Deep learning-based smoke sensing chip dustproof air permeability dynamic identification method

Abstract

The invention discloses a smoke sensing chip dustproof air permeability dynamic identification method based on deep learning, and the method comprises the following steps: S1, collecting signal data of a smoke sensing chip, and carrying out the preprocessing; s2, constructing a WaveNet network based on gating convolution, and performing multi-scale feature extraction on the standardized signal pair set; s3, inputting the multi-scale feature map sequence into a Siamese network to generate fusion difference representation; s4, performing similarity calculation on the fusion difference representation, and performing measurement by adopting a weighted combination function; s5, comparing the similarity score with a threshold function, and outputting a corresponding dustproof air permeability state label; and S6, executing sliding window clustering analysis, identifying a stable state segment and a change state segment, and performing network feedback on an identification result for parameter updating. Through a deep learning dynamic identification method, real-time accurate monitoring of the dustproof air permeability of the smoke sensing chip is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of smoke sensor chip design, and in particular to a method for dynamically identifying the dust-proof and air-permeability performance of a smoke sensor chip based on deep learning. Background Art

[0002] With the continuous development of intelligent hardware technology, smoke sensor chips, as a key component of fire detection, are widely used in various security monitoring systems. The performance of smoke sensor chips directly affects the response speed and accuracy of the fire alarm system. Especially in complex environments, how to maintain its stability and reliability is an urgent problem to be solved. Smoke sensor chips are usually used to detect changes in smoke particles in the air and determine whether a fire has occurred by sensing environmental factors such as airflow, temperature, and humidity. However, over time, the dustproof and breathable performance of the smoke sensor chip will be affected by changes in environmental conditions and dust accumulation, causing its sensor function to gradually degrade, thereby affecting the accuracy of smoke detection. In order to ensure the normal operation of smoke sensor chips in various environments, dustproof and breathable properties are key technical indicators. However, the existing technology has certain defects in detecting the dustproof and breathable performance of smoke sensor chips.

[0003] Existing methods for monitoring the dust and air permeability of smoke sensor chips mainly rely on manual inspection and regular maintenance. These methods usually require regular disassembly and cleaning of the chip, or the use of simple testing equipment to check whether it is working properly. These methods are not only inefficient, but also cannot achieve real-time and comprehensive status monitoring of the smoke sensor chip. Once the dustproof layer of the chip is blocked, resulting in a decrease in air permeability, false alarms or missed alarms may occur, posing a huge safety hazard to the fire warning system. In addition, existing technologies are difficult to cope with changes in factors such as dust and humidity in the environment, because these factors will have a dynamic impact on the performance of the smoke sensor chip, and traditional methods often cannot capture these changes in time. Therefore, how to provide a method that can monitor the dust and air permeability of smoke sensor chips in real time and dynamically is the key to improving the overall performance and accuracy of smoke sensor chips.

[0004] Currently, some detection methods based on traditional machine learning have been applied to the performance monitoring of smoke sensor chips. These methods typically rely on rule-based feature extraction and model prediction, collecting and analyzing environmental data to determine the operating status of the smoke sensor chip. However, these methods often face two problems. First, they are based on static rules or simple statistical models and cannot flexibly cope with the complexity brought about by environmental changes. Second, these methods have poor adaptability to different environmental conditions and are easily affected by external interference, resulting in performance fluctuations. Although these methods can work in some specific environments, they are often unable to adapt to changing environments, especially the impact of dynamic changes in dust, humidity, air pressure, etc. on chip performance. Therefore, these existing technologies cannot effectively meet the needs of dynamic monitoring of the dust and air permeability performance of smoke sensor chips.

[0005] Furthermore, with the rapid development of deep learning technology, a growing number of researchers are exploring the use of deep neural networks to solve complex problems, particularly in areas such as image processing and signal processing. Deep learning has demonstrated remarkable success in a variety of fields by automatically extracting complex features. For example, in smoke sensor chip monitoring, deep learning technology can automatically extract meaningful features and make accurate predictions based on these features by training on large amounts of sensor data. However, existing deep learning methods mostly focus on processing data such as images or voice, and lack effective solutions for dynamically identifying sensor data, particularly the dust and air permeability performance of smoke sensor chips.

