic compression function." Logarithmic compression functions, such as log(1+x) or log(x) (when x>0), are characterized by their ability to non-linearly compress a large range of input values ​​into a relatively small output range, while preserving the relative order and differences between input values. This effectively prevents the activation function from saturating due to excessively large inputs, ensuring that the "adjusted activation function input" is "mapped within the effective processing limit of subsequent processing modules," thereby avoiding information loss and maintaining the model's sensitivity.

[0185] When an input is determined to be an "extremely low value," it is "non-linearly expanded" using a "preset exponential expansion function." Exponential expansion functions, such as exp(x)⁻¹ or exp(x), are characterized by their ability to non-linearly expand a small range of input values ​​to a relatively large output range, thereby enhancing their discriminative power within the activation function. This effectively prevents the activation function from ignoring information due to excessively small inputs, ensuring that the "adjusted activation function input" is "mapped above the effective processing lower limit of subsequent processing modules," thus improving the model's ability to perceive weak signals.

[0186] When the input is determined to be a "non-extreme value," it is adjusted using a "linear scaling function." Linear scaling functions are typically implemented through simple multiplication and addition operations, such as y = ax + b. Their purpose is to map the input value proportionally to the target range while maintaining its linearity. This ensures that the "adjusted activation function input" maintains stable and predictable behavior within the "effective processing range of subsequent processing modules," avoiding unnecessary nonlinear distortion and thus maintaining the model's accuracy within the normal range.

[0187] In some preferred embodiments, it is assumed that during the laser detection of methane leaks, after preprocessing and bias calculation, the input values ​​of the activation function may be distributed in a wide range from -1000 to 10000. To optimize the performance of the activation function, a lower threshold (e.g., -100) and an upper threshold (e.g., 1000) can be set.

[0188] Specifically:

[0189] When the activation function input before adjustment is 1500 (an extremely high value), the system will determine it as such. In this case, a logarithmic compression function can be used, such as y = log10(x + 1 - 1000) + 1000 (this is an example function; it can be adjusted according to actual needs), to compress it to a smaller range, for example, compressing 1500 to 1000.7. This way, even if the original input value is very large, after compression, it can be mapped to within the effective processing limit of subsequent processing modules, avoiding activation function saturation.

[0190] When the input to the activation function before adjustment is -200 (an extremely low value), the system will determine it as such. In this case, an exponential expansion function can be used, such as y = exp((x + 100) / 50) - 100 (this is an example function; it can be adjusted according to actual needs), to expand it to a larger range, for example, expanding -200 to -98.6. This way, even a weak signal can be amplified, causing it to produce a more significant response in the activation function, mapping it above the effective processing lower limit of subsequent processing modules.

[0191] When the activation function input before adjustment is 500 (a non-extreme value), the system will determine that it is a non-extreme value. In this case, a linear scaling function can be used, such as y = 0.8 * x + 100 (this is an example function; it can be adjusted according to actual needs), to adjust it to 500. This way, data within the normal range can maintain its linear relationship, ensuring the accuracy of subsequent processing and keeping it within the effective processing range of subsequent processing modules.

[0192] This segmented dynamic adjustment ensures that the most appropriate processing is obtained regardless of the range of the input value, thereby maximizing the effectiveness of the activation function.

[0193] refer to Figure 2 This application proposes a laser detection system for methane leaks based on industrial vision, applied to a laser detection method for methane leaks based on industrial vision. The system includes:

[0194] The monitoring module is used to monitor the laser signal intensity and capture the signal intensity drop event when the laser signal intensity drops below a preset intensity threshold.

[0195] The extraction module extracts the speed, amplitude, and fluctuation frequency of the signal intensity decrease for signal intensity decrease events, thereby obtaining the laser attenuation mode characteristics;

[0196] The analysis module is used to analyze the color, transparency, shape, internal texture, edges, and movement trajectory of objects within the laser path area to obtain visual features;

[0197] The discrimination module is used to correlate laser attenuation mode features with visual features, and based on the correlation results, to determine whether the signal intensity decrease is caused by absorption by the target gas or by scattering by non-target aerosols, and obtain the discrimination result.

