A room differential pressure monitoring error correction method based on self-supervised learning

Abstract

The application discloses a room differential pressure monitoring error correction method based on self-supervised learning, which comprises the following steps: step one, collecting differential pressure monitoring data and environment running state data; step two, calculating differential pressure gradient enhancement features through space-time gradient feature analysis; step three, fusing the differential pressure monitoring data, the differential pressure gradient enhancement features and the environment running state data; step four, constructing a differential pressure behavior self-learning constraint task; step five, constructing a differential pressure behavior feature model through an improved Koopman neural network, and training and optimizing the differential pressure behavior feature model based on the differential pressure behavior self-learning constraint task; step six, loading the trained differential pressure behavior feature model to perform differential pressure prediction and determine whether there is a differential pressure monitoring error; and step seven, performing differential pressure correction through a neighborhood fusion compensation method. Through the improved Koopman neural network, the application improves the accuracy of room differential pressure monitoring and the reliability of error correction.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring and industrial data analysis technology, and in particular to a method for correcting room pressure difference monitoring errors based on self-supervised learning. Background Technology

[0002] With the increasing demands for air cleanliness and airflow organization in cleanrooms, hospital negative pressure wards, biological laboratories, and pharmaceutical production workshops, room differential pressure monitoring systems are playing an increasingly important role in environmental safety control. Room differential pressure is a crucial parameter for ensuring airflow in a predetermined direction, preventing contamination diffusion, and maintaining cleanliness levels. It requires continuous monitoring of the differential pressure at each monitoring point using differential pressure sensors and real-time feedback to the environmental control system. In practical applications, existing differential pressure monitoring methods mainly rely on single sensor measurements or simple threshold judgments for monitoring and alarms, and still generally suffer from the following problems:

[0003] Differential pressure sensors are susceptible to factors such as temperature changes, airflow disturbances, equipment vibration, and sensor aging during long-term operation, leading to drift, instantaneous fluctuations, or abnormal jumps in measurement data. This reduces the stability and reliability of the monitoring data, thus affecting the accuracy of differential pressure monitoring results. Spatial airflow coupling exists between multiple monitoring points, and the differential pressure changes at each point exhibit significant spatiotemporal correlations. However, traditional monitoring systems typically analyze only the differential pressure at a single point independently, failing to effectively utilize the differential pressure gradient relationships between rooms for comprehensive judgment, resulting in insufficient ability to identify abnormal differential pressure changes. Existing methods, when processing differential pressure time series data with nonlinear and non-stationary characteristics, often employ simple filtering or empirical rules for error correction, lacking the ability to model the dynamic evolution of the differential pressure system. Under complex operating conditions, it is difficult to accurately distinguish between normal operating disturbances and actual monitoring errors, thereby reducing the reliability and intelligence level of the differential pressure monitoring system.

[0004] Therefore, how to provide a room pressure differential monitoring error correction method based on self-supervised learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for correcting room differential pressure monitoring errors based on self-supervised learning. This invention performs spatiotemporal gradient feature analysis on differential pressure monitoring data and constructs a differential pressure behavior feature vector sequence by combining it with environmental operating status data. An improved Koopman neural network is then used to establish a differential pressure behavior feature model, enabling dynamic learning and prediction of room differential pressure variation patterns. Furthermore, the deviation between the predicted differential pressure results and the actual monitoring data is used to determine the differential pressure monitoring error, and a neighborhood fusion compensation method is used to correct abnormal monitoring values, thereby obtaining more accurate differential pressure monitoring results. This invention fully utilizes self-supervised learning, spatiotemporal feature modeling, and Koopman dynamic modeling techniques to achieve intelligent analysis and error correction of differential pressure monitoring data, possessing advantages such as high differential pressure monitoring accuracy, strong error identification capability, and good system stability.

[0006] A method for correcting room pressure differential monitoring errors based on self-supervised learning according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Collect differential pressure monitoring data and environmental operation status data at each monitoring point;

[0008] Step 2: Based on differential pressure monitoring data, calculate the differential pressure gradient enhancement characteristics through spatiotemporal gradient feature analysis;

[0009] Step 3: Combine and fuse differential pressure monitoring data, differential pressure gradient enhancement features, and environmental operation status data to obtain a differential pressure behavior feature vector sequence;

[0010] Step 4: Construct a self-learning constraint task for pressure difference behavior based on the feature vector sequence of pressure difference behavior;

[0011] Step 5: Construct a differential pressure behavior feature model using an improved Koopman neural network, and train and optimize the differential pressure behavior feature model based on a differential pressure behavior self-learning constraint task to obtain the trained differential pressure behavior feature model; the improved Koopman neural network includes a differential pressure feature encoding network, a topological constraint feature fusion module, a Koopman dynamics operator layer, a state update module, a differential pressure prediction decoding network, and a feature reconstruction branch;

[0012] Step 6: Input the real-time generated differential pressure behavior feature vector sequence into the trained differential pressure behavior feature model to predict differential pressure and determine whether there is a differential pressure monitoring error;

[0013] Step 7: If there is a differential pressure monitoring error, the differential pressure is corrected by the neighborhood fusion compensation method to obtain the corrected differential pressure monitoring value, and the corrected differential pressure monitoring value is output to the room environment monitoring system.

[0014] Optionally, step one specifically includes:

[0015] Differential pressure sensors are deployed between the target room and adjacent rooms, between the target room and the corridor, or between adjacent functional areas. Differential pressure monitoring values ​​at each sampling time point are collected according to a preset sampling cycle to form differential pressure monitoring data.

[0016] Simultaneously collect environmental operating status data corresponding to differential pressure monitoring data, including access control opening and closing status, fan operating status, and air conditioning system operating status.

[0017] Optionally, step two specifically includes:

[0018] The pressure gradient enhancement features include pressure difference time difference features, pressure difference change rate features, pressure difference spatial difference features, and pressure difference trend offset features;

[0019] Based on differential pressure monitoring data, the difference between differential pressure monitoring values ​​at adjacent sampling time points is calculated at the same monitoring point to obtain differential pressure time difference characteristics. The ratio of differential pressure time difference to the time difference between adjacent sampling time points is also calculated to obtain differential pressure change rate characteristics.

[0020] Based on the spatial connection between each monitoring point, the difference in pressure difference monitoring values ​​between adjacent monitoring points is calculated at the same sampling time point to obtain the spatial difference characteristics of pressure difference.

