WO2024021253A1 - 一种用于六氟丁二烯制备的智能化有毒有害气体报警系统 - Google Patents

一种用于六氟丁二烯制备的智能化有毒有害气体报警系统 Download PDF

Info

Publication number
WO2024021253A1
WO2024021253A1 PCT/CN2022/119276 CN2022119276W WO2024021253A1 WO 2024021253 A1 WO2024021253 A1 WO 2024021253A1 CN 2022119276 W CN2022119276 W CN 2022119276W WO 2024021253 A1 WO2024021253 A1 WO 2024021253A1
Authority
WO
WIPO (PCT)
Prior art keywords
matrix
topological
feature matrix
toxic
feature
Prior art date
Application number
PCT/CN2022/119276
Other languages
English (en)
French (fr)
Inventor
张前臻
朱军伟
傅晓腾
张鸿铨
张奎
Original Assignee
福建省杭氟电子材料有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 福建省杭氟电子材料有限公司 filed Critical 福建省杭氟电子材料有限公司
Publication of WO2024021253A1 publication Critical patent/WO2024021253A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G08—SIGNALLING
    • G08B—SIGNALLING OR CALLING SYSTEMS; ORDER TELEGRAPHS; ALARM SYSTEMS
    • G08B21/00—Alarms responsive to a single specified undesired or abnormal condition and not otherwise provided for
    • G08B21/02—Alarms for ensuring the safety of persons
    • G08B21/12—Alarms for ensuring the safety of persons responsive to undesired emission of substances, e.g. pollution alarms
    • G08B21/14—Toxic gas alarms
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A50/00—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE in human health protection, e.g. against extreme weather
    • Y02A50/20—Air quality improvement or preservation, e.g. vehicle emission control or emission reduction by using catalytic converters
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/02—Total factory control, e.g. smart factories, flexible manufacturing systems [FMS] or integrated manufacturing systems [IMS]

