Method for identifying spatial features of amc sensor array and detecting leakage point based on deep learning

Abstract

The application discloses an AMC sensor array space feature recognition and leakage point detection method based on deep learning, and relates to the technical field of sensors.The method comprises the following steps: performing space correlation analysis on a pretreatment result to extract a space feature signal reflecting the space distribution characteristics of a gas; using an empirical mode decomposition algorithm to decompose the denoised space feature to obtain a plurality of intrinsic mode components, and generating a reconstruction signal based on the intrinsic mode components; using a frame windowing technology to perform frame processing on the reconstruction signal, calculating the energy value and zero-crossing rate of the corresponding frame based on the frame processing result, combining the extracted space feature signal, constructing a leakage perception model, and generating a leakage probability distribution sequence; and inputting a time sequence window sequence into a self-encoder, combining a knowledge distillation mechanism and a comparative analysis technology, and performing gas leakage point detection.The application realizes accurate extraction and denoising processing of the space feature signal, improves the signal quality, and further improves the detection accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sensors, in particular to an AMC sensor array spatial feature recognition and leakage point detection method based on deep learning. BACKGROUND

[0002] With the rapid development of advanced manufacturing, semiconductor production, new energy and other industries, the demand for monitoring airborne molecular contaminants (AMC) in the environment is increasing. AMC mainly includes volatile organic compounds (VOCs), acid-base gases and metal vapors, etc. Even a small amount of leakage may have a serious impact on product quality, equipment life, and even production safety. Therefore, it is of great industrial value to achieve high-precision and real-time leakage detection of AMC.

[0003] Currently, AMC leakage detection mainly uses multi-sensor array technology as the core, which constructs a spatial monitoring network by deploying multiple gas sensors in the target area to realize continuous perception of the environment. However, traditional detection methods often rely on single-point threshold alarms or simple statistical analysis, which is difficult to accurately distinguish between real leakage signals and environmental background noise, production equipment electromagnetic interference, sensor drift and other non-leakage disturbances, resulting in high false alarm rate and missed alarm rate. In addition, the diffusion process of AMC is affected by complex factors such as airflow field, environmental temperature, and on-site obstacles, and its spatial propagation presents significant characteristics such as nonlinearity, time-varying and diffusion delay. It is more difficult to accurately locate the leakage source position by simply relying on independent analysis of single-channel signals.

[0004] In recent years, deep learning technology has shown strong advantages in time series signal processing and pattern recognition, and convolutional neural networks, recurrent neural networks and graph neural networks have been gradually applied to multi-sensor data fusion and anomaly detection scenarios. However, most existing methods independently process time and space dimensions, and fail to fully exploit the dynamic characteristics of the spatio-temporal coupling in the AMC diffusion process. At the same time, the model training relies on a large number of labeled samples, while the leakage samples are scarce in actual industrial environments, limiting the model generalization ability. In addition, the existing leakage detection system generally uses a fixed threshold judgment mechanism, which cannot adapt to dynamic changes in environmental conditions, resulting in sensitivity imbalance problems in different operating stages.

[0005] There is no effective solution to the problems in the related art. SUMMARY

[0006] To solve the problems in the related art, the present application proposes an AMC sensor array spatial feature recognition and leakage point detection method based on deep learning to overcome the above technical problems existing in the prior art.

[0007] To this end, the specific technical solutions adopted by the present application are as follows:

[0008] The AMC sensor array spatial feature recognition and leakage point detection method based on deep learning comprises:

[0009] S1, preprocessing the multi-channel gas concentration signals synchronously collected by the sensor array, and combining the spatial topology structure of the sensor, performing spatial correlation analysis on the preprocessing result, and extracting spatial feature signals reflecting the spatial distribution characteristics of the gas;

[0010] S2, denoising the extracted spatial feature signals using wavelet transform, and decomposing the denoised spatial features using empirical mode decomposition algorithm to obtain a plurality of intrinsic mode components, and generating a reconstructed signal based on each intrinsic mode component;

[0011] S3, using frame windowing technology to perform frame processing on the reconstructed signal, and calculating the energy value and zero-crossing rate of the corresponding frame based on the frame processing result, combining the extracted spatial feature signals, constructing a leakage perception model, and generating a leakage probability distribution sequence;

[0012] S4, based on the generated leakage probability distribution sequence, establishing a time sequence window sequence, and inputting the time sequence window sequence into a double auto-encoder, combining a knowledge distillation mechanism and a comparative analysis technology, and performing gas leakage point detection.

[0013] Further, the wavelet transform is used to denoise the extracted spatial feature signals, and the empirical mode decomposition algorithm is used to decompose the denoised spatial features to obtain a plurality of intrinsic mode components, and the reconstructed signal is generated based on each intrinsic mode component, comprising:

[0014] S21, using wavelet transform to perform multi-scale decomposition on the extracted spatial feature signals to obtain multi-scale sub-signals, and using adaptive local threshold denoising technology to denoise each scale sub-signal to obtain denoised spatial feature signals;

[0015] S22, using empirical mode decomposition algorithm to decompose the denoised spatial feature signals into a plurality of intrinsic mode components, and extracting multi-dimensional features of each intrinsic mode component;

[0016] S23, inputting the extracted multi-dimensional features into a preset attention network, learning the importance of each intrinsic mode component in the reconstruction of the spatial feature signal through the attention mechanism, and outputting the attention weight corresponding to each intrinsic mode component;

[0017] S24, inputting the intrinsic mode component and the corresponding attention weight as input conditions into the pre-trained generative adversarial network, and through the generator in the adversarial network, the multi-scale component information is fused to reconstruct the spatial feature signal, and the reconstructed signal is generated.