[0021] Set the sliding time window length, calculate the average value of the differential pressure monitoring value of each monitoring point within the sliding time window length to obtain the differential pressure sliding mean, calculate the difference between the differential pressure monitoring value of each monitoring point at the current sampling time point and the corresponding differential pressure sliding mean to obtain the differential pressure trend offset characteristics.

[0022] Optionally, step three specifically includes:

[0023] Align the differential pressure monitoring data, differential pressure gradient enhancement features, and environmental operation status data according to the sampling time point and monitoring point dimension;

[0024] At each sampling time point, the differential pressure monitoring value, differential pressure time difference characteristics, differential pressure change rate characteristics, differential pressure spatial difference characteristics, differential pressure trend offset characteristics, access control opening and closing status, fan operation status, and air conditioning system operation status of each monitoring point are spliced ​​together according to the monitoring point dimension, and a differential pressure behavior feature vector sequence is formed according to the time step order.

[0025] Optionally, step four specifically includes:

[0026] The pressure difference behavior self-learning constraint task includes a pressure difference time series prediction task, a pressure difference change consistency constraint task, and a pressure difference data reconstruction task.

[0027] Set the prediction time window length, take the pressure difference behavior feature vector sequence within the prediction time window length as the input sample, and take the data at the t-th sampling time point as the prediction target;

[0028] The pressure difference time series prediction task is as follows: take the pressure difference monitoring value corresponding to the t-th sampling time point as the prediction target to obtain the predicted pressure difference value; calculate the mean square error between the predicted pressure difference value and the pressure difference monitoring value to obtain the pressure difference time series prediction loss.

[0029] The pressure difference change consistency constraint task is as follows: take the pressure difference spatial difference feature at the t-th sampling time point as the prediction target to obtain the predicted pressure difference spatial difference feature; calculate the mean square error between the predicted pressure difference spatial difference feature and the pressure difference spatial difference feature to obtain the pressure difference change consistency constraint loss.

[0030] The differential pressure data reconstruction task is as follows: input the differential pressure behavior feature vector at the t-th sampling time point into the reconstruction branch to obtain the reconstructed differential pressure behavior feature vector, and calculate the mean square error between the reconstructed differential pressure behavior feature vector and the differential pressure behavior feature vector to obtain the differential pressure data reconstruction loss.

[0031] Three loss weighting coefficients are set to weight and fuse the pressure difference time series prediction loss, pressure difference change consistency constraint loss, and pressure difference data reconstruction loss to obtain the total loss function of pressure difference behavior.

[0032] Optionally, step five specifically includes:

[0033] Define an embedding mapping space, and map the pressure difference behavior feature vector sequence to the embedding mapping space through a linear transformation to obtain the mapped feature vector sequence;

[0034] The pressure difference feature encoding network generates a sequence of potential state vectors through multi-scale feature fusion mapping.

[0035] In the topology constraint feature fusion module, a topology relation matrix of monitoring points is constructed. Specifically, based on the spatial connection relationship between each monitoring point, if there is an airflow connection between monitoring point i and monitoring point j, the value of the element in the i-th row and j-th column of the topology relation matrix of monitoring points is set to 1, otherwise it is set to 0.

[0036] Define the latent state mapping weight matrix, the topological relation mapping weight matrix, and the bias vector;

[0037] The potential state vector at the current sampling time point is mapped to a potential state mapping vector through the potential state mapping weight matrix, and then mapped to a topology mapping vector through the topology mapping weight matrix and the topology mapping matrix of the monitoring points.

[0038] The potential state mapping vector, topological relation mapping vector, and bias vector at the current sampling time point are added element by element to obtain the topological constraint state vector at the current sampling time point.

[0039] The topological constraint state vector at the current sampling time point is input into the Koopman dynamics operator layer. The potential state evolution of the topological constraint state vector is performed through the trainable Koopman dynamics operator matrix to obtain the evolved potential state vector at the next sampling time point.

[0040] In the state update module, the evolved potential state vector at the next sampling time point is weighted and fused with the topological constraint state vector at the current sampling time point to obtain the predicted potential state vector at the next sampling time point.

[0041] In the differential pressure prediction decoding network, the predicted potential state vector at the next sampling time point is used to generate the predicted differential pressure value through the differential pressure value decoding function, and the predicted differential pressure spatial difference feature is generated through the differential pressure spatial difference decoding function.

[0042] The differential pressure decoding function adopts a three-layer fully connected layer structure, with ReLU functions used between the fully connected layers; the differential pressure spatial differential decoding function adopts a two-layer fully connected layer structure, with ReLU functions used between the fully connected layers, and topological masking is applied to the output features of the fully connected layers based on the topological relationship matrix of the monitoring points.

[0043] The potential state vector at the current sampling time point is input into the feature reconstruction branch, and the reconstructed pressure difference behavior feature vector is obtained through a fully connected layer, a ReLU function, and a fully connected layer.

[0044] The total loss function of differential pressure behavior is used as the training objective function to iteratively train the improved Koopman neural network until the total loss function of differential pressure behavior reaches the set convergence threshold or training rounds, thus obtaining the trained differential pressure behavior feature model.

[0045] Optionally, the pressure difference feature encoding network generates a sequence of latent state vectors through multi-scale feature fusion mapping, specifically including:

[0046] The pressure difference feature coding network includes a short-scale feature extraction branch, a medium-scale feature extraction branch, and a long-scale feature extraction branch;

[0047] The short-scale feature extraction branch uses one-dimensional convolution to perform convolution operation on the mapped feature vector sequence, extracts the local change features between adjacent sampling time points, and outputs a short-scale feature vector sequence.

[0048] The mesoscale feature extraction branch uses a gated recurrent unit to perform temporal modeling on the mapped feature vector sequence, extracts dynamic change features across sampling time points, and outputs a mesoscale feature vector sequence.

[0049] The long-scale feature extraction branch uses a Transformer encoder to perform global temporal correlation modeling on the mapped feature vector sequence, extracts trend change features, and outputs a long-scale feature vector sequence.

[0050] The short-scale feature vector sequence, the medium-scale feature vector sequence, and the long-scale feature vector sequence are concatenated along the feature dimension to obtain a multi-scale feature vector sequence;

[0051] A sequence of potential state vectors is generated from a sequence of multi-scale feature vectors through linear mapping.