[0006] Therefore, how to provide a dynamic identification method for the dust-proof and breathable performance of smoke sensor chips based on deep learning is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0007] One objective of this invention is to propose a deep learning-based method for dynamically identifying the dustproof and breathable properties of smoke sensor chips. This method leverages WaveNet and Siamese network technologies, describing in detail a deep learning algorithm that combines a WaveNet network based on gated convolution with a cross-scale alignment module. By dynamically monitoring, extracting features, and calculating similarity from smoke sensor chip signal data under different environmental conditions, this method achieves real-time identification and feedback on the dustproof and breathable properties of smoke sensor chips. This method boasts strong real-time performance, high accuracy, and adaptability, effectively improving the reliability and early warning capabilities of smoke sensor chips in complex environments.

[0008] According to an embodiment of the present invention, a method for dynamically identifying the dust-proof and air-permeability performance of a smoke sensor chip based on deep learning includes the following steps:

[0009] S1. Collect signal data from the smoke sensor chip under different environmental conditions, perform preprocessing, and construct a standardized signal pair set;

[0010] S2. Construct a WaveNet network based on gated convolution, use different receptive field configurations to perform multi-scale feature extraction on the standardized signal pair set, and generate a multi-scale feature map sequence;

[0011] S3. Input the multi-scale feature map sequence into a Siamese network to generate a fused difference representation. The Siamese network includes two parameter-sharing feature extraction channels and a cross-scale alignment module;

[0012] S4, performing similarity calculation on the fused difference representation, measuring it using a weighted combination function based on Euclidean distance, cosine similarity, and Mahalanobis distance, and outputting a similarity score;

[0013] S5. Compare the similarity score with a threshold function and output a corresponding dustproof and breathable performance status label, wherein the threshold function is constructed based on the signal pair distribution characteristics and updated during the training phase;

[0014] S6. Associate the dustproof and breathable performance status label with the timestamp, perform sliding window cluster analysis, identify stable state segments and changing state segments, and use the identification results as network feedback for parameter update.

[0015] Optionally, the signal data includes air flow rate, micro-pressure difference change, particulate matter concentration change, time series response value and ambient temperature and humidity information.

[0016] Optionally, the preprocessing includes data cleaning, time series segmentation, wavelet transformation and Z-score normalization.

[0017] Optionally, the cross-scale alignment module pairs and fuses feature maps of different scales based on an attention mechanism and an inter-channel correlation weighting method, and outputs a fused difference representation.

[0018] Optionally, the dustproof and breathable performance status label is generated by a learnable threshold function trained based on the similarity score distribution. The threshold function is jointly optimized with the Siamese network through the backpropagation mechanism during the training phase to divide the similarity score into corresponding "normal state", "mildly degraded state" or "severely degraded state" intervals.

[0019] Optionally, the S2 specifically includes:

[0020] S21. Construct a WaveNet network based on gated convolution, which consists of N sequentially stacked residual blocks. Each residual block contains a pair of parallel causal dilated convolution channels and increases the dilation rate r with an exponential growth. n =2 n Construct, where n is the layer index, satisfying n∈{0,1,...,N-1};

[0021] S22. Based on different receptive field configurations, align the time dimension of each signal pair in the standardized signal pair set, unify the sequence length, and construct an embedding vector X;

[0022] S23. Input the embedding vector X into the WaveNet network, pass it through all residual blocks in turn, perform multi-scale feature extraction, and calculate the gated convolution features in each residual block:

[0023]

[0024] Among them, Z (n) represents the gated convolution feature of the n-th layer residual block, X (n) represents the embedding vector of the residual block of the nth layer, are the filter convolution kernel and the gated convolution kernel, respectively. tanh(·) is the hyperbolic tangent function, σ is the Sigmoid function, * represents the one-dimensional causal dilated convolution operation, and ⊙ is the element-by-element multiplication operation.

[0025] S24. Concatenate the gated convolution features of all residual blocks in the channel dimension to obtain a multi-scale feature map sequence F.

[0026] Optionally, the S3 specifically includes:

[0027] S31. Decompose the multi-scale feature map sequence F into two sub-map sequences F a and F b , and are respectively input into two feature extraction channels with shared parameters in the Siamese network, each of which contains A convolutional layers and B fully connected layers, where each convolutional layer uses a convolution kernel of size k for feature extraction;

[0028] S32. In the Siamese network, high-dimensional feature mapping is performed by parameter sharing and layer-by-layer transmission to the fully connected layer, respectively. a and F b Perform feature extraction to generate the feature vector Z of each input signal a and Z b ;

[0029] S33, through the cross-scale alignment module, the feature vector Z a and Z b Perform pairwise fusion to obtain fusion difference representation:

[0030]

[0031] Among them, Z a [i],Z b [i] represents the eigenvector Z a and Zb The i-th element of a [i],σ b [i] is the standard deviation corresponding to the i-th element, α j and β j is the attention weight in the cross-scale alignment module, represents the inter-channel correlation of the j-th scale feature, ⊙ represents element-wise multiplication, D represents the fusion difference representation, C ′ is the dimension of the feature vector, and L represents the sequence length.

