A Kalman filtering method and system for airtightness detection equipment

Abstract

The present invention relates to the technical field of airtightness detection, and discloses a Kalman filtering method and system for airtightness detection equipment. A Kalman filtering method for airtightness detection equipment includes: using a graph structure model to model a multi-detection point airtightness detection system to generate a system topology representation; applying an edge functionality verification algorithm to automatically identify the dynamic topology structure of the detection network and output a real-time topology relationship matrix; using a continuous-time neural differential equation model to represent the system state evolution to generate a continuous-time domain state representation; using a time series prediction model to analyze the historical topology change pattern and predict the future topology evolution trajectory; applying a continuous deep Kalman filtering algorithm to perform real-time estimation and update of the system state and output a high-precision airtightness detection result; the present invention can adapt to the dynamic changes of the detection network topology structure in the industrial environment and improve the accuracy of airtightness detection.

Description

Technical Field

[0001] The present invention relates to the technical field of airtightness detection, and more specifically, it relates to a Kalman filtering method and system for airtightness detection equipment. Background Art

[0002] In large industrial systems, airtightness detection is an important link to ensure production safety. Especially in high-speed fluid systems such as chemical and petroleum industries, any minor leakage may lead to serious safety accidents and economic losses. Currently, airtightness detection systems generally adopt a multi-detection point distributed deployment method to capture possible leakage points in the system.

[0003] Existing airtightness detection technologies mainly include independent detection methods based on traditional Kalman filtering, which independently process and analyze the data of each detection point. However, this method has the following main technical problems: it cannot effectively utilize the spatio-temporal correlation between detection points, resulting in insufficient detection accuracy, especially in the case of minor leakage, it cannot accurately locate the leakage point; it cannot adapt to the dynamic changes of the detection network topology in the industrial environment, and operations such as valve switching and pipeline switching often lead to false alarms; it cannot accurately model and capture the rapid topology changes in high-speed gas flow systems in the continuous time domain, resulting in system response lag.

[0004] Therefore, there is a need for an airtightness detection method that can fully utilize the spatio-temporal correlation of multi-detection points, adapt to the dynamic changes of network topology, and achieve accurate modeling in the continuous time domain, so as to improve the detection accuracy and system reliability. Summary of the Invention

[0005] The present invention provides a Kalman filtering method and system for airtightness detection equipment, which solves the technical problems in the related art of being unable to effectively utilize the spatio-temporal correlation of multi-detection points, being unable to adapt to the dynamic changes of network topology, and being unable to achieve accurate modeling in the continuous time domain.

[0006] The present invention provides a Kalman filtering method for airtightness detection equipment, including:

[0007] Modeling the multi-detection point airtightness detection system using a graph structure model to generate a system topology representation;

[0008] Based on the system topology representation, applying an edge functional verification algorithm to automatically identify the dynamic topology structure of the detection network and output a real-time topology relationship matrix;

[0009] Combining the real-time topology relationship matrix, using a continuous-time neural differential equation model to represent the system state evolution and generate a continuous-time domain state representation;

[0010] Based on the real-time topological relation matrix and the time-domain state representation, use the time series prediction model to analyze the historical topological change patterns and predict the future topological evolution trajectory;

[0011] Integrate the continuous time-domain state representation and the predicted future topological evolution trajectory, apply the continuous deep Kalman filtering algorithm to perform real-time estimation and update of the system state, and output high-precision airtightness detection results.

[0012] In a preferred embodiment, the steps of modeling the multi-detection-point airtightness detection system using the graph structure model include:

[0013] Model the airtightness detection system as a graph structure, where the nodes represent the detection points and the edges represent the physical connection relationships;

[0014] Use the graph convolutional network to extract the spatial correlation features between the detection points;

[0015] Combine the gated recurrent unit to capture the time dependence of the detection point states;

[0016] Output the spatio-temporal feature representation of the detection system.

[0017] In a preferred embodiment, the steps of applying the edge functional verification algorithm to automatically identify the dynamic topological structure of the detection network include:

[0018] Automatically discover the real physical connection relationships based on the pressure wave propagation characteristics;

[0019] Introduce the graph attention algorithm to adaptively adjust the influence weights between the nodes;

[0020] Construct the weighted adjacency matrix;

[0021] Output the real-time updated topological relation matrix.

