Ethylene oxide alarm meter test method

Abstract

The present application relates to the technical field of ethylene oxide alarm detector detection, and disclose a method for detecting ethylene oxide alarm instrument, comprising collecting the original voltage readings of each sensor in the absence of ethylene oxide environment, and carrying out normalization processing; the normalized data at each time is constructed as a node set, and a symmetric distance matrix is generated based on the Euclidean distance; the historical minimum spanning tree is constructed by a self-defined algorithm, a baseline model is formed, and the total weight is calculated; real-time acquisition of new data, after normalization, the minimum incremental weight is calculated with the historical node, and compared with the adaptive threshold to determine whether to trigger the alarm. The system structure is expressed by the graph model, which makes the alarm judgment more robust and adaptable; the incremental weight discrimination mechanism realizes the quantification of the abnormal degree, and the adaptive threshold enhances the sensitivity of the system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ethylene oxide alarm detection, in particular to an ethylene oxide alarm detection method. BACKGROUND

[0002] Ethylene oxide (EO) is an important industrial disinfectant and sterilizing agent, widely used in medical devices, pharmaceuticals and electronics industry. However, EO has strong volatility, high toxicity, flammable and explosive safety hazards, once leaked, not only poses a serious threat to the health of operating personnel, but also may cause fire or explosion accidents. Therefore, the rapid and accurate monitoring of EO becomes the key to ensure the safety of production and use.

[0003] Currently, the commonly used EO alarm on the market is mainly based on metal oxide semiconductor (MOS) sensor, photo ionization detector (PID) or infrared absorption technology. MOS sensor detects gas by measuring the change of surface resistance of metal oxide material, has the advantages of low cost and sensitivity to multiple gases, but is easily affected by temperature and humidity changes and sensor aging, resulting in baseline drift and sensitivity decline, which requires frequent calibration; PID ionizes gas molecules by high-voltage electric field, detects the size of downstream ion current, has high sensitivity, but is expensive, cross-sensitive to water vapor and other volatile organic compounds (VOCs); the infrared absorption method relies on the absorption peak of EO to specific infrared band, has good anti-interference ability, but requires high stability of instrument optical path and light source, and has relatively large volume and power consumption, which is difficult to deploy in portable or distributed scenarios. For single sensor technology, traditional alarm methods mostly use fixed threshold judgment, i.e. a voltage or concentration threshold is set in advance, and alarm is triggered when the real-time monitoring value exceeds the threshold. This method lacks adaptive adjustment to baseline drift and environmental background changes, often leading to two problems: one is that when the sensor ages or the temperature and humidity change causes the baseline to move up, the fixed threshold is easy to trigger false alarm; the other is that when the sensor sensitivity decreases, the fixed threshold may not detect the initial leakage of concentration rise, which has the risk of missing alarm. To improve the monitoring reliability, some domestic or imported EO alarms introduce multi-sensor redundancy design, which connects multiple groups of the same or different types of sensors in parallel, and integrates the gas concentration through averaging, weighting or majority voting, but its fusion logic still relies on static weighting coefficients or manual experience setting, cannot adaptively adjust the threshold according to the sensor drift history, and is difficult to extract feature change patterns from multi-dimensional data, resulting in high false alarm rate and missing alarm rate. At the algorithm level, some studies propose methods based on sliding window statistics, derivative mutation, conventional machine learning models (such as support vector machine, decision tree) for secondary processing of sensor signals, but most of them rely on large-scale training data or complex feature extraction collected on the ground, which is not suitable for real-time and online deployment; at the same time, existing algorithms usually contain hyperparameters, weight coefficients or deep models that are difficult to tune, lack the ability of lightweight online implementation in hardware resource limited scenarios.

