AMC sensor array spatial feature recognition and leakage point detection method based on deep learning

Abstract

The invention discloses an AMC sensor array spatial feature recognition and leakage point detection method based on deep learning, and relates to the technical field of sensors, and the method comprises the steps: carrying out the spatial correlation analysis of a preprocessing result, and extracting a spatial feature signal reflecting the spatial distribution characteristics of gas; decomposing the denoised spatial features by using an empirical mode decomposition algorithm to obtain a plurality of intrinsic mode components, and generating a reconstruction signal based on each intrinsic mode component; performing framing processing on the reconstructed signal by using a framing windowing technology, calculating an energy value and a zero-crossing rate of a corresponding frame based on a framing processing result, constructing a leakage sensing model in combination with the extracted spatial feature signal, and generating a leakage probability distribution sequence; and inputting the time sequence window sequence into a self-encoder, and performing gas leakage point detection in combination with a knowledge distillation mechanism and a comparative analysis technology. According to the invention, accurate extraction and de-noising processing of the spatial feature signals are realized, the signal quality is improved, and the detection precision is further improved.

Description

Technical Field

[0001] This invention relates to the field of sensor technology, and more specifically, to a method for spatial feature recognition and leakage point detection of AMC sensor arrays based on deep learning. Background Technology

[0002] With the rapid development of advanced manufacturing, semiconductor production, and new energy industries, the demand for monitoring airborne molecular contaminants (AMCs) in the environment is increasing. AMCs mainly include volatile organic compounds (VOCs), acidic and alkaline gases, and metal vapors. Even minute leaks can have serious impacts on product quality, equipment lifespan, and even production safety. Therefore, achieving high-precision, real-time leak detection of AMCs has significant industrial value.

[0003] Currently, AMC leak detection primarily relies on multi-sensor array technology. This involves deploying multiple gas sensors in the target area to construct a spatial monitoring network, enabling continuous environmental sensing. However, traditional detection methods often depend on single-point threshold alarms or simple statistical analysis, making it difficult to accurately distinguish between actual leak signals and non-leakage disturbances such as environmental background noise, electromagnetic interference from production equipment, and sensor drift. This results in persistently high false alarm and false negative rates. Furthermore, the diffusion process of AMC is influenced by complex factors such as airflow fields, ambient temperature, and on-site obstacles. Its spatial propagation exhibits significant characteristics such as nonlinearity, time-varying nature, and diffusion delay. Simply relying on independent analysis of single-channel signals makes it even more difficult to accurately locate the leak source.

[0004] In recent years, deep learning technology has demonstrated strong advantages in temporal signal processing and pattern recognition. Models such as 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 process the temporal and spatial dimensions independently, failing to fully exploit the dynamic characteristics of spatiotemporal coupling during AMC diffusion. Furthermore, model training often relies on a large number of labeled samples, while leak samples are scarce in real-world industrial environments, limiting the model's generalization ability. In addition, existing leak detection systems generally employ fixed threshold judgment mechanisms, which cannot adapt to dynamic changes in environmental conditions, leading to sensitivity imbalances at different operational stages.

[0005] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0006] To address the problems in related technologies, this invention proposes a deep learning-based method for spatial feature recognition and leakage point detection of AMC sensor arrays, thereby overcoming the aforementioned technical problems in existing related technologies.

[0007] Therefore, the specific technical solution adopted by the present invention is as follows: A deep learning-based method for spatial feature recognition and leak point detection of AMC sensor arrays, comprising: S1. Preprocess the multi-channel gas concentration signals synchronously acquired by the sensor array, and combine the spatial topology of the sensor to perform spatial correlation analysis on the preprocessing results and extract spatial feature signals that reflect the spatial distribution characteristics of the gas. S2. Denoise the extracted spatial feature signal using wavelet transform, and decompose the denoised spatial features using empirical mode decomposition algorithm to obtain several intrinsic mode components. Generate a reconstructed signal based on each intrinsic mode component. S3. The reconstructed signal is processed by framing and windowing techniques, and the energy value and zero-crossing rate of the corresponding frame are calculated based on the framing processing results. Combined with the extracted spatial feature signals, a leakage sensing model is constructed to generate a leakage probability distribution sequence. S4. Based on the generated leakage probability distribution sequence, establish a time-series window sequence and input the time-series window sequence into a dual autoencoder. Combine the knowledge distillation mechanism and comparative analysis technology to perform gas leak point detection.