Definitions

  • the present invention relates to the field of intelligent gas monitoring, and more specifically, to an intelligent toxic and harmful gas alarm system for the preparation of hexafluorobutadiene and an alarm method thereof.
  • fluorine-containing electronic gases account for about 30% and are mainly used as etchants and cleaning agents.
  • PFCs perfluoroalkane
  • Hexafluorobutadiene has become one of the best substitutes for traditional fluorine-containing electron gases due to its excellent performance in all aspects. It is a monomer for preparing a variety of fluoropolymer materials and is also a green, environmentally friendly and efficient dry etching gas. , has attracted great attention from domestic and foreign scholars in recent years.
  • hexafluorobutadiene is a flammable, toxic, colorless and odorless gas. After it is mixed with air, when the concentration reaches 7%, there is an immediate danger of combustion and explosion. Moreover, when inhaled into the body, it can cause harm to the human body and may cause respiratory irritation, coughing, dizziness, anesthesia, irregular heartbeat and negative kidney effects.
  • Embodiments of the present application provide an intelligent toxic and harmful gas alarm system for the preparation of hexafluorobutadiene and an alarm method thereof, by deploying multiple toxic and harmful gas monitors in the preparation site of hexafluorobutadiene To collect gas concentration values at multiple locations at multiple time points, and use a deep neural network model to extract implicit dynamic correlation features for multiple gas concentration values. At the same time, topological features are also used to perform feature extraction. spatial domain mapping to take into account more feature information during classification, thereby improving the classification effect. In this way, toxic and harmful gases in the preparation site can be accurately monitored to ensure the safety of personnel in the preparation site.
  • an intelligent toxic and harmful gas alarm system for the preparation of hexafluorobutadiene which includes: a gas monitoring data acquisition module, used to obtain the gas monitoring system deployed in hexafluorobutadiene in a predetermined topology pattern. Gas concentration values at multiple predetermined time points collected by multiple toxic and harmful gas monitors in the diene preparation site; a single sample gas data encoding module used to combine multiple predetermined time points collected by each of the toxic and harmful gas monitors.
  • the gas concentration value of the point is passed through a time series encoder containing a one-dimensional convolution layer to obtain the time series feature vector of the measurement data corresponding to each of the toxic and harmful gas monitors; a multi-sample gas data correlation encoding module is used to convert the corresponding to The measurement data time series feature vectors of each toxic and harmful gas monitor are arranged into a two-dimensional feature matrix and then passed through the first convolutional neural network as a filter to obtain the measurement data correlation feature matrix; the sensor topology matrix construction unit is used to obtain all The topological matrix of multiple toxic and harmful gas monitors, wherein the characteristic value of each position on the off-diagonal position in the topological matrix is the distance between the corresponding two toxic and harmful gas monitors, and the pair of pairs of toxic and harmful gas monitors in the topological matrix The eigenvalues of each position on the corner position are zero; the topological matrix encoding module is used to pass the topological matrix through the second convolutional neural network as a filter to obtain the topological feature matrix; the topological
  • the high-dimensional topological information is mapped to the high-dimensional feature space of the measurement data associated feature matrix to obtain a classification feature matrix; and an alarm result generation module is used to pass the classification feature matrix through a classifier to obtain a classification result, the The classification result is used to indicate whether an alarm prompt is generated.
  • the single sample gas data encoding module includes: an input vector construction unit for converting the multiple data collected by each of the toxic and harmful gas monitors.
  • the gas concentration values at a predetermined time point are arranged according to the time dimension into a one-dimensional input vector corresponding to each of the toxic and hazardous gas monitors; a fully connected encoding unit is used to use the fully connected layer of the timing encoder according to the following formula Fully connected encoding is performed on the input vector to extract high-dimensional hidden features of the eigenvalues at each position in the input vector, where the formula is: where X is the input vector, Y is the output vector, W is the weight matrix, and B is the bias vector, Represents matrix multiplication; a one-dimensional convolution coding unit, used to use the one-dimensional convolution layer of the temporal encoder to perform one-dimensional convolution coding on the input vector according to the following formula to extract the position of each position in the input vector High-dimensional implicit correlation features between eigenvalues, where the formula is:
  • a is the width of the convolution kernel in the x direction
  • F is the convolution kernel parameter vector
  • G is the local vector matrix that operates with the convolution kernel function
  • w is the size of the convolution kernel.
  • the multi-sample gas data correlation coding module is further used to: each layer of the first convolutional neural network is in the forward direction of the layer.
  • the input data are separately: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension on the convolution feature map to obtain a pooled feature map; and, perform a pooling feature map on the convolution feature map.
  • the pooled feature map performs nonlinear activation to obtain an activation feature map; wherein, the output of the last layer of the first convolutional neural network is the measurement data associated feature matrix, and the first layer of the first convolutional neural network The input to the layer is the two-dimensional feature matrix.
  • the topological matrix encoding module is further used: each layer of the second convolutional neural network performs The input data is subjected to convolution processing, mean pooling processing along the channel dimension and activation processing to generate the topological feature matrix from the last layer of the second convolutional neural network, wherein The input to the first layer is the topology matrix.
  • the topological feature correction module includes: an exponential operation unit for calculating the eigenvalues of each position in the topological feature matrix as powers.
  • the natural exponential function value of The exponential function value is subtracted from the reciprocal of the eigenvalue of the position in the topological feature matrix and then minus one to obtain the constraint value corresponding to the eigenvalue of each position in the topological feature matrix; and, a structured understanding unit, used to calculate the The logarithmic function value of the absolute value of the constraint value corresponding to the eigenvalue at each position in the topological feature matrix is calculated to obtain the corrected topological feature matrix.
  • the fusion module is further used to: matrix-matrix the measured data correlation characteristic matrix and the corrected topological characteristic matrix according to the following formula Multiply to map the high-dimensional topological information of the corrected topological feature matrix into the high-dimensional feature space of the measured data associated feature matrix to obtain the classification feature matrix; wherein, the formula is:
  • M represents the classification feature matrix
  • M 1 represents the measurement data associated feature matrix
  • M 2 represents the corrected topological feature matrix
  • the alarm result generation module is further used: the classifier processes the classification feature matrix according to the following formula to generate a classification result , where the formula is: softmax ⁇ (W n ,B n ):...:(W 1 ,B 1 )
  • an alarm method for an intelligent toxic and hazardous gas alarm system for the preparation of hexafluorobutadiene includes:
  • the gas concentration value at the predetermined time point is passed through a time series encoder including a one-dimensional convolution layer to obtain the measurement data time series feature vector corresponding to each of the toxic and harmful gas monitors;
  • the time series feature vectors of the measurement data are arranged into a two-dimensional feature matrix and then passed through the first convolutional neural network as a filter to obtain the measurement data correlation feature matrix;
  • the topological matrix of the multiple toxic and harmful gas monitors is obtained, wherein, the The eigenvalues of each position on the non-diagonal position in the topology matrix are the distances between the corresponding two toxic and harmful gas monitors, and the eigenvalues of each position on the diagonal position in the topology matrix are zero;
  • the topology The matrix is passed through the second convolutional neural network as a filter to obtain a topological feature matrix; the feature distribution correction is performed
  • the classification feature matrix is passed through a classifier to obtain a classification result, and the classification result is used to indicate whether an alarm prompt is generated.
  • the gas concentration values at multiple predetermined time points collected by each of the toxic and harmful gas monitors are passed through a one-dimensional convolution layer.
  • the time series encoder is used to obtain the measurement data time series feature vector corresponding to each of the toxic and harmful gas monitors, including: arranging the gas concentration values at multiple predetermined time points collected by each of the toxic and harmful gas monitors according to the time dimension as Corresponding to the one-dimensional input vector of each toxic and hazardous gas monitor; use the fully connected layer of the temporal encoder to fully connect the input vector with the following formula to extract the position of each position in the input vector High-dimensional latent features of eigenvalues, where the formula is: where X is the input vector, Y is the output vector, W is the weight matrix, and B is the bias vector, represents matrix multiplication; use the one-dimensional convolution layer of the temporal encoder to perform one-dimensional convolution encoding on the input vector with the following
  • a is the width of the convolution kernel in the x direction
  • F is the convolution kernel parameter vector
  • G is the local vector matrix that operates with the convolution kernel function
  • w is the size of the convolution kernel.
  • the time series feature vectors of the measurement data corresponding to each of the toxic and harmful gas monitors are arranged into a two-dimensional feature matrix and then passed
  • the first convolutional neural network as a filter to obtain the measurement data correlation feature matrix includes: each layer of the first convolutional neural network performs on the input data in the forward pass of the layer: convolution on the input data Processing to obtain a convolution feature map; performing mean pooling based on local channel dimensions on the convolution feature map to obtain a pooled feature map; and performing nonlinear activation on the pooled feature map to obtain an activation feature map;
  • the output of the last layer of the first convolutional neural network is the measurement data associated feature matrix
  • the input of the first layer of the first convolutional neural network is the two-dimensional feature matrix.
  • each layer of the second convolutional neural network performs convolution processing on the input data in the forward transmission of the layer, along the Mean pooling processing and activation processing of the channel dimension are used to generate the topological feature matrix by the last layer of the second convolutional neural network, wherein the input of the first layer of the second convolutional neural network is the topological matrix.
  • performing characteristic distribution correction on the topological characteristic matrix to obtain the corrected topological characteristic matrix includes: calculating the topological characteristic matrix The natural exponential function value whose eigenvalue at each position in the topological feature matrix is the power of the natural exponential function value; calculate the reciprocal of the eigenvalue of each position in the topological feature matrix; subtract the natural exponential function value whose power is the eigenvalue of each position in the topological feature matrix.
  • the reciprocal of the eigenvalue of the position in the topological feature matrix is subtracted by one to obtain the constraint value corresponding to the eigenvalue of each position in the topological feature matrix; and, calculating the constraint value corresponding to the eigenvalue of each position in the topological feature matrix.
  • the logarithmic function value of the absolute value of the constraint value is used to obtain the corrected topological feature matrix.
  • the measured data correlation characteristic matrix and the corrected topology characteristic matrix are matrix multiplied to obtain the corrected topology
  • Mapping the high-dimensional topological information of the feature matrix into the high-dimensional feature space of the measured data associated feature matrix to obtain the classification feature matrix includes: performing the following formula on the measured data associated feature matrix and the corrected topological feature matrix Matrix multiplication is used to map the high-dimensional topological information of the corrected topological feature matrix into the high-dimensional feature space of the measured data associated feature matrix to obtain the classification feature matrix; wherein, the formula is:
  • M represents the classification feature matrix
  • M 1 represents the measurement data associated feature matrix
  • M 2 represents the corrected topological feature matrix
  • the classification feature matrix is passed through a classifier to obtain a classification result, including: the classifier classifies the classification according to the following formula
  • the feature matrix is processed to generate a classification result, where the formula is: softmax ⁇ (W n ,B n ):...:(W 1 ,B 1 )
  • the classification feature matrix is projected as a vector, W 1 to W n are the weight matrices of the fully connected layers of each layer, and B 1 to B n represent the bias matrices of the fully connected layers of each layer.
  • the intelligent toxic and harmful gas alarm system and alarm method for the preparation of hexafluorobutadiene deploy multiple toxic and harmful gases in the preparation site of hexafluorobutadiene.
  • Monitors are used to collect gas concentration values at multiple locations at multiple time points, and a deep neural network model is used to extract implicit dynamic correlation features for multiple gas concentration values.
  • topological features are also used to Carry out spatial domain mapping of features to take into account more feature information during classification, thereby improving the classification effect. In this way, toxic and harmful gases in the preparation site can be accurately monitored to ensure the safety of personnel in the preparation site.
  • Figure 1 is an application scenario diagram of an intelligent toxic and hazardous gas alarm system for the preparation of hexafluorobutadiene according to an embodiment of the present application.
  • Figure 2 is a block diagram of an intelligent toxic and harmful gas alarm system for the preparation of hexafluorobutadiene according to an embodiment of the present application.
  • Figure 3 is a block diagram of a topological feature correction module in an intelligent toxic and hazardous gas alarm system for hexafluorobutadiene preparation according to an embodiment of the present application.
  • Figure 4 is a flow chart of an alarm method of an intelligent toxic and harmful gas alarm system for hexafluorobutadiene preparation according to an embodiment of the present application.
  • Figure 5 is a schematic structural diagram of an alarm method of an intelligent toxic and harmful gas alarm system for hexafluorobutadiene preparation according to an embodiment of the present application.
  • fluorine-containing electronic gases account for about 30% of the global electronic gas market and are mainly used as etchants and cleaning agents.
  • PFCs perfluoroalkane
  • Hexafluorobutadiene has become one of the best substitutes for traditional fluorine-containing electron gases due to its excellent performance in all aspects. It is a monomer for preparing a variety of fluoropolymer materials and is also a green, environmentally friendly and efficient dry etching gas. , has attracted great attention from domestic and foreign scholars in recent years.
  • hexafluorobutadiene is a flammable, toxic, colorless and odorless gas. After it is mixed with air, when the concentration reaches 7%, there is an immediate danger of combustion and explosion. Moreover, when inhaled into the body, it can cause harm to the human body and may cause respiratory irritation, coughing, dizziness, anesthesia, irregular heartbeat and negative kidney effects.
  • the inventor of the present application found that when using gas sensors to monitor toxic and harmful gases (mainly hexafluorobutadiene gas), the distribution of toxic and harmful gases at various locations in the place to be monitored is uneven. Therefore, if If a separate gas sensor is used for gas monitoring, the gas concentration value it may detect does not exceed the safety threshold, but the gas concentration elsewhere in the site may have exceeded the predetermined threshold. Secondly, a single sensor may also cause errors in measurement due to failure. Therefore, there are potential safety hazards in using a single gas sensor to monitor gas concentration. Moreover, any gas sensor has its own system error. Therefore, even if a single sensor itself does not have a fault, it is unreasonable to rely on the data of a single sensor as a basis for monitoring.