[0004] To this end, the present application aims to propose an ethylene oxide alarm detection method, which uses the behavior of sensors in a non-toxic environment as a "historical normal reference" to construct a set of quantifiable, traceable, and dynamically evolving baseline models, and in subsequent real-time monitoring, uses the minimum spanning tree incremental change as a sensitive criterion to dynamically determine whether an abnormal gas leak has occurred. SUMMARY

[0005] The present application provides an ethylene oxide alarm detection method, which promotes the solution to the problems mentioned in the background art.

[0006] The present application provides the following technical solutions: an ethylene oxide alarm detection method, comprising:

[0007] In an ethylene oxide-free environment, the original voltage readings of each sensor are collected in time series, the historical average value and the maximum deviation are calculated, and the readings of each sensor are normalized based on this;

[0008] The normalized data vectors at each historical time are constructed as nodes, and a node set is generated;

[0009] The similarity between nodes is calculated according to the Euclidean distance, and a symmetric distance matrix is constructed;

[0010] According to the distance matrix, a historical minimum spanning tree is constructed using a self-defined method, a historical baseline model is obtained, and the total weight is calculated;

[0011] The alarm threshold is dynamically calculated according to the minimum distance of the node joining the minimum spanning tree;

[0012] Real-time sensor data is collected at the target detection time, and normalized according to the established historical calibration parameters;

[0013] The minimum incremental weight between the current normalized vector and the historical node is calculated;

[0014] The increment is compared with the adaptive threshold, if it exceeds the threshold, it is determined as an ethylene oxide leak and an alarm is triggered, otherwise it remains normal.

[0015] Optionally, the original voltage readings of each sensor are collected in time series in an ethylene oxide-free environment, the historical average value and the maximum deviation are calculated, and the readings of each sensor are normalized based on this, specifically comprising:

[0016] In an ethylene oxide-free environment, the original readings of the sensors are collected in time set {t i |i=1,2,...,P} in turn;

[0017] For each sensor j∈{1,2,...,M}, record Sj (t i ) ; wherein, M is the total number of parallel sensors; S j (t i ) is the original analog voltage reading of sensor j at time t i ;

[0018] calculating the historical average voltage of sensor j

[0019] calculating the historical maximum absolute deviation of sensor j

[0020] calculating the normalized reading of sensor j at any time t

[0021] Optionally, the normalized data vector of each historical time is constructed as a node, and a node set is generated, specifically including:

[0022] for each historical time t i , the normalized vector is set as X i = [x1(t i ), x2(t i ),..., x M (t i )] T ;

[0023] the historical node set is set as wherein, element v i corresponds to vector X i ; is a mapping relationship, indicating that the node and the vector correspond one by one.

[0024] Optionally, the similarity between nodes is calculated according to the Euclidean distance, and a symmetric distance matrix is constructed, specifically including:

[0025] the edge set of the node complete graph is set as E = {(i, k) | 1≤i

[0026] the Euclidean distance between nodes v i and v k is calculated

[0027] all D i,k are stored in the symmetric matrix [D i,k ] P×P .

[0028] Optionally, the historical minimum spanning tree is constructed by a self-defined method according to the distance matrix, a historical baseline model is obtained, and the total weight is calculated, specifically including:

[0029] Set {D i,k Sort |(i,k)∈E} in ascending order, denoted as Among them, D (m) The m-th smallest distance;

[0030] Initialize edge set Initialize connected group set

[0031] For m=1 to Perform the following steps:

[0032] S110, Take the m-th smallest distance and its corresponding node pair (i m ,k m ),satisfy

[0033] S120, in the set Search for containing and The two groups are denoted as C. a and C b ;

[0034] S130, if C a ≠C b :

[0035] S131, move the edge (i m ,k m Add to T0;

[0036] S132, Update Group

[0037] S140. Repeat steps S110 to S130 until |T0| = P-1;

[0038] Calculate the baseline total weight

[0039] Optionally, the step of dynamically calculating the alarm threshold based on the minimum distance for a node to join the minimum spanning tree specifically includes:

[0040] When a node q = {2, 3, ..., P} is added, calculate the minimum distance to the nearest existing node:

[0041]

[0042] Set the adaptive alarm threshold to

[0043] Optionally, the step of collecting real-time sensor data at the target detection moment and normalizing it according to predetermined historical calibration parameters specifically includes:

[0044] At the real-time monitoring moment t * Collecting the original readings S of each sensor j (t * );

[0045] Applying the existing calibration normalization:

[0046] X * =[x1(t * ),x2(t * ),...,x M (t * )] T ; wherein X * is a new normalized vector.