[0008] Furthermore, wavelet transform is used to denoise the extracted spatial feature signal, and empirical mode decomposition algorithm is applied to decompose the denoised spatial features to obtain several intrinsic mode components. The reconstructed signal is generated based on each intrinsic mode component, including: S21. Wavelet transform is used to decompose the extracted spatial feature signal into multi-scale sub-signals, and adaptive local threshold denoising technology is used to denoise the sub-signals at each scale to obtain the denoised spatial feature signal. S22. Using the empirical mode decomposition algorithm, the denoised spatial feature signal is decomposed into several intrinsic mode components, and the multidimensional features of each intrinsic mode component are extracted. S23. Input the extracted multidimensional features into the preset attention network, learn the importance of each intrinsic modal component in the reconstruction of spatial feature signals through the attention mechanism, and output the attention weights corresponding to each intrinsic modal component. S24. The intrinsic modal components and their corresponding attention weights are used as input conditions and fed into a pre-trained generative adversarial network. The multi-scale component information is fused through the generator in the adversarial network to reconstruct the spatial feature signal and generate a reconstructed signal.

[0009] Furthermore, the reconstructed signal is processed by framing and windowing techniques, and the energy value and zero-crossing rate of the corresponding frame are calculated based on the framing results. Combined with the extracted spatial feature signals, a leakage sensing model is constructed, generating a leakage probability distribution sequence including: S31. Use the sliding window technique to perform frame-by-frame processing on the reconstructed signal, generate time-series signal segments, apply a window function to each frame of time-series signal segments, calculate the energy value of the corresponding frame, and establish an energy sequence in chronological order. S32. The energy sequence is centered and amplitude constrained. The zero-crossing rate of each frame is calculated based on the processing results. The zero-crossing rate of each frame is combined with the corresponding energy value to construct the energy-zero ratio. The energy-zero ratio sequence is generated in chronological order. S33. Combining the extracted spatial feature signals with the temporal dynamic features of the zero-energy ratio sequence, a leakage sensing model is constructed, and the leakage sensing model is optimized using a contrastive clustering loss algorithm. S34. Input the generated zero-energy ratio sequence into the optimized leakage sensing model and extract the feature embedding vectors corresponding to each frame of time-series signal segments. S35. By calculating the similarity distance between the feature embedding vector and the preset leakage class cluster center, output the leakage probability of each frame of time-series signal segment, and generate a leakage probability distribution sequence in time order.

[0010] Furthermore, by combining the extracted spatial feature signals with the temporal dynamic features of the zero-energy ratio sequence, a leakage sensing model is constructed, and the leakage sensing model is optimized using a contrastive clustering loss algorithm, including: S331. Based on the extracted spatial feature signals, calculate the Euclidean distance between each sensor, and construct a spatial adjacency matrix with each sensor as a node. S332. Perform time-domain and frequency-domain analysis on the zero-energy ratio sequence, extract time dynamic features based on the analysis results, and generate a spatiotemporal joint feature tensor by combining the constructed spatial adjacency matrix. S333. Combining the spatiotemporal joint feature tensor with a deep learning framework, a leakage perception model is constructed, and the leakage perception model is optimized using a contrastive clustering loss algorithm.

[0011] Furthermore, by combining the spatiotemporal joint feature tensor with a deep learning framework, a leakage-aware model is constructed, and the leakage-aware model is optimized using a contrastive clustering loss algorithm, including: S3331. Based on the spatial topology and temporal dynamic information in the spatiotemporal joint feature tensor, and combined with a deep learning framework, a leakage perception model based on graph convolutional networks and temporal attention mechanism is constructed. S3332. Collect sample data containing leakage and non-leakage states from a preset database, and use information recognition technology to extract structured event information from the sample data; S3333. Input the extracted event information into the constructed leakage perception model, and process the input event information in a coordinated manner through graph convolution operation and temporal attention mechanism to output the event representation vector of each sample data. S3334. Based on 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 the preset contrast 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. S3335. Based on the similarity and difference analysis results, the parameters of the leakage sensing model are updated using the backpropagation algorithm to optimize the leakage sensing model.

[0012] Furthermore, based on the generated leakage probability distribution sequence, a time-series window sequence is established and input into a dual autoencoder. Combining a knowledge distillation mechanism and comparative analysis technology, gas leak point detection is performed, including: S41. Based on the constructed leakage probability distribution sequence, calculate the energy-to-zero ratio and the corresponding mean energy-to-zero ratio of each frame of the gas concentration signal, and dynamically determine the adaptive time window length based on the current time to generate the corresponding time window sequence. S42. Input the time-series window sequence into the dual autoencoder to generate a fusion weight strategy, and combine it with the knowledge distillation mechanism to construct a comprehensive loss function to generate the adaptive leakage judgment threshold at the current time. S43. Collect real-time gas concentration signals, calculate the corresponding real-time zero energy ratio, and add the real-time zero energy ratio to the time-series window sequence in chronological order. Simultaneously remove the earliest mean zero energy ratio in the time-series window sequence to update the time-series window sequence. S44. Input the updated timing window sequence into the dual autoencoder to obtain the anomaly score, and compare the anomaly score with the adaptive leak judgment threshold to determine the gas leak point based on the comparison result.