[0198] The early warning output module outputs corresponding early warning information based on the judgment results.

[0199] Specifically, the monitoring module can be understood as the hardware and software components responsible for real-time acquisition of laser signals. This module typically includes a laser receiver, such as a photodiode or avalanche photodiode, which converts the received laser signal into an electrical signal. These electrical signals are then sent to a signal processing unit that continuously monitors the signal strength. When a decrease in signal strength is detected, and the decrease exceeds a preset intensity threshold, the monitoring module triggers a signal strength decrease event and captures it. This preset intensity threshold can be calibrated based on factors such as environmental conditions, laser power, and detection distance.

[0200] The extraction module is configured to receive signal strength decrease events from the monitoring module. Once the event is captured, the extraction module immediately analyzes the laser signal data before and after the event. Specifically, the module determines the rate of signal strength decrease by calculating the rate of change of signal strength over time; determines the magnitude of the decrease by comparing the peak signal strength before and after the decrease; and identifies the signal fluctuation frequency by performing spectral analysis on the signal waveform. Through these analyses, a series of laser attenuation mode characteristics that characterize the laser attenuation properties can be obtained.

[0201] In practical applications, the parsing module is designed to process data from industrial vision systems. This module captures real-time video or images within a laser path area using an image sensor (such as a CCD or CMOS camera). Subsequently, image processing algorithms are applied to this visual data to identify and parse various visual attributes of objects within the area. For example, the color of an object can be obtained through color space analysis (such as RGB, HSV); transparency can be obtained through image transmittance or opacity analysis; shape can be identified through edge detection, contour extraction, and shape matching algorithms; internal texture can be extracted through texture analysis algorithms (such as Gabor filters, LBP); edges can be detected using operators such as Canny and Sobel; and movement trajectories can be analyzed using target tracking algorithms (such as Kalman filtering, optical flow). Through these processes, comprehensive visual features can be obtained.

[0202] Furthermore, the discrimination module receives laser attenuation mode features from the extraction module and visual features from the analysis module. The core function of this module is to perform correlation analysis on the features of these two modalities. The correlation process can employ various machine learning or deep learning models, such as support vector machines, neural networks, or decision trees, which are trained to identify the relationship between different feature combinations and the absorption of a specific gas (e.g., methane) or the scattering of non-target aerosols. Based on the results of the correlation analysis, the discrimination module can distinguish whether the signal intensity decrease is due to a genuine leak event caused by the absorption of the target gas (methane) or interference caused by the scattering of non-target aerosols such as dust, water mist, or smoke. This generates a clear discrimination result.

[0203] In addition, the early warning output module is configured to receive the judgment results generated by the judgment module. Once the judgment result indicates a methane leak or an anomaly requiring attention, this module will immediately generate and output corresponding early warning information. Early warning information can include various forms such as audible and visual alarms, SMS notifications, emails, and SCADA system integration messages, and can be displayed on the operator interface, providing key information such as the leak location and severity, so that operators can take timely countermeasures.

[0204] The content disclosed above is only a preferred and feasible embodiment of the present invention, and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent technical changes made based on the content of the present invention specification and drawings are included within the scope of protection of the present invention. Furthermore, the elements therein can be updated as technology develops.

Claims

1. A laser detection method for methane leaks based on industrial vision, characterized in that, The method includes the following steps: Monitor the laser signal intensity and capture the signal intensity drop event when the laser signal intensity drops below a preset intensity threshold; For signal strength degradation events, the speed, amplitude, and fluctuation frequency of signal strength degradation are extracted to obtain laser attenuation mode characteristics; Simultaneously, the color, transparency, shape, internal texture, edges, and movement trajectory of objects within the laser path area are analyzed to obtain visual features; The laser attenuation mode features are correlated with visual features, and based on the correlation results, it is determined whether the signal intensity decrease is caused by absorption by the target gas or by scattering by non-target aerosols, thus obtaining the discrimination result; Based on the judgment results, the corresponding warning information is output.