[0047] Optionally, the calculation of the minimum incremental weight between the current normalized vector and the historical node specifically comprises:

[0048] For all historical nodes i={1, 2,..., P}, the Euclidean distance between the new vector X * and the historical vector X i is calculated:

[0049]

[0050] Extracting the minimum increment

[0051] Optionally, the increment is compared with an adaptive threshold value, if the threshold value is exceeded, it is determined that ethylene oxide is leaked and an alarm is triggered, otherwise, the normal state is maintained, and the specific steps comprise:

[0052] The alarm judgment function is set as:

[0053]

[0054] Wherein, when Alarm=1, the ethylene oxide leakage alarm is triggered; when Alarm=0, the normal monitoring state is maintained.

[0055] The present application has the following beneficial effects:

[0056] 1. Raw voltage readings from sensors at multiple times in an ethylene oxide-free environment are collected. The historical average value and maximum absolute deviation of each sensor are extracted to normalize all data. This normalization method is more engineering-significant than traditional "standard deviation normalization" or "maximum / minimum scaling" because it eliminates the absolute influence of voltage drift between sensors and improves comparability across time periods and batches. Instead of simply relying on maximum and minimum value standards, a statistical steady-state model is established based on historical behavior, giving the sensor signals uniform dimensional characteristics. This "baseline construction" method effectively solves the problems of zero-point drift and amplitude fluctuations caused by environmental factors, aging, or temperature, laying a solid data foundation for subsequent graph modeling and dynamic analysis.

[0057] 2. By converting normalized sensor vectors into nodes in a graph structure over time and constructing a complete set of historical nodes, a graph structure representation of time-series data is achieved. This approach changes the traditional "point-by-point comparison or window sliding" approach in sequence analysis, introducing graph structure modeling, which helps capture global relationships and spatial trends. It not only endows historical moments with topological properties but also provides a structural foundation for subsequently extracting "normal trajectories" using minimum spanning trees. Compared to traditional time-series regression or single-point trend detection, the graph method is more holistic, avoids the problem of local overfitting, and is easily extended to deep algorithms such as graph neural networks.

[0058] 3. Euclidean distance is used to quantify the similarity between every two historical nodes, and a symmetric matrix is ​​constructed for subsequent graph algorithms. Its form of "symmetric distance matrix + complete graph assumption" comprehensively and evenly considers the relationships between historical nodes, rather than relying solely on fixed adjacent nodes. Furthermore, the use of Euclidean distance closely complements the normalization operation mentioned earlier, ensuring comparability across all dimensions. Compared to correlation coefficients or Dynamic Time Warping (DTW) methods, this approach is more robust under small sample conditions, has lower algorithm complexity, and is easy to deploy, making it particularly suitable for rapid implementation in embedded hardware scenarios.

[0059] 4. Using the Minimum Spanning Tree (MST) algorithm based on the distance matrix, an acyclic connected subgraph connecting all historical nodes is autonomously constructed, and its total weight is defined as the "normal model baseline weight." The key technical advantage of this method lies in modeling "normal behavior" as a low-redundancy tree structure, ensuring the shortest connections between nodes and global reachability, thus forming a "topological skeleton" representing the steady-state behavior of the system. This approach avoids the subjectivity of fitting functions or statistical models, achieving an objective expression of historical normal data. Compared to existing traditional methods that model using "threshold fitting," the MST model not only considers both global and local aspects but also possesses resistance to local perturbations, making it suitable for stable modeling of multivariable, weakly coupled systems.

