Chiller failure prediction method based on rejection and calibration mechanism

Abstract

A kind of robust fault prediction method of water chilling unit based on rejection and calibration mechanism, in offline stage, the historical operation data of water chilling unit under typical fault category is collected and training set is constructed, for training deep learning diagnosis model and automatic encoder model;In online stage, whether real-time operation data needs calibration is judged first by trained automatic encoder model, then real-time fault prediction and prediction calibration are carried out by deep learning diagnosis model, the prediction of model confidence is filtered out by rejection mechanism in the application, and the reliability of fault diagnosis result is further guaranteed by calibration mechanism, and the fault diagnosis performance under variable working condition and variable control scene is significantly improved.

Description

Technical Field

[0001] This invention relates to a technology in the field of big data analysis and fault diagnosis of building air conditioning systems, specifically a method for predicting faults in chiller units based on rejection and calibration mechanisms. Background Technology

[0002] As a critical refrigeration device in energy systems, chiller units require essential health management and fault diagnosis. In recent years, deep learning-based fault diagnosis technology for chiller units has garnered significant attention. However, existing deep learning-based fault diagnosis technologies generally lack robustness and cannot provide accurate diagnostic results under varying operating conditions. Once the operating conditions of the chiller unit exceed the range covered by the training set, the deep learning model will provide incorrect diagnostic results. Building energy systems have a wide range of operating conditions and complex control logic. If the diagnostic capabilities of deep learning models under varying operating conditions cannot be improved, system maintenance personnel will be overwhelmed by a large amount of misdiagnostic information, leading to the failure of the entire fault diagnosis system. Summary of the Invention

[0003] This invention addresses the shortcomings of existing chiller unit fault diagnosis technologies, which cannot predict faults under varying operating conditions. Furthermore, existing models often misinterpret changes in operating conditions as fault occurrences, leading to frequent false alarms, once the actual operating conditions of the equipment exceed the range of the training data. To address these shortcomings, a robust fault prediction method for chiller units based on rejection and calibration mechanisms is proposed. The rejection mechanism filters out predictions based on model confidence, while the calibration mechanism further ensures the reliability of fault diagnosis results. This significantly improves fault diagnosis performance under varying operating conditions and control scenarios.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a robust fault prediction method for chiller units based on a rejection and calibration mechanism. In the offline phase, historical operating data of the chiller unit under typical fault categories is collected and a training set is constructed for training a deep learning diagnostic model and an autoencoder model. In the online phase, the trained autoencoder model first determines whether real-time operating data requires calibration, and then the deep learning diagnostic model performs real-time fault prediction and prediction calibration. Specifically, it includes:

[0006] Step 1) Fault data collection and preprocessing: By conducting fault experiments, the operating data of the chiller unit under typical fault categories are collected, and the collected data is filtered for steady state and outliers are removed to obtain a cleaned dataset; the dataset is then divided into training set, validation set and test set.

[0007] Step 2) Construct a deep learning diagnostic model: Using the training dataset, construct a deep learning classifier. The output of this classifier is the probability of each fault category occurring under the current operating state of the unit, and the confidence threshold is determined simultaneously. After the diagnostic model is trained, determine the model's confidence index, and simultaneously determine the confidence threshold of the diagnostic model using a validation dataset;

[0008] The fault categories include: refrigerant leakage, refrigerant overcharging, evaporator scaling, and reduced water flow.

[0009] The deep learning diagnostic model includes an input layer, five hidden layers, and an output layer. The input layer extracts the current fault symptoms of the equipment based on the information from the temperature and pressure sensors of the chiller unit. Then, the features are transformed through the five hidden layers. Finally, the output layer outputs the probability that the current equipment belongs to each fault category.

[0010] The hidden layers contain 128 neurons per layer.

[0011] Step 3) Constructing the autoencoder model: Using the training dataset, train an autoencoder model in parallel to learn the data distribution of the training data; the input of the autoencoder is the fault-sensitive features, and the output is the fault-sensitive features reconstructed by the decoder. Simultaneously, determine the reconstruction error threshold: After the autoencoder is trained, determine the threshold for the autoencoder reconstruction error using a validation dataset.

[0012] The autoencoder model includes an encoder and a decoder, wherein the encoder reduces the original features to a lower-dimensional feature, and the decoder reconstructs the original features from the lower-dimensional feature.

[0013] Step 4) Online Fault Prediction: The real-time data from the chiller unit is first processed by an autoencoder model to determine if its reconstruction error exceeds a threshold. If it does, the data is rejected and awaits further calibration. If the reconstruction error does not exceed the threshold, the data is entered into the classifier for fault prediction. At this point, if the confidence level of the prediction result is less than the threshold, the data is also rejected and awaits further calibration. If the confidence level is higher than the threshold, the diagnostic result is recorded, and the data is entered into the result pool.