[0013] Furthermore, based on the constructed leakage probability distribution sequence, the energy-to-zero ratio and the corresponding mean energy-to-zero ratio of each frame of the gas concentration signal are calculated. Then, using the current time as a reference, the adaptive time window length is dynamically determined, generating the corresponding time-series window sequence, including: S411. Based on the constructed leakage probability distribution sequence, calculate the energy-to-zero ratio of the gas concentration signal in each frame, and select the maximum and minimum energy-to-zero ratios from the energy-to-zero ratios. Calculate the average energy-to-zero ratio of all frames based on the selection results. S412. Based on the calculated mean of the zero-energy ratio, set the minimum window length and the maximum window length, and using the current time as the benchmark, start from the minimum window length and calculate the variance of the zero-energy ratio sequence within different window lengths in sequence. S413. Based on the variance of the zero-energy sequence of each window length, dynamically select the window length whose variance first exceeds the preset stability threshold as the adaptive time window length, and generate the corresponding time window sequence based on the adaptive time window length.

[0014] Furthermore, the temporal window sequence is input into the dual autoencoder to generate a fusion weight strategy, and a comprehensive loss function is constructed by combining it with a knowledge distillation mechanism to generate the adaptive leakage judgment threshold for the current time step, including: S421. Construct a spatial and channel parallel dual autoencoder, input the temporal window sequence into the dual autoencoder, and output the corresponding spatial feature vector and channel feature vector; 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 the weight allocation mechanism. S423. Based on the generated fusion weight strategy, the spatial feature vector and the channel feature vector are weighted and fused to obtain the fused feature vector. The fused feature vector is then input into the preset knowledge distillation network, and the distillation loss function is constructed using the distribution difference between the teacher model and the student model. S424. Using the reconstruction error of the dual autoencoder architecture as the structural loss function and the cosine loss function as the constraint, combined with the constructed distillation loss function, a comprehensive loss function is established, and the corresponding comprehensive loss value is calculated. S425. The comprehensive loss value is used as the anomaly assessment value at the current moment. Combined with the preset dynamic adjustment coefficient, the anomaly assessment value is dynamically scaled to generate the adaptive leakage judgment threshold at the current moment.

[0015] Furthermore, the information entropy of each dimension in the spatial feature vector and channel feature vector are calculated separately, and a fusion weight strategy is generated by combining the weight allocation mechanism, including: S4221. Align the spatial feature vector and the channel feature vector in terms of dimensions, and use a standardization algorithm to map the spatial feature vector and the channel feature vector in each dimension to a preset standard range. S4222. Based on the standardized spatial feature vector and channel feature vector, calculate the information entropy of each dimension and normalize the information entropy. S4223. Based on the normalization results of the information entropy of each dimension, assign initial fusion weights to each dimension, and construct an objective function based on the assignment results to generate a fusion weight strategy.

[0016] Furthermore, based on the normalization results of the information entropy of each dimension, initial fusion weights are assigned to each dimension, and an objective function is constructed based on the assignment results. The fusion weight strategy includes: S42231. The normalized results of information entropy in each dimension are transformed into the utility values ​​of the information contribution of spatial feature vectors and channel feature vectors in each dimension, and initial fusion weights are assigned to each dimension. S42232. Based on the assigned initial fusion weights, construct an objective function with the goal of maximizing the Nash product, and solve the objective function using a mathematical analytical algorithm to obtain the weighted solution; S42233. Analyze the obtained weighted solution and adjust the expression form of the utility value in the objective function based on the analysis results to generate a fusion weight strategy.

[0017] The beneficial effects of this invention are as follows: 1. This invention achieves accurate extraction and denoising of spatial feature signals by fusing the spatial topology of sensor arrays with deep learning technology; at the same time, combining wavelet transform and empirical mode decomposition techniques can suppress noise interference, improve signal quality, and further improve detection accuracy.

[0018] 2. This invention, through the synergistic effect of frame-by-frame windowing technology and a leak detection model, combined with real-time calculated energy values ​​and zero-crossing rates, can detect subtle changes in gas concentration and improve sensitivity to early leaks. At the same time, by integrating an autoencoder and a knowledge distillation mechanism, it can accurately generate leak judgment thresholds, thereby improving the sensitivity and accuracy of gas leak detection. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of a deep learning-based method for spatial feature identification and leak point detection of an AMC sensor array according to an embodiment of the present invention. Detailed Implementation

[0021] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0022] According to an embodiment of the present invention, a method for spatial feature identification and leakage point detection of AMC sensor array based on deep learning is provided.

[0023] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a deep learning-based method for spatial feature identification and leakage point detection of an AMC sensor array includes: S1. Preprocess the multi-channel gas concentration signals synchronously acquired by the sensor array, and combine the spatial topology of the sensor to perform spatial correlation analysis on the preprocessing results, and extract spatial feature signals that reflect the spatial distribution characteristics of the gas.

[0024] It should be further noted that preprocessing includes denoising, standardization, and missing value imputation. Spatial correlation analysis employs methods such as cross-correlation analysis and principal component analysis to quantify the correlation between sensors. For example, sensors with high correlation are considered to have detected concentration fluctuations in the same gas region, while sensors with low correlation detect gas distributions at different locations. Furthermore, spatial characteristic signals include spatial average concentration, spatial concentration gradient, spatial covariance, spatial patterns extracted by spatial correlation analysis, spatial correlation, energy distribution of spatial patterns, and cluster structure of spatial distributions.