[0060] 5、In the process of gradually adding new nodes to the MST, record the minimum edge weight of the nearest neighbor, and dynamically calculate the critical threshold value required for alarm. The threshold value is based on the maximum value of all "nearest neighbor weights" in history, with the characteristics of behavior-driven and strong adaptability. The core innovation is to break the static strategy of "preset value threshold" and build a "dynamic threshold under data evolution", which significantly improves the sensitivity and robustness of the model. Compared with the traditional "experience threshold", this method is more intelligent, does not rely on expert setting parameters, has high self-learning and migration ability, and is especially suitable for gas alarm which has nonlinear change and unpredictable data distribution in industrial scene.

[0061] 6、In the new data acquisition stage, the existing calibration parameters are used for real-time normalization, and the Euclidean distance matching is performed with the historical nodes to extract the minimum incremental weight, which represents the "graph cost" required for the new node to be integrated into the normal model. The innovation lies in quantifying the "degree of acceptance of new input by the system" as the edge weight increment in the graph structure. This "edge weight-behavior deviation" mapping mechanism is an important breakthrough that distinguishes this scheme from traditional classification detection. It not only solves the discrimination difficulty caused by data uncertainty, but also makes the alarm behavior more interpretable (i.e., which step deviates? How much does it deviate?). This is significantly better than the "black box decision" based on probability or classifier model.

[0062] 7、By comparing the increment calculated in real time with the adaptive threshold through a simple and clear alarm function, the triggering or ignoring of abnormal state is realized. The innovation of this step is not in its formula form, but in the "closed-loop mechanism": from historical modeling, adaptive threshold, to real-time incremental evaluation and final alarm decision, forming a complete closed-loop system driven by data, stable model and rapid response. Compared with the isolation of the existing system "each detection = one independent judgment", this method realizes the concept of "current state = perturbation detection of historical model", making the alarm behavior traceable and interpretable, and greatly improving user trust and industrial deployment efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 The flowchart of the present application. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0065] Embodiment, refer to Figure 1 , the detection method of the ethylene oxide alarm instrument, comprising:

[0066] In the absence of ethylene oxide environment, according to the time sequence of each sensor acquisition of the original voltage readings, calculate its historical average and maximum deviation, and based on this to each sensor readings for normalization processing;

[0067] The normalized data vector of each historical moment is constructed as a node, and a node set is generated;

[0068] According to the Euclidean distance, the similarity between nodes is calculated, and a symmetric distance matrix is constructed;

[0069] According to the distance matrix, a custom method is used to construct the historical minimum spanning tree, and the total weight of the historical baseline model is calculated;

[0070] According to the minimum distance of the node joining the minimum spanning tree, the alarm threshold is dynamically calculated;

[0071] In the target detection moment, real-time sensor data is collected, and normalized according to the established historical calibration parameters;

[0072] The minimum incremental weight between the current normalized vector and the historical node is calculated;

[0073] The increment is compared with the adaptive threshold value, if it exceeds the threshold value, it is determined as ethylene oxide leakage and triggers an alarm, otherwise it maintains normal state.

[0074] By constructing a complete detection process, including historical data acquisition and normalization, node construction, distance matrix establishment, minimum spanning tree model, adaptive threshold setting, real-time data acquisition, incremental weight calculation and alarm decision, the entire detection scheme has integrity, logical closed loop and dynamic adaptability. The problem of not being able to adapt to environmental drift, sensor aging and fixed threshold false alarm in traditional method is solved. Especially for ethylene oxide, which is volatile and changes rapidly in concentration, it is difficult for the fixed threshold method to ensure sensitivity and stability. This method expresses the structure through the graph model, which significantly enhances the robustness of abnormal judgment. The historical behavior model can be dynamically updated, allowing the system to adapt to new working environment and equipment state; the incremental weight can directly quantify "whether to deviate from normal", and the adaptive threshold setting enhances the immunity of the model to environmental changes; a whole chain methodology from historical modeling to real-time decision-making is constructed, which is suitable for various high-risk scenarios such as industry, medical treatment and laboratory.