[0018] Further, the frame division and windowing technology is applied to frame division processing of the reconstructed signal, and the energy value and zero-crossing rate of the corresponding frame are calculated based on the frame division processing result, the spatial feature signal is combined to construct a leakage perception model, and a leakage probability distribution sequence is generated, including:

[0019] S31, the frame division and windowing technology is applied to frame division processing of the reconstructed signal, time sequence signal segments are generated, and a window function is applied to each frame of time sequence signal segment, the energy value of the corresponding frame is calculated, and an energy sequence is established in time sequence;

[0020] S32, the energy sequence is subjected to centering and amplitude constraint processing, the zero-crossing rate of each frame is calculated based on the processing result, and each frame of zero-crossing rate is combined with the corresponding energy value to construct an energy-zero ratio, and an energy-zero ratio sequence is generated in time sequence;

[0021] S33, the spatial feature signal and the time dynamic feature of the energy-zero ratio sequence are combined to construct a leakage perception model, and the contrast clustering loss algorithm is used to optimize the leakage perception model;

[0022] S34, the generated energy-zero ratio sequence is input into the optimized leakage perception model, and the feature embedding vector corresponding to each frame of time sequence signal segment is extracted;

[0023] S35, the similarity distance between the feature embedding vector and the preset leakage class clustering center is calculated, the leakage probability of each frame of time sequence signal segment is output, and a leakage probability distribution sequence is generated in time sequence.

[0024] Further, the spatial feature signal and the time dynamic feature of the energy-zero ratio sequence are combined to construct a leakage perception model, and the contrast clustering loss algorithm is used to optimize the leakage perception model, including:

[0025] S331, based on the extracted spatial feature signal, the Euclidean distance between each sensor is calculated, and each sensor is taken as a node to construct a spatial adjacency matrix;

[0026] S332, the time domain and frequency domain analysis is performed on the energy-zero ratio sequence, the time dynamic feature is extracted based on the analysis result, and the spatial adjacency matrix is combined to generate a space-time joint feature tensor;

[0027] S333, the space-time joint feature tensor and the deep learning framework are combined to construct a leakage perception model, and the contrast clustering loss algorithm is used to optimize the leakage perception model.

[0028] Further, the space-time joint feature tensor and the deep learning framework are combined to construct a leakage perception model, and the contrast clustering loss algorithm is used to optimize the leakage perception model, including:

[0029] S3331, based on the spatial topology and time dynamic information in the spatio-temporal joint feature tensor, a leakage perception model based on graph convolution network and time series attention mechanism is constructed by combining a deep learning framework;

[0030] S3332, sample data containing leakage and non-leakage states are collected from a preset database, and structured event information is extracted from the sample data using information recognition technology;

[0031] S3333, the extracted event information is input into the constructed leakage perception model, and the input event information is cooperatively processed through graph convolution operation and time series attention mechanism, and the event representation vector of each sample data is output;

[0032] S3334, according to the leakage and non-leakage states of the sample data, sample data similar to the event representation vector are selected as target samples from the sample data, and a preset contrast loss function is used to analyze the similarity between the sample data and the target samples and the difference between the sample data and the remaining sample data;

[0033] S3335, based on the similarity and difference analysis results, the parameters of the leakage perception model are updated using the back propagation algorithm to optimize the leakage perception model.

[0034] Further, based on the generated leakage probability distribution sequence, a time series window sequence is established, and the time series window sequence is input into a double autoencoder, combined with a knowledge distillation mechanism and a contrast analysis technology, to perform gas leakage point detection, including:

[0035] S41, according to the constructed leakage probability distribution sequence, the energy zero ratio of each frame of gas concentration signal and the corresponding energy zero ratio average value are calculated, and the adaptive time window length is dynamically determined based on the current time as the reference, and the corresponding time series window sequence is generated;

[0036] S42, input the time series window sequence into the double autoencoder to generate a fusion weight strategy, and construct a comprehensive loss function combined with a knowledge distillation mechanism to generate an adaptive leakage judgment threshold at the current time;

[0037] S43, collect real-time gas concentration signals, calculate the corresponding real-time energy zero ratio, and add the real-time energy zero ratio to the time series window sequence in time sequence, and remove the earliest energy zero ratio average value in the time series window sequence to update the time series window sequence;

[0038] S44, input the updated time series window sequence into the double autoencoder to obtain an anomaly score, and compare the anomaly score with the adaptive leakage judgment threshold, and determine the gas leakage point based on the comparison result.