[0052] Optionally, step six specifically includes:

[0053] Load the trained differential pressure behavior feature model, and use the real-time generated differential pressure behavior feature vector sequence to generate predicted differential pressure value and predicted differential pressure spatial difference feature;

[0054] Simultaneously collect the actual differential pressure monitoring values ​​at the corresponding sampling time points;

[0055] Based on the spatial connection relationship between each monitoring point, the spatial difference characteristics of the actual pressure difference between adjacent monitoring points are calculated;

[0056] Calculate the difference between the predicted differential pressure value and the actual differential pressure monitoring value to obtain the differential pressure prediction deviation at each monitoring point;

[0057] The difference between the predicted spatial differential pressure characteristics and the actual spatial differential pressure characteristics is calculated to obtain the spatial gradient error of each monitoring point.

[0058] The average value of the differential pressure prediction deviation at each monitoring point is calculated as the average value of the differential pressure prediction error at the current sampling time point, and the average value of the spatial gradient error at each monitoring point is calculated as the average value of the spatial gradient error at the current sampling time point.

[0059] The mean of differential pressure prediction error and the mean of spatial gradient error are weighted and fused to obtain the differential pressure monitoring error judgment value.

[0060] Set a differential pressure monitoring error threshold. If the differential pressure monitoring error judgment value is greater than the differential pressure monitoring error threshold, it is determined that there is a differential pressure monitoring error at the current sampling time point.

[0061] Optionally, step seven specifically includes:

[0062] Obtain neighboring monitoring points that are spatially connected to the current monitoring point, and obtain the spatial difference characteristics of the pressure difference between the current monitoring point and the neighboring monitoring points;

[0063] Add the pressure difference monitoring value of the current monitoring point to the pressure difference spatial difference characteristics of each neighboring monitoring point, and divide by the total number of neighboring monitoring points to obtain the estimated value of the neighborhood pressure difference;

[0064] The predicted differential pressure value at the current monitoring point is weighted and fused with the estimated differential pressure value in the neighborhood to obtain the fused differential pressure value.

[0065] Calculate the difference between the current differential pressure monitoring value and the fused differential pressure estimate at the current monitoring point to obtain the differential pressure correction value;

[0066] Set a differential pressure correction factor, multiply the differential pressure correction factor by the differential pressure correction value to obtain the correction compensation term; calculate the difference between the differential pressure monitoring value and the correction compensation term to obtain the corrected differential pressure monitoring value.

[0067] The beneficial effects of this invention are:

[0068] This invention analyzes the spatiotemporal gradient features of differential pressure monitoring data to construct a differential pressure gradient enhancement feature that includes differential pressure temporal difference features, differential pressure change rate features, differential pressure spatial difference features, and differential pressure trend offset features. The differential pressure monitoring data, differential pressure gradient enhancement features, and environmental operating status data are then fused to form a differential pressure behavior feature vector sequence that characterizes the variation patterns of room differential pressure. By constructing a differential pressure time series prediction task, a differential pressure change consistency constraint task, and a differential pressure data reconstruction task, a self-learning constraint task for differential pressure behavior is formed, enabling the differential pressure behavior feature model to continuously learn the spatiotemporal variation patterns of the differential pressure system under unlabeled data conditions. Furthermore, this invention uses an improved Koopman neural network to model the dynamic evolution relationship of differential pressure in the latent space and introduces spatial connectivity between monitoring points through a topological constraint feature fusion module, improving the ability of the differential pressure behavior feature model to express differential pressure change patterns under complex operating conditions. During the monitoring phase, the deviation between the predicted differential pressure value and the actual differential pressure monitoring value, as well as the spatial gradient error, are jointly judged to accurately identify differential pressure monitoring errors. Abnormal monitoring data are then corrected using a neighborhood fusion compensation method, thereby obtaining more reliable differential pressure monitoring results. Attached Figure Description

[0069] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0070] Figure 1 This is a schematic diagram of a room pressure difference monitoring error correction method based on self-supervised learning proposed in this invention;

[0071] Figure 2This invention presents an improved Koopman neural network structure and training flowchart for a room pressure differential monitoring error correction method based on self-supervised learning.

[0072] Figure 3 This is a flowchart of the differential pressure monitoring error determination and neighborhood fusion compensation correction process in a room differential pressure monitoring error correction method based on self-supervised learning proposed in this invention. Detailed Implementation

[0073] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0074] refer to Figures 1-3 A method for correcting room pressure differential monitoring errors based on self-supervised learning includes the following steps:

[0075] Step 1: Collect differential pressure monitoring data and environmental operation status data at each monitoring point;

[0076] Step 2: Based on differential pressure monitoring data, calculate the differential pressure gradient enhancement characteristics through spatiotemporal gradient feature analysis;

[0077] Step 3: Combine and fuse differential pressure monitoring data, differential pressure gradient enhancement features, and environmental operation status data to obtain a differential pressure behavior feature vector sequence;

[0078] Step 4: Construct a self-learning constraint task for pressure difference behavior based on the feature vector sequence of pressure difference behavior;

[0079] Step 5: Construct a differential pressure behavior feature model using an improved Koopman neural network, and train and optimize the differential pressure behavior feature model based on a differential pressure behavior self-learning constraint task to obtain the trained differential pressure behavior feature model; wherein, the improved Koopman neural network includes a differential pressure feature encoding network, a topological constraint feature fusion module, a Koopman dynamics operator layer, a state update module, a differential pressure prediction decoding network, and a feature reconstruction branch;

[0080] Step 6: Input the real-time generated differential pressure behavior feature vector sequence into the trained differential pressure behavior feature model to predict differential pressure and determine whether there is a differential pressure monitoring error;

[0081] Step 7: If there is a differential pressure monitoring error, the differential pressure is corrected by the neighborhood fusion compensation method to obtain the corrected differential pressure monitoring value, and the corrected differential pressure monitoring value is output to the room environment monitoring system.

[0082] In this embodiment, step one specifically includes:

[0083] Differential pressure sensors are deployed between the target room and adjacent rooms, between the target room and the corridor, or between adjacent functional areas. Differential pressure monitoring values ​​at each sampling time point are collected according to a preset sampling cycle to form differential pressure monitoring data.

[0084] Simultaneously collect environmental operation status data corresponding to differential pressure monitoring data, including access control opening and closing status, fan operation status, and air conditioning system operation status.