[0075] The in the absence of ethylene oxide environment, according to the time sequence of each sensor acquisition of the original voltage readings, calculate its historical average and maximum deviation, and based on this to each sensor readings for normalization processing, specifically includes:

[0076] In the absence of ethylene oxide environment, according to the time sequence of each sensor acquisition of the original voltage readings, calculate its historical average and maximum deviation, and based on this to each sensor readings for normalization processing, specifically includes: i{1,2,...,P} sequentially collect sensor raw readings;

[0077] Record S j (t i ) for each sensor j∈{1,2,...,M}; j (t i ) is the raw analog voltage reading of sensor j at time t i ;

[0078] Obtain historical baseline data for subsequent statistics and normalization;

[0079] Calculate the historical average voltage of sensor j Remove static offset, centralize sensor readings;

[0080] Calculate the historical maximum absolute deviation of sensor j Quantify the amplitude of the reading fluctuation, which is the denominator for normalization;

[0081] Calculate the normalized reading of sensor j at any time t

[0082] Map the reading to [-1,1] to eliminate the difference in dimension and amplitude.

[0083] By collecting raw voltage data in the absence of ethylene oxide in time series, and sequentially calculating the average value and maximum absolute deviation, the sensor data is normalized to complete the construction of the unified standard of historical data. The problem of multi-dimensional data incomparability caused by different sensor models, parameter drift and different dimensions is solved; especially for the zero drift and gain change of long-term used sensors, conventional linear calibration is difficult to completely handle, while the scheme constructs a unified data semantic space by statistical reference. The unified dimension expression of different sensor data is realized, the stability of the subsequent modeling stage is improved; the interference of external environment and hardware difference on the result is eliminated, so that the system can run stably in the field complex environment for a long time. As a preprocessing step, normalization also ensures the mathematical effectiveness of the subsequent Euclidean distance and incremental calculation.

[0084] The normalized data vector of each historical time is constructed as a node, and a node set is generated, which specifically includes:

[0085] For each historical time t i , set the normalized vector as X i =[x1(t i ),x2(t i ),...,x M (t i )] T ; integrate multi-sensor data to form a vector representation that can calculate the distance;

[0086] Set the historical node set as Where element v i The corresponding vector X i ; is a mapping relationship, indicating that the node and the vector are one-to-one corresponding; the vector is mapped into a graph node, preparing for the input of a graph algorithm.

[0087] By encapsulating the normalized sensor data vector at each moment as a node in the graph model, a complete historical node set is formed, and the node is kept one-to-one mapping with the original time series data. The problem that the original data cannot participate in the graph model or vector calculation is solved. By "nodalizing" the time series data, the time series information has a structural representation capability, which is converted into an input format suitable for graph theory processing. While maintaining data continuity, the data has a topological structure expression capability, which can be used for graph algorithm modeling; the node set is convenient for subsequent combination with the graph distance matrix for global pattern analysis; the calculation efficiency and scalability of the system in high-dimensional multivariate monitoring scenarios are improved, providing a data basis for further introduction of graph neural networks or clustering analysis algorithms.

[0088] The similarity between nodes is calculated according to the Euclidean distance, and a symmetric distance matrix is constructed, specifically including:

[0089] The edge set of the node complete graph is set as E={(i,k)|1≤i

[0090] The Euclidean distance between nodes v i and v k is calculated to measure the similarity of nodes and provide a measure for the weight of the minimum spanning tree;

[0091] All D i,k are stored in the symmetric matrix [D i,k ] P×P ; unified storage for quick search and sorting.