[0025] S2. Denoise the extracted spatial feature signal using wavelet transform, and decompose the denoised spatial features using empirical mode decomposition algorithm to obtain several intrinsic mode components. Generate a reconstructed signal based on each intrinsic mode component.

[0026] In this optional embodiment, wavelet transform is used to denoise the extracted spatial feature signal, and empirical mode decomposition algorithm is applied to decompose the denoised spatial features to obtain several intrinsic mode components. The reconstructed signal is generated based on each intrinsic mode component, including: S21. Wavelet transform is used to decompose the extracted spatial feature signal into multi-scale sub-signals, and adaptive local threshold denoising technology is used to denoise the sub-signals at each scale to obtain the denoised spatial feature signal.

[0027] S22. Using the empirical mode decomposition algorithm, the denoised spatial feature signal is decomposed into several intrinsic mode components, and the multidimensional features of each intrinsic mode component are extracted.

[0028] It should be further explained that multidimensional features include time-domain features, frequency-domain features, time-frequency features, higher-order statistical features, and statistical features. For example, time-domain features include mean, standard deviation, kurtosis, skewness, extrema, and number of zero intersections; frequency-domain features include spectral center, spectral width, spectral entropy, and maximum frequency component; time-frequency features include time-frequency energy and instantaneous frequency; higher-order statistical features include higher-order cumulants and fractal dimension; and statistical features include autoregressive coefficients and correlation coefficients.

[0029] S23. Input the extracted multidimensional features into the preset attention network, learn the importance of each intrinsic modal component in the reconstruction of spatial feature signals through the attention mechanism, and output the attention weights corresponding to each intrinsic modal component.

[0030] S24. The intrinsic modal components and their corresponding attention weights are used as input conditions and fed into a pre-trained generative adversarial network. The multi-scale component information is fused through the generator in the adversarial network to reconstruct the spatial feature signal and generate a reconstructed signal.

[0031] S3. The reconstructed signal is processed by framing and windowing techniques, and the energy value and zero-crossing rate of the corresponding frame are calculated based on the framing results. Combined with the extracted spatial feature signals, a leakage sensing model is constructed to generate a leakage probability distribution sequence.

[0032] In this optional embodiment, the reconstructed signal is processed by framing and windowing techniques, and the energy value and zero-crossing rate of the corresponding frame are calculated based on the framing results. Combined with the extracted spatial feature signals, a leakage sensing model is constructed, and a leakage probability distribution sequence is generated, including: S31. The reconstructed signal is processed by frame segmentation using sliding window technology to generate time sequence signal segments. A window function is applied to each time sequence signal segment to calculate the energy value of the corresponding frame and establish an energy sequence in chronological order.

[0033] It should be noted that window functions include rectangular windows, Hamming windows, Hanning windows, Blackman windows, etc.

[0034] S32. The energy sequence is centered and amplitude constrained. The zero-crossing rate of each frame is calculated based on the processing results. The zero-crossing rate of each frame is combined with the corresponding energy value to construct the energy-zero ratio. The energy-zero ratio sequence is generated in chronological order.

[0035] It should be further explained that the energy sequence is centered and amplitude-constrained. Based on the processing results, the zero-crossing rate of each frame is calculated, and the zero-crossing rate of each frame is combined with the corresponding energy value to construct the energy-to-zero ratio. The energy-to-zero ratio sequence is generated in chronological order, specifically including: Centering the energy sequence involves subtracting the average value of the sequence from each energy value, thereby eliminating the DC component of the signal and ensuring that the energy sequence accurately reflects the dynamic changes of the signal. For example, the centered energy sequence E t ′ can be represented as: Among them, E t It is the original energy sequence, the energy value of the t-th frame, μ EThe mean of the energy sequence is used as the basis for calculation. The amplitude of the centered energy sequence is constrained to limit its amplitude to a reasonable range to avoid distortion of the analysis results due to extreme energy values. For example, the maximum and minimum values ​​of the energy sequence are compressed to a predetermined range [-1, 1]. Based on the energy sequence after amplitude constraint, each frame of the 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 compared with the energy value. By calculating the zero-crossing ratio of each frame and arranging these values ​​in chronological order, a zero-crossing ratio sequence representing the change of the signal over time is obtained.

[0036] S33. By combining the extracted spatial feature signals with the temporal dynamic features of the zero-ratio energy sequence, a leakage sensing model is constructed, and the leakage sensing model is optimized using a contrastive clustering loss algorithm.

[0037] In this optional embodiment, a leakage sensing model is constructed by combining the extracted spatial feature signals with the temporal dynamic features of the zero-energy ratio sequence, and the leakage sensing model is optimized using a contrastive clustering loss algorithm, including: S331. Based on the extracted spatial feature signals, calculate the Euclidean distance between each sensor, and construct a spatial adjacency matrix with each sensor as a node.