  • toxic and harmful gases mainly hexafluorobutadiene gas
  • multiple toxic and harmful gas monitors are deployed in a hexafluorobutadiene preparation site in a predetermined topological pattern to collect gas concentration values at multiple predetermined time points. It should be understood that considering that the deployment space areas of toxic and harmful gas monitors in the hexafluorobutadiene preparation site are connected, this will cause the gas concentration value to have a dynamic change pattern, that is, each location will have a dynamic change pattern. The gas concentration at the deployment point of the toxic and harmful gas monitor will diffuse according to changes in time, for example, from an area with higher concentration to an area with lower concentration.
  • the gas concentration values at multiple predetermined time points collected by each of the toxic and harmful gas monitors are further passed through a time series encoder containing a one-dimensional convolution layer. Processing is performed to obtain measurement data time series feature vectors corresponding to each of the toxic and harmful gas monitors.
  • the temporal encoder consists of alternately arranged fully connected layers and one-dimensional convolutional layers, which extract the correlation of the gas concentration values of each deployment point in the temporal dimension through one-dimensional convolutional encoding.
  • Features and high-dimensional latent features of the gas concentration values at each deployment point are extracted through fully connected encoding.
  • time series feature vectors of the measurement data corresponding to each of the toxic and harmful gas monitors are arranged into a two-dimensional feature matrix to integrate the dynamic change feature information of the gas concentration values at the deployment points of each of the toxic and harmful gas monitors. , and then process this change feature through the first convolutional neural network as a filter to extract the global dynamic features of the gas concentration values at multiple predetermined time points collected by each of the toxic and harmful gas monitors. expressed to obtain the measurement data correlation characteristic matrix.
  • the topological matrix of the multiple toxic and harmful gas monitors is obtained according to a predetermined topological pattern.
  • the non-diagonal positions in the topological matrix The eigenvalue of each position is the distance between the two corresponding toxic and harmful gas monitors, and the eigenvalue of each position on the diagonal position in the topological matrix is zero.
  • the topological matrix is passed through the second convolutional neural network as a filter to perform deep mining of topological features to obtain a topological feature matrix.
  • the measurement data associated feature matrix and the topological feature matrix are matrix multiplied to map the high-dimensional topological information of the topological feature matrix into the high-dimensional feature space of the measured data associated feature matrix, and then The feature information of the two is fused to perform classification to obtain a classification result indicating whether an alarm prompt is generated.
  • the topological feature matrix is multiplied by the measured data associated feature matrix for feature fusion, since the topological feature matrix expresses the topological features of the sensor location and does not contain the numerical features of the measured data, this results in the The obtained classification feature matrix has a weak constraint on the classification target of the time series numerical correlation features of the measurement data, so there may be a problem of poor classification effect.
  • class condition boundary constraints are first applied to the topological feature matrix:
  • m i,j is the eigenvalue of each position of the topological eigenmatrix.
  • the class condition boundary constraint performs boundary constraints on the features by performing a rule-based structured understanding of the feature values and the class conditions to which they belong, so as to avoid the feature value set from being in the classification target due to the out-of-distribution characteristics of the set. Excessive fragmentation of the decision-making area in the domain to obtain robust conditional class boundaries, thereby improving the constraint of the classification feature matrix on the classification target by improving the convergence within the class condition boundary of the topological feature matrix itself, and then improving the constraint of the classification feature matrix on the classification target. Improve the classification effect.
  • this application proposes an intelligent toxic and harmful gas alarm system for the preparation of hexafluorobutadiene, which includes: a gas monitoring data acquisition module, used to obtain information from hexafluorobutadiene deployed in a predetermined topology pattern.
  • the gas concentration value is passed through a time series encoder containing a one-dimensional convolution layer to obtain the time series feature vector of the measurement data corresponding to each of the toxic and hazardous gas monitors; a multi-sample gas data correlation encoding module is used to convert the time series feature vector corresponding to each of the toxic and hazardous gas monitors.
  • the time series feature vectors of the measurement data of the toxic and harmful gas monitor are arranged into a two-dimensional feature matrix and then passed through the first convolutional neural network as a filter to obtain the measurement data correlation feature matrix; the sensor topology matrix construction unit is used to obtain the multi-dimensional feature matrix.
  • a topological matrix of toxic and harmful gas monitors wherein the eigenvalues of each position on the off-diagonal position in the topological matrix are the distances between the corresponding two toxic and harmful gas monitors, and the diagonal lines in the topological matrix
  • the eigenvalue of each position on the position is zero;
  • a topological matrix encoding module is used to pass the topological matrix through the second convolutional neural network as a filter to obtain a topological feature matrix;
  • a topological feature correction module is used to correct the topological feature
  • the feature matrix performs feature distribution correction to obtain the corrected topological feature matrix;
  • the fusion module is used to perform matrix multiplication of the measured data associated feature matrix and the corrected topological feature matrix to obtain the high value of the corrected topological feature matrix.
  • the dimensional topological information is mapped to the high-dimensional feature space of the measurement data associated feature matrix to obtain a classification feature matrix; and, an alarm result generation module is used to pass the classification feature matrix through a classifier to obtain a classification result.
  • the result is used to indicate whether an alarm prompt is generated.
  • Figure 1 illustrates an application scenario diagram of an intelligent toxic and harmful gas alarm system for hexafluorobutadiene preparation according to an embodiment of the present application.
  • FIG 1 in this application scenario, first, multiple toxic and hazardous gas monitors (for example, as shown in Figure 1) deployed in the preparation site of hexafluorobutadiene (for example, H as shown in Figure 1)
  • the topology matrix is obtained with a predetermined topology pattern of T1-Tn) illustrated in Figure 1, and multiple predetermined times are collected by multiple toxic and harmful gas monitors deployed in the preparation site of hexafluorobutadiene in a predetermined topology pattern. point gas concentration value.
  • the obtained gas concentration values at multiple predetermined time points collected by the multiple toxic and hazardous gas monitors and the topological matrix of the multiple toxic and hazardous gas monitors are input to the equipment deployed for the preparation of hexafluorobutadiene.
  • a server with an intelligent toxic and harmful gas alarm algorithm for example, the server S as shown in Figure 1
  • the server can use the intelligent toxic and harmful gas alarm algorithm for the preparation of hexafluorobutadiene to
  • the gas concentration values at multiple predetermined time points collected by multiple toxic and hazardous gas monitors and the topology matrices of the multiple toxic and hazardous gas monitors are processed to generate a classification result indicating whether an alarm prompt is generated.
  • Figure 2 illustrates a block diagram of an intelligent toxic and harmful gas alarm system for hexafluorobutadiene preparation according to an embodiment of the present application.
  • the intelligent toxic and harmful gas alarm system 200 for hexafluorobutadiene preparation according to the embodiment of the present application includes: a gas monitoring data collection module 210, used to obtain the data collected from six gases deployed in a predetermined topology pattern. The gas concentration values at multiple predetermined time points collected by multiple toxic and hazardous gas monitors in the fluorobutadiene preparation site; the single sample gas data encoding module 220 is used to combine the multiple toxic and hazardous gas monitors collected by each of the toxic and hazardous gas monitors.
  • the gas concentration values at each predetermined time point are passed through a time series encoder including a one-dimensional convolution layer to obtain the measurement data time series feature vector corresponding to each of the toxic and harmful gas monitors;
  • the multi-sample gas data correlation encoding module 230 is used to The measured data time series feature vectors corresponding to each of the toxic and hazardous gas monitors are arranged into a two-dimensional feature matrix and then passed through the first convolutional neural network as a filter to obtain the measured data correlation feature matrix;
  • the sensor topology matrix construction unit 240 used to obtain the topological matrix of the multiple toxic and harmful gas monitors, wherein the characteristic value of each position on the off-diagonal position in the topological matrix is the distance between the corresponding two toxic and harmful gas monitors, so The eigenvalues of each position on the diagonal position in the topological matrix are zero;
  • the topological matrix encoding module 250 is used to pass the topological matrix through the second convolutional neural network as a filter to obtain the topological feature matrix;
  • the classification result is obtained through the classifier, and the classification result is used to indicate whether an alarm prompt is generated.
  • the gas monitoring data acquisition module 210 and the single sample gas data encoding module 220 are used to obtain multiple gases deployed in a hexafluorobutadiene preparation site in a predetermined topological pattern.
  • Gas concentration values at multiple predetermined time points collected by each toxic and hazardous gas monitor, and passing the gas concentration values at multiple predetermined time points collected by each toxic and hazardous gas monitor through a time series encoder including a one-dimensional convolution layer To obtain the measurement data time series feature vector corresponding to each of the toxic and harmful gas monitors.
  • multiple toxic and hazardous gas monitors are deployed in a hexafluorobutadiene preparation site in a predetermined topological pattern to collect gas concentration values at multiple predetermined time points. It should be understood that considering that the deployment space areas of toxic and harmful gas monitors in the hexafluorobutadiene preparation site are connected, this will cause the gas concentration value to have a dynamic change pattern, that is, each location will have a dynamic change pattern.
  • the gas concentration at the deployment point of the toxic and harmful gas monitor will diffuse according to changes in time, for example, from an area with higher concentration to an area with lower concentration.
  • the gas concentration values at multiple predetermined time points collected by each of the toxic and harmful gas monitors are further passed through a time series encoder containing a one-dimensional convolution layer. Processing is performed to obtain measurement data time series feature vectors corresponding to each of the toxic and harmful gas monitors.
  • the temporal encoder consists of alternately arranged fully connected layers and one-dimensional convolutional layers, which extract the correlation of the gas concentration values of each deployment point in the temporal dimension through one-dimensional convolutional encoding.
  • Features and high-dimensional latent features of the gas concentration values at each deployment point are extracted through fully connected encoding.
  • the single sample gas data encoding module includes: an input vector construction unit for converting the gas concentration values at multiple predetermined time points collected by each of the toxic and harmful gas monitors according to The time dimension is arranged as a one-dimensional input vector corresponding to each of the toxic and harmful gas monitors; a fully connected encoding unit is used to use the fully connected layer of the timing encoder to perform fully connected encoding on the input vector using the following formula To extract the high-dimensional implicit features of the eigenvalues of each position in the input vector, where the formula is: where X is the input vector, Y is the output vector, W is the weight matrix, and B is the bias vector, Represents matrix multiplication; a one-dimensional convolution coding unit, used to use the one-dimensional convolution layer of the temporal encoder to perform one-dimensional convolution coding on the input vector according to the following formula to extract the position of each position in the input vector High-dimensional implicit correlation features between eigenvalues, where the formula is:
  • a is the width of the convolution kernel in the x direction
  • F is the convolution kernel parameter vector
  • G is the local vector matrix that operates with the convolution kernel function
  • w is the size of the convolution kernel.
  • the multi-sample gas data correlation encoding module 230 is used to arrange the time series feature vectors of the measurement data corresponding to each of the toxic and harmful gas monitors into a two-dimensional feature matrix and then pass The first convolutional neural network is used as a filter to obtain the measurement data associated feature matrix. That is, in the technical solution of the present application, further, the measurement data time series feature vectors corresponding to each of the toxic and harmful gas monitors are arranged into a two-dimensional feature matrix to integrate each of the toxic and harmful gas monitors.
  • the dynamic change characteristic information of the gas concentration value at the deployment point is processed through the first convolutional neural network as a filter to extract multiple predetermined times collected by each toxic and harmful gas monitor.
  • the global dynamic characteristic representation of the gas concentration value of the point is used to obtain the measurement data correlation characteristic matrix.
  • the multi-sample gas data correlation coding module is further used to perform: each layer of the first convolutional neural network on the input data in the forward transmission of the layer:
  • the input data is subjected to convolution processing to obtain a convolution feature map;
  • the convolution feature map is subjected to mean pooling based on the local channel dimension to obtain a pooled feature map; and, the pooled feature map is nonlinearly activated to obtain Obtain activation feature map;
  • the output of the last layer of the first convolutional neural network is the measurement data associated feature matrix
  • the input of the first layer of the first convolutional neural network is the two-dimensional feature matrix.
  • the sensor topology matrix construction unit 240 and the topology matrix encoding module 250 are used to obtain the topology matrices of the multiple toxic and harmful gas monitors, wherein in the topology matrix
  • the eigenvalues of each position on the non-diagonal position are the distances between the corresponding two toxic and harmful gas monitors.
  • the eigenvalues of each position on the diagonal position in the topological matrix are zero, and the topological matrix is passed
  • the second convolutional neural network is used as a filter to obtain the topological feature matrix.
  • the topological matrix of the multiple toxic and harmful gas monitors is obtained according to a predetermined topological pattern, here , the eigenvalue of each position on the off-diagonal position in the topological matrix is the distance between the corresponding two toxic and harmful gas monitors, and the eigenvalue of each position on the diagonal position in the topological matrix is zero. Then, the topological matrix is passed through the second convolutional neural network as a filter to perform deep mining of topological features to obtain a topological feature matrix.
  • each layer of the second convolutional neural network performs convolution processing, mean pooling processing along the channel dimension and activation processing on the input data in the forward pass of the layer to be processed by the The last layer of the second convolutional neural network generates the topological feature matrix, wherein the input of the first layer of the second convolutional neural network is the topological matrix.