[0092] By calculating the Euclidean distance between the vectors of each pair of nodes, a complete similarity measurement graph between nodes is constructed, and the results are arranged into a symmetric matrix for unified management. The problem of unclear similarity quantization between data is solved. Traditional methods usually classify and judge based on fixed thresholds or coarse-grained intervals, which is difficult to refine the difference degree. Using Euclidean distance provides a fine and continuous abnormal measurement method. Through the matrix structure, quick indexing and sorting can improve the efficiency and accuracy of subsequent minimum spanning tree construction; the symmetric structure is naturally adapted to graph theory and network analysis algorithms; the unified similarity evaluation mechanism also avoids the problem that different algorithms have different processing methods for different nodes.

[0093] The historical minimum spanning tree is constructed by using a self-defined method according to the distance matrix to obtain a historical baseline model, and a total weight thereof is calculated, specifically including:

[0094] The set {D i,k |(i,k)∈E} is sorted in ascending order, denoted as Wherein, D (m) is the mth smallest distance; the merging order is determined, and the smallest weight edge is preferentially selected;

[0095] The edge set T0 is initialized The connected component set C is initialized The empty tree and the independent component set are established, and the merging is prepared;

[0096] For m=1 to The following steps are executed:

[0097] S110, the mth smallest distance and the corresponding node pair (i m ,k m ) are taken, satisfying

[0098] S120, two groups containing and in the set are found, denoted as C a and C b ;

[0099] S130, if C a ≠C b :

[0100] S131, the edge (i m ,k m ) is added to T0;

[0101] S132, the group C

[0102] S140, steps S110 to S130 are repeated until |T0|=P-1;

[0103] The minimum spanning tree without loops and connected is constructed;

[0104] The total weight of the baseline is calculated The historical baseline MST weight is recorded as a comparison reference.

[0105] The Kruskal algorithm or its variants are used to select the shortest edge from the distance matrix to construct a connected and acyclic graph, complete the structural modeling of normal behavior, and record the total edge weight of the generated tree as a benchmark. The problem of how to abstract the "normal data structure" under the assumption of no model is solved. Traditional models rely on formulas or expert rules for modeling, lacking an interpretable structure. This step uses graph theory to construct a naturally generated, non-redundant normal pattern. It forms a skeleton expression of the system's historical state and has the ability to resist local disturbances; the sparsity of the structure brings higher computational efficiency and abnormal separability; the total weight of the historical MST as a stable reference value provides a quantitative evaluation standard for subsequent incremental detection.

[0106] The alarm threshold is dynamically calculated according to the minimum distance of the node joining the minimum spanning tree, specifically including:

[0107] When q = {2, 3,..., P} nodes join, the minimum value of the nearest distance from the existing nodes is calculated:

[0108] The minimum edge weight of the qth historical node integrated into the existing MST is measured;

[0109] The adaptive alarm threshold is set as The maximum "nearest neighbor distance" in history is taken as the alarm threshold.

[0110] By recording the minimum edge weight when each new node joins the historical MST, and taking the historical maximum edge weight as the adaptive threshold for alarm judgment, the system sensitivity is dynamically adjusted. The problem of difficult static threshold setting and easy failure is solved. In actual use, the system state may change with temperature, time, humidity, etc., and a fixed threshold value is easy to cause frequent false alarms or missed alarms. The system alarm logic can adjust itself with the change of historical data structure; the dynamic threshold avoids artificial limitation and improves the algorithm's ability to adapt to complex environments; the threshold based on graph structure enhances the system's sensitivity to structural variation and responds more intelligently.

[0111] The real-time sensor data is collected at the target detection time, and normalized according to the established historical calibration parameters, specifically including:

[0112] At the real-time monitoring time t * Collect the original readings S j (t * ) of each sensor; obtain the current monitoring data;

[0113] Apply existing calibration normalization:

[0114] X * = [x1(t * ), x2(t * ),..., xM (t * )] T ; wherein X * is a new normalized vector; real-time data is incorporated into the historical benchmark in the same dimension.