[0039] Further, according to the constructed leakage probability distribution sequence, the energy zero ratio of each frame of the gas concentration signal is calculated, and the corresponding energy zero ratio mean value is calculated, and the adaptive time window length is dynamically determined based on the current time as the benchmark, and the corresponding time sequence window sequence is generated, including:

[0040] S411, according to the constructed leakage probability distribution sequence, the energy zero ratio of each frame of the gas concentration signal is calculated, and the maximum energy zero ratio and the minimum energy zero ratio are selected from the energy zero ratio, and the energy zero ratio mean value of all frames is calculated based on the selected result;

[0041] S412, based on the calculated energy zero ratio mean value, the minimum window length and the maximum window length are set, and the variance of the energy zero ratio sequence in different window lengths is calculated from the minimum window length as the benchmark;

[0042] S413, according to the variance of the energy zero ratio sequence in each window length, the window length whose variance first exceeds the preset stability threshold is dynamically selected as the adaptive time window length, and the corresponding time sequence window sequence is generated based on the adaptive time window length.

[0043] Further, the time sequence window sequence is input into the double autoencoder to generate a fusion weight strategy, and a comprehensive loss function is constructed by combining a knowledge distillation mechanism to generate an adaptive leakage judgment threshold at the current time, including:

[0044] S421, a double autoencoder with spatial and channel parallelism is constructed, the time sequence window sequence is input into the double autoencoder, and the corresponding spatial feature vector and channel feature vector are output;

[0045] S422, the information entropy of each dimension in the spatial feature vector and the channel feature vector is calculated respectively, and a fusion weight strategy is generated by combining a weight distribution mechanism;

[0046] S423, according to the generated fusion weight strategy, the spatial feature vector and the channel feature vector are weighted and fused to obtain a fusion feature vector, and the fusion feature vector is input into a preset knowledge distillation network, and a distillation loss function is constructed by using the distribution difference between the teacher model and the student model;

[0047] S424, the reconstruction error of the double autoencoder architecture is taken as a structural loss function, the cosine loss function is taken as a constraint, the distillation loss function is constructed, and the corresponding comprehensive loss value is calculated;

[0048] S425, the comprehensive loss value is taken as the abnormal evaluation value at the current time, the abnormal evaluation value is dynamically scaled by combining a preset dynamic adjustment coefficient, and an adaptive leakage judgment threshold at the current time is generated.

[0049] Further, the information entropy of each dimension in the spatial feature vector and the channel feature vector is calculated respectively, and a fusion weight strategy is generated by combining a weight distribution mechanism, including:

[0050] S4221, the spatial feature vector and the channel feature vector are dimensionally aligned, and the spatial feature vector and the channel feature vector of each dimension are mapped to a preset standard interval by using a standardization algorithm;

[0051] S4222, based on the standardized spatial feature vector and the channel feature vector, the information entropy of each dimension is calculated respectively, and the information entropy is normalized;

[0052] S4223, according to the normalization result of each dimension information entropy, an initial fusion weight is assigned to each dimension, and a target function is constructed based on the assignment result to generate a fusion weight strategy.

[0053] Further, according to the normalization result of each dimension information entropy, an initial fusion weight is assigned to each dimension, and a target function is constructed based on the assignment result to generate a fusion weight strategy, including:

[0054] S42231, the normalization result of each dimension information entropy is converted into the utility value of the information contribution degree of the spatial feature vector and the channel feature vector in each dimension, and an initial fusion weight is assigned to each dimension;

[0055] S42232, based on the assigned initial fusion weight, a target function is constructed with the goal of maximizing Nash product, and the target function is solved by a mathematical analysis algorithm to obtain a weight solution;

[0056] S42233, the obtained weight solution is analyzed, and based on the analysis result, the expression form of the utility value in the target function is adjusted to generate a fusion weight strategy.

[0057] The beneficial effects of the present application are:

[0058] 1. The present application realizes accurate extraction and denoising processing of spatial feature signals by fusing the spatial topology structure of the sensor array and the deep learning technology. At the same time, the noise interference can be suppressed and the signal quality can be improved by combining wavelet transform and empirical mode decomposition technology, further improving the detection accuracy.

[0059] 2. The present application can perceive the weak change of gas concentration by the synergistic effect of frame windowing technology and leakage perception model, combined with the real-time calculated energy value and zero-crossing rate, which can improve the sensitivity to early leakage. At the same time, the self-encoder and knowledge distillation mechanism are fused to accurately generate a leakage judgment threshold, thereby improving the sensitivity and accuracy of gas leakage point detection. BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0061] Figure 1 is a flowchart of a deep learning-based AMC sensor array spatial feature recognition and leakage point detection method according to an embodiment of the present application. DETAILED DESCRIPTION

[0062] In order to further illustrate the embodiments, the present application provides drawings which are part of the disclosure of the present application, mainly used to illustrate the embodiments, and can explain the operating principle of the embodiments in conjunction with the related description of the specification. Those skilled in the art should understand other possible embodiments and advantages of the present application by referring to these contents.

[0063] According to an embodiment of the present application, a deep learning-based AMC sensor array spatial feature recognition and leakage point detection method is provided.

[0064] The present application will be further described in conjunction with the drawings and specific embodiments, as shown in Figure 1 According to the deep learning-based AMC sensor array spatial feature recognition and leakage point detection method according to an embodiment of the present application, the method comprises:

[0065] S1, pre-processing the multi-channel gas concentration signals synchronously collected by the sensor array, and combining the spatial topology structure of the sensor, performing spatial correlation analysis on the pre-processing result, and extracting spatial feature signals reflecting the spatial distribution characteristics of the gas.