[0085] In this embodiment, step two specifically includes:

[0086] The characteristics of pressure gradient enhancement include pressure gradient time difference characteristics, pressure gradient change rate characteristics, pressure gradient spatial difference characteristics, and pressure gradient trend offset characteristics;

[0087] Based on differential pressure monitoring data, the difference between differential pressure monitoring values ​​at adjacent sampling time points is calculated at the same monitoring point to obtain differential pressure time difference characteristics. The ratio of differential pressure time difference to the time difference between adjacent sampling time points is also calculated to obtain differential pressure change rate characteristics.

[0088] Based on the spatial connection between each monitoring point, the difference in pressure difference monitoring values ​​between adjacent monitoring points is calculated at the same sampling time point to obtain the spatial difference characteristics of pressure difference.

[0089] Set the sliding time window length, calculate the average value of the differential pressure monitoring value of each monitoring point within the sliding time window length to obtain the differential pressure sliding mean, calculate the difference between the differential pressure monitoring value of each monitoring point at the current sampling time point and the corresponding differential pressure sliding mean to obtain the differential pressure trend offset characteristics.

[0090] In this embodiment, step three specifically includes:

[0091] Align the differential pressure monitoring data, differential pressure gradient enhancement features, and environmental operation status data according to the sampling time point and monitoring point dimension;

[0092] At each sampling time point, the differential pressure monitoring value, differential pressure time difference characteristics, differential pressure change rate characteristics, differential pressure spatial difference characteristics, differential pressure trend offset characteristics, access control opening and closing status, fan operation status, and air conditioning system operation status of each monitoring point are spliced ​​together according to the monitoring point dimension, and a differential pressure behavior feature vector sequence is formed according to the time step order.

[0093] In this embodiment, step four specifically includes:

[0094] The pressure differential behavior self-learning constraint task includes the pressure differential time series prediction task, the pressure differential change consistency constraint task, and the pressure differential data reconstruction task;

[0095] Set the prediction time window length, take the pressure difference behavior feature vector sequence within the prediction time window length as the input sample, and take the data at the t-th sampling time point as the prediction target;

[0096] The pressure difference time series prediction task is as follows: take the pressure difference monitoring value corresponding to the t-th sampling time point as the prediction target to obtain the predicted pressure difference value; calculate the mean square error between the predicted pressure difference value and the pressure difference monitoring value to obtain the pressure difference time series prediction loss.

[0097] The pressure difference change consistency constraint task is as follows: take the pressure difference spatial difference feature at the t-th sampling time point as the prediction target to obtain the predicted pressure difference spatial difference feature; calculate the mean square error between the predicted pressure difference spatial difference feature and the pressure difference spatial difference feature to obtain the pressure difference change consistency constraint loss.

[0098] The differential pressure data reconstruction task is as follows: input the differential pressure behavior feature vector at the t-th sampling time point into the reconstruction branch to obtain the reconstructed differential pressure behavior feature vector, and calculate the mean square error between the reconstructed differential pressure behavior feature vector and the differential pressure behavior feature vector to obtain the differential pressure data reconstruction loss.

[0099] Three loss weighting coefficients are set to weight and fuse the pressure difference time series prediction loss, pressure difference change consistency constraint loss, and pressure difference data reconstruction loss to obtain the total loss function of pressure difference behavior.

[0100] In this embodiment, step five specifically includes:

[0101] Define an embedding mapping space, and map the pressure difference behavior feature vector sequence to the embedding mapping space through a linear transformation to obtain the mapped feature vector sequence;

[0102] Pressure difference feature encoding network generates a sequence of latent state vectors through multi-scale feature fusion mapping;

[0103] In the topology constraint feature fusion module, a topology relation matrix of monitoring points is constructed. Specifically, based on the spatial connection relationship between each monitoring point, if there is an airflow connection between monitoring point i and monitoring point j, the value of the element in the i-th row and j-th column of the topology relation matrix of monitoring points is set to 1, otherwise it is set to 0.

[0104] Define the latent state mapping weight matrix, the topological relation mapping weight matrix, and the bias vector;

[0105] The potential state vector at the current sampling time point is mapped to a potential state mapping vector through the potential state mapping weight matrix, and then mapped to a topology mapping vector through the topology mapping weight matrix and the topology mapping matrix of the monitoring points.

[0106] The potential state mapping vector, topological relation mapping vector, and bias vector at the current sampling time point are added element by element to obtain the topological constraint state vector at the current sampling time point.

[0107] The topological constraint state vector at the current sampling time point is input into the Koopman dynamics operator layer. The potential state evolution of the topological constraint state vector is performed through the trainable Koopman dynamics operator matrix to obtain the evolved potential state vector at the next sampling time point.

[0108] In the state update module, the evolved potential state vector at the next sampling time point is weighted and fused with the topological constraint state vector at the current sampling time point to obtain the predicted potential state vector at the next sampling time point.

[0109] In the differential pressure prediction decoding network, the predicted potential state vector at the next sampling time point is used to generate the predicted differential pressure value through the differential pressure value decoding function, and the predicted differential pressure spatial difference feature is generated through the differential pressure spatial difference decoding function.

[0110] The differential pressure decoding function adopts a three-layer fully connected layer structure, with ReLU functions used between the fully connected layers; the differential pressure spatial differential decoding function adopts a two-layer fully connected layer structure, with ReLU functions used between the fully connected layers, and topological masking is applied to the output features of the fully connected layers based on the topological relationship matrix of the monitoring points.

[0111] The potential state vector at the current sampling time point is input into the feature reconstruction branch, and the reconstructed pressure difference behavior feature vector is obtained through a fully connected layer, a ReLU function, and a fully connected layer.

[0112] The total loss function of differential pressure behavior is used as the training objective function to iteratively train the improved Koopman neural network until the total loss function of differential pressure behavior reaches the set convergence threshold or training rounds, thus obtaining the trained differential pressure behavior feature model.

[0113] In this embodiment, the pressure difference feature encoding network generates a sequence of latent state vectors through multi-scale feature fusion mapping, specifically including:

[0114] The pressure difference feature coding network includes short-scale feature extraction branches, medium-scale feature extraction branches, and long-scale feature extraction branches;

[0115] Among them, the short-scale feature extraction branch uses one-dimensional convolution to perform convolution operation on the mapped feature vector sequence, extracts the local change features between adjacent sampling time points, and outputs a short-scale feature vector sequence.