[0115] By collecting real-time raw voltage readings and normalizing based on historical statistical parameters, the new data is compared and analyzed with the historical data under the same standard. This solves the problem of direct comparison of old and new data, ensures data distribution consistency and scale uniformity; at the same time, it makes real-time monitoring feasible to compare with historical behavior. Avoid misjudgment caused by inconsistent data amplitude; improve the efficiency of legal verification of new data before entering the MST model; ensure the mathematical correctness and abnormal detection effect of subsequent incremental analysis.

[0116] The minimum incremental weight between the current normalized vector and the historical node is calculated, specifically including:

[0117] For all historical nodes i={1, 2,..., P}, calculate the new vector X * and the Euclidean distance between the historical vector X i :

[0118] Measure the similarity between new data and each historical node;

[0119] Extract the minimum increment Determine the minimum weight increment required for the new node to integrate into the MST.

[0120] By calculating the Euclidean distance between the new normalized vector and each historical node, and extracting the minimum value as the incremental weight, it is determined whether the new data belongs to the normal behavior set boundary. This solves the problem that the traditional method cannot quantify whether the current state is consistent with the history; especially in the scene where there is no label or cannot preset the abnormal model, the incremental weight provides a simple and effective abnormality measurement method. It provides a fast response mechanism, which can make a preliminary judgment only by the minimum increment; after integrating with the MST structure, the incremental weight has the meaning of "deviation from the center of the structure", which improves the system interpretability; low complexity, fast calculation, suitable for deployment on resource-constrained devices.

[0121] The increment is compared with the adaptive threshold value, if it exceeds the threshold value, it is determined that there is an ethylene oxide leakage and an alarm is triggered, otherwise it remains normal, specifically including:

[0122] The alarm judgment function is set as:

[0123]

[0124] When Alarm=1, the ethylene oxide leakage alarm is triggered; when Alarm=0, the normal monitoring state is maintained; according to the comparison of the increment and the threshold value, it is quickly decided whether to alarm.

[0125] By comparing the increment weight with the adaptive threshold value, if the threshold value is exceeded, the alarm is triggered, realizing the binary decision mechanism based on the structural deviation degree. The problem of unclear or no basis for alarm triggering is solved. The system alarm is no longer dependent on the hard-coded threshold, but comes from the structural analysis results of the historical behavior model. The alarm has quantifiable, traceable and interpretable basis; the false alarm rate is reduced while the response ability to real abnormal events is enhanced; it can be quickly deployed and compatible with various upper systems, realizing rapid linkage and feedback in industrial field.

[0126] It should be noted that in this paper, relational terms such as first and second are used merely to distinguish one entity or action from another, without necessarily requiring or implying that there is any such actual relationship or order between such entities or actions. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus including a series of elements includes not only those elements, but also other elements not explicitly listed, or inherent to such a process, method, article or apparatus.

[0127] The above is only the preferred embodiment of the present application, it should be noted that for those skilled in the art, without departing from the technical principles of the present application, a number of improvements and refinements can be made, which should also be considered as the protection scope of the present application.

Claims

1. A method of detecting an ethylene oxide alarm, characterized in that, The method comprises the following steps: In the absence of ethylene oxide environment, the original analog voltage readings of each sensor are collected in time sequence, the historical average value and the historical maximum absolute deviation are calculated, and the original analog voltage readings of each sensor are normalized based on the historical average value and the historical maximum absolute deviation; The normalized data vectors of each historical time are constructed as nodes, and a node set is generated; The similarity between nodes is calculated according to the Euclidean distance, and a symmetric distance matrix is constructed; According to the distance matrix, a historical minimum spanning tree is constructed by using a self-defined method, a historical baseline model is obtained, and the total weight is calculated; The adaptive alarm threshold is dynamically calculated according to the minimum distance of the node joining the minimum spanning tree; Real-time sensor data is collected at the target detection time, and normalized according to the predetermined historical calibration parameters; The minimum incremental weight between the current normalized vector and the historical node is calculated; The increment is compared with the adaptive alarm threshold, if the increment exceeds the adaptive alarm threshold, it is determined that there is an ethylene oxide leakage and an alarm is triggered, otherwise, the normal state is maintained.