2. The laser detection method for methane leakage based on industrial vision as described in claim 1, characterized in that, The method also includes the following steps: The laser attenuation mode characteristics and visual characteristics were standardized to obtain the original evidence. Configure an evidence weight set for complex aerosol scenarios. The weight set is set based on the discriminative power of different features in distinguishing between methane leakage and complex aerosol interference. For each characteristic piece of evidence in the original evidence, calculate its strength of evidence for the two possibilities of methane leakage and complex aerosol interference. The strength of evidence is calculated using a preset range of characteristic values ​​and the corresponding strength function. By combining the evidence weight set for complex aerosol scenarios, the evidence strength of all features and their corresponding weights are comprehensively calculated to obtain two overall confidence scores for methane leakage and complex aerosol interference. When both overall confidence scores are lower than the preset high confidence threshold, or when the difference between the two overall confidence scores is less than the preset discrimination margin, it is determined to be a fuzzy discrimination and the enhanced analysis mode is triggered. The enhanced analysis mode includes increasing the laser signal sampling frequency, extending the sampling time, increasing the video capture frame rate, and performing higher resolution local image acquisition. At the same time, the system outputs preliminary judgment results with low confidence indicators to the operator and displays the main features that lead to fuzzy judgments.

3. The laser detection method for methane leakage based on industrial vision as described in claim 2, characterized in that, The steps to trigger enhanced analytics mode include: Real-time monitoring of processor load, memory usage, and data transfer bandwidth provides information on system resource utilization. Based on the difference between the overall confidence scores of the system resource usage and the fuzzy discrimination, the data acquisition parameters of the augmentation analysis are dynamically adjusted to obtain the augmented laser signal and visual image data. The acquired laser signals and visual image data are compressed and feature dimensionality reduced in real time to obtain preprocessed and dimensionality-reduced enhanced feature data. Using a multi-threaded or parallel processing architecture, data acquisition, preprocessing, and feature extraction tasks are assigned to different processing cores, which helps to improve processing speed. An event-triggered transmission mechanism is used to transmit preprocessed and dimensionality-reduced enhanced feature data. Using the preprocessed and dimensionality-reduced enhanced feature data, cross-modal information association and discrimination are re-performed.

4. The laser detection method for methane leakage based on industrial vision as described in claim 2, characterized in that, The steps for standardizing laser attenuation mode features and visual features to obtain raw evidence include: Continuously monitor the ambient light intensity, temperature, and air humidity in the laser path area to obtain environmental parameters; Real-time analysis of the historical fluctuation range and current trend of environmental parameters yields dynamic information on environmental parameters; Based on the identified types of interference, the typical dynamic range of the laser attenuation mode characteristics and visual characteristics of that type of interference under different environmental conditions is obtained. Based on environmental parameters, dynamic information of environmental parameters, and typical dynamic range, the parameters of the standardized process are adaptively adjusted. Based on the adaptively adjusted standardized processing parameters, the laser attenuation mode features and visual features are standardized to obtain the original evidence.

5. The laser detection method for methane leakage based on industrial vision as described in claim 2, characterized in that, The steps for configuring the evidence weight set for complex aerosol scenarios include: Continuously monitor the ambient light intensity, temperature, air humidity, and air pressure in the laser path area to obtain environmental monitoring parameters; The composition, morphology, and physical properties of composite aerosols are analyzed in real time using an industrial vision system to obtain aerosol characteristics. Based on environmental monitoring parameters and aerosol characteristics, the discriminative power of each feature in distinguishing between methane leakage and complex aerosol interference was evaluated. Based on the discriminative power of the assessment, the set of evidence weights for complex aerosol scenarios is dynamically adjusted.