  • the topological feature correction module 260 is used to perform feature distribution correction on the topological feature matrix to obtain a corrected topological feature matrix. It should be understood that in the technical solution of the present application, in this way, the measurement data associated feature matrix and the topological feature matrix are matrix multiplied to map the high-dimensional topological information of the topological feature matrix to the The measurement data is associated with the high-dimensional feature space of the feature matrix, and then the feature information of the two is merged for classification, and the classification result indicating whether an alarm prompt is generated can be obtained.
  • the obtained classification feature matrix has a weak constraint on the classification target of the time-series numerical correlation features of the measurement data, and thus may have a problem of poor classification effect. Therefore, before performing feature fusion, it is necessary to The above-mentioned topological characteristic matrix performs class condition boundary constraints.
  • the topological feature correction module includes: first, calculating the natural exponential function value with the eigenvalue of each position in the topological feature matrix as the power. Next, the reciprocal of the eigenvalues of each position in the topological feature matrix is calculated. Then, subtract the reciprocal of the eigenvalue of the position in the topological feature matrix from the natural exponential function value raised to the power of the eigenvalue of each position in the topological feature matrix and then subtract one to obtain the value of each position in the topological feature matrix.
  • the constraint value corresponding to the eigenvalue.
  • the logarithmic function value of the absolute value of the constraint value corresponding to the eigenvalue at each position in the topological feature matrix is calculated to obtain the corrected topological feature matrix.
  • the formula for class condition boundary constraints on the topological feature matrix is:
  • the class condition boundary constraint performs boundary constraints on features through a rule-based structured understanding of feature values and the class conditions to which they belong, so as to avoid the set of feature values from being affected by the out-of-distribution characteristics of the set. Excessive fragmentation of the decision-making area within the classification target domain to obtain robust conditional class boundaries, thereby improving the constraint of the classification feature matrix on the classification target by improving the convergence within the class condition boundary of the topological feature matrix itself , thereby improving the classification effect.
  • FIG. 3 illustrates a block diagram of a topological feature correction module in an intelligent toxic and harmful gas alarm system for hexafluorobutadiene preparation according to an embodiment of the present application.
  • the topological feature correction module 260 includes: an exponential operation unit 261, used to calculate the natural exponential function value with the eigenvalue of each position in the topological feature matrix as the power; a reciprocal operation unit 262, using To calculate the reciprocal of the eigenvalues of each position in the topological feature matrix; the constraint value calculation unit 263 is used to subtract the topological feature matrix from the natural exponential function value with the eigenvalue of each position in the topological feature matrix as the power.
  • the fusion module 270 and the alarm result generation module 280 are used to perform matrix multiplication of the measurement data associated feature matrix and the corrected topological feature matrix to obtain the The high-dimensional topological information of the corrected topological feature matrix is mapped into the high-dimensional feature space of the measured data associated feature matrix to obtain a classification feature matrix, and the classification feature matrix is passed through a classifier to obtain a classification result.
  • the classification result Used to indicate whether an alarm prompt is generated. That is to say, in the technical solution of the present application, after obtaining the corrected topological feature matrix, the measured data associated feature matrix and the corrected topological feature matrix are further matrix multiplied to obtain the corrected topological feature matrix.
  • the high-dimensional topological information of the topological feature matrix is mapped to the high-dimensional feature space of the measurement data associated feature matrix, thereby obtaining a classification feature matrix for classification, so as to obtain a classification result indicating whether an alarm prompt is generated.
  • the classifier processes the classification feature matrix with the following formula to generate a classification result, where the formula is: softmax ⁇ (W n ,B n ):...:(W 1 ,B 1 )
  • the fusion module is further configured to perform matrix multiplication of the measured data associated feature matrix and the corrected topology feature matrix using the following formula to obtain the corrected topology
  • the high-dimensional topological information of the feature matrix is mapped to the high-dimensional feature space of the measurement data associated feature matrix to obtain the classification feature matrix; wherein, the formula is:
  • M represents the classification feature matrix
  • M 1 represents the measurement data associated feature matrix
  • M 2 represents the corrected topological feature matrix
  • the intelligent toxic and harmful gas alarm system 200 for hexafluorobutadiene preparation based on the embodiment of the present application is clarified, which monitors toxic and harmful gases by deploying multiple toxic and harmful gases in the hexafluorobutadiene preparation site.
  • the instrument is used to collect gas concentration values at multiple locations at multiple time points, and a deep neural network model is used to extract implicit dynamic correlation features for multiple gas concentration values.
  • topological features are also used to extract The spatial domain mapping of features can take into account more feature information during classification, thereby improving the classification effect. In this way, toxic and harmful gases in the preparation site can be accurately monitored to ensure the safety of personnel in the preparation site.
  • the intelligent toxic and harmful gas alarm system 200 for the preparation of hexafluorobutadiene can be implemented in various terminal equipment, such as the intelligent toxic and harmful gas alarm system for the preparation of hexafluorobutadiene. Gas alarm algorithm server, etc.
  • the intelligent toxic and harmful gas alarm system 200 for hexafluorobutadiene preparation according to the embodiment of the present application can be integrated into the terminal device as a software module and/or hardware module.
  • the intelligent toxic and hazardous gas alarm system 200 for the preparation of hexafluorobutadiene can be a software module in the operating system of the terminal device, or can be an application program developed for the terminal device; of course , the intelligent toxic and harmful gas alarm system 200 for the preparation of hexafluorobutadiene can also be one of the many hardware modules of the terminal equipment.
  • the intelligent toxic and hazardous gas alarm system 200 for hexafluorobutadiene preparation and the terminal device can also be separate devices, and the intelligent toxic and hazardous gas alarm system 200 for hexafluorobutadiene preparation can also be separate devices.
  • the toxic and harmful gas alarm system 200 can be connected to the terminal device through a wired and/or wireless network, and transmit interactive information according to an agreed data format.
  • Figure 4 illustrates a flow chart of the alarm method of the intelligent toxic and hazardous gas alarm system for the preparation of hexafluorobutadiene.
  • the alarm method of the intelligent toxic and harmful gas alarm system for hexafluorobutadiene preparation includes the step: S110, obtaining the information from hexafluorobutadiene deployed in a predetermined topology pattern.
  • Figure 5 illustrates a schematic structural diagram of an alarm method of an intelligent toxic and harmful gas alarm system for hexafluorobutadiene preparation according to an embodiment of the present application.
  • the network architecture of the alarm method of the intelligent toxic and harmful gas alarm system for the preparation of hexafluorobutadiene first, multiple data collected by each of the toxic and harmful gas monitors are obtained.
  • the gas concentration value at a predetermined time point (for example, P1 as shown in Figure 5) is passed through a temporal encoder (for example, E as shown in Figure 5) including a one-dimensional convolution layer to obtain the corresponding toxic gas concentration value for each toxic gas.
  • the measurement data time series feature vector of the harmful gas monitor (for example, VF1 as shown in Figure 5); then, the measurement data time series feature vectors corresponding to each of the toxic and harmful gas monitors are arranged into a two-dimensional feature matrix (for example, MF1 as shown in Figure 5) is passed through the first convolutional neural network (for example, CNN1 as shown in Figure 5) as a filter to obtain the measurement data correlation feature matrix (for example, as shown in Figure 5 MF2 as shown); then, the obtained topology matrix (for example, P2 as shown in Figure 5) is passed through the second convolutional neural network as a filter (for example, CNN2 as shown in Figure 5) To obtain the topological feature matrix (for example, M1 as shown in Figure 5); then, perform feature distribution correction on the topological feature matrix to obtain the corrected topological feature matrix (for example, M2 as shown in Figure 5); Then, the measured data associated feature matrix and the corrected topological feature matrix are matrix multiplied to map the high-dimensional topological information of the corrected topological feature matrix to
  • classification feature matrix for example, MF as shown in Figure 5
  • classification feature matrix for example, MF as shown in Figure 5
  • classification feature matrix for example, MF as shown in Figure 5
  • classifier for example, the classifier as shown in Figure 5
  • steps S110 and S120 gas concentration values at multiple predetermined time points collected by multiple toxic and harmful gas monitors deployed in a predetermined topological pattern in the preparation site of hexafluorobutadiene are obtained, and The gas concentration values at multiple predetermined time points collected by each of the toxic and harmful gas monitors are passed through a time series encoder including a one-dimensional convolution layer to obtain the measurement data time series feature vector corresponding to each of the toxic and harmful gas monitors.
  • multiple toxic and hazardous gas monitors are deployed in a hexafluorobutadiene preparation site in a predetermined topological pattern to collect gas concentration values at multiple predetermined time points. It should be understood that considering that the deployment space areas of toxic and harmful gas monitors in the hexafluorobutadiene preparation site are connected, this will cause the gas concentration value to have a dynamic change pattern, that is, each location will have a dynamic change pattern.
  • the gas concentration at the deployment point of the toxic and harmful gas monitor will diffuse according to changes in time, for example, from an area with higher concentration to an area with lower concentration.
  • the gas concentration values at multiple predetermined time points collected by each of the toxic and harmful gas monitors are further passed through a time series encoder containing a one-dimensional convolution layer. Processing is performed to obtain measurement data time series feature vectors corresponding to each of the toxic and harmful gas monitors.
  • the temporal encoder consists of alternately arranged fully connected layers and one-dimensional convolutional layers, which extract the correlation of the gas concentration values of each deployment point in the temporal dimension through one-dimensional convolutional encoding.
  • Features and high-dimensional latent features of the gas concentration values at each deployment point are extracted through fully connected encoding.
  • step S130 the time series feature vectors of the measurement data corresponding to each of the toxic and hazardous gas monitors are arranged into a two-dimensional feature matrix and then passed through the first convolutional neural network as a filter to obtain the measurement data Associated feature matrix. That is, in the technical solution of the present application, further, the measurement data time series feature vectors corresponding to each of the toxic and harmful gas monitors are arranged into a two-dimensional feature matrix to integrate each of the toxic and harmful gas monitors.
  • the dynamic change characteristic information of the gas concentration value at the deployment point is processed through the first convolutional neural network as a filter to extract multiple predetermined times collected by each toxic and harmful gas monitor.
  • the global dynamic characteristic representation of the gas concentration value of the point is used to obtain the measurement data correlation characteristic matrix.
  • a topological matrix of the plurality of toxic and harmful gas monitors is obtained, wherein the characteristic value of each position on the off-diagonal position in the topological matrix is the corresponding two toxic and harmful gas monitors.
  • the distance between gas monitors, the eigenvalue of each position on the diagonal position in the topological matrix is zero, and the topological matrix is passed through the second convolutional neural network as a filter to obtain the topological feature matrix.
  • the topological matrix of the multiple toxic and harmful gas monitors is obtained according to a predetermined topological pattern, here , the eigenvalue of each position on the off-diagonal position in the topological matrix is the distance between the corresponding two toxic and harmful gas monitors, and the eigenvalue of each position on the diagonal position in the topological matrix is zero. Then, the topological matrix is passed through the second convolutional neural network as a filter to perform deep mining of topological features to obtain a topological feature matrix.
  • each layer of the second convolutional neural network performs convolution processing, mean pooling processing along the channel dimension and activation processing on the input data in the forward pass of the layer to be processed by the The last layer of the second convolutional neural network generates the topological feature matrix, wherein the input of the first layer of the second convolutional neural network is the topological matrix.
  • step S160 feature distribution correction is performed on the topological feature matrix to obtain a corrected topological feature matrix.
  • the measurement data associated feature matrix and the topological feature matrix are matrix multiplied to map the high-dimensional topological information of the topological feature matrix to the
  • the measurement data is associated with the high-dimensional feature space of the feature matrix, and then the feature information of the two is merged for classification, and the classification result indicating whether an alarm prompt is generated can be obtained.
  • the obtained classification feature matrix has a weak constraint on the classification target of the time-series numerical correlation features of the measurement data, and thus may have a problem of poor classification effect. Therefore, before performing feature fusion, it is necessary to The above-mentioned topological characteristic matrix performs class condition boundary constraints.
  • steps S170 and S180 matrix multiplication is performed on the measured data associated feature matrix and the corrected topological feature matrix to map the high-dimensional topological information of the corrected topological feature matrix to the corrected topological feature matrix.
  • the measurement data is associated with the high-dimensional feature space of the feature matrix to obtain a classification feature matrix, and the classification feature matrix is passed through a classifier to obtain a classification result.
  • the classification result is used to indicate whether an alarm prompt is generated. That is to say, in the technical solution of the present application, after obtaining the corrected topological feature matrix, the measured data associated feature matrix and the corrected topological feature matrix are further matrix multiplied to obtain the corrected topological feature matrix.
  • the high-dimensional topological information of the topological feature matrix is mapped to the high-dimensional feature space of the measurement data associated feature matrix, thereby obtaining a classification feature matrix for classification, so as to obtain a classification result indicating whether an alarm prompt is generated.
  • the classifier processes the classification feature matrix with the following formula to generate a classification result, where the formula is: softmax ⁇ (W n ,B n ):...:(W 1 ,B 1 )
  • the alarm method of the intelligent toxic and harmful gas alarm system for the preparation of hexafluorobutadiene is clarified, which is achieved by deploying multiple toxic and harmful gases in the preparation site of hexafluorobutadiene.
  • the gas monitor collects gas concentration values at multiple locations at multiple time points, and uses a deep neural network model to extract implicit dynamic correlation features for multiple gas concentration values.
  • topological features are also used To perform spatial domain mapping of features to take into account more feature information during classification, thereby improving the classification effect. In this way, toxic and harmful gases in the preparation site can be accurately monitored to ensure the safety of personnel in the preparation site.
  • each component or each step can be decomposed and/or recombined. These decompositions and/or recombinations shall be considered equivalent versions of this application.