[0022] In a preferred embodiment, the steps of using the continuous time neural differential equation model to represent the system state evolution include:

[0023] Construct the continuous time differential equation of the system state evolution;

[0024] Define the neural network structure, adopting a combined structure of graph convolution and multi-layer perceptron;

[0025] Solve the differential equation through the adaptive numerical integration method to obtain the state representation in the continuous time domain.

[0026] In a preferred embodiment, the steps of using the time series prediction model to analyze the historical topological change patterns include:

[0027] Construct the spatio-temporal sequence prediction model;

[0028] Design a prediction function structure and adopt a spatio-temporal attention network architecture;

[0029] Adjust the model parameters in advance based on the predicted topological changes;

[0030] Output the predicted topological structure and the pre-adjusted model parameters.

[0031] In a preferred embodiment, the step of using the continuous-depth Kalman filter algorithm to perform real-time estimation and update of the system state includes:

[0032] Represent the system dynamics model using a stochastic differential equation;

[0033] Design an observation equation;

[0034] Dynamically generate the Kalman filter state transition matrix through the graph attention algorithm;

[0035] Execute the continuous-discrete extended Kalman filter algorithm, including a prediction step and an update step;

[0036] Output the filtered system state estimate value and the anomaly detection result.

[0037] In a preferred embodiment, the graph convolutional network includes an input layer, an intermediate layer, and an output layer, where:

[0038] The input layer receives the original node features;

[0039] The intermediate layer contains multiple hidden units and uses an activation function to enhance the non-linear expression ability;

[0040] The output layer maps the features to the hidden state space.

[0041] In a preferred embodiment, the step of automatically discovering the true physical connection relationship based on the pressure wave propagation characteristics includes:

[0042] Analyze the time-delay correlation of the pressure signals between the detection points;

[0043] Judge whether there is a physical connection between the detection points according to whether the correlation exceeds a preset threshold.

[0044] In a preferred embodiment, the spatio-temporal attention network includes:

[0045] An encoder for encoding the graph structure into a hidden state vector;

[0046] An attention module for calculating the importance weights of the graph structures at different time points;

[0047] A recurrent neural network for processing sequence information;

[0048] A decoder for decoding the hidden state into a predicted graph structure.

[0049] In a preferred embodiment, an airtightness detection system applicable to a dynamic topology network, which is used to execute a Kalman filtering method for an airtightness detection device, includes:

[0050] A multi-detection point graph structure modeling module, which is used to model the airtightness detection system and generate a system topology representation;

[0051] A dynamic topology identification module, which is used to automatically identify the dynamic topology structure of the detection network by applying an edge functionality verification algorithm;

[0052] A continuous-time state representation module, which is used to represent the system state evolution by using a continuous-time neural differential equation model;

[0053] A topology prediction module, which is used to analyze the historical topology change pattern by using a time series prediction model and predict the future topology evolution trajectory;

[0054] A continuous-depth Kalman filtering module, which is used to perform real-time estimation and update of the system state and output high-precision airtightness detection results.

[0055] The beneficial effects of the present invention are as follows:

[0056] By modeling the structural relationship between detection points, the collaborative filtering of the entire system is realized, the detection accuracy is improved, and the leakage point positioning accuracy is improved;

[0057] It can adapt to the dynamic changes of the detection network topology structure in the industrial environment in real time and shorten the response time; by representing the system state evolution through a continuous-time differential equation, the rapid topology changes in a high-speed gas flow system are accurately captured, and the model prediction error is reduced;

[0058] It can predict the future topology evolution trajectory, adjust the model parameters in advance, and improve the system stability. Description of the Drawings

[0059] Figure 1 It is a flowchart of a Kalman filtering method for an airtightness detection device according to the present invention. Detailed Embodiments

[0060] Now, the subject matter described herein will be discussed with reference to exemplary embodiments. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein. Without departing from the protection scope of the content of this specification, changes can be made to the functions and arrangements of the elements discussed. Each example can omit, substitute, or add various processes or components as needed. Additionally, the features described in some examples can also be combined in other examples.

[0061] In at least one embodiment of the present invention, a hermeticity detection method applicable to a dynamic topology network is disclosed, as follows Figure 1 shown, including the following steps:

[0062] Step 1, use a graph structure model to model the multi-detection point hermeticity detection system to generate a system topology representation;

[0063] The specific steps are as follows:

[0064] 1.1, model the hermeticity detection system as a graph structure , where the node set represents the set of detection points, and the edge set represents the physical connection relationship. For any detection points , if there is a physical connection between them, then the edge . Each detection point has a feature vector , including sensor measurement values such as pressure, temperature, and flow rate.