2. The method of claim 1, wherein the ethylene oxide alarm instrument is a gas chromatograph. The method for collecting the original analog voltage readings of each sensor in time sequence, calculating the historical average value and the historical maximum absolute deviation, and normalizing the original analog voltage readings of each sensor based on the historical average value and the historical maximum absolute deviation in the absence of ethylene oxide environment, comprises the following steps: In the absence of ethylene oxide, at the time set The original analog voltage readings of each sensor are collected in sequence; For each sensor Record ;in, This represents the total number of sensors connected in parallel. For sensors At any moment The original analog voltage reading; Computing sensor Historical average voltage ; Computing sensor Historical maximum absolute deviation ; Computing sensor At any instant Normalized reading .

3. The method of claim 2, wherein the ethylene oxide alarm instrument is a gas chromatograph. The method for constructing the normalized data vectors of each historical time as nodes and generating a node set, comprises the following steps: For each historical time instant , the normalized vector is set as ; The set of historical nodes is set as ; wherein the element corresponds to the vector ; is a mapping relationship, indicating that the node and the vector are one-to-one.

4. The method of claim 3, wherein the ethylene oxide alarm instrument is a gas chromatograph. The method for calculating the similarity between nodes according to the Euclidean distance and constructing a symmetric distance matrix, comprises the following steps: The set of edges of the complete graph of nodes is set to one undirected edge for each pair of nodes; Computing node With Euclidean distance between ; All stored in symmetric matrices ​ 5. The method of claim 4, wherein the ethylene oxide alarm instrument is a Model 1000 Ethylene Oxide Monitor manufactured by the Fisher and Porter Company of Warminster, Pennsylvania. The method for constructing a historical minimum spanning tree according to the distance matrix, obtaining a historical baseline model, and calculating the total weight by using a self-defined method, comprises the following steps: Collecting sets In ascending order, denoted as ; where, is the smallest distance; initializing the edge set , initializing the connected component set ; To To the following steps are performed: S110, take the first Small distance and its corresponding node pair , meet ; S120, in the set find two packets containing and , denoted as and ; S130, if : S131, the edge joining ; S132, updating the packet ; S140, repeating steps S110 to S130 until ; Calculate baseline total weight .

6. The method of claim 5, wherein the ethylene oxide alarm instrument is a Model 1000 Ethylene Oxide Monitor manufactured by the Fisher and Porter Company of Warminster, Pennsylvania. The method for dynamically calculating the adaptive alarm threshold according to the minimum distance of the node joining the minimum spanning tree, comprises the following steps: When Upon node join, compute the minimum value of the closest distance to the existing nodes: ； Setting adaptive alarm threshold to .

7. The method of claim 6, wherein the ethylene oxide alarm instrument is a Model 1000 Ethylene Oxide Monitor manufactured by the Fisher and Porter Company of Warminster, Pennsylvania. The method for collecting real-time sensor data at the target detection time and normalizing according to the predetermined historical calibration parameters, comprises the following steps: At the moment of real-time monitoring Collecting raw analog voltage readings from each sensor ; The application of existing calibration normalization: , ; wherein, is the new normalized vector.

8. The method of claim 7, wherein the ethylene oxide alarm instrument is a gas chromatograph. The method for calculating the minimum incremental weight between the current normalized vector and the historical node, comprises the following steps: For all historical nodes , compute new vector Euclidean distance from historical vector ​ ； extracting minimum increments .

9. The method of claim 8, wherein the ethylene oxide alarm instrument is a gas chromatograph. The method for comparing the increment with the adaptive alarm threshold, if the increment exceeds the adaptive alarm threshold, it is determined that there is an ethylene oxide leakage and an alarm is triggered, otherwise, the normal state is maintained, comprises the following steps: The alarm judgment function is set as ； When an ethylene oxide leak alarm is triggered; and when normal monitoring status is maintained.

Images