Landscapes

  • Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Physics & Mathematics (AREA)
  • Toxicology (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Evolutionary Computation (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Environmental & Geological Engineering (AREA)
  • Business, Economics & Management (AREA)
  • Emergency Management (AREA)
  • Emergency Alarm Devices (AREA)

Abstract

一种用于六氟丁二烯制备的智能化有毒有害气体报警系统, 涉及气体智能监测的领域，其通过在六氟丁二烯的制备场所内部署多个有毒有害气体监测仪来采集在多个时间点下的多个位置的气体浓度值，并使用深度神经网络模型来对多个所述气体浓度值进行隐含的动态关联特征提取，同时，还使用拓扑特征来进行特征的空间域映射，以在分类时兼顾到更多的特征信息，进而提高分类的效果。这样，可以对制备场所内的有毒有害气体进行准确地监测以确保制备场所内的人员安全。

Description

一种用于六氟丁二烯制备的智能化有毒有害气体报警系统 技术领域
本发明涉及气体智能监测的领域，且更为具体地，涉及一种用于六氟丁二烯制备的智能化有毒有害气体报警系统及其报警方法。
背景技术
在全球电子气体市场上，含氟电子气体约占30％，主要用作蚀刻剂和清洗剂等。当前广泛使用的全氟烷烃类(PFCs)化合物虽然不破坏臭氧层，但在《京都议定书》中被认定为较强的温室气体。随着人们对环境要求的不断提高，传统含氟电子气体的使用将会受到极大的限制。因此需要寻找新型的环保型含氟电子气体。
六氟丁二烯凭借其各方面的优异性能成为传统含氟电子气体的最佳替代品之一，它是制备多种含氟聚合物材料的单体，还是一种绿色环保的高效干蚀刻气体，近年来已引起国内外学者的高度关注。
但是，六氟丁二烯是一种易燃、有毒、无色、无味的气体，其与空气混合后，浓度达到7％时，有立即燃烧和爆炸的危险。并且，被吸入体内后，对人体会产生危害，可能导致呼吸系统刺激、咳嗽、昏眩、麻醉、心律不齐和负面的肾脏影响。
因此，在六氟丁二烯的制备过程中，期待对制备场所内的有毒有害气体进行监测以确保制备场所内的人员安全。
发明内容
为了解决上述技术问题，提出了本申请。本申请的实施例提供了一种用于六氟丁二烯制备的智能化有毒有害气体报警系统及其报警方法，其通过在六氟丁二烯的制备场所内部署多个有毒有害气体监测仪来采集在多个时间点下的多个位置的气体浓度值，并使用深度神经网络模型来对多个所述气体浓度值进行隐含的动态关联特征提取，同时，还使用拓扑特征来进行特征的空间域映射，以在分类时兼顾到更多的特征信息，进而提高分类的效果。这样，可以对制备场所内的有毒有害气体进行准确地监测以确保制备场所内的人员安全。
根据本申请的一个方面，提供了一种用于六氟丁二烯制备的智能化有毒有害气体报警系统，其包括：气体监测数据采集模块，用于获取由以预定拓扑样式部署于六氟丁二烯的制备场所内的多个有毒有害气体监测仪采集的多个预定时间点的气体浓度值；单样本气体数据编码模块，用于将各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值通过包含一维卷积层的时序编码器以得到对应于各个所述有毒有害气体监测仪的测量数据时序特征向量；多样本气体数据关联编码模块，用于将所述对应于各个所述有毒有害气体监测仪的测量数据时序特征向量排列为二维特征矩阵后通过作为过滤器的第一卷积神经网络以得到测量数据关联特征矩阵；传感器拓扑矩阵构造单元，用于获取所述多个有毒有害气体监测仪的拓扑矩阵，其中，所述拓扑矩阵中非对角线位置上各个位置的特征值为相应两个有毒有害气体监测仪之间的距离，所述拓扑矩阵中对角线位置上各个位置的特征值为零；拓扑矩阵编码模块，用于将所述拓扑矩阵通过作为过滤器的第二卷积神经网络以得到拓扑特征矩阵；拓扑特征校正模块，用于对所述拓扑特征矩阵进行特征分布校正以得到校正后拓扑特征矩阵；融合模块，用于将所述测量数据关联特征矩阵与所述校正后拓扑特征矩阵进行矩阵相乘以将所述校正后拓扑特征矩阵的高维拓扑信息映射到所述测量数据关联特征矩阵的高维特征空间中以得到分类特征矩阵；以及报警结果生成模块，用于将所述分类特征矩阵通过分类器以得到分类结果，所述分类结果用于表示是否产生报警提示。
在上述用于六氟丁二烯制备的智能化有毒有害气体报警系统中，所述单样本气体数据编码模块，包括：输入向量构造单元，用于将各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值按照时间维度排列为对应于各个所述有毒有害气体监测仪的一维的输入向量；全连接编码单元，用于使用所述时序编码器的全连接层以如下公式对所述输入向量进行全连接编码以提取出所述输入向量中各个位置的特征值的高维隐含特征，其中，所述公式为：
Figure PCTCN2022119276-appb-000001
其中X是所述输入向量，Y是输出向量，W是权重矩阵，B是偏置向量，
Figure PCTCN2022119276-appb-000002
表示矩阵乘；一维卷积编码单元，用于使用所述时序编码器的一维卷积层以如下公式对所述输入向量进行一维卷积编码以提取出所述输入向量中各个位置的特征值间的高维隐含关联特征，其中，所述公式为：
Figure PCTCN2022119276-appb-000003
其中，a为卷积核在x方向上的宽度、F为卷积核参数向量、G为与卷积核函数运算的局部向量矩阵，w为卷积核的尺寸。
在上述用于六氟丁二烯制备的智能化有毒有害气体报警系统中，所述多样本气体数据关联编码模 块，进一步用于：所述第一卷积神经网络的各层在层的正向传递中对输入数据分别进行：对输入数据进行卷积处理以得到卷积特征图；对所述卷积特征图进行基于局部通道维度的均值池化以得到池化特征图；以及，对所述池化特征图进行非线性激活以得到激活特征图；其中，所述第一卷积神经网络的最后一层的输出为所述测量数据关联特征矩阵，所述第一卷积神经网络的第一层的输入为所述二维特征矩阵。
在上述用于六氟丁二烯制备的智能化有毒有害气体报警系统中，所述拓扑矩阵编码模块，进一步用于：所述第二卷积神经网络的各层在层的正向传递中对输入数据进行卷积处理、沿通道维度的均值池化处理和激活处理以由所述第二卷积神经网络的最后一层生成所述拓扑特征矩阵，其中，所述第二卷积神经网络的第一层的输入为所述拓扑矩阵。
在上述用于六氟丁二烯制备的智能化有毒有害气体报警系统中，所述拓扑特征校正模块，包括：指数运算单元，用于计算以所述拓扑特征矩阵中各个位置的特征值为幂的自然指数函数值；倒数运算单元，用于计算所述拓扑特征矩阵中各个位置的特征值的倒数；约束值计算单元，用于以所述拓扑特征矩阵中各个位置的特征值为幂的自然指数函数值减去所述拓扑特征矩阵中该位置的特征值的倒数再减一以得到所述拓扑特征矩阵中各个位置的特征值对应的约束值；以及，结构化理解单元，用于计算所述拓扑特征矩阵中各个位置的特征值对应的约束值的绝对值的对数函数值以得到所述校正后拓扑特征矩阵。
在上述用于六氟丁二烯制备的智能化有毒有害气体报警系统中，所述融合模块，进一步用于：以如下公式将所述测量数据关联特征矩阵与所述校正后拓扑特征矩阵进行矩阵相乘以将所述校正后拓扑特征矩阵的高维拓扑信息映射到所述测量数据关联特征矩阵的高维特征空间中以得到所述分类特征矩阵；其中，所述公式为：
Figure PCTCN2022119276-appb-000004
其中M表示所述分类特征矩阵，M 1表示所述所述测量数据关联特征矩阵，M 2表示所述校正后拓扑特征矩阵，
Figure PCTCN2022119276-appb-000005
表示矩阵相乘。
在上述用于六氟丁二烯制备的智能化有毒有害气体报警系统中，所述报警结果生成模块，进一步用于：所述分类器以如下公式对所述分类特征矩阵进行处理以生成分类结果，其中，所述公式为：softmax{(W n,B n):…:(W 1,B 1)|Project(F)}，其中Project(F)表示将所述分类特征矩阵投影为向量，W 1至W n为各层全连接层的权重矩阵，B 1至B n表示各层全连接层的偏置矩阵。
根据本申请的另一方面，一种用于六氟丁二烯制备的智能化有毒有害气体报警系统的报警方法，其包括：
获取由以预定拓扑样式部署于六氟丁二烯的制备场所内的多个有毒有害气体监测仪采集的多个预定时间点的气体浓度值；将各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值通过包含一维卷积层的时序编码器以得到对应于各个所述有毒有害气体监测仪的测量数据时序特征向量；将所述对应于各个所述有毒有害气体监测仪的测量数据时序特征向量排列为二维特征矩阵后通过作为过滤器的第一卷积神经网络以得到测量数据关联特征矩阵；获取所述多个有毒有害气体监测仪的拓扑矩阵，其中，所述拓扑矩阵中非对角线位置上各个位置的特征值为相应两个有毒有害气体监测仪之间的距离，所述拓扑矩阵中对角线位置上各个位置的特征值为零；将所述拓扑矩阵通过作为过滤器的第二卷积神经网络以得到拓扑特征矩阵；对所述拓扑特征矩阵进行特征分布校正以得到校正后拓扑特征矩阵；将所述测量数据关联特征矩阵与所述校正后拓扑特征矩阵进行矩阵相乘以将所述校正后拓扑特征矩阵的高维拓扑信息映射到所述测量数据关联特征矩阵的高维特征空间中以得到分类特征矩阵；以及
将所述分类特征矩阵通过分类器以得到分类结果，所述分类结果用于表示是否产生报警提示。
在上述用于六氟丁二烯制备的智能化有毒有害气体报警系统的报警方法中，将各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值通过包含一维卷积层的时序编码器以得到对应于各个所述有毒有害气体监测仪的测量数据时序特征向量，包括：将各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值按照时间维度排列为对应于各个所述有毒有害气体监测仪的一维的输入向量；使用所述时序编码器的全连接层以如下公式对所述输入向量进行全连接编码以提取出所述输入向量中各个位置的特征值的高维隐含特征，其中，所述公式为：
Figure PCTCN2022119276-appb-000006
其中X是所述输入向量，Y是输出向量，W是权重矩阵，B是偏置向量，
Figure PCTCN2022119276-appb-000007
表示矩阵乘；使用所述时序编码器的一维卷积层以如下公式对所述输入向量进行一维卷积编码以提取出所述输入向量中各个位置的特征值间的高维隐含关联特征，其中，所述公式为：
Figure PCTCN2022119276-appb-000008
其中，a为卷积核在x方向上的宽度、F为卷积核参数向量、G为与卷积核函数运算的局部向量 矩阵，w为卷积核的尺寸。
在上述用于六氟丁二烯制备的智能化有毒有害气体报警系统的报警方法中，将所述对应于各个所述有毒有害气体监测仪的测量数据时序特征向量排列为二维特征矩阵后通过作为过滤器的第一卷积神经网络以得到测量数据关联特征矩阵，包括：所述第一卷积神经网络的各层在层的正向传递中对输入数据分别进行：对输入数据进行卷积处理以得到卷积特征图；对所述卷积特征图进行基于局部通道维度的均值池化以得到池化特征图；以及，对所述池化特征图进行非线性激活以得到激活特征图；其中，所述第一卷积神经网络的最后一层的输出为所述测量数据关联特征矩阵，所述第一卷积神经网络的第一层的输入为所述二维特征矩阵。
在上述用于六氟丁二烯制备的智能化有毒有害气体报警系统的报警方法中，所述第二卷积神经网络的各层在层的正向传递中对输入数据进行卷积处理、沿通道维度的均值池化处理和激活处理以由所述第二卷积神经网络的最后一层生成所述拓扑特征矩阵，其中，所述第二卷积神经网络的第一层的输入为所述拓扑矩阵。
在上述用于六氟丁二烯制备的智能化有毒有害气体报警系统的报警方法中，对所述拓扑特征矩阵进行特征分布校正以得到校正后拓扑特征矩阵，包括：计算以所述拓扑特征矩阵中各个位置的特征值为幂的自然指数函数值；计算所述拓扑特征矩阵中各个位置的特征值的倒数；以所述拓扑特征矩阵中各个位置的特征值为幂的自然指数函数值减去所述拓扑特征矩阵中该位置的特征值的倒数再减一以得到所述拓扑特征矩阵中各个位置的特征值对应的约束值；以及，计算所述拓扑特征矩阵中各个位置的特征值对应的约束值的绝对值的对数函数值以得到所述校正后拓扑特征矩阵。
在上述用于六氟丁二烯制备的智能化有毒有害气体报警系统的报警方法中，将所述测量数据关联特征矩阵与所述校正后拓扑特征矩阵进行矩阵相乘以将所述校正后拓扑特征矩阵的高维拓扑信息映射到所述测量数据关联特征矩阵的高维特征空间中以得到分类特征矩阵，包括：以如下公式将所述测量数据关联特征矩阵与所述校正后拓扑特征矩阵进行矩阵相乘以将所述校正后拓扑特征矩阵的高维拓扑信息映射到所述测量数据关联特征矩阵的高维特征空间中以得到所述分类特征矩阵；其中，所述公式为：
Figure PCTCN2022119276-appb-000009
其中M表示所述分类特征矩阵，M 1表示所述所述测量数据关联特征矩阵，M 2表示所述校正后拓扑特征矩阵，
Figure PCTCN2022119276-appb-000010
表示矩阵相乘。
在上述用于六氟丁二烯制备的智能化有毒有害气体报警系统的报警方法中，将所述分类特征矩阵通过分类器以得到分类结果，包括：所述分类器以如下公式对所述分类特征矩阵进行处理以生成分类结果，其中，所述公式为：softmax{(W n,B n):…:(W 1,B 1)|Project(F)}，其中Project(F)表示将所述分类特征矩阵投影为向量，W 1至W n为各层全连接层的权重矩阵，B 1至B n表示各层全连接层的偏置矩阵。
与现有技术相比，本申请提供的用于六氟丁二烯制备的智能化有毒有害气体报警系统及其报警方法，其通过在六氟丁二烯的制备场所内部署多个有毒有害气体监测仪来采集在多个时间点下的多个位置的气体浓度值，并使用深度神经网络模型来对多个所述气体浓度值进行隐含的动态关联特征提取，同时，还使用拓扑特征来进行特征的空间域映射，以在分类时兼顾到更多的特征信息，进而提高分类的效果。这样，可以对制备场所内的有毒有害气体进行准确地监测以确保制备场所内的人员安全。
附图说明
通过结合附图对本申请实施例进行更详细的描述，本申请的上述以及其他目的、特征和优势将变得更加明显。附图用来提供对本申请实施例的进一步理解，并且构成说明书的一部分，与本申请实施例一起用于解释本申请，并不构成对本申请的限制。在附图中，相同的参考标号通常代表相同部件或步骤。
图1为根据本申请实施例的用于六氟丁二烯制备的智能化有毒有害气体报警系统的应用场景图。
图2为根据本申请实施例的用于六氟丁二烯制备的智能化有毒有害气体报警系统的框图。
图3为根据本申请实施例的用于六氟丁二烯制备的智能化有毒有害气体报警系统中拓扑特征校正模块的框图。
图4为根据本申请实施例的用于六氟丁二烯制备的智能化有毒有害气体报警系统的报警方法的流程图。
图5为根据本申请实施例的用于六氟丁二烯制备的智能化有毒有害气体报警系统的报警方法的架构示意图。
具体实施方式
下面，将参考附图详细地描述根据本申请的示例实施例。显然，所描述的实施例仅仅是本申请的一部分实施例，而不是本申请的全部实施例，应理解，本申请不受这里描述的示例实施例的限制。
场景概述