[0032] Optionally, the S4 specifically includes:

[0033] S41. Calculate the similarity of the fused difference representation D, using a weighted combination function based on Euclidean distance, cosine similarity, and Mahalanobis distance for measurement. The weighted combination function is:

[0034]

[0035] Among them, S represents the similarity score, w1, w2, w3 represent the weighting coefficients, and D eu represents the Euclidean distance, D cos represents the cosine similarity, D ma represents the Mahalanobis distance, γ1, γ2, and γ3 are hyperparameters that adjust the influence of Euclidean distance, cosine similarity, and Mahalanobis distance, respectively, and exp(·) represents the natural exponential function with e as the base;

[0036] S42. Calculate the Euclidean distance D eu , specifically:

[0037]

[0038] Among them, D a [i],D b [i] are fused difference representation D a and D b The i-th element of C ′ is the dimension of the feature vector, where D a ,D b ∈D;

[0039] S43, calculate cosine similarity D cos , specifically:

[0040]

[0041] Among them, D a [i],D b [i] is the fusion difference representation D a and D b The i-th element of ;

[0042] S44. Calculate Mahalanobis distance D ma , specifically:

[0043]

[0044] Among them, Σ is the fusion difference representation D a and D b The covariance matrix of (D a -D b ) is the difference vector of the fused difference representation, Σ -1 represents the inverse of the covariance matrix.

[0045] Optionally, the S5 specifically includes:

[0046] S51, comparing the similarity score S with a learnable threshold function constructed during the training phase, wherein the threshold function is dynamically updated according to the distribution characteristics of the signal pair and is jointly optimized with the Siamese network through a backpropagation mechanism during the training phase;

[0047] S52 , based on the comparison result between the similarity score S and the threshold function, the similarity score is divided into three intervals: “normal state”, “mildly degraded state” or “severely degraded state”, and the corresponding dustproof and breathable performance state label is output.

[0048] Optionally, the S6 specifically includes:

[0049] S61. Associating the dustproof and breathable performance status labels with corresponding timestamps to form a time series dataset.

[0050] S62. Performing a sliding window cluster analysis on the time series data set, where the length and step size of the sliding window are set according to the distribution of historical data to divide the data into stable state segments and changing state segments;

[0051] S63. Based on the cluster analysis results, stable state segments and changing state segments are identified, and these identification results are input into the WaveNet network and the Siamese network as feedback to update the network parameters, thereby optimizing the network's ability to identify the dustproof and breathable performance status.

[0052] The beneficial effects of the present invention are:

[0053] First, the present invention innovatively proposes a dynamic identification method for the dust-proof and breathable performance of smoke sensor chips. Compared with traditional static monitoring methods, the multi-scale feature extraction capability based on the WaveNet network can effectively capture subtle changes in smoke sensor chips under different environmental conditions, especially the impact of factors such as dust and humidity on chip performance, greatly improving the real-time and accuracy of monitoring.

[0054] Secondly, this invention uses a Siamese network for similarity calculation and dynamic feedback. By fusing feature maps at different scales, it can accurately determine changes in the smoke sensor chip's air permeability during operation, thereby promptly identifying the degree of dust barrier blockage. This method enables continuous monitoring and dynamic updating of smoke sensor chip performance, eliminating the periodic inspections and manual maintenance required by traditional methods. This improves maintenance efficiency and reduces labor costs. It also identifies chip performance degradation in real time, providing early warnings and ensuring the system takes timely action when potential problems arise.