[0065] 1.2, use a graph convolutional network to extract the spatial correlation features between detection points. For the node feature matrix at time (where is the number of detection points, is the feature dimension), the graph convolution operation is defined as:

[0066] ;

[0067] where, represents the node hidden representation of the th layer, that is, the node features after the graph convolution operation, is the adjacency matrix with self-connection added, is the original adjacency matrix, is the order identity matrix, is the angle matrix, represents the degree of node , , represents that the input of the initial layer is equal to the original node feature matrix, is the learnable weight matrix of the th layer, represents the sigmod activation function.

[0068] In this embodiment, the graph convolutional network used consists of three layers: an input layer, an intermediate layer, and an output layer. Among them, the input layer receives the original node features, and the dimension is ; The middle layer contains 64 hidden units, using the ReLU activation function to enhance the non-linear expression ability; the output layer maps the features to -dimensional hidden state space.

[0069] The specific applications of the graph convolutional network in the airtightness detection scenario include:

[0070] Effectively extract features such as the pressure difference and temperature correlation between adjacent detection points;

[0071] According to the physical connection relationship, automatically learn the attenuation law of the pressure wave transmitted between detection points;

[0072] In the complex pipe network structure of a chemical plant, it can automatically identify the topological relationship differences caused by valve state changes.

[0073] 1.3, combined with the gated recurrent unit (GRU) to capture the time dependence of the detection point state. For the th node, the GRU update formula is:

[0074] ;

[0075] ;

[0076] ;

[0077] ;

[0078] Among them, is the reset gate, is the update gate, is the candidate hidden state, is the node at time 's hidden state, is the node at time 's hidden state, is the node at time 's input feature, represents element-wise multiplication, , , are respectively used for the calculation of the reset gate, update gate and candidate hidden state, , , are respectively the bias vectors of the reset gate, update gate and candidate hidden state, represents the sigmod activation function, is the hyperbolic tangent activation function.

[0079] The GRU network in this embodiment is configured as a single-layer structure, with the hidden state dimension set to 32, and the input is the node feature at the current moment concatenated with the hidden state at the previous moment . In the airtightness detection system, the specific applications of the GRU network include:

[0080] Capturing the time-series variation laws of parameters such as pressure and temperature at the detection points, such as identifying the propagation delay of pressure waves in the oil pipeline system;

[0081] Learning the short-term and long-term memory characteristics of system parameters under different operating conditions (such as flow rate changes, temperature fluctuations);

[0082] In a high-speed gas flow system, being able to distinguish the pressure fluctuations caused by normal operation from the abnormal pressure drop patterns caused by leakage.

[0083] 1.4, outputting the spatio-temporal feature representation of the detection system , where is the hidden state dimension. This representation simultaneously contains the system topology structure information and the time evolution characteristics of the node states.

[0084] Step 2, based on the system topology representation, applying the edge functional verification algorithm to automatically identify the dynamic topology structure of the detection network and output the real-time topology relationship matrix;

[0085] The specific steps are as follows:

[0086] 2.1, automatically discovering the real physical connection relationship based on the pressure wave propagation characteristics. For any two detection points and , by analyzing the time-delay correlation of the pressure signals to determine whether there is a physical connection between them:

[0087] ;

[0088] where, represents the time-delay correlation of the pressure signal, and are the pressure values of the detection points and at time , and are the means of the detection points and respectively, is the observation window size, is the time-delay parameter. If ( is the preset threshold), then it is considered that the detection points and There is a physical connection between them.

[0089] 2.2, Introduce the graph attention algorithm to adaptively adjust the influence weights between nodes. For each pair of connected nodes , calculate the attention coefficient:

[0090] ;

[0091] ;

[0092] Among them, is the unnormalized attention score, is the normalized attention coefficient, , are the attention weight matrix and the feature transformation matrix respectively, represents the vector concatenation operation, represents node 's neighbor set, LeakyReLU is an activation function with a small negative slope part, , are the hidden state representations of points and respectively, represents the exponential function.

[0093] 2.3, Construct the weighted adjacency matrix , where the element represents the influence intensity of node on node at time .