[0116] The mesoscale feature extraction branch uses a gated recurrent unit to perform temporal modeling on the mapped feature vector sequence, extracts dynamic change features across sampling time points, and outputs a mesoscale feature vector sequence.

[0117] The long-scale feature extraction branch uses a Transformer encoder to perform global temporal correlation modeling on the mapped feature vector sequence, extracts trend change features, and outputs a long-scale feature vector sequence.

[0118] For example, in a sterile preparation production workshop of a pharmaceutical company, the system deploys differential pressure sensors between rooms and continuously collects differential pressure data with a sampling period of 2 seconds. For the collected differential pressure time series data, the short-scale feature extraction branch analyzes the pressure difference changes between adjacent sampling time points through one-dimensional convolution to extract local change features. For example, when a worker briefly opens the door between the buffer room and the clean area, the pressure difference between the two sampling time points may drop instantaneously from 10 Pa to 7 Pa and recover within a few seconds. This type of instantaneous fluctuation reflects the rapid change pattern of the pressure difference in a very short time, i.e., local change features. The mesoscale feature extraction branch performs time series modeling on data from several consecutive sampling time points through a gated loop unit to extract dynamic change features. For example, during the start-up and shutdown of the air conditioning fan or the adjustment of the air volume, the pressure difference may gradually rise from 9 Pa to 12 Pa and remain stable within tens of seconds to several minutes. This type of pressure difference process that gradually changes across multiple time steps reflects the dynamic change features caused by changes in the system's operating state. The long-scale feature extraction branch uses a Transformer encoder to globally correlate and model the pressure difference sequence over a longer time range to extract trend change features. For example, after the equipment has been running continuously for several hours, due to sensor temperature drift or long-term changes in airflow, the pressure difference may gradually shift slowly from a stable 10 Pa to 12 Pa or 8 Pa. This type of feature, which shows a continuous trend over a longer time scale, is called trend change feature. By simultaneously modeling these three types of features, the multi-scale dynamic changes of the room pressure difference system can be more comprehensively reflected.

[0119] The short-scale feature vector sequence, the medium-scale feature vector sequence, and the long-scale feature vector sequence are concatenated along the feature dimension to obtain a multi-scale feature vector sequence;

[0120] A sequence of potential state vectors is generated from a sequence of multi-scale feature vectors through linear mapping.

[0121] In this invention, the improved Koopman neural network maintains the same overall modeling concept and basic structure as the original Koopman neural network. The core idea of ​​the original Koopman neural network is to map a nonlinear dynamic system to a high-dimensional latent space using a neural network. In this latent space, linear Koopman dynamic operators are used to linearly evolve the system state, thereby achieving dynamic modeling and prediction of complex nonlinear systems. Its basic structure includes a state encoding module, a Koopman dynamic operator module, and a state decoding module. Specifically, the state encoding module maps the original input data into a latent state vector; the Koopman dynamic operator module models the temporal evolution of the latent state using a trainable linear operator matrix; and the state decoding module maps the evolved latent state back to the original data space to obtain the prediction result. Furthermore, to improve model stability, the original Koopman neural network typically also includes a state update mechanism and a reconstruction branch to ensure that the latent space representation retains the dynamic characteristics of the system.

[0122] Based on the original Koopman neural network structure, this invention makes structural improvements to the Koopman neural network. First, a multi-scale feature fusion structure is introduced in the encoding stage, constructing three feature extraction branches: short-scale, medium-scale, and long-scale. The short-scale branch extracts local temporal variation features through one-dimensional convolution, the medium-scale branch extracts dynamic features of the time series through gated recurrent units, and the long-scale branch extracts global trend features through a Transformer encoder, thereby generating richer latent state vectors. Second, a topological constraint feature fusion module is added in the latent state modeling stage. By constructing a topological relationship matrix of monitoring points, the spatial connectivity between rooms is introduced into the latent state calculation process, so that the latent state includes not only temporal dynamic information but also spatial correlation information. Furthermore, a pressure differential spatial differential decoding network is designed in the prediction stage. Through topological mask constraints on the output structure, the prediction results maintain a reasonable spatial gradient relationship. Through the above structural design, the improved Koopman neural network enhances its ability to express complex spatiotemporal features.

[0123] Through the aforementioned structural improvements, the improved Koopman neural network can simultaneously capture multi-scale temporal features of pressure differential changes and spatial topological relationships between rooms, thereby significantly enhancing the model's ability to model dynamic pressure differential behavior. The multi-scale feature encoding structure improves the model's ability to identify short-term disturbances, periodic changes, and long-term trend changes, making the latent state representation more stable and comprehensive. The topological constraint feature fusion module strengthens the modeling of spatial correlations between monitoring points, avoiding prediction results that do not conform to actual airflow propagation patterns. The spatial differential pressure decoding mechanism further ensures the consistency of prediction results in spatial gradients. Therefore, compared to the original Koopman neural network, the improved Koopman neural network shows significant improvements in pressure differential prediction accuracy, anomaly detection capability, and model stability, making it more suitable for room pressure differential monitoring and error correction tasks in complex environments.

[0124] In this embodiment, step six specifically includes:

[0125] Load the trained differential pressure behavior feature model, and use the real-time generated differential pressure behavior feature vector sequence to generate predicted differential pressure value and predicted differential pressure spatial difference feature;

[0126] Simultaneously collect the actual differential pressure monitoring values ​​at the corresponding sampling time points;

[0127] Based on the spatial connection relationship between each monitoring point, the spatial difference characteristics of the actual pressure difference between adjacent monitoring points are calculated;

[0128] Calculate the difference between the predicted differential pressure value and the actual differential pressure monitoring value to obtain the differential pressure prediction deviation at each monitoring point;

[0129] The difference between the predicted spatial differential pressure characteristics and the actual spatial differential pressure characteristics is calculated to obtain the spatial gradient error of each monitoring point.

[0130] The average value of the differential pressure prediction deviation at each monitoring point is calculated as the average value of the differential pressure prediction error at the current sampling time point, and the average value of the spatial gradient error at each monitoring point is calculated as the average value of the spatial gradient error at the current sampling time point.

[0131] The mean value of differential pressure prediction error is obtained by weighted fusion with the mean value of spatial gradient error.

[0132] Set a differential pressure monitoring error threshold. If the differential pressure monitoring error judgment value is greater than the differential pressure monitoring error threshold, it is determined that there is a differential pressure monitoring error at the current sampling time point.