[0066] It should be noted that the pre-processing includes denoising, standardization, missing value filling, etc.; wherein the spatial correlation analysis adopts cross-correlation analysis, principal component analysis, etc., to quantify the correlation between sensors, for example, sensors with high correlation can be considered to detect the concentration fluctuation of the same gas region, while sensors with low correlation can perceive the gas distribution at different positions. In addition, the spatial feature signals include spatial average concentration, spatial concentration gradient, spatial covariance, spatial mode extracted by spatial correlation analysis, spatial correlation, energy distribution of spatial mode, cluster structure of spatial distribution, etc.

[0067] S2, denoising the extracted spatial feature signals using wavelet transform, and decomposing the denoised spatial features using empirical mode decomposition algorithm to obtain a plurality of intrinsic mode components, and generating a reconstructed signal based on each intrinsic mode component.

[0068] In the optional embodiment, the extracted spatial feature signal is denoised by using wavelet transform, and the denoised spatial feature is decomposed by using an empirical mode decomposition algorithm to obtain a plurality of intrinsic mode components, and a reconstruction signal is generated based on the intrinsic mode components, including:

[0069] S21, the extracted spatial feature signal is multi-scale decomposed by using wavelet transform to obtain multi-scale sub-signals, and each scale sub-signal is denoised by using an adaptive local threshold denoising technology to obtain a denoised spatial feature signal.

[0070] S22, the denoised spatial feature signal is decomposed into a plurality of intrinsic mode components by using an empirical mode decomposition algorithm, and multi-dimensional features of each intrinsic mode component are extracted.

[0071] It should be noted that the multi-dimensional features include time domain features, frequency domain features, time-frequency features, high-order statistical features, statistical features, etc.; for example, the time domain features include mean, standard deviation, kurtosis, skewness, extreme value, and zero crossing point number, the frequency domain features include spectral center, spectral width, spectral entropy, and maximum frequency component, the time-frequency features include time-frequency energy and instantaneous frequency, the high-order statistical features include high-order cumulants and fractal dimension, and the statistical features include autoregressive coefficients and correlation coefficients.

[0072] S23, the extracted multi-dimensional features are input into a preset attention network, the importance of each intrinsic mode component in the reconstruction of the spatial feature signal is learned through an attention mechanism, and an attention weight corresponding to each intrinsic mode component is output.

[0073] S24, the intrinsic mode components and the corresponding attention weights are input into a pre-trained generative adversarial network as input conditions, and the spatial feature signal is reconstructed by fusing multi-scale component information in the generator of the adversarial network to generate a reconstruction signal.

[0074] S3, the reconstruction signal is frame-processed by using a frame windowing technology, the energy value and the zero-crossing rate of the corresponding frame are calculated based on the frame processing result, a leakage perception model is constructed in combination with the extracted spatial feature signal, and a leakage probability distribution sequence is generated.

[0075] In the optional embodiment, the reconstruction signal is frame-processed by using a frame windowing technology, the energy value and the zero-crossing rate of the corresponding frame are calculated based on the frame processing result, a leakage perception model is constructed in combination with the extracted spatial feature signal, and a leakage probability distribution sequence is generated, including:

[0076] S31, the reconstruction signal is frame-processed by using a sliding window technology to generate a time sequence signal segment, a window function is applied to each frame of the time sequence signal segment, the energy value of the corresponding frame is calculated, and an energy sequence is established in time sequence.

[0077] It needs to be added that the window function includes rectangular window, Hamming window, Hann window, Blackman window, etc.

[0078] S32, centering and amplitude constraint processing are performed on the energy sequence, the zero-crossing rate of each frame is calculated based on the processing result, each frame zero-crossing rate is combined with the corresponding energy value, the energy zero ratio value is constructed, and the energy zero ratio sequence is generated in time sequence.

[0079] It needs to be added that the centering and amplitude constraint processing are performed on the energy sequence, the zero-crossing rate of each frame is calculated based on the processing result, each frame zero-crossing rate is combined with the corresponding energy value, the energy zero ratio value is constructed, and the energy zero ratio sequence is generated in time sequence, which specifically includes:

[0080] The energy sequence is centered, that is, each energy value is subtracted from the average value of the sequence, so as to eliminate the direct current component of the signal and ensure that the energy sequence reflects the dynamic change of the signal, for example, the energy sequence E t ′ after centering can be represented as: ; Wherein E t is the original energy sequence, the energy value of the t frame, μ E is the average value of the energy sequence; The amplitude of the energy sequence after centering is constrained to a reasonable range to avoid distortion of the analysis result due to extreme energy value, for example, the maximum and minimum values of the energy sequence are compressed to a predetermined range [-1, 1]; Based on the amplitude-constrained energy sequence, each frame signal is converted into a binary signal, that is, it is determined whether the signal crosses zero, and the number of times the signal crosses zero is counted to obtain the zero-crossing rate of each frame; The zero-crossing rate of each frame is multiplied by the energy value to obtain the energy zero ratio value of each frame, and these values are arranged in time sequence to obtain an energy zero ratio sequence representing the change of the signal with time.

[0081] S33, combining the extracted spatial feature signal and the time dynamic feature of the energy zero ratio sequence, constructing a leakage perception model, and optimizing the leakage perception model by using contrast clustering loss algorithm.