如前所述，在全球电子气体市场上，含氟电子气体约占30％，主要用作蚀刻剂和清洗剂等。当前广泛使用的全氟烷烃类(PFCs)化合物虽然不破坏臭氧层，但在《京都议定书》中被认定为较强的温室气体。随着人们对环境要求的不断提高，传统含氟电子气体的使用将会受到极大的限制。因此需要寻找新型的环保型含氟电子气体。
六氟丁二烯凭借其各方面的优异性能成为传统含氟电子气体的最佳替代品之一，它是制备多种含氟聚合物材料的单体，还是一种绿色环保的高效干蚀刻气体，近年来已引起国内外学者的高度关注。
但是，六氟丁二烯是一种易燃、有毒、无色、无味的气体，其与空气混合后，浓度达到7％时，有立即燃烧和爆炸的危险。并且，被吸入体内后，对人体会产生危害，可能导致呼吸系统刺激、咳嗽、昏眩、麻醉、心律不齐和负面的肾脏影响。
因此，在六氟丁二烯的制备过程中，期待对制备场所内的有毒有害气体进行监测以确保制备场所内的人员安全。
相应地，本申请发明人发现在利用气体传感器进行有毒有害气体(主要是六氟丁二烯气体)监测时，由于有毒有害气体在待监测场所内各个位置的分布是不均匀的，因此，如果采用一个单独的气体传感器进行气体监测，其可能监测到的气体浓度值未超过安全阈值，但场所内其他位置的气体浓度有可能已超过预定阈值。其次，单个传感器也有可能因故障而测量产生误差，因此，采用单个气体传感器来进行气体浓度监测存在安全隐患。并且，任何气体传感器都有自身系统误差，因此，即便单个传感器自身不存在故障，依赖单个传感器的数据作为监测依据是不合理的。
基于此，在本申请的技术方案中，通过以预定的拓扑样式来在六氟丁二烯的制备场所内部署多个有毒有害气体监测仪，以采集多个预定时间点的气体浓度值。应可以理解，考虑到所述六氟丁二烯的制备场所内有毒有害气体监测仪的部署空间区域是连通的，这样会使得所述气体浓度值具有动态性的变化规律，也就是，各个所述有毒有害气体监测仪部署点的气体浓度会根据时间的变化发生扩散，例如会从浓度较高的地区扩散至浓度较低的地区。因此为了更为充分地提取这种动态性的隐含变化规律，进一步将各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值通过包含一维卷积层的时序编码器中进行处理，以得到对应于各个所述有毒有害气体监测仪的测量数据时序特征向量。在一个具体示例中，所述时序编码器由交替设置的全连接层和一维卷积层组成，其通过一维卷积编码提取出各个部署点的所述气体浓度值在时序维度上的关联特征和通过全连接编码提取各个部署点的所述气体浓度值的高维隐含特征。
进一步地，将所述对应于各个所述有毒有害气体监测仪的测量数据时序特征向量排列为二维特征矩阵，以整合各个所述有毒有害气体监测仪部署点的气体浓度值的动态变化特征信息，再将这种变化特征通过作为过滤器的第一卷积神经网络中进行处理，以提取出各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值的全局性的动态特征表示，以得到测量数据关联特征矩阵。
由于考虑到这种扩散规律也与空间的特征分布有关，因此，根据预定的拓扑样式来获取所述多个有毒有害气体监测仪的拓扑矩阵，这里，所述拓扑矩阵中非对角线位置上各个位置的特征值为相应两个有毒有害气体监测仪之间的距离，所述拓扑矩阵中对角线位置上各个位置的特征值为零。然后，将所述拓扑矩阵通过作为过滤器的第二卷积神经网络中进行拓扑特征的深层挖掘，以得到拓扑特征矩阵。
这样，将所述测量数据关联特征矩阵与所述拓扑特征矩阵进行矩阵相乘，以将所述拓扑特征矩阵的高维拓扑信息映射到所述测量数据关联特征矩阵的高维特征空间中，进而融合这两者的特征信息来进行分类，以获得用于表示是否产生报警提示的分类结果。但是，在将所述拓扑特征矩阵乘以所述测量数据关联特征矩阵以进行特征融合时，由于所述拓扑特征矩阵表达传感器位置的拓扑特征，而不包含测量数据的数值特征，这就导致所获得的所述分类特征矩阵在针对测量数据的时序数值关联特征的分类目标上存在约束性弱的问题，从而可能存在分类效果不佳的问题。
基于此，在进行特征融合之前，首先对所述拓扑特征矩阵进行类条件边界约束：
Figure PCTCN2022119276-appb-000011
Figure PCTCN2022119276-appb-000012
其中m i,j是所述拓扑特征矩阵的每个位置的特征值。
这里，该所述类条件边界约束通过对特征值及其所属的类条件进行基于规则的结构化理解，来进行特征的边界约束，以避免特征值集合由于集合的分布外特性而导致在分类目标域内的决策区域的过度碎片化，以获得稳健的条件化的类边界，从而通过提升所述拓扑特征矩阵本身的类条件边界内的收敛性来提高分类特征矩阵在分类目标上的约束性，进而提高分类的效果。
基于此，本申请提出了一种用于六氟丁二烯制备的智能化有毒有害气体报警系统，其包括：气体监测数据采集模块，用于获取由以预定拓扑样式部署于六氟丁二烯的制备场所内的多个有毒有害气体监测仪采集的多个预定时间点的气体浓度值；单样本气体数据编码模块，用于将各个所述有毒有害气 体监测仪采集的多个预定时间点的气体浓度值通过包含一维卷积层的时序编码器以得到对应于各个所述有毒有害气体监测仪的测量数据时序特征向量；多样本气体数据关联编码模块，用于将所述对应于各个所述有毒有害气体监测仪的测量数据时序特征向量排列为二维特征矩阵后通过作为过滤器的第一卷积神经网络以得到测量数据关联特征矩阵；传感器拓扑矩阵构造单元，用于获取所述多个有毒有害气体监测仪的拓扑矩阵，其中，所述拓扑矩阵中非对角线位置上各个位置的特征值为相应两个有毒有害气体监测仪之间的距离，所述拓扑矩阵中对角线位置上各个位置的特征值为零；拓扑矩阵编码模块，用于将所述拓扑矩阵通过作为过滤器的第二卷积神经网络以得到拓扑特征矩阵；拓扑特征校正模块，用于对所述拓扑特征矩阵进行特征分布校正以得到校正后拓扑特征矩阵；融合模块，用于将所述测量数据关联特征矩阵与所述校正后拓扑特征矩阵进行矩阵相乘以将所述校正后拓扑特征矩阵的高维拓扑信息映射到所述测量数据关联特征矩阵的高维特征空间中以得到分类特征矩阵；以及，报警结果生成模块，用于将所述分类特征矩阵通过分类器以得到分类结果，所述分类结果用于表示是否产生报警提示。
图1图示了根据本申请实施例的用于六氟丁二烯制备的智能化有毒有害气体报警系统的应用场景图。如图1所示，在该应用场景中，首先，通过部署于六氟丁二烯的制备场所(例如，如图1中所示意的H)内的多个有毒有害气体监测仪(例如，如图1中所示意的T1-Tn)的预定拓扑样式来获取拓扑矩阵，并且通过以预定拓扑样式部署于六氟丁二烯的制备场所内的多个有毒有害气体监测仪采集的多个预定时间点的气体浓度值。然后，将获得的所述多个有毒有害气体监测仪采集的多个预定时间点的气体浓度值以及所述多个有毒有害气体监测仪的拓扑矩阵输入至部署有用于六氟丁二烯制备的智能化有毒有害气体报警算法的服务器中(例如，如图1中所示意的服务器S)，其中，所述服务器能够以用于六氟丁二烯制备的智能化有毒有害气体报警算法对所述多个有毒有害气体监测仪采集的多个预定时间点的气体浓度值以及所述多个有毒有害气体监测仪的拓扑矩阵进行处理，以生成用于表示是否产生报警提示的分类结果。
在介绍了本申请的基本原理之后，下面将参考附图来具体介绍本申请的各种非限制性实施例。
示例性系统
图2图示了根据本申请实施例的用于六氟丁二烯制备的智能化有毒有害气体报警系统的框图。如图2所示，根据本申请实施例的用于六氟丁二烯制备的智能化有毒有害气体报警系统200，包括：气体监测数据采集模块210，用于获取由以预定拓扑样式部署于六氟丁二烯的制备场所内的多个有毒有害气体监测仪采集的多个预定时间点的气体浓度值；单样本气体数据编码模块220，用于将各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值通过包含一维卷积层的时序编码器以得到对应于各个所述有毒有害气体监测仪的测量数据时序特征向量；多样本气体数据关联编码模块230，用于将所述对应于各个所述有毒有害气体监测仪的测量数据时序特征向量排列为二维特征矩阵后通过作为过滤器的第一卷积神经网络以得到测量数据关联特征矩阵；传感器拓扑矩阵构造单元240，用于获取所述多个有毒有害气体监测仪的拓扑矩阵，其中，所述拓扑矩阵中非对角线位置上各个位置的特征值为相应两个有毒有害气体监测仪之间的距离，所述拓扑矩阵中对角线位置上各个位置的特征值为零；拓扑矩阵编码模块250，用于将所述拓扑矩阵通过作为过滤器的第二卷积神经网络以得到拓扑特征矩阵；拓扑特征校正模块260，用于对所述拓扑特征矩阵进行特征分布校正以得到校正后拓扑特征矩阵；融合模块270，用于将所述测量数据关联特征矩阵与所述校正后拓扑特征矩阵进行矩阵相乘以将所述校正后拓扑特征矩阵的高维拓扑信息映射到所述测量数据关联特征矩阵的高维特征空间中以得到分类特征矩阵；以及，报警结果生成模块280，用于将所述分类特征矩阵通过分类器以得到分类结果，所述分类结果用于表示是否产生报警提示。
具体地，在本申请实施例中，所述气体监测数据采集模块210和所述单样本气体数据编码模块220，用于获取由以预定拓扑样式部署于六氟丁二烯的制备场所内的多个有毒有害气体监测仪采集的多个预定时间点的气体浓度值，并将各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值通过包含一维卷积层的时序编码器以得到对应于各个所述有毒有害气体监测仪的测量数据时序特征向量。如前所述，应可以理解，在利用气体传感器进行有毒有害气体(主要是六氟丁二烯气体)监测时，由于所述有毒有害气体在待监测场所内各个位置的分布是不均匀的，因此，如果采用一个单独的气体传感器进行气体监测，其可能监测到的气体浓度值未超过安全阈值，但场所内其他位置的气体浓度有可能已超过预定阈值。其次，单个传感器也有可能因故障而测量产生误差，因此，采用所述单个气体传感器来进行气体浓度监测存在安全隐患。并且，任何气体传感器都有自身系统误差，因此，即便所述单个传感器自身不存在故障，依赖所述单个传感器的数据作为监测依据是不合理的。
因此，在本申请的技术方案中，通过以预定的拓扑样式来在六氟丁二烯的制备场所内部署多个有毒有害气体监测仪，以采集多个预定时间点的气体浓度值。应可以理解，考虑到所述六氟丁二烯的制备场所内有毒有害气体监测仪的部署空间区域是连通的，这样会使得所述气体浓度值具有动态性的变 化规律，也就是，各个所述有毒有害气体监测仪部署点的气体浓度会根据时间的变化发生扩散，例如会从浓度较高的地区扩散至浓度较低的地区。因此为了更为充分地提取这种动态性的隐含变化规律，进一步将各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值通过包含一维卷积层的时序编码器中进行处理，以得到对应于各个所述有毒有害气体监测仪的测量数据时序特征向量。在一个具体示例中，所述时序编码器由交替设置的全连接层和一维卷积层组成，其通过一维卷积编码提取出各个部署点的所述气体浓度值在时序维度上的关联特征和通过全连接编码提取各个部署点的所述气体浓度值的高维隐含特征。
更具体地，在本申请实施例中，所述单样本气体数据编码模块，包括：输入向量构造单元，用于将各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值按照时间维度排列为对应于各个所述有毒有害气体监测仪的一维的输入向量；全连接编码单元，用于使用所述时序编码器的全连接层以如下公式对所述输入向量进行全连接编码以提取出所述输入向量中各个位置的特征值的高维隐含特征，其中，所述公式为：
Figure PCTCN2022119276-appb-000013
其中X是所述输入向量，Y是输出向量，W是权重矩阵，B是偏置向量，
Figure PCTCN2022119276-appb-000014
表示矩阵乘；一维卷积编码单元，用于使用所述时序编码器的一维卷积层以如下公式对所述输入向量进行一维卷积编码以提取出所述输入向量中各个位置的特征值间的高维隐含关联特征，其中，所述公式为：
Figure PCTCN2022119276-appb-000015
其中，a为卷积核在x方向上的宽度、F为卷积核参数向量、G为与卷积核函数运算的局部向量矩阵，w为卷积核的尺寸。
具体地，在本申请实施例中，所述多样本气体数据关联编码模块230，用于将所述对应于各个所述有毒有害气体监测仪的测量数据时序特征向量排列为二维特征矩阵后通过作为过滤器的第一卷积神经网络以得到测量数据关联特征矩阵。也就是，在本申请的技术方案中，进一步地，将所述对应于各个所述有毒有害气体监测仪的测量数据时序特征向量排列为二维特征矩阵，以整合各个所述有毒有害气体监测仪部署点的气体浓度值的动态变化特征信息，再将这种变化特征通过作为过滤器的第一卷积神经网络中进行处理，以提取出各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值的全局性的动态特征表示，以得到测量数据关联特征矩阵。
更具体地，在本申请实施例中，所述多样本气体数据关联编码模块，进一步用于：所述第一卷积神经网络的各层在层的正向传递中对输入数据分别进行：对输入数据进行卷积处理以得到卷积特征图；对所述卷积特征图进行基于局部通道维度的均值池化以得到池化特征图；以及，对所述池化特征图进行非线性激活以得到激活特征图；其中，所述第一卷积神经网络的最后一层的输出为所述测量数据关联特征矩阵，所述第一卷积神经网络的第一层的输入为所述二维特征矩阵。
具体地，在本申请实施例中，所述传感器拓扑矩阵构造单元240和所述拓扑矩阵编码模块250，用于获取所述多个有毒有害气体监测仪的拓扑矩阵，其中，所述拓扑矩阵中非对角线位置上各个位置的特征值为相应两个有毒有害气体监测仪之间的距离，所述拓扑矩阵中对角线位置上各个位置的特征值为零，并将所述拓扑矩阵通过作为过滤器的第二卷积神经网络以得到拓扑特征矩阵。应可以理解，考虑到由于这种扩散规律也与空间的特征分布有关，因此，在本申请的技术方案中，根据预定的拓扑样式来获取所述多个有毒有害气体监测仪的拓扑矩阵，这里，所述拓扑矩阵中非对角线位置上各个位置的特征值为相应两个有毒有害气体监测仪之间的距离，所述拓扑矩阵中对角线位置上各个位置的特征值为零。然后，将所述拓扑矩阵通过作为过滤器的第二卷积神经网络中进行拓扑特征的深层挖掘，以得到拓扑特征矩阵。相应地，在一个具体示例中，所述第二卷积神经网络的各层在层的正向传递中对输入数据进行卷积处理、沿通道维度的均值池化处理和激活处理以由所述第二卷积神经网络的最后一层生成所述拓扑特征矩阵，其中，所述第二卷积神经网络的第一层的输入为所述拓扑矩阵。
具体地，在本申请实施例中，所述拓扑特征校正模块260，用于对所述拓扑特征矩阵进行特征分布校正以得到校正后拓扑特征矩阵。应可以理解，在本申请的技术方案中，这样，再将所述测量数据关联特征矩阵与所述拓扑特征矩阵进行矩阵相乘，以将所述拓扑特征矩阵的高维拓扑信息映射到所述测量数据关联特征矩阵的高维特征空间中，进而融合这两者的特征信息来进行分类，就可以获得用于表示是否产生报警提示的分类结果。但是，考虑到在将所述拓扑特征矩阵乘以所述测量数据关联特征矩阵以进行特征融合时，由于所述拓扑特征矩阵表达传感器位置的拓扑特征，而不包含测量数据的数值特征，这就导致所获得的所述分类特征矩阵在针对测量数据的时序数值关联特征的分类目标上存在约束性弱的问题，从而可能存在分类效果不佳的问题，因此，在进行特征融合之前，需要对所述拓扑特征矩阵进行类条件边界约束。
更具体地，在本申请实施例中，所述拓扑特征校正模块，包括：首先，计算以所述拓扑特征矩阵 中各个位置的特征值为幂的自然指数函数值。接着，计算所述拓扑特征矩阵中各个位置的特征值的倒数。然后，以所述拓扑特征矩阵中各个位置的特征值为幂的自然指数函数值减去所述拓扑特征矩阵中该位置的特征值的倒数再减一以得到所述拓扑特征矩阵中各个位置的特征值对应的约束值。最后，计算所述拓扑特征矩阵中各个位置的特征值对应的约束值的绝对值的对数函数值以得到所述校正后拓扑特征矩阵。相应地，在一个具体示例中，对所述拓扑特征矩阵进行类条件边界约束的公式为：