[0038] S332. Perform time-domain and frequency-domain analysis on the zero-ratio energy sequence, extract time dynamic features based on the analysis results, and generate a spatiotemporal joint feature tensor by combining the constructed spatial adjacency matrix.

[0039] S333. Combining the spatiotemporal joint feature tensor with a deep learning framework, a leakage perception model is constructed, and the leakage perception model is optimized using a contrastive clustering loss algorithm.

[0040] In this optional embodiment, a leakage-aware model is constructed by combining the spatiotemporal joint feature tensor with a deep learning framework, and the leakage-aware model is optimized using a contrastive clustering loss algorithm, including: S3331. Based on the spatial topology and temporal dynamic information in the spatiotemporal joint feature tensor, and combined with a deep learning framework, a leakage perception model based on graph convolutional networks and temporal attention mechanism is constructed. S3332. Collect sample data containing leakage and non-leakage states from a preset database, and use information recognition technology to extract structured event information from the sample data; S3333. Input the extracted event information into the constructed leakage perception model, and process the input event information in a coordinated manner through graph convolution operation and temporal attention mechanism to output the event representation vector of each sample data. S3334. Based on 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 the preset contrast 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. S3335. Based on the similarity and difference analysis results, the parameters of the leakage sensing model are updated using the backpropagation algorithm to optimize the leakage sensing model.

[0041] S34. Input the generated zero-energy ratio sequence into the optimized leakage sensing model and extract the feature embedding vectors corresponding to each frame of time-series signal segments.

[0042] S35. By calculating the similarity distance between the feature embedding vector and the preset leakage class cluster center, output the leakage probability of each frame of time-series signal segment, and generate a leakage probability distribution sequence in time order.

[0043] It should be further explained that the leakage cluster centers are representative embedding vectors obtained by performing unsupervised or supervised clustering analysis on historical leakage samples, such as K-means, GMM, or Gaussian kernel-based clustering. Each cluster center represents a leakage pattern or behavioral feature, used to subsequently determine the similarity between new samples and leakage patterns. Furthermore, similarity distance can be calculated using common metrics such as Euclidean distance, cosine similarity, or Mahalanobis distance.

[0044] S4. Based on the generated leakage probability distribution sequence, establish a time-series window sequence and input the time-series window sequence into a dual autoencoder. Combine the knowledge distillation mechanism and comparative analysis technology to perform gas leak point detection.

[0045] In this optional embodiment, 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 dual autoencoder. Combining a knowledge distillation mechanism and comparative analysis technology, gas leak point detection is performed, including: S41. Based on the constructed leakage probability distribution sequence, calculate the energy-to-zero ratio and the corresponding mean energy-to-zero ratio of each frame of the gas concentration signal, and dynamically determine the adaptive time window length based on the current time to generate the corresponding time window sequence.

[0046] In this optional embodiment, based on the constructed leakage probability distribution sequence, the energy-to-zero ratio and the corresponding mean energy-to-zero ratio of each frame of the gas concentration signal are calculated, and the adaptive time window length is dynamically determined based on the current time, generating the corresponding time-series window sequence, including: S411. Based on the constructed leakage probability distribution sequence, calculate the energy-to-zero ratio of the gas concentration signal in each frame, and select the maximum and minimum energy-to-zero ratios from the energy-to-zero ratios. Calculate the average energy-to-zero ratio of all frames based on the selection results. S412. Based on the calculated mean of the zero-energy ratio, set the minimum window length and the maximum window length, and using the current time as the benchmark, start from the minimum window length and calculate the variance of the zero-energy ratio sequence within different window lengths in sequence. S413. Based on the variance of the zero-energy sequence of each window length, dynamically select the window length whose variance first exceeds the preset stability threshold as the adaptive time window length, and generate the corresponding time window sequence based on the adaptive time window length.

[0047] S42. Input the temporal window sequence into the dual autoencoder to generate a fusion weight strategy, and combine it with the knowledge distillation mechanism to construct a comprehensive loss function to generate the adaptive leakage judgment threshold at the current moment.

[0048] In this optional embodiment, the temporal window sequence is input into a dual autoencoder to generate a fusion weight strategy, and a comprehensive loss function is constructed by combining a knowledge distillation mechanism to generate the adaptive leakage judgment threshold for the current time step, including: S421. Construct a spatial and channel parallel dual autoencoder, input the temporal window sequence into the dual autoencoder, and output the corresponding spatial feature vector and channel feature vector.

[0049] 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 the weight allocation mechanism.

[0050] In this optional embodiment, 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 the weight allocation mechanism, including: S4221. Align the spatial feature vectors and channel feature vectors by dimensions, and use a standardization algorithm to map the spatial feature vectors and channel feature vectors of each dimension to a preset standard range.

[0051] S4222. Based on the standardized spatial feature vector and channel feature vector, calculate the information entropy of each dimension and normalize the information entropy.