[0055] Finally, this invention boasts strong adaptability and robustness. Through network optimization and iterative parameter updates, the system can self-adjust to varying environmental conditions, enabling it to cope with changes in climate, pollution, dust, and other factors. During dynamic monitoring, the system not only accurately determines the chip's operating status but also automatically updates the threshold function based on changing environmental information. This ensures stable and efficient operation in a variety of environments, significantly improving the reliability and safety of the smoke sensor chip. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0057] Figure 1 This is a flowchart of the method for dynamically identifying the dust-proof and air-permeability performance of a smoke sensor chip based on deep learning proposed in the present invention;

[0058] Figure 2 This is a schematic diagram of the Siamese network structure of the deep learning-based dynamic identification method for smoke sensor chip dustproof and breathable performance proposed in the present invention;

[0059] Figure 3 This is a schematic diagram of the similarity calculation of the dynamic identification method of the dustproof and breathable performance of the smoke sensor chip based on deep learning proposed in the present invention. DETAILED DESCRIPTION

[0060] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0061] refer to Figure 1-3 The dynamic identification method of the dustproof and breathable performance of the smoke sensor chip based on deep learning includes the following steps:

[0062] S1. Collect signal data from the smoke sensor chip under different environmental conditions, perform preprocessing, and construct a standardized signal pair set;

[0063] S2. Construct a WaveNet network based on gated convolution, use different receptive field configurations to perform multi-scale feature extraction on the standardized signal pair set, and generate a multi-scale feature map sequence;

[0064] S3. Input the multi-scale feature map sequence into a Siamese network to generate a fused difference representation. The Siamese network includes two parameter-sharing feature extraction channels and a cross-scale alignment module;

[0065] S4, performing similarity calculation on the fused difference representation, measuring it using a weighted combination function based on Euclidean distance, cosine similarity, and Mahalanobis distance, and outputting a similarity score;

[0066] S5. Compare the similarity score with a threshold function and output a corresponding dustproof and breathable performance status label, wherein the threshold function is constructed based on the signal pair distribution characteristics and updated during the training phase;

[0067] S6. Associate the dustproof and breathable performance status label with the timestamp, perform sliding window cluster analysis, identify stable state segments and changing state segments, and use the identification results as network feedback for parameter update.

[0068] This invention provides a dynamic identification method for the dustproof and breathable performance of smoke sensor chips. This method collects real-time signal data from smoke sensor chips under different environmental conditions and accurately determines the chip's dustproof and breathable performance through multi-scale feature extraction and similarity calculation. Compared to traditional static monitoring methods, this method can dynamically reflect changes in chip performance, thereby improving the operational stability and safety of smoke sensor chips in complex environments.

[0069] In this embodiment, the signal data includes air flow rate, micro-pressure difference change, particle concentration change, time series response value and ambient temperature and humidity information.

[0070] This invention comprehensively captures all factors affecting the smoke sensor chip's dust and air permeability performance by collecting multiple signal data, including airflow rate, micro-pressure differential changes, particle concentration changes, time series response values, and ambient temperature and humidity information. This multi-dimensional data collection method more accurately reflects the dynamic changes in chip performance and improves the ability to identify dust and air permeability performance status.

[0071] In this embodiment, the preprocessing includes data cleaning, time series segmentation, wavelet transformation and Z-score normalization.

[0072] In the preprocessing step, the present invention uses a variety of techniques, including data cleaning, time series segmentation, wavelet transform, and Z-score normalization, to effectively preprocess the collected signal data. These processing methods can remove noise, eliminate irrelevant factors, and standardize the data, thereby improving the training effect and accuracy of the subsequent deep learning model.

[0073] In this embodiment, the cross-scale alignment module pairs and fuses feature maps of different scales based on the attention mechanism and the inter-channel correlation weighting method, and outputs a fused difference representation.

[0074] The cross-scale alignment module of this invention utilizes an attention mechanism and a weighted approach based on inter-channel correlation, effectively pairing and fusing feature maps at different scales. This method enhances the correlation between features, thereby better capturing subtle changes in the smoke sensor chip under different environments and improving recognition accuracy and robustness.

[0075] In this embodiment, the dustproof and breathable performance status label is generated by a learnable threshold function trained based on the similarity score distribution. The threshold function is jointly optimized with the Siamese network through the backpropagation mechanism during the training phase to divide the similarity score into corresponding "normal state", "mildly degraded state" or "severely degraded state" intervals.

[0076] This method dynamically adjusts the classification of dustproof and breathable performance status labels using a learnable threshold function trained based on similarity score distribution. This method, combined with backpropagation and Siamese network optimization, enables more precise classification of similarity scores into three intervals: "normal," "mildly degraded," and "severely degraded," thereby improving the system's accuracy and adaptability.