Figure PCTCN2022119276-appb-000016
其中m i,j是所述拓扑特征矩阵的每个位置的特征值。应可以理解，该所述类条件边界约束通过对特征值及其所属的类条件进行基于规则的结构化理解，来进行特征的边界约束，以避免特征值集合由于集合的分布外特性而导致在分类目标域内的决策区域的过度碎片化，以获得稳健的条件化的类边界，从而通过提升所述拓扑特征矩阵本身的类条件边界内的收敛性来提高分类特征矩阵在分类目标上的约束性，进而提高分类的效果。
图3图示了根据本申请实施例的用于六氟丁二烯制备的智能化有毒有害气体报警系统中拓扑特征校正模块的框图。如图3所示，所述拓扑特征校正模块260，包括：指数运算单元261，用于计算以所述拓扑特征矩阵中各个位置的特征值为幂的自然指数函数值；倒数运算单元262，用于计算所述拓扑特征矩阵中各个位置的特征值的倒数；约束值计算单元263，用于以所述拓扑特征矩阵中各个位置的特征值为幂的自然指数函数值减去所述拓扑特征矩阵中该位置的特征值的倒数再减一以得到所述拓扑特征矩阵中各个位置的特征值对应的约束值；以及，结构化理解单元264，用于计算所述拓扑特征矩阵中各个位置的特征值对应的约束值的绝对值的对数函数值以得到所述校正后拓扑特征矩阵。
具体地，在本申请实施例中，所述融合模块270和所述报警结果生成模块280，用于将所述测量数据关联特征矩阵与所述校正后拓扑特征矩阵进行矩阵相乘以将所述校正后拓扑特征矩阵的高维拓扑信息映射到所述测量数据关联特征矩阵的高维特征空间中以得到分类特征矩阵，并将所述分类特征矩阵通过分类器以得到分类结果，所述分类结果用于表示是否产生报警提示。也就是，在本申请的技术方案中，在获得所述校正后拓扑特征矩阵后，进一步将所述测量数据关联特征矩阵与所述校正后拓扑特征矩阵进行矩阵相乘，以将所述校正后拓扑特征矩阵的高维拓扑信息映射到所述测量数据关联特征矩阵的高维特征空间中，从而得到分类特征矩阵来进行分类，以得到用于表示是否产生报警提示的分类结果。相应地，在一个具体示例中，所述分类器以如下公式对所述分类特征矩阵进行处理以生成分类结果，其中，所述公式为：softmax{(W n,B n):…:(W 1,B 1)|Project(F)}，其中Project(F)表示将所述分类特征矩阵投影为向量，W 1至W n为各层全连接层的权重矩阵，B 1至B n表示各层全连接层的偏置矩阵。
更具体地，在本申请实施例中，所述融合模块，进一步用于：以如下公式将所述测量数据关联特征矩阵与所述校正后拓扑特征矩阵进行矩阵相乘以将所述校正后拓扑特征矩阵的高维拓扑信息映射到所述测量数据关联特征矩阵的高维特征空间中以得到所述分类特征矩阵；其中，所述公式为：
Figure PCTCN2022119276-appb-000017
Figure PCTCN2022119276-appb-000018
其中M表示所述分类特征矩阵，M 1表示所述所述测量数据关联特征矩阵，M 2表示所述校正后拓扑特征矩阵，
Figure PCTCN2022119276-appb-000019
表示矩阵相乘。
综上，基于本申请实施例的所述用于六氟丁二烯制备的智能化有毒有害气体报警系统200被阐明，其通过在六氟丁二烯的制备场所内部署多个有毒有害气体监测仪来采集在多个时间点下的多个位置的气体浓度值，并使用深度神经网络模型来对多个所述气体浓度值进行隐含的动态关联特征提取，同时，还使用拓扑特征来进行特征的空间域映射，以在分类时兼顾到更多的特征信息，进而提高分类的效果。这样，可以对制备场所内的有毒有害气体进行准确地监测以确保制备场所内的人员安全。
如上所述，根据本申请实施例的用于六氟丁二烯制备的智能化有毒有害气体报警系统200可以实现在各种终端设备中，例如用于六氟丁二烯制备的智能化有毒有害气体报警算法的服务器等。在一个示例中，根据本申请实施例的用于六氟丁二烯制备的智能化有毒有害气体报警系统200可以作为一个软件模块和/或硬件模块而集成到终端设备中。例如，该用于六氟丁二烯制备的智能化有毒有害气体报警系统200可以是该终端设备的操作系统中的一个软件模块，或者可以是针对于该终端设备所开发的一个应用程序；当然，该用于六氟丁二烯制备的智能化有毒有害气体报警系统200同样可以是该终端设备的众多硬件模块之一。
替换地，在另一示例中，该用于六氟丁二烯制备的智能化有毒有害气体报警系统200与该终端设备也可以是分立的设备，并且该用于六氟丁二烯制备的智能化有毒有害气体报警系统200可以通过有线和/或无线网络连接到该终端设备，并且按照约定的数据格式来传输交互信息。
示例性方法
图4图示了用于六氟丁二烯制备的智能化有毒有害气体报警系统的报警方法的流程图。如图4所示，根据本申请实施例的用于六氟丁二烯制备的智能化有毒有害气体报警系统的报警方法，包括步骤：S110，获取由以预定拓扑样式部署于六氟丁二烯的制备场所内的多个有毒有害气体监测仪采集的多个预定时间点的气体浓度值；S120，将各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值通过包含一维卷积层的时序编码器以得到对应于各个所述有毒有害气体监测仪的测量数据时序特征向量；S130，将所述对应于各个所述有毒有害气体监测仪的测量数据时序特征向量排列为二维特征矩阵后通过作为过滤器的第一卷积神经网络以得到测量数据关联特征矩阵；S140，获取所述多个有毒有害气体监测仪的拓扑矩阵，其中，所述拓扑矩阵中非对角线位置上各个位置的特征值为相应两个有毒有害气体监测仪之间的距离，所述拓扑矩阵中对角线位置上各个位置的特征值为零；S150，将所述拓扑矩阵通过作为过滤器的第二卷积神经网络以得到拓扑特征矩阵；S160，对所述拓扑特征矩阵进行特征分布校正以得到校正后拓扑特征矩阵；S170，将所述测量数据关联特征矩阵与所述校正后拓扑特征矩阵进行矩阵相乘以将所述校正后拓扑特征矩阵的高维拓扑信息映射到所述测量数据关联特征矩阵的高维特征空间中以得到分类特征矩阵；以及，S180，将所述分类特征矩阵通过分类器以得到分类结果，所述分类结果用于表示是否产生报警提示。
图5图示了根据本申请实施例的用于六氟丁二烯制备的智能化有毒有害气体报警系统的报警方法的架构示意图。如图5所示，在所述用于六氟丁二烯制备的智能化有毒有害气体报警系统的报警方法的网络架构中，首先，将获得的各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值(例如，如图5中所示意的P1)通过包含一维卷积层的时序编码器(例如，如图5中所示意的E)以得到对应于各个所述有毒有害气体监测仪的测量数据时序特征向量(例如，如图5中所示意的VF1)；接着，将所述对应于各个所述有毒有害气体监测仪的测量数据时序特征向量排列为二维特征矩阵(例如，如图5中所示意的MF1)后通过作为过滤器的第一卷积神经网络(例如，如图5中所示意的CNN1)以得到测量数据关联特征矩阵(例如，如图5中所示意的MF2)；然后，将获得的所述拓扑矩阵(例如，如图5中所示意的P2)通过作为过滤器的第二卷积神经网络(例如，如图5中所示意的CNN2)以得到拓扑特征矩阵(例如，如图5中所示意的M1)；接着，对所述拓扑特征矩阵进行特征分布校正以得到校正后拓扑特征矩阵(例如，如图5中所示意的M2)；然后，将所述测量数据关联特征矩阵与所述校正后拓扑特征矩阵进行矩阵相乘以将所述校正后拓扑特征矩阵的高维拓扑信息映射到所述测量数据关联特征矩阵的高维特征空间中以得到分类特征矩阵(例如，如图5中所示意的MF)；以及，最后，将所述分类特征矩阵通过分类器(例如，如图5中所示意的分类器)以得到分类结果，所述分类结果用于表示是否产生报警提示。
更具体地，在步骤S110和S120中，获取由以预定拓扑样式部署于六氟丁二烯的制备场所内的多个有毒有害气体监测仪采集的多个预定时间点的气体浓度值，并将各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值通过包含一维卷积层的时序编码器以得到对应于各个所述有毒有害气体监测仪的测量数据时序特征向量。应可以理解，在利用气体传感器进行有毒有害气体(主要是六氟丁二烯气体)监测时，由于所述有毒有害气体在待监测场所内各个位置的分布是不均匀的，因此，如果采用一个单独的气体传感器进行气体监测，其可能监测到的气体浓度值未超过安全阈值，但场所内其他位置的气体浓度有可能已超过预定阈值。其次，单个传感器也有可能因故障而测量产生误差，因此，采用所述单个气体传感器来进行气体浓度监测存在安全隐患。并且，任何气体传感器都有自身系统误差，因此，即便所述单个传感器自身不存在故障，依赖所述单个传感器的数据作为监测依据是不合理的。
因此，在本申请的技术方案中，通过以预定的拓扑样式来在六氟丁二烯的制备场所内部署多个有毒有害气体监测仪，以采集多个预定时间点的气体浓度值。应可以理解，考虑到所述六氟丁二烯的制备场所内有毒有害气体监测仪的部署空间区域是连通的，这样会使得所述气体浓度值具有动态性的变化规律，也就是，各个所述有毒有害气体监测仪部署点的气体浓度会根据时间的变化发生扩散，例如会从浓度较高的地区扩散至浓度较低的地区。因此为了更为充分地提取这种动态性的隐含变化规律，进一步将各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值通过包含一维卷积层的时序编码器中进行处理，以得到对应于各个所述有毒有害气体监测仪的测量数据时序特征向量。在一个具体示例中，所述时序编码器由交替设置的全连接层和一维卷积层组成，其通过一维卷积编码提取出各个部署点的所述气体浓度值在时序维度上的关联特征和通过全连接编码提取各个部署点的所述气体浓度值的高维隐含特征。
更具体地，在步骤S130中，将所述对应于各个所述有毒有害气体监测仪的测量数据时序特征向量排列为二维特征矩阵后通过作为过滤器的第一卷积神经网络以得到测量数据关联特征矩阵。也就是，在本申请的技术方案中，进一步地，将所述对应于各个所述有毒有害气体监测仪的测量数据时序特征向量排列为二维特征矩阵，以整合各个所述有毒有害气体监测仪部署点的气体浓度值的动态变化特征 信息，再将这种变化特征通过作为过滤器的第一卷积神经网络中进行处理，以提取出各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值的全局性的动态特征表示，以得到测量数据关联特征矩阵。
更具体地，在步骤S140和步骤S150中，获取所述多个有毒有害气体监测仪的拓扑矩阵，其中，所述拓扑矩阵中非对角线位置上各个位置的特征值为相应两个有毒有害气体监测仪之间的距离，所述拓扑矩阵中对角线位置上各个位置的特征值为零，并将所述拓扑矩阵通过作为过滤器的第二卷积神经网络以得到拓扑特征矩阵。应可以理解，考虑到由于这种扩散规律也与空间的特征分布有关，因此，在本申请的技术方案中，根据预定的拓扑样式来获取所述多个有毒有害气体监测仪的拓扑矩阵，这里，所述拓扑矩阵中非对角线位置上各个位置的特征值为相应两个有毒有害气体监测仪之间的距离，所述拓扑矩阵中对角线位置上各个位置的特征值为零。然后，将所述拓扑矩阵通过作为过滤器的第二卷积神经网络中进行拓扑特征的深层挖掘，以得到拓扑特征矩阵。相应地，在一个具体示例中，所述第二卷积神经网络的各层在层的正向传递中对输入数据进行卷积处理、沿通道维度的均值池化处理和激活处理以由所述第二卷积神经网络的最后一层生成所述拓扑特征矩阵，其中，所述第二卷积神经网络的第一层的输入为所述拓扑矩阵。
更具体地，在步骤S160中，对所述拓扑特征矩阵进行特征分布校正以得到校正后拓扑特征矩阵。应可以理解，在本申请的技术方案中，这样，再将所述测量数据关联特征矩阵与所述拓扑特征矩阵进行矩阵相乘，以将所述拓扑特征矩阵的高维拓扑信息映射到所述测量数据关联特征矩阵的高维特征空间中，进而融合这两者的特征信息来进行分类，就可以获得用于表示是否产生报警提示的分类结果。但是，考虑到在将所述拓扑特征矩阵乘以所述测量数据关联特征矩阵以进行特征融合时，由于所述拓扑特征矩阵表达传感器位置的拓扑特征，而不包含测量数据的数值特征，这就导致所获得的所述分类特征矩阵在针对测量数据的时序数值关联特征的分类目标上存在约束性弱的问题，从而可能存在分类效果不佳的问题，因此，在进行特征融合之前，需要对所述拓扑特征矩阵进行类条件边界约束。
更具体地，在步骤S170和步骤S180中，将所述测量数据关联特征矩阵与所述校正后拓扑特征矩阵进行矩阵相乘以将所述校正后拓扑特征矩阵的高维拓扑信息映射到所述测量数据关联特征矩阵的高维特征空间中以得到分类特征矩阵，并将所述分类特征矩阵通过分类器以得到分类结果，所述分类结果用于表示是否产生报警提示。也就是，在本申请的技术方案中，在获得所述校正后拓扑特征矩阵后，进一步将所述测量数据关联特征矩阵与所述校正后拓扑特征矩阵进行矩阵相乘，以将所述校正后拓扑特征矩阵的高维拓扑信息映射到所述测量数据关联特征矩阵的高维特征空间中，从而得到分类特征矩阵来进行分类，以得到用于表示是否产生报警提示的分类结果。
相应地，在一个具体示例中，所述分类器以如下公式对所述分类特征矩阵进行处理以生成分类结果，其中，所述公式为：softmax{(W n,B n):…:(W 1,B 1)|Project(F)}，其中Project(F)表示将所述分类特征矩阵投影为向量，W 1至W n为各层全连接层的权重矩阵，B 1至B n表示各层全连接层的偏置矩阵。
综上，基于本申请实施例的所述用于六氟丁二烯制备的智能化有毒有害气体报警系统的报警方法被阐明，其通过在六氟丁二烯的制备场所内部署多个有毒有害气体监测仪来采集在多个时间点下的多个位置的气体浓度值，并使用深度神经网络模型来对多个所述气体浓度值进行隐含的动态关联特征提取，同时，还使用拓扑特征来进行特征的空间域映射，以在分类时兼顾到更多的特征信息，进而提高分类的效果。这样，可以对制备场所内的有毒有害气体进行准确地监测以确保制备场所内的人员安全。
以上结合具体实施例描述了本申请的基本原理，但是，需要指出的是，在本申请中提及的优点、优势、效果等仅是示例而非限制，不能认为这些优点、优势、效果等是本申请的各个实施例必须具备的。另外，上述公开的具体细节仅是为了示例的作用和便于理解的作用，而非限制，上述细节并不限制本申请为必须采用上述具体的细节来实现。
本申请中涉及的器件、装置、设备、系统的方框图仅作为例示性的例子并且不意图要求或暗示必须按照方框图示出的方式进行连接、布置、配置。如本领域技术人员将认识到的，可以按任意方式连接、布置、配置这些器件、装置、设备、系统。诸如“包括”、“包含”、“具有”等等的词语是开放性词汇，指“包括但不限于”，且可与其互换使用。这里所使用的词汇“或”和“和”指词汇“和/或”，且可与其互换使用，除非上下文明确指示不是如此。这里所使用的词汇“诸如”指词组“诸如但不限于”，且可与其互换使用。
还需要指出的是，在本申请的装置、设备和方法中，各部件或各步骤是可以分解和/或重新组合的。这些分解和/或重新组合应视为本申请的等效方案。
提供所公开的方面的以上描述以使本领域的任何技术人员能够做出或者使用本申请。对这些方面的各种修改对于本领域技术人员而言是非常显而易见的，并且在此定义的一般原理可以应用于其他方面而不脱离本申请的范围。因此，本申请不意图被限制到在此示出的方面，而是按照与在此公开的原 理和新颖的特征一致的最宽范围。
为了例示和描述的目的已经给出了以上描述。此外，此描述不意图将本申请的实施例限制到在此公开的形式。尽管以上已经讨论了多个示例方面和实施例，但是本领域技术人员将认识到其某些变型、修改、改变、添加和子组合。