[0082] In this optional embodiment, combining the extracted spatial feature signal and the time dynamic feature of the energy zero ratio sequence, constructing a leakage perception model, and optimizing the leakage perception model by using contrast clustering loss algorithm includes:

[0083] S331, based on the extracted spatial feature signal, calculating the Euclidean distance between each sensor, and constructing a spatial adjacency matrix by taking each sensor as a node.

[0084] S332, time domain and frequency domain analysis are performed on the energy zero ratio sequence, time dynamic features are extracted based on the analysis result, and a space-time joint feature tensor is generated by combining the constructed spatial adjacency matrix.

[0085] S333, combine the spatio-temporal joint feature tensor with the deep learning framework to construct a leakage perception model, and optimize the leakage perception model using a contrastive clustering loss algorithm.

[0086] In this optional embodiment, combining the spatio-temporal joint feature tensor with the deep learning framework to construct a leakage perception model, and optimizing the leakage perception model using a contrastive clustering loss algorithm includes:

[0087] S3331, based on the spatial topology and temporal dynamic information in the spatio-temporal joint feature tensor, combine the deep learning framework to construct a leakage perception model based on a graph convolution network and a time series attention mechanism;

[0088] S3332, collect sample data containing leakage and non-leakage states from a pre-set database, and use information recognition technology to extract structured event information from the sample data;

[0089] S3333, input the extracted event information into the constructed leakage perception model, and cooperatively process the input event information through graph convolution operation and time series attention mechanism, and output the event representation vector of each sample data;

[0090] S3334, according to the leakage and non-leakage states of the sample data, select sample data similar to the event representation vector from the sample data as target samples, and use a pre-set contrastive loss function to analyze the similarity between the sample data and the target samples and the difference between the sample data and the remaining sample data;

[0091] S3335, based on the similarity and difference analysis results, use a backpropagation algorithm to update the parameters of the leakage perception model to optimize the leakage perception model.

[0092] S34, input the generated zero ratio sequence into the optimized leakage perception model to extract the feature embedding vector corresponding to each frame of time series signal segment.

[0093] S35, by calculating the similarity distance between the feature embedding vector and the pre-set leakage class clustering center, output the leakage probability of each frame of time series signal segment, and generate a leakage probability distribution sequence in time order.

[0094] It needs to be supplemented that the leakage class clustering center is a representative embedding vector obtained by unsupervised or supervised clustering analysis of historical leakage samples, such as K-means, GMM or clustering based on Gaussian kernel function; each clustering center represents a leakage mode or behavior feature, which is used for subsequent similarity judgment between new samples and leakage modes. In addition, the similarity distance calculation can use common measurement methods such as Euclidean distance, cosine similarity or Mahalanobis distance.

[0095] S4, based on the generated sequence of leakage probability distributions, a sequence of timing windows is established, and the sequence of timing windows is input to the dual autoencoder, combined with the knowledge distillation mechanism and the comparative analysis technology, to perform gas leakage point detection.

[0096] In this optional embodiment, based on the generated sequence of leakage probability distributions, a sequence of timing windows is established, and the sequence of timing windows is input to the dual autoencoder, combined with the knowledge distillation mechanism and the comparative analysis technology, to perform gas leakage point detection, which includes:

[0097] S41, according to the constructed sequence of leakage probability distributions, the energy-to-zero ratio of each frame of the gas concentration signal and the corresponding energy-to-zero ratio average are calculated, and the adaptive time window length is dynamically determined based on the current time as the reference, and the corresponding timing window sequence is generated.

[0098] In this optional embodiment, according to the constructed sequence of leakage probability distributions, the energy-to-zero ratio of each frame of the gas concentration signal and the corresponding energy-to-zero ratio average are calculated, and the adaptive time window length is dynamically determined based on the current time as the reference, and the corresponding timing window sequence is generated, which includes:

[0099] S411, according to the constructed sequence of leakage probability distributions, the energy-to-zero ratio of each frame of the gas concentration signal is calculated, and the maximum energy-to-zero ratio and the minimum energy-to-zero ratio are selected from the energy-to-zero ratio, and the energy-to-zero ratio average of all frames is calculated based on the selected results;

[0100] S412, based on the calculated energy-to-zero ratio average, the minimum window length and the maximum window length are set, and starting from the minimum window length, the variances of the energy-to-zero ratio sequences in different window lengths are calculated in turn based on the current time as the reference;

[0101] S413, according to the variances of the energy-to-zero ratio sequences in each window length, the window length at which the variance first exceeds the preset stability threshold is dynamically selected as the adaptive time window length, and the corresponding timing window sequence is generated based on the adaptive time window length.

[0102] S42, input the timing window sequence to the dual autoencoder to generate a fusion weight strategy, and construct a comprehensive loss function combined with the knowledge distillation mechanism to generate an adaptive leakage judgment threshold at the current time.

[0103] In this optional embodiment, the timing window sequence is input to the dual autoencoder to generate a fusion weight strategy, and a comprehensive loss function is constructed combined with the knowledge distillation mechanism to generate an adaptive leakage judgment threshold at the current time, which includes:

[0104] S421, a dual autoencoder with spatial and channel parallelism is constructed, the timing window sequence is input to the dual autoencoder, and the corresponding spatial feature vector and channel feature vector are output.