[0133] In this embodiment, step seven specifically includes:

[0134] Obtain neighboring monitoring points that are spatially connected to the current monitoring point, and obtain the spatial difference characteristics of the pressure difference between the current monitoring point and the neighboring monitoring points;

[0135] Add the pressure difference monitoring value of the current monitoring point to the pressure difference spatial difference characteristics of each neighboring monitoring point, and divide by the total number of neighboring monitoring points to obtain the estimated value of the neighborhood pressure difference;

[0136] The predicted differential pressure value at the current monitoring point is weighted and fused with the estimated differential pressure value in the neighborhood to obtain the fused differential pressure value.

[0137] Calculate the difference between the current differential pressure monitoring value and the fused differential pressure estimate at the current monitoring point to obtain the differential pressure correction value;

[0138] Set a differential pressure correction factor, multiply the differential pressure correction factor by the differential pressure correction value to obtain the correction compensation term; calculate the difference between the differential pressure monitoring value and the correction compensation term to obtain the corrected differential pressure monitoring value.

[0139] In this invention, the room differential pressure system exhibits significant spatial coupling characteristics during actual operation. That is, the pressure difference changes between adjacent rooms typically show a stable spatial gradient relationship. When a sensor at a monitoring point experiences drift, noise interference, or transient anomalies, its measured value often deviates from the airflow distribution pattern commonly reflected by surrounding monitoring points. Therefore, by acquiring neighboring monitoring points spatially connected to the current monitoring point and fusing the data from these neighboring monitoring points using the spatial difference characteristics of the pressure difference, a neighboring differential pressure estimate reflecting the true spatial airflow relationship can be constructed as a reference benchmark for the current monitoring point's pressure difference. Weighted fusing this neighboring estimate with the model-predicted differential pressure value allows for the simultaneous use of data-driven prediction and spatial structure information, improving the reliability of the estimation results. Finally, by calculating and compensating for the deviation between the current monitored value and the fused estimate, errors caused by sensor drift or local anomalies can be effectively eliminated. This ensures that the corrected differential pressure monitoring value conforms to the dynamic changes of the differential pressure system while maintaining a reasonable spatial gradient relationship, thereby improving the overall accuracy and stability of the room differential pressure monitoring system.

[0140] Example 1: To verify the feasibility of this invention in practice, the method of this invention was applied to a clean environment differential pressure monitoring system in a GMP aseptic preparation production workshop of a pharmaceutical company. The aseptic filling workshop of this pharmaceutical company consists of multiple functional areas, including corridors, buffer rooms, clean areas, and isolation areas. A stable differential pressure gradient needs to be maintained between different areas to ensure that air flows from areas with high cleanliness to areas with low cleanliness, thereby preventing the diffusion of contaminants. A total of 12 differential pressure monitoring points are deployed inside the workshop. Differential pressure sensors are installed in the walls or passageways between adjacent rooms. The clean environment differential pressure monitoring system collects differential pressure data at a sampling cycle of once every 2 seconds, and simultaneously collects environmental operating status data such as access control status, air supply fan operating status, and air conditioning system operating status.

[0141] During daily operation, cleanroom differential pressure monitoring systems are frequently affected by factors such as personnel entry and exit, equipment vibration, changes in air conditioning volume, and sensor drift due to long-term use. This can cause short-term fluctuations or long-term deviations in differential pressure data at some monitoring points. For example, after three months of continuous operation, the differential pressure readings at some monitoring points may show a slow drift of approximately 1 Pa to 3 Pa. This drift may not trigger a traditional threshold alarm, but it can alter the differential pressure gradient relationship, thereby affecting the reliability of cleanroom control.

[0142] In the implementation of the method of this invention, differential pressure monitoring data is collected in real time by differential pressure sensors deployed between each room, and the opening and closing status of access control, fan operation status, and air conditioning system operation status are also acquired. Subsequently, spatiotemporal gradient feature analysis is performed on the collected differential pressure data to form a differential pressure behavior feature vector sequence, and a self-learning constraint task for differential pressure behavior is constructed based on the differential pressure behavior feature vector sequence.

[0143] During the model training phase, the clean environment differential pressure monitoring system uses three months of historical differential pressure monitoring data as training samples to establish a differential pressure behavior feature model through an improved Koopman neural network. After training, the real-time generated differential pressure behavior feature vector sequence is input into the differential pressure behavior feature model for differential pressure prediction. The system then uses the deviation between the predicted and actual monitored values, along with spatial gradient error, to jointly determine whether there is a differential pressure monitoring error at the current time step. If an error exists, it is corrected using a neighborhood fusion compensation method, and the corrected differential pressure monitoring value is fed back to the clean environment differential pressure monitoring system.

[0144] To verify the effectiveness of the present invention, the method of the present invention was compared with the original monitoring method based on the original Koopman neural network, the moving average filtering method, and the Kalman filtering method. The monitoring results were statistically analyzed during a 30-day continuous production cycle. The comparison indicators included the root mean square error of differential pressure monitoring, differential pressure drift recognition rate, differential pressure abnormality detection accuracy, spatial gradient consistency error, system false alarm rate, system missed detection rate, and differential pressure stability index. The comparison results are shown in Table 1.

[0145] Table 1. Performance Comparison of Different Differential Pressure Monitoring Error Correction Methods

[0146] Comparison indicators Original monitoring methods Moving mean filtering method Kalman filtering method Method of the present invention Mean square error of differential pressure monitoring (Pa²) 4.62 3.28 2.16 0.93 Differential pressure drift recognition rate (%) 48.3 57.4 68.7 92.5 Accuracy rate of pressure difference abnormality detection (%) 63.7 71.2 80.6 95.8 Spatial gradient consistency error (Pa) 3.15 2.44 1.72 0.64 System false alarm rate (%) 9.6 8.1 6.7 2.3 System false negative rate (%) 17.4 14.9 10.2 3.1 Differential pressure stability index 0.61 0.69 0.77 0.92

[0147] As shown in Table 1, the method of this invention has significant advantages in differential pressure monitoring accuracy, anomaly identification capability, and system stability. Regarding the mean square error of differential pressure monitoring, the method of this invention is only 0.93 Pa², significantly lower than the 4.62 Pa² of the original monitoring method, indicating that this invention can effectively reduce the impact of sensor noise and environmental disturbances on the monitoring results, thereby improving the accuracy of differential pressure monitoring data. In terms of differential pressure drift identification rate, the method of this invention reaches 92.5%, significantly higher than the moving average filtering method and the Kalman filtering method, indicating that this invention can more accurately identify long-term sensor drift or abnormal changes. Regarding the accuracy of differential pressure anomaly detection, the method of this invention reaches 95.8%, significantly higher than other comparative methods, indicating that after modeling the differential pressure behavior pattern through self-supervised learning and the improved Koopman neural network, the clean environment differential pressure monitoring system can more accurately distinguish between normal operating condition fluctuations and actual monitoring errors.