Claims (10)

  1. 一种用于六氟丁二烯制备的智能化有毒有害气体报警系统，其特征在于，包括：气体监测数据采集模块，用于获取由以预定拓扑样式部署于六氟丁二烯的制备场所内的多个有毒有害气体监测仪采集的多个预定时间点的气体浓度值；单样本气体数据编码模块，用于将各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值通过包含一维卷积层的时序编码器以得到对应于各个所述有毒有害气体监测仪的测量数据时序特征向量；多样本气体数据关联编码模块，用于将所述对应于各个所述有毒有害气体监测仪的测量数据时序特征向量排列为二维特征矩阵后通过作为过滤器的第一卷积神经网络以得到测量数据关联特征矩阵；传感器拓扑矩阵构造单元，用于获取所述多个有毒有害气体监测仪的拓扑矩阵，其中，所述拓扑矩阵中非对角线位置上各个位置的特征值为相应两个有毒有害气体监测仪之间的距离，所述拓扑矩阵中对角线位置上各个位置的特征值为零；拓扑矩阵编码模块，用于将所述拓扑矩阵通过作为过滤器的第二卷积神经网络以得到拓扑特征矩阵；拓扑特征校正模块，用于对所述拓扑特征矩阵进行特征分布校正以得到校正后拓扑特征矩阵；融合模块，用于将所述测量数据关联特征矩阵与所述校正后拓扑特征矩阵进行矩阵相乘以将所述校正后拓扑特征矩阵的高维拓扑信息映射到所述测量数据关联特征矩阵的高维特征空间中以得到分类特征矩阵；以及报警结果生成模块，用于将所述分类特征矩阵通过分类器以得到分类结果，所述分类结果用于表示是否产生报警提示。
  2. 根据权利要求1所述的用于六氟丁二烯制备的智能化有毒有害气体报警系统，其中，所述单样本气体数据编码模块，包括：输入向量构造单元，用于将各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值按照时间维度排列为对应于各个所述有毒有害气体监测仪的一维的输入向量；全连接编码单元，用于使用所述时序编码器的全连接层以如下公式对所述输入向量进行全连接编码以提取出所述输入向量中各个位置的特征值的高维隐含特征，其中，所述公式为：
    Figure PCTCN2022119276-appb-100001
    其中X是所述输入向量，Y是输出向量，W是权重矩阵，B是偏置向量，
    Figure PCTCN2022119276-appb-100002
    表示矩阵乘；一维卷积编码单元，用于使用所述时序编码器的一维卷积层以如下公式对所述输入向量进行一维卷积编码以提取出所述输入向量中各个位置的特征值间的高维隐含关联特征，其中，所述公式为：
    Figure PCTCN2022119276-appb-100003
    其中，a为卷积核在x方向上的宽度、F为卷积核参数向量、G为与卷积核函数运算的局部向量矩阵，w为卷积核的尺寸。
  3. 根据权利要求2所述的用于六氟丁二烯制备的智能化有毒有害气体报警系统，其中，所述多样本气体数据关联编码模块，进一步用于：所述第一卷积神经网络的各层在层的正向传递中对输入数据分别进行：对输入数据进行卷积处理以得到卷积特征图；对所述卷积特征图进行基于局部通道维度的均值池化以得到池化特征图；以及对所述池化特征图进行非线性激活以得到激活特征图；其中，所述第一卷积神经网络的最后一层的输出为所述测量数据关联特征矩阵，所述第一卷积神经网络的第一层的输入为所述二维特征矩阵。
  4. 根据权利要求3所述的用于六氟丁二烯制备的智能化有毒有害气体报警系统，其中，所述拓扑矩阵编码模块，进一步用于：所述第二卷积神经网络的各层在层的正向传递中对输入数据进行卷积处理、沿通道维度的均值池化处理和激活处理以由所述第二卷积神经网络的最后一层生成所述拓扑特征矩阵，其中，所述第二卷积神经网络的第一层的输入为所述拓扑矩阵。
  5. 根据权利要求4所述的用于六氟丁二烯制备的智能化有毒有害气体报警系统，其中，所述拓扑特征校正模块，包括：指数运算单元，用于计算以所述拓扑特征矩阵中各个位置的特征值为幂的自然指数函数值；倒数运算单元，用于计算所述拓扑特征矩阵中各个位置的特征值的倒数；约束值计算单元，用于以所述拓扑特征矩阵中各个位置的特征值为幂的自然指数函数值减去所述拓扑特征矩阵中该位置的特征值的倒数再减一以得到所述拓扑特征矩阵中各个位置的特征值对应的约束值；以及
    结构化理解单元，用于计算所述拓扑特征矩阵中各个位置的特征值对应的约束值的绝对值的对数函数值以得到所述校正后拓扑特征矩阵。
  6. 根据权利要求5所述的用于六氟丁二烯制备的智能化有毒有害气体报警系统，其中，所述融合模块，进一步用于：以如下公式将所述测量数据关联特征矩阵与所述校正后拓扑特征矩阵进行矩阵相乘以将所述校正后拓扑特征矩阵的高维拓扑信息映射到所述测量数据关联特征矩阵的高维特征空间中以得到所述分类特征矩阵；其中，所述公式为：
    Figure PCTCN2022119276-appb-100004
    其中M表示所述分类特征矩阵，M 1表示所述所述测量数据关联特征矩阵，M 2表示所述校正后拓扑特征矩阵，
    Figure PCTCN2022119276-appb-100005
    表示矩阵相乘。
  7. 根据权利要求6所述的用于六氟丁二烯制备的智能化有毒有害气体报警系统，其中，所述报警结果生成模块，进一步用于：所述分类器以如下公式对所述分类特征矩阵进行处理以生成分类结果， 其中，所述公式为：softmax{(W n,B n):…:(W 1,B 1)|Project(F)}，其中Project(F)表示将所述分类特征矩阵投影为向量，W 1至W n为各层全连接层的权重矩阵，B 1至B n表示各层全连接层的偏置矩阵。
  8. 一种用于六氟丁二烯制备的智能化有毒有害气体报警系统的报警方法，其特征在于，包括：
    获取由以预定拓扑样式部署于六氟丁二烯的制备场所内的多个有毒有害气体监测仪采集的多个预定时间点的气体浓度值；将各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值通过包含一维卷积层的时序编码器以得到对应于各个所述有毒有害气体监测仪的测量数据时序特征向量；将所述对应于各个所述有毒有害气体监测仪的测量数据时序特征向量排列为二维特征矩阵后通过作为过滤器的第一卷积神经网络以得到测量数据关联特征矩阵；获取所述多个有毒有害气体监测仪的拓扑矩阵，其中，所述拓扑矩阵中非对角线位置上各个位置的特征值为相应两个有毒有害气体监测仪之间的距离，所述拓扑矩阵中对角线位置上各个位置的特征值为零；将所述拓扑矩阵通过作为过滤器的第二卷积神经网络以得到拓扑特征矩阵；对所述拓扑特征矩阵进行特征分布校正以得到校正后拓扑特征矩阵；将所述测量数据关联特征矩阵与所述校正后拓扑特征矩阵进行矩阵相乘以将所述校正后拓扑特征矩阵的高维拓扑信息映射到所述测量数据关联特征矩阵的高维特征空间中以得到分类特征矩阵；以及将所述分类特征矩阵通过分类器以得到分类结果，所述分类结果用于表示是否产生报警提示。
  9. 根据权利要求8所述的用于六氟丁二烯制备的智能化有毒有害气体报警系统的报警方法，其中，将各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值通过包含一维卷积层的时序编码器以得到对应于各个所述有毒有害气体监测仪的测量数据时序特征向量，包括：将各个所述有毒有害气体监测仪采集的多个预定时间点的气体浓度值按照时间维度排列为对应于各个所述有毒有害气体监测仪的一维的输入向量；使用所述时序编码器的全连接层以如下公式对所述输入向量进行全连接编码以提取出所述输入向量中各个位置的特征值的高维隐含特征，其中，所述公式为：
    Figure PCTCN2022119276-appb-100006
    Figure PCTCN2022119276-appb-100007
    其中X是所述输入向量，Y是输出向量，W是权重矩阵，B是偏置向量，
    Figure PCTCN2022119276-appb-100008
    表示矩阵乘；使用所述时序编码器的一维卷积层以如下公式对所述输入向量进行一维卷积编码以提取出所述输入向量中各个位置的特征值间的高维隐含关联特征，其中，所述公式为：
    Figure PCTCN2022119276-appb-100009
    其中，a为卷积核在x方向上的宽度、F为卷积核参数向量、G为与卷积核函数运算的局部向量矩阵，w为卷积核的尺寸。
  10. 根据权利要求9所述的用于六氟丁二烯制备的智能化有毒有害气体报警系统的报警方法，其中，对所述拓扑特征矩阵进行特征分布校正以得到校正后拓扑特征矩阵，包括：计算以所述拓扑特征矩阵中各个位置的特征值为幂的自然指数函数值；计算所述拓扑特征矩阵中各个位置的特征值的倒数；以所述拓扑特征矩阵中各个位置的特征值为幂的自然指数函数值减去所述拓扑特征矩阵中该位置的特征值的倒数再减一以得到所述拓扑特征矩阵中各个位置的特征值对应的约束值；以及
    计算所述拓扑特征矩阵中各个位置的特征值对应的约束值的绝对值的对数函数值以得到所述校正后拓扑特征矩阵。
PCT/CN2022/119276 2022-07-28 2022-09-16 一种用于六氟丁二烯制备的智能化有毒有害气体报警系统 WO2024021253A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210897022.7 2022-07-28
CN202210897022.7A CN115346345B (zh) 2022-07-28 2022-07-28 一种用于六氟丁二烯制备的智能化有毒有害气体报警系统

Publications (1)

Publication Number Publication Date
WO2024021253A1 true WO2024021253A1 (zh) 2024-02-01

Family

ID=83950222

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/119276 WO2024021253A1 (zh) 2022-07-28 2022-09-16 一种用于六氟丁二烯制备的智能化有毒有害气体报警系统

Country Status (2)

Country Link
CN (1) CN115346345B (zh)
WO (1) WO2024021253A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117787143A (zh) * 2024-02-23 2024-03-29 国网天津市电力公司城东供电分公司 变电站毒害气体的气流分布场重构方法及系统、电子设备

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115932140A (zh) * 2022-11-16 2023-04-07 福建省杭氟电子材料有限公司 一种电子级六氟丁二烯的质检系统及其方法
CN115903705B (zh) * 2022-11-30 2023-07-14 福建省杭氟电子材料有限公司 用于电子级六氟丁二烯制备的生产管理控制系统
CN116935586B (zh) * 2023-03-29 2024-06-04 福建天甫电子材料有限公司 用于硝酸的制备的智能化有毒有害气体报警系统
CN116106457B (zh) * 2023-04-13 2023-06-27 天津海河标测技术检测有限公司 空气采样检测一体化装置

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20050156730A1 (en) * 2004-01-08 2005-07-21 Maple Chase Company System and method for controlling ignition sources and ventilating systems during high carbon monoxide conditions
CN111127825A (zh) * 2019-12-25 2020-05-08 深圳供电局有限公司 环境预测方法和装置、电子设备
CN113947739A (zh) * 2021-10-15 2022-01-18 深圳市铠硕达科技有限公司 基于物联网的小区安全管理监控系统及其监控方法
CN114647198A (zh) * 2022-03-09 2022-06-21 深圳市经纬纵横科技有限公司 基于物联网的智能家居控制方法、系统和电子设备

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP3962361B2 (ja) * 2003-06-27 2007-08-22 インターナショナル・ビジネス・マシーンズ・コーポレーション 位相判断装置、分解可能形状生成装置、構造メッシュ生成装置、位相判断方法、分解可能形状生成方法、および該位相判断方法を実行させるためのコンピュータ実行可能なプログラム、分解可能形状生成方法を実行させるためのコンピュータ実行可能なプログラム並びに構造メッシュ生成システム
CN202795675U (zh) * 2012-08-13 2013-03-13 长城金点定位测控(北京）有限公司 气体报警器
US11995557B2 (en) * 2017-05-30 2024-05-28 Kuano Ltd. Tensor network machine learning system
WO2021120834A1 (zh) * 2020-07-09 2021-06-24 平安科技（深圳）有限公司 基于生物识别的手势识别方法、装置、计算机设备及介质
KR102229617B1 (ko) * 2020-08-20 2021-03-17 홍성희 스마트 iot 인공지능 가스경보 시스템
CN112907522B (zh) * 2021-02-03 2023-07-28 北京遥感设备研究所 一种智能红外气体泄漏监测装置及监测方法
CN114615641A (zh) * 2022-03-03 2022-06-10 上海土蜂科技有限公司 高空平台设备协同管理方法、系统和电子设备

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20050156730A1 (en) * 2004-01-08 2005-07-21 Maple Chase Company System and method for controlling ignition sources and ventilating systems during high carbon monoxide conditions
CN111127825A (zh) * 2019-12-25 2020-05-08 深圳供电局有限公司 环境预测方法和装置、电子设备
CN113947739A (zh) * 2021-10-15 2022-01-18 深圳市铠硕达科技有限公司 基于物联网的小区安全管理监控系统及其监控方法
CN114647198A (zh) * 2022-03-09 2022-06-21 深圳市经纬纵横科技有限公司 基于物联网的智能家居控制方法、系统和电子设备

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117787143A (zh) * 2024-02-23 2024-03-29 国网天津市电力公司城东供电分公司 变电站毒害气体的气流分布场重构方法及系统、电子设备

Also Published As

Publication number Publication date
CN115346345B (zh) 2023-06-06
CN115346345A (zh) 2022-11-15

Similar Documents

Publication Publication Date Title
WO2024021253A1 (zh) 一种用于六氟丁二烯制备的智能化有毒有害气体报警系统
WO2024011732A1 (zh) 六氟丁二烯储放场所的气体监测系统及其监测方法
WO2023226228A1 (zh) 用于电子级双氧水制备的智能化有毒有害气体报警系统
CN105992977B (zh) 成套设备的诊断装置和诊断方法
CN112133052B (zh) 核电厂图像火灾探测方法
CN115013298B (zh) 污水泵的实时性能在线监测系统及其监测方法
Lestari et al. Fire hotspots detection system on CCTV videos using you only look once (YOLO) method and tiny YOLO model for high buildings evacuation
CN108830215B (zh) 基于人员骨架信息的危险行为识别方法
TWI498770B (zh) 一種異常應用程式的識別方法及裝置
CN115375691A (zh) 基于图像的半导体扩散纸源缺陷检测系统及其方法
CN111986436B (zh) 一种基于紫外、深度神经网络的综合火焰检测方法
CN108596062B (zh) 基于深度学习的人脸图片实时高光去除方法及装置
CN106980827A (zh) 一种基于航拍图像的识别输电线路中鸟巢的方法
CN116051506B (zh) 聚合物锂离子电池的智能化生产系统及其方法
CN115259089B (zh) 用于电子级氢氟酸制备的生产管理控制系统及其控制方法
CN115099684B (zh) 企业安全生产管理系统及其管理方法
CN115456047A (zh) 基于人工智能的海上风机结构状态监测系统及其方法
CN115324843A (zh) 基于监测数据的风力发电机组的故障诊断系统及其方法
CN113947739A (zh) 基于物联网的小区安全管理监控系统及其监控方法
CN117269644A (zh) 电流互感器的线路故障监测系统及其方法
CN107818135A (zh) 一种基于灰色关联法的维诺图电力大数据异常检测方法
CN113343123B (zh) 一种生成对抗多关系图网络的训练方法和检测方法
WO2011030662A1 (ja) 被災範囲推定装置及びプログラム
CN105898222B (zh) 一种图像采集方法及图像采集系统
WO2024103470A1 (zh) 自动扶梯生产用性能检测系统及其方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22952694

Country of ref document: EP

Kind code of ref document: A1