[0105] S422、Calculate the information entropy of each dimension in the spatial feature vector and the channel feature vector respectively, and generate a fusion weight strategy combined with the weight distribution mechanism.

[0106] In this optional embodiment, calculating the information entropy of each dimension in the spatial feature vector and the channel feature vector respectively, and generating a fusion weight strategy combined with the weight distribution mechanism includes:

[0107] S4221, Align the dimensions of the spatial feature vector and the channel feature vector, and map the spatial feature vector and the channel feature vector of each dimension to a pre-set standard interval using a standardization algorithm.

[0108] S4222, Based on the standardized spatial feature vector and the channel feature vector, calculate the information entropy of each dimension respectively, and normalize the information entropy.

[0109] S4223, According to the normalization result of each dimension information entropy, assign an initial fusion weight to each dimension, and based on the assignment result, construct a target function to generate a fusion weight strategy.

[0110] In this optional embodiment, according to the normalization result of each dimension information entropy, assign an initial fusion weight to each dimension, and based on the assignment result, construct a target function to generate a fusion weight strategy includes:

[0111] S42231, Convert the normalization result of each dimension information entropy into the utility value of the information contribution of the spatial feature vector and the channel feature vector in each dimension, and assign an initial fusion weight to each dimension;

[0112] S42232, Based on the assigned initial fusion weight, construct a target function with the goal of maximizing Nash product, and solve the target function through mathematical analysis algorithm to obtain a weight solution;

[0113] S42233, Analyze the obtained weight solution, and based on the analysis result, adjust the expression form of the utility value in the target function to generate a fusion weight strategy.

[0114] S423, According to the generated fusion weight strategy, weight fusion is performed on the spatial feature vector and the channel feature vector to obtain a fusion feature vector, and the fusion feature vector is input into a pre-set knowledge distillation network to construct a distillation loss function using the distribution difference between the teacher model and the student model.

[0115] S424, Take the reconstruction error of the double auto-encoder architecture as a structural loss function, take the cosine loss function as a constraint, combine the constructed distillation loss function, and establish a comprehensive loss function, and calculate the corresponding comprehensive loss value.

[0116] S425, take the integrated loss value as the abnormality evaluation value at the current moment, combine the preset dynamic adjustment coefficient to dynamically scale the abnormality evaluation value, and generate an adaptive leakage judgment threshold at the current moment.

[0117] It should be noted that the dynamic adjustment coefficient is adjusted according to historical data, current state and changes in external environment, and can be determined by historical average value, sliding window or adaptive method. For example, the dynamic adjustment coefficient can be adaptively adjusted according to the loss mean, variance and other statistical characteristics in the past period of time; assuming that the adjustment coefficient is a scaling factor of the integrated loss value at time t, then the adjusted integrated loss value L scaled can be expressed as: ; wherein D t is a coefficient adaptively adjusted based on the current moment environment or state information. In addition, by using the dynamic adjustment coefficient to generate the adaptive threshold, the judgment standard can be automatically adjusted according to different environments, so that the detection result is more accurate and flexible; the adaptive leakage judgment threshold can be generated by combining a preset threshold with the dynamic adjustment coefficient, so that it can be adjusted according to the change of the operating environment.

[0118] S43, collect real-time gas concentration signals, calculate corresponding real-time energy-zero ratio values, and add the real-time energy-zero ratio values to the time sequence window sequence in time sequence, and remove the earliest energy-zero ratio mean value in the time sequence window sequence to update the time sequence window sequence.

[0119] S44, input the updated time sequence window sequence into the double autoencoder to obtain an abnormality score, and compare the abnormality score with the adaptive leakage judgment threshold to determine a gas leakage point based on the comparison result.

[0120] It should be noted that inputting the updated time sequence window sequence into the double autoencoder to obtain an abnormality score, and comparing the abnormality score with the adaptive leakage judgment threshold to determine a gas leakage point based on the comparison result specifically includes:

[0121] The updated time sequence window sequence is input as input data into the double autoencoder model for processing. For normal signal data, the encoder can well compress and reconstruct the input, while for abnormal data such as leakage signals, the reconstruction error will increase due to the deviation of its features from the normal mode. By monitoring the reconstruction error of each window, an abnormality score can be calculated, which is a value representing the abnormality degree of the time sequence window; compare the abnormality score with the adaptive leakage judgment threshold, that is:

[0122] ;

[0123] In the formula, Leak represents a gas leakage point; St representing the time instant t anomaly score; T adaptive representing the adaptive leakage decision threshold for the corresponding time instant; if S t exceeds T adaptive i.e. indicating that there is a possibility of gas leakage at the current time instant, the system will flag it as a leakage point ( Leak =1); otherwise, if the anomaly score is below the threshold, it is flagged as a non-leakage point ( Leak =0).