[0052] S4223. Based on the normalization results of the information entropy of each dimension, assign initial fusion weights to each dimension, and construct an objective function based on the assignment results to generate a fusion weight strategy.

[0053] In this optional embodiment, based on the normalization result of the information entropy of each dimension, initial fusion weights are assigned to each dimension, and an objective function is constructed based on the assignment result to generate the fusion weight strategy, including: S42231. The normalized results of information entropy in each dimension are transformed into the utility values ​​of the information contribution of spatial feature vectors and channel feature vectors in each dimension, and initial fusion weights are assigned to each dimension. S42232. Based on the assigned initial fusion weights, construct an objective function with the goal of maximizing the Nash product, and solve the objective function using a mathematical analytical algorithm to obtain the weighted solution; S42233. Analyze the obtained weighted solution and adjust the expression form of the utility value in the objective function based on the analysis results to generate a fusion weight strategy.

[0054] S423. Based on the generated fusion weight strategy, the spatial feature vector and the channel feature vector are weighted and fused to obtain the fused feature vector. The fused feature vector is then input into the preset knowledge distillation network, and the distillation loss function is constructed using the distribution difference between the teacher model and the student model.

[0055] S424. Using the reconstruction error of the dual autoencoder architecture as the structural loss function and the cosine loss function as the constraint, combined with the constructed distillation loss function, a comprehensive loss function is established, and the corresponding comprehensive loss value is calculated.

[0056] S425. The comprehensive loss value is used as the anomaly assessment value at the current moment. Combined with the preset dynamic adjustment coefficient, the anomaly assessment value is dynamically scaled to generate the adaptive leakage judgment threshold at the current moment.

[0057] It should be further explained that the dynamic adjustment coefficient is adjusted based on historical data, the current state, and changes in the external environment. It can be determined using historical averages, a sliding window, or adaptive methods. For example, the dynamic adjustment coefficient can be adaptively adjusted based on statistical characteristics such as the mean and variance of losses over a past period. Assuming the adjustment coefficient is a scaling factor on the overall loss value at time t, then the adjusted overall loss value L... scaled It can be represented as: ; where D t These coefficients are adaptively adjusted based on the current environmental or state information. Furthermore, by using dynamically adjusted coefficients to generate adaptive thresholds, the judgment criteria can be automatically adjusted according to different environments, making the detection results more accurate and flexible. The adaptive leakage judgment threshold can be generated by combining a preset threshold with dynamically adjusted coefficients, thus enabling adjustment according to changes in the operating environment.

[0058] S43. Acquire real-time gas concentration signals, calculate the corresponding real-time energy-to-zero ratio, and add the real-time energy-to-zero ratio to the time-series window sequence in chronological order. Simultaneously remove the earliest average energy-to-zero ratio in the time-series window sequence to update the time-series window sequence.

[0059] S44. Input the updated timing window sequence into the dual autoencoder to obtain the anomaly score, and compare the anomaly score with the adaptive leak judgment threshold to determine the gas leak point based on the comparison result.

[0060] It should be further explained that the updated time-series window sequence is input into the dual autoencoder to obtain anomaly scores. These anomaly scores are then compared with an adaptive leak detection threshold. Based on the comparison results, the gas leak points are specifically determined as follows: The updated time-series window sequence is used as input data and fed into a dual autoencoder model for processing. For normal signal data, the encoder can effectively compress and reconstruct the input. However, for anomalous data, such as leaky signals, the reconstruction error increases due to deviations from the normal pattern. By monitoring the reconstruction error of each window, an anomaly score can be calculated, representing the degree of anomalousness of that time-series window. The anomaly score is then compared with an adaptive leakage detection threshold. ; In the formula, Leaks Indicates the gas leak point; S t Indicates time t Abnormal scores; T adaptive This represents the adaptive leakage judgment threshold at the corresponding time; if S t Exceeded T adaptive This indicates that there is a possibility of gas leakage at this moment, and the system will mark it as a leak point. Leaks =1); conversely, if the anomaly score is below the threshold, it is marked as a non-leakage point ( =1). Leaks =0).

[0061] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for spatial feature recognition and leakage point detection of AMC sensor arrays based on deep learning, characterized in that, The method includes: S1. Preprocess the multi-channel gas concentration signals synchronously acquired by the sensor array, and combine the spatial topology of the sensor to perform spatial correlation analysis on the preprocessing results and extract spatial feature signals that reflect the spatial distribution characteristics of the gas. S2. Denoise the extracted spatial feature signal using wavelet transform, and decompose the denoised spatial features using empirical mode decomposition algorithm to obtain several intrinsic mode components. Generate a reconstructed signal based on each intrinsic mode component. S3. The reconstructed signal is processed by framing and windowing techniques, and the energy value and zero-crossing rate of the corresponding frame are calculated based on the framing processing results. Combined with the extracted spatial feature signals, a leakage sensing model is constructed to generate a leakage probability distribution sequence. S4. Based on the generated leakage probability distribution sequence, establish a time-series window sequence and input the time-series window sequence into a dual autoencoder. Combine the knowledge distillation mechanism and comparative analysis technology to perform gas leak point detection.