[0077] In this embodiment, S2 specifically includes:

[0078] S21. Construct a WaveNet network based on gated convolution, which consists of N sequentially stacked residual blocks. Each residual block contains a pair of parallel causal dilated convolution channels and increases the dilation rate r with an exponential growth. n =2 n Construct, where n is the layer index, satisfying n∈{0,1,...,N-1};

[0079] S22. Based on different receptive field configurations, align the time dimension of each signal pair in the standardized signal pair set, unify the sequence length, and construct an embedding vector X;

[0080] S23. Input the embedding vector X into the WaveNet network, pass it through all residual blocks in turn, perform multi-scale feature extraction, and calculate the gated convolution features in each residual block:

[0081]

[0082] Among them, Z (n) represents the gated convolution feature of the n-th layer residual block, X (n) represents the embedding vector of the residual block of the nth layer, are the filter convolution kernel and the gated convolution kernel, respectively. tanh(·) is the hyperbolic tangent function, σ is the Sigmoid function, * represents the one-dimensional causal dilated convolution operation, and ⊙ is the element-by-element multiplication operation.

[0083] S24. Concatenate the gated convolution features of all residual blocks in the channel dimension to obtain a multi-scale feature map sequence F.

[0084] This paper combines multiple residual blocks and dilated convolutions within a WaveNet network, employing a gated convolution-based feature extraction method for multi-scale feature extraction. This design effectively captures the complex characteristics of smoke sensor chips under varying environmental conditions, improving the ability to integrate multi-scale information and enhancing the model's expressiveness and generalization capabilities.

[0085] In this embodiment, S3 specifically includes:

[0086] S31. Decompose the multi-scale feature map sequence F into two sub-map sequences F a and F b , and are respectively input into two feature extraction channels with shared parameters in the Siamese network, each of which contains A convolutional layers and B fully connected layers, where each convolutional layer uses a convolution kernel of size k for feature extraction;

[0087] S32. In the Siamese network, high-dimensional feature mapping is performed by parameter sharing and layer-by-layer transmission to the fully connected layer, respectively. a and F b Perform feature extraction to generate the feature vector Z of each input signal a and Z b ;

[0088] S33, through the cross-scale alignment module, the feature vector Z a and Z b Perform pairwise fusion to obtain fusion difference representation:

[0089]

[0090] Among them, Z a [i],Zb [i] represents the eigenvector Z a and Z b The i-th element of a [i],σ b [i] is the standard deviation corresponding to the i-th element, α j and β j is the attention weight in the cross-scale alignment module, represents the inter-channel correlation of the j-th scale feature, ⊙ represents element-wise multiplication, D represents the fusion difference representation, C ′ is the dimension of the feature vector, and L represents the sequence length.

[0091] This method implements high-dimensional feature mapping and comparative analysis by inputting multi-scale feature map sequences into a Siamese network for parameter-sharing feature extraction channel processing. A cross-scale alignment module pairs and fuses feature vectors to generate a fused difference representation. This process enables the system to accurately identify and compare the status of smoke sensor chips in different environments, improving the accuracy of dynamic recognition.

[0092] In this embodiment, the S4 specifically includes:

[0093] S41. Calculate the similarity of the fused difference representation D, using a weighted combination function based on Euclidean distance, cosine similarity, and Mahalanobis distance for measurement. The weighted combination function is:

[0094]

[0095] Among them, S represents the similarity score, w1, w2, w3 represent the weighting coefficients, and D eu represents the Euclidean distance, D cos represents the cosine similarity, D ma represents the Mahalanobis distance, γ1, γ2, and γ3 are hyperparameters that adjust the influence of Euclidean distance, cosine similarity, and Mahalanobis distance, respectively, and exp(·) represents the natural exponential function with e as the base;

[0096] S42. Calculate the Euclidean distance D eu , specifically:

[0097]

[0098] Among them, D a [i],D b [i] are fused difference representation D a and D b The i-th element of C ′ is the dimension of the feature vector, where D a ,D b ∈D;

[0099] S43, calculate cosine similarity D cos , specifically:

[0100]

[0101] Among them, D a [i],D b [i] is the fusion difference representation D a and D b The i-th element of ;

[0102] S44. Calculate Mahalanobis distance D ma , specifically:

[0103]

[0104] Among them, Σ is the fusion difference representation D a and D b The covariance matrix of (D a -D b ) is the difference vector of the fused difference representation, Σ -1 represents the inverse of the covariance matrix.

[0105] The present invention uses a weighted combination of Euclidean distance, cosine similarity, and Mahalanobis distance in its similarity calculation process. This method comprehensively considers the differences between different features and adjusts the influence of each distance metric using a natural exponential function. This complex weighted combination makes similarity calculation more accurate and effectively improves the ability to identify the dustproof and breathable performance status of smoke sensor chips.