[0148] Furthermore, the spatial gradient consistency error of this invention is only 0.64 Pa, significantly lower than other comparative methods, indicating that by introducing the spatial topological relationship between monitoring points and the spatial differential pressure characteristics, this invention can effectively maintain a reasonable differential pressure gradient relationship between each monitoring point. Regarding the system false alarm rate and system missed detection rate, the method of this invention is 2.3% and 3.1% respectively, both significantly lower than other methods, demonstrating that this invention not only improves the accuracy of anomaly detection but also reduces unnecessary alarms and missed detections. In addition, the differential pressure stability index of this invention reaches 0.92, significantly higher than the comparative methods, indicating that this invention can maintain more stable and reliable differential pressure monitoring performance during long-term operation.

[0149] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for correcting room differential pressure monitoring errors based on self-supervised learning, characterized in that, Includes the following steps: Step 1: Collect differential pressure monitoring data and environmental operation status data at each monitoring point; Step 2: Based on differential pressure monitoring data, calculate the differential pressure gradient enhancement characteristics through spatiotemporal gradient feature analysis; Step 3: Combine and fuse differential pressure monitoring data, differential pressure gradient enhancement features, and environmental operation status data to obtain a differential pressure behavior feature vector sequence; Step 4: Construct a self-learning constraint task for pressure difference behavior based on the feature vector sequence of pressure difference behavior; Step 5: Construct a differential pressure behavior feature model using an improved Koopman neural network, and train and optimize the differential pressure behavior feature model based on a differential pressure behavior self-learning constraint task to obtain the trained differential pressure behavior feature model; the improved Koopman neural network includes a differential pressure feature encoding network, a topological constraint feature fusion module, a Koopman dynamics operator layer, a state update module, a differential pressure prediction decoding network, and a feature reconstruction branch; Step 6: Input the real-time generated differential pressure behavior feature vector sequence into the trained differential pressure behavior feature model to predict differential pressure and determine whether there is a differential pressure monitoring error; Step 7: If there is a differential pressure monitoring error, the differential pressure is corrected by the neighborhood fusion compensation method to obtain the corrected differential pressure monitoring value, and the corrected differential pressure monitoring value is output to the room environment monitoring system.

2. The method for correcting room pressure difference monitoring errors based on self-supervised learning according to claim 1, characterized in that, Step one specifically includes: Differential pressure sensors are deployed between the target room and adjacent rooms, between the target room and the corridor, or between adjacent functional areas. Differential pressure monitoring values ​​at each sampling time point are collected according to a preset sampling cycle to form differential pressure monitoring data. Simultaneously collect environmental operating status data corresponding to differential pressure monitoring data, including access control opening and closing status, fan operating status, and air conditioning system operating status.

3. The method for correcting room pressure difference monitoring errors based on self-supervised learning according to claim 1, characterized in that, Step two specifically includes: The pressure gradient enhancement features include pressure difference time difference features, pressure difference change rate features, pressure difference spatial difference features, and pressure difference trend offset features; Based on differential pressure monitoring data, the difference between differential pressure monitoring values ​​at adjacent sampling time points is calculated at the same monitoring point to obtain differential pressure time difference characteristics. The ratio of differential pressure time difference to the time difference between adjacent sampling time points is also calculated to obtain differential pressure change rate characteristics. Based on the spatial connection between each monitoring point, the difference in pressure difference monitoring values ​​between adjacent monitoring points is calculated at the same sampling time point to obtain the spatial difference characteristics of pressure difference. Set the sliding time window length, calculate the average value of the differential pressure monitoring value of each monitoring point within the sliding time window length to obtain the differential pressure sliding mean, calculate the difference between the differential pressure monitoring value of each monitoring point at the current sampling time point and the corresponding differential pressure sliding mean to obtain the differential pressure trend offset characteristics.

4. The method for correcting room pressure difference monitoring errors based on self-supervised learning according to claim 1, characterized in that, Step three specifically includes: Align the differential pressure monitoring data, differential pressure gradient enhancement features, and environmental operation status data according to the sampling time point and monitoring point dimension; At each sampling time point, the differential pressure monitoring value, differential pressure time difference characteristics, differential pressure change rate characteristics, differential pressure spatial difference characteristics, differential pressure trend offset characteristics, access control opening and closing status, fan operation status, and air conditioning system operation status of each monitoring point are spliced ​​together according to the monitoring point dimension, and a differential pressure behavior feature vector sequence is formed according to the time step order.

5. The method for correcting room pressure difference monitoring errors based on self-supervised learning according to claim 1, characterized in that, Step four specifically includes: The pressure difference behavior self-learning constraint task includes a pressure difference time series prediction task, a pressure difference change consistency constraint task, and a pressure difference data reconstruction task. Set the prediction time window length, take the pressure difference behavior feature vector sequence within the prediction time window length as the input sample, and take the data at the t-th sampling time point as the prediction target; The pressure difference time series prediction task is as follows: take the pressure difference monitoring value corresponding to the t-th sampling time point as the prediction target to obtain the predicted pressure difference value; calculate the mean square error between the predicted pressure difference value and the pressure difference monitoring value to obtain the pressure difference time series prediction loss. The pressure difference change consistency constraint task is as follows: take the pressure difference spatial difference feature at the t-th sampling time point as the prediction target to obtain the predicted pressure difference spatial difference feature; calculate the mean square error between the predicted pressure difference spatial difference feature and the pressure difference spatial difference feature to obtain the pressure difference change consistency constraint loss. The differential pressure data reconstruction task is as follows: input the differential pressure behavior feature vector at the t-th sampling time point into the reconstruction branch to obtain the reconstructed differential pressure behavior feature vector, and calculate the mean square error between the reconstructed differential pressure behavior feature vector and the differential pressure behavior feature vector to obtain the differential pressure data reconstruction loss. Three loss weighting coefficients are set to weight and fuse the pressure difference time series prediction loss, pressure difference change consistency constraint loss, and pressure difference data reconstruction loss to obtain the total loss function of pressure difference behavior.