[0094] 2.4, Output the topologic relation matrix that is updated in real time, where represents the topologic relation matrix at time , represents the weighted adjacency matrix at time , represents the node feature matrix at time . This matrix synthesizes the physical connection relationship and node state characteristics of the system and can reflect the dynamic changes of the system topology structure.

[0095] Step 3, Combine the real-time topologic relation matrix and use the continuous-time neural differential equation model to represent the system state evolution and generate the continuous-time domain state representation;

[0096] The specific steps are as follows:

[0097] 3.1, Construct the continuous-time differential equation for the system state evolution. For the system state at time , its evolution law is expressed as:

[0098] ;

[0099] Among them, represents how the system state changes continuously over time, is the node hidden state matrix, is the time-varying graph structure, is a neural network with parameter , describing the state change rate.

[0100] 3.2. Define the structure of the neural network . Adopt a combined structure of graph convolution and multi-layer perceptron:

[0101] ;

[0102] Among them is a parameterized neural network function, represents the set of learnable parameters of the neural network, represents the system state vector at time , represents the changing graph structure at time , is the time variable, GCN represents the graph convolutional network, MLP represents the multi-layer perceptron, is the time-dependent coefficient.

[0103] The neural differential equation network in this embodiment is specifically implemented as:

[0104] The GCN part adopts a 2-layer structure, with 32 hidden units in each layer, and uses the Tanh activation function to ensure smooth and differentiable output;

[0105] The MLP part includes a 3-layer fully connected network, and the dimensions of the hidden layers are 64, 32, and 32 respectively, and the Tanh activation function is also used.

[0106] The time-dependent coefficient is set to 0.01 to balance the time sensitivity of the system.

[0107] In the airtightness detection system, the specific applications of this network include:

[0108] Real-time tracking of pressure wave propagation in the complex pipe network of the petrochemical plant, and accurately modeling the continuous change of the system state;

[0109] Adapting to topological mutations caused by operations such as rapid opening and closing of valves, and continuously capturing the transition state of the system;

[0110] In high-speed gas flow scenarios, it is possible to distinguish normal operating condition fluctuations from the slow pressure drop caused by minor leaks.

[0111] 3.3, Solve the differential equation through an adaptive numerical integration method to obtain the state representation in the continuous time domain. For the time interval , use the 5th-order Runge-Kutta method to solve:

[0112] ;

[0113] where represents the system state vector at the termination time , represents the system state vector at the initial time , denotes the definite integral operation over the time interval , is the neural network function with parameter , representing the instantaneous rate of change of the system state, represents the system state vector at time , represents the change diagram structure at time , represents the infinitesimal time increment.

[0114] 3.4, Output the system state representation in the continuous time domain , which can accurately capture the continuous state changes in the high-speed gas flow system and overcome the information loss problem caused by discrete time step updates.

[0115] Step 4, Based on the real-time topological relationship matrix and the time domain state representation, use the time series prediction model to analyze the historical topological change patterns and predict the future topological evolution trajectory;

[0116] The specific steps are as follows:

[0117] 4.1, Construct a spatio-temporal sequence prediction model. This model takes the topological structure sequence at the past time points as the input and predicts the topological structure at the future time , where , , , the first, second, and third respectively represent the historical topological structure sequences at times , , , , represents the earliest historical time point, represents the historical time window size, denotes the time step, denotes the current time point.

[0118] 4.2, Design the prediction function with the structure of a spatio-temporal attention network architecture:

[0119] ;

[0120] Among them, denotes the topological structure matrix of the future moment predicted by the model , , , respectively denote the topological structure matrices of the current moment , the previous moment , and the earliest considered historical moment ; denotes the time step, denotes the size of the historical time window, denotes the current time point;

[0121] It is implemented as a recurrent attention network:

[0122] ;

[0123] ;

[0124] ;

[0125] Among them, denotes the initial hidden state vector, denotes the hidden state vector after processing the th historical time point, denotes the topological structure matrix of the future moment predicted by the model, denotes the encoder function, denotes the long short-term memory network unit, denotes the attention mechanism function, denotes the system topological structure at the historical moment , denotes the time step, denotes the index of the historical time point, denotes the decoder function, denotes the hidden state vector after processing the th historical time point.

[0126] The spatio-temporal attention network in this embodiment is specifically implemented as:

[0127] The Encoder uses a 3-layer graph convolutional network to encode the graph structure into a hidden state vector;

[0128] The Attention module uses scaled dot-product attention to calculate the importance weights of the graph structure at different time points;

[0129] The LSTM is configured as a 2-layer structure, with each layer containing 128 hidden units;

[0130] The Decoder uses a 3-layer transposed graph convolutional network to decode the hidden state into the predicted graph structure.