2. The method for spatial feature recognition and leakage point detection of AMC sensor array based on deep learning according to claim 1, characterized in that, The process involves denoising the extracted spatial feature signal using wavelet transform and then decomposing the denoised spatial features using an empirical mode decomposition algorithm to obtain several intrinsic mode components. The generation of the reconstructed signal based on these intrinsic mode components includes: S21. Wavelet transform is used to decompose the extracted spatial feature signal into multi-scale sub-signals, and adaptive local threshold denoising technology is used to denoise the sub-signals at each scale to obtain the denoised spatial feature signal. S22. Using the empirical mode decomposition algorithm, the denoised spatial feature signal is decomposed into several intrinsic mode components, and the multidimensional features of each intrinsic mode component are extracted. S23. Input the extracted multidimensional features into the preset attention network, learn the importance of each intrinsic modal component in the reconstruction of spatial feature signals through the attention mechanism, and output the attention weights corresponding to each intrinsic modal component. S24. The intrinsic modal components and their corresponding attention weights are used as input conditions and fed into a pre-trained generative adversarial network. The multi-scale component information is fused through the generator in the adversarial network to reconstruct the spatial feature signal and generate a reconstructed signal.

3. The method for spatial feature recognition and leakage point detection of AMC sensor array based on deep learning according to claim 1, characterized in that, The process involves using frame-segmentation and windowing techniques to segment the reconstructed signal, calculating the energy value and zero-crossing rate of the corresponding frame based on the segmentation results, and combining this with the extracted spatial feature signals to construct a leakage sensing model and generate a leakage probability distribution sequence, including: S31. Use the sliding window technique to perform frame-by-frame processing on the reconstructed signal, generate time-series signal segments, apply a window function to each frame of time-series signal segments, calculate the energy value of the corresponding frame, and establish an energy sequence in chronological order. S32. The energy sequence is centered and amplitude constrained. The zero-crossing rate of each frame is calculated based on the processing results. The zero-crossing rate of each frame is combined with the corresponding energy value to construct the energy-zero ratio. The energy-zero ratio sequence is generated in chronological order. S33. Combining the extracted spatial feature signals with the temporal dynamic features of the zero-energy ratio sequence, a leakage sensing model is constructed, and the leakage sensing model is optimized using a contrastive clustering loss algorithm. S34. Input the generated zero-energy ratio sequence into the optimized leakage sensing model and extract the feature embedding vectors corresponding to each frame of time-series signal segments. S35. By calculating the similarity distance between the feature embedding vector and the preset leakage class cluster center, output the leakage probability of each frame of time-series signal segment, and generate a leakage probability distribution sequence in time order.

4. The method for spatial feature recognition and leakage point detection of AMC sensor array based on deep learning according to claim 3, characterized in that, The process of constructing a leakage sensing model by combining the extracted spatial feature signals with the temporal dynamic features of the zero-energy ratio sequence, and optimizing the leakage sensing model using a contrastive clustering loss algorithm, includes: S331. Based on the extracted spatial feature signals, calculate the Euclidean distance between each sensor, and construct a spatial adjacency matrix with each sensor as a node. S332. Perform time-domain and frequency-domain analysis on the zero-energy ratio sequence, extract time dynamic features based on the analysis results, and generate a spatiotemporal joint feature tensor by combining the constructed spatial adjacency matrix. S333. Combining the spatiotemporal joint feature tensor with a deep learning framework, a leakage perception model is constructed, and the leakage perception model is optimized using a contrastive clustering loss algorithm.

5. The method for spatial feature recognition and leakage point detection of AMC sensor array based on deep learning according to claim 4, characterized in that, The method of constructing a leakage perception model by combining spatiotemporal joint feature tensors with a deep learning framework, and optimizing the leakage perception model using a contrastive clustering loss algorithm, includes: S3331. Based on the spatial topology and temporal dynamic information in the spatiotemporal joint feature tensor, and combined with a deep learning framework, a leakage perception model based on graph convolutional networks and temporal attention mechanism is constructed. S3332. Collect sample data containing leakage and non-leakage states from a preset database, and use information recognition technology to extract structured event information from the sample data; S3333. Input the extracted event information into the constructed leakage perception model, and process the input event information in a coordinated manner through graph convolution operation and temporal attention mechanism to output the event representation vector of each sample data. S3334. Based on 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 the preset contrast 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. S3335. Based on the similarity and difference analysis results, the parameters of the leakage sensing model are updated using the backpropagation algorithm to optimize the leakage sensing model.