[0106] In this embodiment, the S5 specifically includes:

[0107] S51, comparing the similarity score S with a learnable threshold function constructed during the training phase, wherein the threshold function is dynamically updated according to the distribution characteristics of the signal pair and is jointly optimized with the Siamese network through a backpropagation mechanism during the training phase;

[0108] S52 , based on the comparison result between the similarity score S and the threshold function, the similarity score is divided into three intervals: “normal state”, “mildly degraded state” or “severely degraded state”, and the corresponding dustproof and breathable performance state label is output.

[0109] This paper uses a dynamic update method for the learnable threshold function constructed during the training phase to compare the similarity score with the threshold function. By combining backpropagation with Siamese network optimization, the threshold interval can be adaptively adjusted, making the state labels more consistent with the actual situation and further improving the robustness and reliability of the system.

[0110] In this embodiment, S6 specifically includes:

[0111] S61. Associating the dustproof and breathable performance status labels with corresponding timestamps to form a time series dataset.

[0112] S62. Performing a sliding window cluster analysis on the time series data set, where the length and step size of the sliding window are set according to the distribution of historical data to divide the data into stable state segments and changing state segments;

[0113] S63. Based on the cluster analysis results, stable state segments and changing state segments are identified, and these identification results are input into the WaveNet network and the Siamese network as feedback to update the network parameters, thereby optimizing the network's ability to identify the dustproof and breathable performance status.

[0114] By associating dustproof and breathable performance status tags with timestamps and performing sliding window cluster analysis, this method can identify stable and changing state segments and feed the identification results back to the network for parameter updates. This method not only continuously monitors the state changes of the smoke sensor chip but also enables self-optimization, improving the system's real-time responsiveness and long-term operational stability.

[0115] Example 1:

[0116] In order to verify the feasibility of the present invention in implementation, the present invention is applied to a dynamic identification system for the dustproof and breathable performance of smoke sensor chips in a complex environment. The system is deployed in an industrial plant. The air environment in the plant is complex. The smoke sensor chip will be affected by various factors such as dust, temperature changes, and humidity fluctuations during long-term operation. This may cause the dustproof layer of the smoke sensor chip to be blocked, thereby affecting its detection accuracy and response time. Traditional monitoring methods usually rely on manual regular inspections and simple performance tests. These methods cannot capture changes in chip performance in real time, nor can they make adaptive adjustments to environmental changes, resulting in frequent false alarms or missed alarms. Therefore, how to monitor and accurately evaluate the dustproof and breathable performance of smoke sensor chips in real time has become the key to improving the reliability and safety of smoke alarm systems.

[0117] In this scenario, the deep learning-based dynamic identification method for the dustproof and breathable performance of smoke sensor chips provided by the present invention can dynamically identify the dustproof and breathable performance status of the chip by collecting signal data such as airflow, micro-pressure difference changes, and particulate matter concentration changes of the smoke sensor chip in real time, using the WaveNet network for multi-scale feature extraction, and then using the Siamese network to achieve signal comparison and similarity calculation. Specifically, the present invention extracts multi-level features of signals such as airflow and micro-pressure difference through the gated convolution structure of the WaveNet network, and combines the Siamese network to measure the similarity of the signal pairs, and finally outputs three dustproof and breathable performance status labels of "normal state", "mild degradation state" or "severe degradation state". This method promptly identifies the degradation of the dustproof and breathable performance of the smoke sensor chip through real-time monitoring and dynamic feedback, and provides accurate maintenance guidance to system maintenance personnel, thereby avoiding the problems of manual inspection lag and misjudgment in traditional methods.

[0118] In the experiment, in order to verify the effectiveness of the present invention, three groups of data under different environmental conditions were selected for experiments. These environmental conditions include high temperature and high humidity, dusty working environments, and ordinary laboratory environments. Under each environment, the operating data of the smoke sensor chip was collected and preprocessed, including data cleaning, time series segmentation, wavelet transform and Z-score normalization. Then, the WaveNet network was used to perform multi-scale feature extraction on these signal data, and the Siamese network was used to generate a fusion difference representation, and finally the similarity score was calculated by the weighted combination function. Based on the threshold function constructed in the training phase, the similarity scores were divided into different dustproof and breathable performance state labels, and a sliding window cluster analysis was performed to identify stable state segments and changing state segments. By comparing with the results of manual inspection, the method of the present invention shows obvious advantages in recognition accuracy and real-time performance.