[0131] In the airtightness detection system, the specific applications of this network include:

[0132] Predicting topological structure changes in the oil pipeline system caused by temperature changes or flow adjustments;

[0133] Identifying and predicting the periodic change patterns of the system state caused by periodic operations (such as timed valve opening and closing);

[0134] In complex industrial systems, anticipating upcoming topological changes in advance, such as pipeline reconstruction caused by production process switching.

[0135] 4.3, Adjusting the model parameters in advance based on the predicted topological changes. For the predicted topological structure , calculate the difference from the current topology :

[0136] ;

[0137] If ( is a preset threshold, represents the Frobenius norm), then trigger the pre-adjustment process of the model parameters, where represents the difference matrix between the predicted topology and the current topology, represents the topology structure matrix at the future time predicted by the model, and represents the topology structure matrix at the current time

[0138] 4.4, Output the predicted topology structure and the pre-adjusted model parameters, realizing the advance prediction and adaptation to topological changes, and improving the stability and accuracy of the system during topological changes.

[0139] Step 5, Integrate the continuous-time domain state representation and the predicted future topological evolution trajectory, and apply the continuous-depth Kalman filter algorithm to estimate and update the system state in real time, and output high-precision airtightness detection results;

[0140] The specific steps are as follows:

[0141] 5.1, Represent the system dynamics model using stochastic differential equations. The evolution law of the system state is expressed as:

[0142] ;

[0143] where, represents the differential change of the system state, is the drift function, representing the deterministic part; is the diffusion function, representing the stochastic part; is the Wiener process, representing the system noise.

[0144] 5.2, Design the observation equation. For the system state at time , the relationship between the observed value is:

[0145] ;

[0146] where, represents the observed value vector at time , is the observation matrix, is the observation noise, following the Gaussian distribution .

[0147] 5.3, Dynamically generate the Kalman filter state transition matrix through the graph attention algorithm. At discrete time points , the state transition matrix is generated through the graph attention convolutional network:

[0148] ;

[0149] where, represents the state transition matrix from time to , is the graph structure at time , is the node hidden state matrix at the previous time, and GATConv represents the graph attention convolution operation.

[0150] 5.4, Execute the continuous-discrete extended Kalman filter algorithm. It includes a prediction step and an update step:

[0151] Prediction step:

[0152] ;

[0153] ;

[0154] Update steps:

[0155] ;

[0156] ;

[0157] ;

[0158] Among them, is the predicted value of the state at time is the prediction error covariance matrix, is the Kalman gain, is the state update value, is the updated error covariance matrix, is the process noise covariance matrix, represents the observation matrix, represents the transpose matrix of represents the actual observed value at time represents the expected observed value calculated based on the predicted state, represents the identity matrix, represents the observation noise covariance matrix.

[0159] The specific implementation of the continuous depth Kalman filtering algorithm in this embodiment is as follows:

[0160] The integration process uses the Dormand-Prince method with adaptive step size control, and the maximum relative error is set to 1e-6;

[0161] The process noise covariance matrix and the observation noise covariance matrix are respectively configured as diagonal matrices, where the diagonal elements of are 0.01, and the diagonal elements of

[0162] In the airtightness detection system, the specific applications of this algorithm include:

[0163] Filtering sensor noise in the multi-detection point system of chemical plants to improve the accuracy of measured values such as pressure, temperature, and flow rate;

[0164] Continuously estimating the state in the high-speed gas flow system and real-time detecting system parameter anomalies caused by small leaks;

[0165] In the complex pipe network structure, accurately locating the leak point through collaborative filtering, especially for the multi-point leak situation in large industrial systems;

[0166] 5.5, output the filtered system state estimation value and the anomaly detection result. By analyzing the deviation from the predicted value to determine whether there is an airtightness anomaly in the system, and combine the graph structure information to locate the specific position where the anomaly occurs.

[0167] Real application example of this embodiment

[0168] This embodiment has been actually applied in the airtightness detection system of the ethylene production unit in a large petrochemical plant. The unit has a complex pipe network structure, including 86 airtightness detection points, which are interconnected by valves and pipes to form a dynamically changing network topology. Due to frequent process switches during production, the valve switch states change at any time, resulting in continuous dynamic changes in the topology of the detection network.