[0124] The above merely describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for AMC sensor array spatial feature recognition and leak point detection based on deep learning, characterized in that, The method comprises: S1, preprocessing the multi-channel gas concentration signals synchronously collected by the sensor array, and combining the spatial topology structure of the sensor, performing spatial correlation analysis on the preprocessing result, and extracting a spatial feature signal reflecting the spatial distribution characteristics of the gas; S2, denoising the extracted spatial feature signal by using wavelet transform, and decomposing the denoised spatial feature by using an empirical mode decomposition algorithm to obtain a plurality of intrinsic mode components, and generating a reconstructed signal based on each intrinsic mode component; S3, using a frame windowing technology to frame process the reconstructed signal, and calculating the energy value and zero-crossing rate of the corresponding frame based on the frame processing result, combining the extracted spatial feature signal, constructing a leakage perception model, and generating a leakage probability distribution sequence; S4, based on the generated leakage probability distribution sequence, establishing a time sequence window sequence, and inputting the time sequence window sequence into a double auto-encoder, combining a knowledge distillation mechanism and a comparative analysis technology, and performing gas leakage point detection.

2. The deep learning based AMC sensor array spatial feature identification and leak point detection method according to claim 1, characterized in that, The wavelet transform is used to denoise the extracted spatial feature signal, and the empirical mode decomposition algorithm is used to decompose the denoised spatial feature to obtain a plurality of intrinsic mode components, and the reconstructed signal is generated based on each intrinsic mode component, which comprises: S21, using wavelet transform to perform multi-scale decomposition on the extracted spatial feature signal to obtain multi-scale sub-signals, and using an adaptive local threshold denoising technology to denoise each scale sub-signal to obtain a denoised spatial feature signal; S22, using an empirical mode decomposition algorithm to decompose the denoised spatial feature signal into a plurality of intrinsic mode components, and extracting multi-dimensional features of each intrinsic mode component; S23, inputting the extracted multi-dimensional features into a preset attention network, learning the importance of each intrinsic mode component in the reconstruction of the spatial feature signal through the attention mechanism, and outputting the attention weight corresponding to each intrinsic mode component; S24, inputting the intrinsic mode component and the corresponding attention weight as input conditions into a pre-trained generative adversarial network, and fusing the multi-scale component information through the generator in the adversarial network to reconstruct the spatial feature signal and generate a reconstructed signal.

3. The deep learning based AMC sensor array spatial feature identification and leak point detection method according to claim 1, characterized in that, The frame windowing technology is used to frame process the reconstructed signal, and the energy value and zero-crossing rate of the corresponding frame are calculated based on the frame processing result, the extracted spatial feature signal is combined, a leakage perception model is constructed, and a leakage probability distribution sequence is generated, which comprises: S31, using a sliding window technology to frame process the reconstructed signal to generate a time sequence signal segment, and applying a window function to each frame of the time sequence signal segment to calculate the energy value of the corresponding frame, and establishing an energy sequence in time sequence; S32, centering and amplitude constraining the energy sequence, calculating the zero-crossing rate of each frame based on the processing result, combining each frame of the zero-crossing rate with the corresponding energy value to construct an energy-zero ratio, and generating an energy-zero ratio sequence in time sequence; S33, combining the extracted spatial feature signal and the time dynamic characteristics of the energy-zero ratio sequence to construct a leakage perception model, and optimizing the leakage perception model by using a comparative clustering loss algorithm. S34, input the generated energy zero ratio sequence to the optimized leakage awareness model, and extract the feature embedding vector corresponding to each frame timing signal segment; S35, by calculating the similarity distance between the feature embedding vector and the preset leakage class clustering center, output the leakage probability of each frame timing signal segment, and generate the leakage probability distribution sequence in time sequence.

4. The deep learning based AMC sensor array spatial feature identification and leak point detection method according to claim 3, characterized in that, The combination of the extracted spatial feature signal and the time dynamic feature of the energy zero ratio sequence constructs a leakage awareness model, and the leakage awareness model is optimized by using a contrast clustering loss algorithm, which comprises: S331, based on the extracted spatial feature signal, calculate the Euclidean distance between each sensor, and construct a spatial adjacency matrix by taking each sensor as a node; S332, perform time domain and frequency domain analysis on the energy zero ratio sequence, extract the time dynamic feature based on the analysis result, and generate a spatio-temporal joint feature tensor combined with the constructed spatial adjacency matrix; S333, combine the spatio-temporal joint feature tensor and the deep learning framework to construct a leakage awareness model, and optimize the leakage awareness model by using a contrast clustering loss algorithm.

5. The deep learning based AMC sensor array spatial feature identification and leak point detection method according to claim 4, characterized in that, The combination of the spatio-temporal joint feature tensor and the deep learning framework to construct a leakage awareness model, and the optimization of the leakage awareness model by using a contrast clustering loss algorithm comprises: S3331, based on the spatial topology and time dynamic information in the spatio-temporal joint feature tensor, combine the deep learning framework to construct a leakage awareness model based on the graph convolution network and the time sequence attention mechanism; S3332, collect sample data containing leakage and non-leakage states from a preset database, and extract structured event information from the sample data using information recognition technology; S3333, input the extracted event information into the constructed leakage awareness model, and output the event representation vector of each sample data by cooperatively processing the input event information through graph convolution operation and time sequence attention mechanism; S3334, according to the leakage and non-leakage states of the sample data, select sample data similar to the event representation vector from the sample data as target sample, and analyze the similarity between the sample data and the target sample and the difference between the sample data and the remaining sample data by using a preset contrast loss function; S3335, based on the similarity and difference analysis results, update the parameters of the leakage awareness model by using a back propagation algorithm to optimize the leakage awareness model.