6. The laser detection method for methane leakage based on industrial vision as described in claim 2, characterized in that, For each characteristic piece of evidence in the original evidence, the strength of evidence for the two possibilities of methane leakage and complex aerosol interference is calculated. The steps for calculating the strength of evidence using a preset characteristic value range and the corresponding strength function include: Continuously monitor the ambient light intensity, temperature, air humidity, and air pressure in the laser path area to obtain environmental monitoring parameters; The composition, morphology, and physical properties of composite aerosols are analyzed in real time using an industrial vision system to obtain aerosol characteristics. Based on environmental monitoring parameters and aerosol characteristics, the range of characteristic values ​​used to calculate the strength of evidence is dynamically updated. At the same time, the shape and slope of the intensity function curve are dynamically adjusted based on environmental monitoring parameters and aerosol characteristics; Based on the dynamically updated eigenvalue range and the dynamically adjusted intensity function, the strength of evidence for each characteristic in the original evidence against the two possibilities of methane leakage and complex aerosol interference is calculated.

7. The laser detection method for methane leakage based on industrial vision as described in claim 2, characterized in that, The steps for calculating the overall confidence scores for methane leakage and complex aerosol interference by comprehensively calculating the evidentiary strength of all features and their corresponding weights include: The evidentiary strength of laser attenuation mode features, the evidentiary strength of visual features, and their corresponding weights; The evidentiary strength of laser attenuation pattern features and visual features were normalized. The evidence strength of the normalized laser attenuation pattern features and the evidence strength of the visual features are initially weighted according to the weights to obtain the preprocessed data. The preprocessed data is input into the first layer of the multilayer sensing network; The first layer input is transformed by a non-linear activation function, and then non-linear transformation and feature combination are performed through the hidden layer to obtain the deep fused features. Based on the deeply fused features, the output layer generates two overall confidence scores for methane leakage and complex aerosol interference.

8. The laser detection method for methane leakage based on industrial vision as described in claim 7, characterized in that, The steps of transforming the first-layer input using a non-linear activation function include: The laser attenuation mode features and visual features in the first layer input are preprocessed to obtain the preprocessed features. The degree of deviation between the preprocessed features and the historical average and standard deviation is calculated to obtain information on the degree of deviation. Based on the deviation information, the input to the activation function is dynamically adjusted to obtain the adjusted activation function input; Based on the dynamic range and rate of change of the preprocessed features, the slope parameter of the activation function is adaptively adjusted to obtain the adjusted slope parameter. By combining the adjusted activation function input and the adjusted slope parameter, the first layer input is transformed using a nonlinear activation function.

9. The laser detection method for methane leakage based on industrial vision as described in claim 8, characterized in that, The steps for dynamically adjusting the input to the activation function include: Determine whether the activation function input before adjustment is an extremely high or extremely low value; When the value is determined to be extremely high, the activation function input before adjustment is non-linearly compressed using a preset logarithmic compression function to obtain the adjusted activation function input. The adjusted activation function input is mapped to the effective processing limit of the subsequent processing module. When the value is determined to be extremely low, the activation function input before adjustment is nonlinearly extended through a preset exponential expansion function to obtain the adjusted activation function input. The adjusted activation function input is mapped to the effective processing lower limit above the subsequent processing module. When the value is determined to be non-extreme, the activation function input before adjustment is adjusted by a linear scaling function to obtain the adjusted activation function input. The adjusted activation function input is kept within the effective processing range of the subsequent processing modules.

10. A methane leak laser detection system based on industrial vision, applied to the methane leak laser detection method based on industrial vision as described in claim 1, characterized in that, The system includes: The monitoring module is used to monitor the laser signal intensity and capture the signal intensity drop event when the laser signal intensity drops below a preset intensity threshold. The extraction module extracts the speed, amplitude, and fluctuation frequency of the signal intensity decrease for signal intensity decrease events, thereby obtaining the laser attenuation mode characteristics; The analysis module is used to analyze the color, transparency, shape, internal texture, edges, and movement trajectory of objects within the laser path area to obtain visual features; The discrimination module is used to correlate laser attenuation mode features with visual features, and based on the correlation results, to determine whether the signal intensity decrease is caused by absorption by the target gas or by scattering by non-target aerosols, and obtain the discrimination result. The early warning output module outputs corresponding early warning information based on the judgment results.

Images