[0169] Previously, the petrochemical plant used the traditional independent Kalman filtering method to monitor each detection point, but there were the following problems in actual applications:

[0170] Unable to accurately locate the leakage position: When there is a small leakage, multiple interconnected detection points show anomalies at the same time, and it is difficult for the independent filtering method to determine the position of the real leakage point;

[0171] High false alarm rate: Topology changes (such as valve switches) often cause sudden changes in sensor data, and the system cannot distinguish fluctuations caused by normal operations from leakage anomalies;

[0172] Response lag: The existing methods cannot effectively handle topology changes across sampling intervals in a high-speed gas flow system, resulting in system response lag.

[0173] The specific implementation of this embodiment in this ethylene unit includes the following examples:

[0174] Example of multi-detection point system modeling:

[0175] In actual applications, we collected the basic information of 86 detection points, including sensor types, position coordinates, and initial connection relationships, and constructed an initial graph structure model. The characteristic information collected at each detection point includes data such as pressure (Bar), temperature (°C), flow rate (m³ / h), and gas component concentration, with a sampling frequency of 10 Hz. Table 1 shows an example of the characteristic information of some core detection points:

[0176] Table 1: Example of characteristic information of some detection points;

[0177]

[0178] The input of the graph convolutional network is a feature matrix of 86×12 dimensions (86 detection points, with 12 feature parameters for each detection point), and the output is a hidden state matrix of 86×32 dimensions, effectively extracting the spatial correlation features between detection points.

[0179] Example of edge functionality verification algorithm:

[0180] In practical applications, this algorithm automatically identifies the true physical connection relationship of the detection network by analyzing the propagation characteristics of pressure waves in the system. We inject a tiny pressure pulse signal (peak value 0.1 Bar, duration 0.5 seconds) at a key position (P01) in the system, and then analyze the time-delay response of the pressure signals at other detection points. The results are shown in Table 2:

[0181] Table 2: Results of time-delay correlation analysis of pressure wave propagation;

[0182]

[0183] Example of topology prediction and dynamic adaptation:

[0184] During the process switching, the system can predict future topology changes based on historical topology change patterns. Table 3 shows the prediction results of the system for the key valve state changes during a typical process switching operation (from normal production mode to low-load mode):

[0185] Table 3: Results of topology prediction and dynamic adaptation example;

[0186]

[0187] Verification of technical effects:

[0188] We select "improvement in detection accuracy" and "improvement in dynamic adaptation ability" as the two most important technical effects of this embodiment for verification and analysis:

[0189] Verification of the improvement effect of detection accuracy:

[0190] To verify the improvement effect of this embodiment in the airtightness detection accuracy, we simulated different degrees of leakage in different areas of the ethylene plant and compared the detection results of the traditional independent Kalman filtering method with this embodiment. Table 4 shows the comparison results of 10 simulated leakage tests:

[0191] Table 4: Results of comparative experiments on airtightness detection accuracy;

[0192]

[0193] As can be seen from Table 4, compared with the traditional method, the detection time of this embodiment is shortened by 74.0% on average, and the positioning accuracy of the leakage point is improved by 71.1%, which fully proves the significant advantages of this embodiment in terms of detection accuracy. Especially for minor leaks (<0.5 L / min), the detection ability of this embodiment is improved more significantly.

[0194] Verification of the improvement effect of dynamic adaptation ability:

[0195] To verify the improvement effect of this embodiment in terms of the ability to adapt to dynamic topology changes, we recorded the system's response to topology changes during the production process switch. Table 5 shows the comparison results between the traditional method and this embodiment in 3 typical process switch scenarios:

[0196] Table 5: Comparison experimental results of the ability to adapt to dynamic topology changes;

[0197]

[0198] As can be seen from Table 5, in the scenario of dynamic topology changes, the response time of this embodiment is shortened by 85.2% on average, and the number of false alarms is reduced from an average of 11.33 times to 0.33 times, which fully verifies the significant improvement of this embodiment in terms of the ability to adapt to dynamic topology structures.

[0199] The embodiments of the present invention have been described above, but these embodiments are not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of this embodiment, those of ordinary skill in the art can also make more forms of equivalent embodiments, all of which fall within the protection scope of this embodiment.