6. The deep learning based AMC sensor array spatial feature identification and leak point detection method according to claim 1, characterized in that, The combination of the spatio-temporal joint feature tensor and the deep learning framework to construct a leakage awareness model, and the optimization of the leakage awareness model by using a contrast clustering loss algorithm comprises: S41, according to the constructed leakage probability distribution sequence, calculate the energy zero ratio value and the corresponding energy zero ratio mean of each frame of gas concentration signal, and dynamically determine the adaptive time window length based on the current time as the reference, and generate the corresponding time window sequence; S42, input the time window sequence into the double auto-encoder to generate a fusion weight strategy, and construct a comprehensive loss function combined with the knowledge distillation mechanism to generate an adaptive leakage judgment threshold at the current time; S43, collect the real-time gas concentration signal, calculate the corresponding real-time energy zero ratio, and add the real-time energy zero ratio to the time sequence window sequence in time sequence, and remove the earliest energy zero ratio mean in the time sequence window sequence to update the time sequence window sequence; S44, input the updated time sequence window sequence into the double autoencoder to obtain an anomaly score, compare the anomaly score with an adaptive leakage judgment threshold, and determine a gas leakage point based on the comparison result.

7. The deep learning based AMC sensor array spatial feature identification and leak point detection method according to claim 6, characterized in that, The energy zero ratio of each frame of the gas concentration signal is calculated according to the constructed leakage probability distribution sequence, and the corresponding time sequence window sequence is generated by dynamically determining the adaptive time window length based on the current time. S411, calculate the energy zero ratio of each frame of the gas concentration signal according to the constructed leakage probability distribution sequence, and select the maximum energy zero ratio and the minimum energy zero ratio from the energy zero ratio, and calculate the energy zero ratio mean of all frames based on the selection result; S412, based on the calculated energy zero ratio mean, set the minimum window length and the maximum window length, and starting from the minimum window length, calculate the variance of the energy zero ratio sequence in different window lengths in turn based on the current time as the reference; S413, dynamically select the window length at which the variance first exceeds the preset stability threshold as the adaptive time window length according to the variance of the energy zero ratio sequence in each window length, and generate the corresponding time sequence window sequence based on the adaptive time window length.

8. The deep learning based AMC sensor array spatial feature identification and leak point detection method according to claim 7, characterized in that, The time sequence window sequence is input into the double autoencoder to generate a fusion weight strategy, and a comprehensive loss function is constructed by combining a knowledge distillation mechanism to generate an adaptive leakage judgment threshold at the current time, which includes: S421, constructing a double autoencoder with spatial and channel parallelism, inputting the time sequence window sequence into the double autoencoder, and outputting corresponding spatial feature vectors and channel feature vectors; S422, calculate the information entropy of each dimension in the spatial feature vector and the channel feature vector respectively, and generate a fusion weight strategy by combining a weight distribution mechanism; S423, according to the generated fusion weight strategy, the spatial feature vector and the channel feature vector are weighted and fused to obtain a fusion feature vector, and the fusion feature vector is input into a preset knowledge distillation network to construct a distillation loss function by using the distribution difference between the teacher model and the student model; S424, taking the reconstruction error of the double autoencoder architecture as a structural loss function, taking the cosine loss function as a constraint, combining the constructed distillation loss function, and establishing a comprehensive loss function, and calculating the corresponding comprehensive loss value; S425, taking the comprehensive loss value as the anomaly evaluation value at the current time, combining the preset dynamic adjustment coefficient, dynamically scaling the anomaly evaluation value, and generating the adaptive leakage judgment threshold at the current time.

9. The deep learning based AMC sensor array spatial feature identification and leak point detection method according to claim 8, characterized in that, The information entropy of each dimension in the spatial feature vector and the channel feature vector is calculated respectively, and a fusion weight strategy is generated by combining a weight distribution mechanism, which includes: S4221, align the dimensions of the spatial feature vector and the channel feature vector, and map the spatial feature vector and the channel feature vector of each dimension to a preset standard interval by using a standardization algorithm; S4222、based on the standardized spatial feature vector and the channel feature vector, information entropy of each dimension is calculated respectively, and the information entropy is normalized; S4223, according to the normalization result of the information entropy of each dimension, an initial fusion weight is assigned to each dimension, and a target function is constructed based on the assignment result to generate a fusion weight strategy.

10. The deep learning based AMC sensor array spatial feature identification and leak point detection method according to claim 9, characterized in that, The initial fusion weight is assigned to each dimension according to the normalization result of the information entropy of each dimension, and the target function is constructed based on the assignment result to generate the fusion weight strategy, including: S42231, the normalization result of the information entropy of each dimension is converted into the utility value of the information contribution degree of the spatial feature vector and the channel feature vector in each dimension, and an initial fusion weight is assigned to each dimension; S42232, based on the assigned initial fusion weight, a target function is constructed with the goal of maximizing Nash product, and the target function is solved by mathematical analysis algorithm to obtain a weight solution; S42233, the obtained weight solution is analyzed, and based on the analysis result, the expression form of the utility value in the target function is adjusted to generate a fusion weight strategy.

Images