[0119] Experimental data demonstrates that the dynamic identification method of the present invention significantly improves the accuracy of monitoring the dustproof and breathable properties of smoke sensor chips. In high-temperature, high-humidity environments, the traditional method has an 18% false alarm rate, but using the present method, that rate drops to 2%. In dusty environments, the false alarm rate drops from 15% to 3%. In standard laboratory settings, the recognition accuracy increases from 75% to 98%. These experimental data demonstrate that the present invention's dynamic identification method for the dustproof and breathable properties of smoke sensor chips offers higher accuracy and a lower error rate, effectively resolving the problem of traditional methods' inability to monitor and adapt to environmental changes in real time.

[0120] To better demonstrate the experimental data, the following table compares the identification results of the smoke sensor chip's dustproof and breathable performance status using the traditional method and the method of the present invention under different environmental conditions:

[0121] Table 1 Comparison results between the method of the present invention and the traditional method

[0122]

[0123] As can be seen from the table, the method of the present invention has significant advantages over traditional methods under various environmental conditions. In particular, in environments with high temperature, high humidity, and high dust content, the false alarm and missed alarm rates are significantly reduced, and the recognition accuracy rate is significantly improved. These data verify the effectiveness and feasibility of the method of the present invention, demonstrating that it can address the shortcomings of traditional monitoring methods in complex environments and provide an efficient, real-time, and accurate solution for dynamically identifying the dustproof and breathable properties of smoke sensor chips.

[0124] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A dynamic identification method for the dustproof and breathable performance of smoke sensor chips based on deep learning, characterized by: The steps include: S1. Collect signal data from the smoke sensor chip under different environmental conditions, perform preprocessing, and construct a standardized signal pair set; S2. Construct a WaveNet network based on gated convolution, use different receptive field configurations to perform multi-scale feature extraction on the standardized signal pair set, and generate a multi-scale feature map sequence; S3. Input the multi-scale feature map sequence into a Siamese network to generate a fused difference representation. The Siamese network includes two parameter-sharing feature extraction channels and a cross-scale alignment module; S4, performing similarity calculation on the fused difference representation, measuring it using a weighted combination function based on Euclidean distance, cosine similarity, and Mahalanobis distance, and outputting a similarity score; S5. Compare the similarity score with a threshold function and output a corresponding dustproof and breathable performance status label, wherein the threshold function is constructed based on the signal pair distribution characteristics and updated during the training phase; S6. Associate the dustproof and breathable performance status label with the timestamp, perform sliding window cluster analysis, identify stable state segments and changing state segments, and use the identification results as network feedback for parameter update.

2. The method for dynamic identification of dustproof and air permeability performance of smoke sensor chips based on deep learning according to claim 1 is characterized in that: The signal data includes air flow rate, micro-pressure difference change, particle concentration change, time series response value and ambient temperature and humidity information.

3. The method for dynamic identification of dustproof and air permeability performance of smoke sensor chips based on deep learning according to claim 1 is characterized in that: The preprocessing includes data cleaning, time series segmentation, wavelet transformation and Z-score normalization.

4. The method for dynamic identification of dustproof and air permeability performance of smoke sensor chips based on deep learning according to claim 1 is characterized in that: The cross-scale alignment module pairs and fuses feature maps of different scales based on the attention mechanism and inter-channel correlation weighting method, and outputs a fused difference representation.

5. The method for dynamic identification of dustproof and air permeability performance of smoke sensor chips based on deep learning according to claim 1 is characterized in that: The dustproof and breathable performance status label is generated by a learnable threshold function trained based on the similarity score distribution. The threshold function is jointly optimized with the Siamese network through the backpropagation mechanism during the training phase to divide the similarity score into corresponding "normal state", "mildly degraded state" or "severely degraded state" intervals.