Claims

1. A Kalman filtering method for airtightness detection equipment, characterized in that It includes the following steps: Model the multi-detection point airtightness detection system using a graph structure model to generate a system topology representation; Based on the system topology representation, apply the edge functional verification algorithm to automatically identify the dynamic topology structure of the detection network and output a real-time topology relationship matrix; Combined with the real-time topology relationship matrix, use the continuous-time neural differential equation model to represent the system state evolution and generate a continuous-time domain state representation; Based on the real-time topology relationship matrix and the time-domain state representation, use the time series prediction model to analyze the historical topology change pattern and predict the future topology evolution trajectory; Integrating the continuous-time domain state representation and the predicted future topology evolution trajectory, apply the continuous deep Kalman filtering algorithm to perform real-time estimation and update of the system state, and output high-precision airtightness detection results, including: Use a stochastic differential equation to represent the system dynamics model; Design an observation equation; Dynamically generate the Kalman filter state transition matrix through the graph attention algorithm; Execute the continuous-discrete extended Kalman filtering algorithm, including the prediction step and the update step; Output the filtered system state estimation value and the anomaly detection result.

2. A Kalman filtering method for an airtightness detection device according to claim 1, characterized in that, The step of modeling the multi-detection point airtightness detection system using the graph structure model includes: Model the airtightness detection system as a graph structure, where nodes represent detection points and edges represent physical connection relationships; Adopt a graph convolutional network to extract the spatial correlation features between detection points; Combine a gated recurrent unit to capture the time dependence of the detection point state; Output the spatio-temporal feature representation of the detection system.

3. A Kalman filtering method for an airtightness detection device according to claim 1, characterized in that, The step of applying the edge functional verification algorithm to automatically identify the dynamic topology structure of the detection network includes: Automatically discover the true physical connection relationship based on the pressure wave propagation characteristics; Introduce the graph attention algorithm to adaptively adjust the influence weights between nodes; Construct a weighted adjacency matrix; Output the real-time updated topology relationship matrix.

4. A Kalman filtering method for an airtightness detection device according to claim 1, characterized in that, The step of using the continuous-time neural differential equation model to represent the system state evolution includes: Construct a continuous-time differential equation for the system state evolution; Define the neural network structure, using a combination structure of graph convolution and a multi-layer perceptron; Solve the differential equation through an adaptive numerical integration method to obtain the state representation in the continuous-time domain.

5. A Kalman filtering method for an airtightness detection device according to claim 1, characterized in that, The step of using the time series prediction model to analyze the historical topology change pattern includes: Construct a spatio-temporal sequence prediction model; Design the prediction function structure, using a spatio-temporal attention network architecture; Adjust the model parameters in advance based on the predicted topology changes; Output the predicted topology structure and the pre-adjusted model parameters.

6. A Kalman filtering method for an airtightness detection device according to claim 2, characterized in that, The graph convolutional network includes an input layer, an intermediate layer, and an output layer, where: The input layer receives the original node features; The intermediate layer contains multiple hidden units and uses an activation function to enhance the non-linear expression ability; The output layer maps the features to the hidden state space.

7. A Kalman filtering method for an airtightness detection device according to claim 3, characterized in that, The step of automatically discovering the true physical connection relationship based on the pressure wave propagation characteristics includes: Analyze the time-delay correlation of the pressure signals between detection points; Judge whether there is a physical connection between detection points according to whether the correlation exceeds a preset threshold.

8. A Kalman filtering method for an airtightness detection device according to claim 5, characterized in that, The spatio-temporal attention network includes: An encoder for encoding the graph structure into a hidden state vector; An attention module for calculating the importance weights of the graph structure at different time points; A recurrent neural network for processing sequence information; A decoder for decoding hidden states into a predicted graph structure.

9. An airtightness detection system applicable to a dynamic topology network, which is used to execute a Kalman filtering method for an airtightness detection device according to any one of claims 1-8, characterized in that, It includes: A multi-detection point graph structure modeling module for modeling an airtightness detection system and generating a system topology representation; A dynamic topology recognition module for automatically recognizing the dynamic topology structure of a detection network by applying an edge functionality verification algorithm; A continuous-time state representation module for representing the system state evolution using a continuous-time neural differential equation model; A topology prediction module for analyzing historical topology change patterns using a time series prediction model and predicting future topology evolution trajectories; A continuous-depth Kalman filtering module for performing real-time estimation and update of the system state and outputting high-precision airtightness detection results.

Images