6. The method for spatial feature recognition and leakage point detection of AMC sensor array based on deep learning according to claim 1, characterized in that, The process of establishing a time-series window sequence based on the generated leakage probability distribution sequence, inputting the time-series window sequence into a dual autoencoder, and combining a knowledge distillation mechanism with comparative analysis technology to perform gas leak point detection includes: S41. Based on the constructed leakage probability distribution sequence, calculate the energy-to-zero ratio and the corresponding mean energy-to-zero ratio of each frame of the gas concentration signal, and dynamically determine the adaptive time window length based on the current time to generate the corresponding time window sequence. S42. Input the time-series window sequence into the dual autoencoder to generate a fusion weight strategy, and combine it with the knowledge distillation mechanism to construct a comprehensive loss function to generate the adaptive leakage judgment threshold at the current time. S43. Collect real-time gas concentration signals, calculate the corresponding real-time zero energy ratio, and add the real-time zero energy ratio to the time-series window sequence in chronological order. Simultaneously remove the earliest mean zero energy ratio in the time-series window sequence to update the time-series window sequence. S44. Input the updated timing window sequence into the dual autoencoder to obtain the anomaly score, and compare the anomaly score with the adaptive leak judgment threshold to determine the gas leak point based on the comparison result.

7. The method for spatial feature recognition and leakage point detection of AMC sensor array based on deep learning according to claim 6, characterized in that, The step of calculating the energy-to-zero ratio and the corresponding mean energy-to-zero ratio of each frame of the gas concentration signal based on the constructed leakage probability distribution sequence, and dynamically determining the adaptive time window length based on the current time, and generating the corresponding time-series window sequence includes: S411. Based on the constructed leakage probability distribution sequence, calculate the energy-to-zero ratio of the gas concentration signal in each frame, and select the maximum and minimum energy-to-zero ratios from the energy-to-zero ratios. Calculate the average energy-to-zero ratio of all frames based on the selection results. S412. Based on the calculated mean of the zero-energy ratio, set the minimum window length and the maximum window length, and using the current time as the benchmark, start from the minimum window length and calculate the variance of the zero-energy ratio sequence within different window lengths in sequence. S413. Based on the variance of the zero-energy sequence of each window length, dynamically select the window length whose variance first exceeds the preset stability threshold as the adaptive time window length, and generate the corresponding time window sequence based on the adaptive time window length.

8. The method for spatial feature recognition and leakage point detection of AMC sensor array based on deep learning according to claim 7, characterized in that, The process of inputting the temporal window sequence into a dual autoencoder to generate a fusion weight strategy, and constructing a comprehensive loss function in conjunction with a knowledge distillation mechanism to generate the adaptive leakage judgment threshold for the current time step includes: S421. Construct a spatial and channel parallel dual autoencoder, input the temporal window sequence into the dual autoencoder, and output the corresponding spatial feature vector and channel feature vector; 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 the weight allocation mechanism. S423. Based on the generated fusion weight strategy, the spatial feature vector and the channel feature vector are weighted and fused to obtain the fused feature vector. The fused feature vector is then input into the preset knowledge distillation network, and the distillation loss function is constructed using the distribution difference between the teacher model and the student model. S424. Using the reconstruction error of the dual autoencoder architecture as the structural loss function and the cosine loss function as the constraint, combined with the constructed distillation loss function, a comprehensive loss function is established, and the corresponding comprehensive loss value is calculated. S425. The comprehensive loss value is used as the anomaly assessment value at the current moment. Combined with the preset dynamic adjustment coefficient, the anomaly assessment value is dynamically scaled to generate the adaptive leakage judgment threshold at the current moment.

9. The method for spatial feature recognition and leakage point detection of AMC sensor array based on deep learning according to claim 8, characterized in that, The step of calculating the information entropy of each dimension in the spatial feature vector and the channel feature vector respectively, and generating a fusion weight strategy by combining the weight allocation mechanism, includes: S4221. Align the spatial feature vector and the channel feature vector in terms of dimensions, and use a standardization algorithm to map the spatial feature vector and the channel feature vector in each dimension to a preset standard range. S4222. Based on the standardized spatial feature vector and channel feature vector, calculate the information entropy of each dimension and normalize the information entropy. S4223. Based on the normalization results of the information entropy of each dimension, assign initial fusion weights to each dimension, and construct an objective function based on the assignment results to generate a fusion weight strategy.

10. The method for spatial feature recognition and leakage point detection of AMC sensor array based on deep learning according to claim 9, characterized in that, The step of assigning initial fusion weights to each dimension based on the normalization result of the information entropy of each dimension, and constructing an objective function based on the assignment result to generate the fusion weight strategy includes: S42231. The normalized results of information entropy in each dimension are transformed into the utility values ​​of the information contribution of spatial feature vectors and channel feature vectors in each dimension, and initial fusion weights are assigned to each dimension. S42232. Based on the assigned initial fusion weights, construct an objective function with the goal of maximizing the Nash product, and solve the objective function using a mathematical analytical algorithm to obtain the weighted solution; S42233. Analyze the obtained weighted solution and adjust the expression form of the utility value in the objective function based on the analysis results to generate a fusion weight strategy.

Images