6. The method for correcting room pressure difference monitoring errors based on self-supervised learning according to claim 1, characterized in that, Step five specifically includes: Define an embedding mapping space, and map the pressure difference behavior feature vector sequence to the embedding mapping space through a linear transformation to obtain the mapped feature vector sequence; The pressure difference feature encoding network generates a sequence of potential state vectors through multi-scale feature fusion mapping. In the topology constraint feature fusion module, a topology relation matrix of monitoring points is constructed. Specifically, based on the spatial connection relationship between each monitoring point, if there is an airflow connection between monitoring point i and monitoring point j, the value of the element in the i-th row and j-th column of the topology relation matrix of monitoring points is set to 1, otherwise it is set to 0. Define the latent state mapping weight matrix, the topological relation mapping weight matrix, and the bias vector; The potential state vector at the current sampling time point is mapped to a potential state mapping vector through the potential state mapping weight matrix, and then mapped to a topology mapping vector through the topology mapping weight matrix and the topology mapping matrix of the monitoring points. The potential state mapping vector, topological relation mapping vector, and bias vector at the current sampling time point are added element by element to obtain the topological constraint state vector at the current sampling time point. The topological constraint state vector at the current sampling time point is input into the Koopman dynamics operator layer. The potential state evolution of the topological constraint state vector is performed through the trainable Koopman dynamics operator matrix to obtain the evolved potential state vector at the next sampling time point. In the state update module, the evolved potential state vector at the next sampling time point is weighted and fused with the topological constraint state vector at the current sampling time point to obtain the predicted potential state vector at the next sampling time point. In the differential pressure prediction decoding network, the predicted potential state vector at the next sampling time point is used to generate the predicted differential pressure value through the differential pressure value decoding function, and the predicted differential pressure spatial difference feature is generated through the differential pressure spatial difference decoding function. The differential pressure decoding function adopts a three-layer fully connected layer structure, with ReLU functions used between the fully connected layers; the differential pressure spatial differential decoding function adopts a two-layer fully connected layer structure, with ReLU functions used between the fully connected layers, and topological masking is applied to the output features of the fully connected layers based on the topological relationship matrix of the monitoring points. The potential state vector at the current sampling time point is input into the feature reconstruction branch, and the reconstructed pressure difference behavior feature vector is obtained through a fully connected layer, a ReLU function, and a fully connected layer. The total loss function of differential pressure behavior is used as the training objective function to iteratively train the improved Koopman neural network until the total loss function of differential pressure behavior reaches the set convergence threshold or training rounds, thus obtaining the trained differential pressure behavior feature model.

7. The method for correcting room pressure difference monitoring errors based on self-supervised learning according to claim 6, characterized in that, The pressure difference feature encoding network generates a sequence of latent state vectors through multi-scale feature fusion mapping, specifically including: The pressure difference feature coding network includes a short-scale feature extraction branch, a medium-scale feature extraction branch, and a long-scale feature extraction branch; The short-scale feature extraction branch uses one-dimensional convolution to perform convolution operation on the mapped feature vector sequence, extracts the local change features between adjacent sampling time points, and outputs a short-scale feature vector sequence. The mesoscale feature extraction branch uses a gated recurrent unit to perform temporal modeling on the mapped feature vector sequence, extracts dynamic change features across sampling time points, and outputs a mesoscale feature vector sequence. The long-scale feature extraction branch uses a Transformer encoder to perform global temporal correlation modeling on the mapped feature vector sequence, extracts trend change features, and outputs a long-scale feature vector sequence. The short-scale feature vector sequence, the medium-scale feature vector sequence, and the long-scale feature vector sequence are concatenated along the feature dimension to obtain a multi-scale feature vector sequence; A sequence of potential state vectors is generated from a sequence of multi-scale feature vectors through linear mapping.

8. The method for correcting room pressure differential monitoring errors based on self-supervised learning according to claim 1, characterized in that, Step six specifically includes: Load the trained differential pressure behavior feature model, and use the real-time generated differential pressure behavior feature vector sequence to generate predicted differential pressure value and predicted differential pressure spatial difference feature; Simultaneously collect the actual differential pressure monitoring values ​​at the corresponding sampling time points; Based on the spatial connection relationship between each monitoring point, the spatial difference characteristics of the actual pressure difference between adjacent monitoring points are calculated; Calculate the difference between the predicted differential pressure value and the actual differential pressure monitoring value to obtain the differential pressure prediction deviation at each monitoring point; The difference between the predicted spatial differential pressure characteristics and the actual spatial differential pressure characteristics is calculated to obtain the spatial gradient error of each monitoring point. The average value of the differential pressure prediction deviation at each monitoring point is calculated as the average value of the differential pressure prediction error at the current sampling time point, and the average value of the spatial gradient error at each monitoring point is calculated as the average value of the spatial gradient error at the current sampling time point. The mean value of differential pressure prediction error is obtained by weighted fusion with the mean value of spatial gradient error. Set a differential pressure monitoring error threshold. If the differential pressure monitoring error judgment value is greater than the differential pressure monitoring error threshold, it is determined that there is a differential pressure monitoring error at the current sampling time point.

9. The method for correcting room pressure difference monitoring errors based on self-supervised learning according to claim 1, characterized in that, Step seven specifically includes: Obtain neighboring monitoring points that are spatially connected to the current monitoring point, and obtain the spatial difference characteristics of the pressure difference between the current monitoring point and the neighboring monitoring points; Add the pressure difference monitoring value of the current monitoring point to the pressure difference spatial difference characteristics of each neighboring monitoring point, and divide by the total number of neighboring monitoring points to obtain the estimated value of the neighborhood pressure difference; The predicted differential pressure value at the current monitoring point is weighted and fused with the estimated differential pressure value in the neighborhood to obtain the fused differential pressure value. Calculate the difference between the current differential pressure monitoring value and the fused differential pressure estimate at the current monitoring point to obtain the differential pressure correction value; Set a differential pressure correction factor, multiply the differential pressure correction factor by the differential pressure correction value to obtain the correction compensation term; calculate the difference between the differential pressure monitoring value and the correction compensation term to obtain the corrected differential pressure monitoring value.

Images