6. The method for dynamic identification of dustproof and air permeability performance of smoke sensor chips based on deep learning according to claim 1 is characterized in that: The S2 specifically includes: S21. Construct a WaveNet network based on gated convolution, which consists of N sequentially stacked residual blocks. Each residual block contains a pair of parallel causal dilated convolution channels and increases the dilation rate r with an exponential growth. n =2 n Construct, where n is the layer index, satisfying n∈{0,1,...,N-1}; S22. Based on different receptive field configurations, align the time dimension of each signal pair in the standardized signal pair set, unify the sequence length, and construct an embedding vector X; S23. Input the embedding vector X into the WaveNet network, pass it through all residual blocks in turn, perform multi-scale feature extraction, and calculate the gated convolution features in each residual block: Among them, Z (n) represents the gated convolution feature of the n-th layer residual block, X (n) represents the embedding vector of the residual block of the nth layer, are the filter convolution kernel and the gated convolution kernel, respectively. tanh(·) is the hyperbolic tangent function, σ is the Sigmoid function, * represents the one-dimensional causal dilated convolution operation, and ⊙ is the element-by-element multiplication operation. S24. Concatenate the gated convolution features of all residual blocks in the channel dimension to obtain a multi-scale feature map sequence F.

7. The method for dynamic identification of dustproof and air permeability performance of smoke sensor chips based on deep learning according to claim 1 is characterized in that: The S3 specifically includes: S31. Decompose the multi-scale feature map sequence F into two sub-map sequences F a and F b , and are respectively input into two feature extraction channels with shared parameters in the Siamese network, each of which contains A convolutional layers and B fully connected layers, where each convolutional layer uses a convolution kernel of size k for feature extraction; S32. In the Siamese network, high-dimensional feature mapping is performed by parameter sharing and layer-by-layer transmission to the fully connected layer, respectively. a and F b Perform feature extraction to generate the feature vector Z of each input signal a and Z b ; S33, through the cross-scale alignment module, the feature vector Z a and Z b Perform pairwise fusion to obtain fusion difference representation: Among them, Z a [i],Z b [i] represents the eigenvector Z a and Z b The i-th element of a [i],σ b [i] is the standard deviation corresponding to the i-th element, α j and β j is the attention weight in the cross-scale alignment module, represents the inter-channel correlation of the j-th scale feature, ⊙ represents element-wise multiplication, D represents the fused difference representation, C′ is the dimension of the feature vector, and L represents the sequence length.

8. The method for dynamic identification of dustproof and air permeability performance of smoke sensor chips based on deep learning according to claim 1 is characterized in that: The S4 specifically includes: S41. Calculate the similarity of the fused difference representation D, using a weighted combination function based on Euclidean distance, cosine similarity, and Mahalanobis distance for measurement. The weighted combination function is: Among them, S represents the similarity score, w1, w2, w3 represent the weighting coefficients, and D eu represents the Euclidean distance, D cos represents the cosine similarity, D ma represents the Mahalanobis distance, γ1, γ2, and γ3 are hyperparameters that adjust the influence of Euclidean distance, cosine similarity, and Mahalanobis distance, respectively, and exp(·) represents the natural exponential function with e as the base; S42. Calculate the Euclidean distance D eu , specifically: Among them, D a [i],D b [i] are fused difference representation D a and D b The i-th element of , C′ is the dimension of the feature vector, where D a ,D b ∈D; S43, calculate cosine similarity D cos , specifically: Among them, D a [i],D b [i] is the fusion difference representation D a and D b The i-th element of ; S44. Calculate Mahalanobis distance D ma , specifically: Among them, Σ is the fusion difference representation D a and D b The covariance matrix of (D a -D b ) is the difference vector of the fused difference representation, Σ -1 represents the inverse of the covariance matrix.

9. The method for dynamic identification of dustproof and air permeability performance of smoke sensor chips based on deep learning according to claim 1, characterized in that: The S5 specifically includes: S51, comparing the similarity score S with a learnable threshold function constructed during the training phase, wherein the threshold function is dynamically updated according to the distribution characteristics of the signal pair and is jointly optimized with the Siamese network through a backpropagation mechanism during the training phase; S52. Based on the comparison result between the similarity score S and the threshold function, the similarity score is divided into three intervals: "normal state", "mildly degraded state" or "severely degraded state", and the corresponding dustproof and breathable performance state label is output.

10. The method for dynamic identification of dustproof and air permeability performance of smoke sensor chips based on deep learning according to claim 1, characterized in that: The S6 specifically includes: S61. Associating the dustproof and breathable performance status labels with corresponding timestamps to form a time series dataset. S62. Performing a sliding window cluster analysis on the time series data set, where the length and step size of the sliding window are set according to the distribution of historical data to divide the data into stable state segments and changing state segments; S63. Based on the cluster analysis results, stable state segments and changing state segments are identified, and these identification results are input into the WaveNet network and the Siamese network as feedback to update the network parameters, thereby optimizing the network's ability to identify the dustproof and breathable performance status.

Images