[0014] Preferably, prediction calibration is performed when the amount of data in the recent rejection pool reaches 500 or the inconsistency of the fault prediction results in the recent result pool is high: by using a partial domain adaptive algorithm, the training set in the offline stage is used as the source domain, and the data in the rejection pool and the result pool are mixed as the target domain. By using a weighted domain adversarial neural network, the difference in data distribution between the source domain and the target domain is reduced, thereby providing more accurate diagnostic results.

[0015] Technical effect

[0016] This invention utilizes an autoencoder to reject fault prediction results and employs a domain-adaptive algorithm to calibrate recent fault prediction results. Compared to existing technologies, the fault diagnosis accuracy of this invention is improved by an average of 18% after calibration; the false alarm rate of fault diagnosis after adopting the rejection mechanism can be reduced to 0% by setting a threshold. Attached Figure Description

[0017] Figure 1 This is a flowchart of the present invention;

[0018] Figure 2 Flowchart for fault sensitivity feature analysis;

[0019] Figure 3 A network architecture diagram of a deep learning diagnostic model;

[0020] Figure 4 This is a network architecture diagram of an autoencoder model;

[0021] Figure 5 A flowchart for fault prediction calibration based on partial domain adaptation;

[0022] Figure 6 for Figure 5 Enlarged view of a portion of the diagram. Detailed Implementation

[0023] like Figure 1 As shown in the figure, this embodiment relates to a method for predicting chiller unit faults based on a rejection and calibration mechanism, including:

[0024] S1. Fault Data Acquisition and Preprocessing: Fault simulation experiments were conducted on the 350kW screw chiller unit under study to collect operating data under typical fault categories. The collected fault categories included refrigerant leakage, reduced evaporator water flow, and reduced condenser water flow. Refrigerant leakage was simulated by artificially reducing the system refrigerant charge, with three fault levels corresponding to 90% (Level 1), 80% (Level 2), and 70% (Level 3) of the nominal charge. Similarly, the other two faults were simulated by reducing the evaporator and condenser water flow, each with two fault levels corresponding to 85% (Level 1) and 70% (Level 2) of the nominal water flow. Operating data was collected under 27 operating conditions by adjusting the chiller unit's cooling load, cooling water inlet temperature, and chilled water outlet temperature. Each condition was operated for at least 30 minutes, with at least 10 minutes of steady-state operation. For the original dataset, steady-state detection and outlier removal are performed to remove dynamic and outlier data to prevent impact on model performance. Steady-state detection uses a steady-state window identification method. A sliding window is used, and the standard deviation of a specified feature within each window is calculated. When the standard deviation is less than a threshold, the data in that window is considered steady-state, and 10% of the data in that window is randomly selected and added to the training dataset. When the standard deviation is higher than the threshold, the process moves directly to the next window. Outlier removal uses the quartile outlier removal method. The 25th and 75th quartiles of a specified feature are calculated, and upper and lower bounds are determined based on these quartiles. When a data point for that feature exceeds these bounds, it is considered an outlier and removed from the dataset. These two methods result in a clean dataset.

[0025] S2, Fault Sensitivity Characteristic Analysis: such as Figure 2 As shown, fault sensitivity feature analysis is performed using the cleaned dataset. First, data from the normal category are selected in the dataset to construct a fault-free model. The inputs to this fault-free model are cooling load, chilled water outlet temperature, and cooling water inlet temperature; the outputs are fault symptom variables, including condensing pressure, compressor power, subcooling, evaporator temperature difference, and condenser temperature difference. The purpose of the fault-free model is to predict the baseline values ​​of the fault symptom variables under given operating conditions. For each fault category in the dataset, the actual values ​​are compared with the baseline values ​​to obtain the residual values. By analyzing the residual values, features sensitive to the fault category under study can be selected. For example, subcooling is highly sensitive to refrigerant leakage faults. Features that are not sensitive are not considered in subsequent model construction. The fault-sensitive features analyzed in this example are: subcooling, condensing pressure, expansion valve opening, evaporator temperature difference, condenser temperature difference, chilled water temperature difference, and cooling water temperature difference, totaling seven features, which are used as inputs for the subsequent model.

[0026] S3, Dataset Partitioning: The cleaned dataset is randomly divided into a training dataset, a validation dataset, and a test dataset in a 7:1:2 ratio. The training set is used to train the model, the validation set is used to determine the threshold for the rejection mechanism metric, and the test set is used to test the final performance of the model.

[0027] S4, construct as follows Figure 3 The deep learning diagnostic model shown uses a training dataset to construct a deep learning classifier. The output of this classifier is the probability of each fault category occurring under the current operating state of the unit. The input of the model is the fault-sensitive features obtained from the analysis in S2.

[0028] The deep learning diagnostic model includes 5 hidden layers, each with 128 neurons. It performs further feature transformation based on fault-sensitive features and finally outputs the probability corresponding to the fault category through a softmax layer, thereby realizing fault diagnosis. The deep learning diagnostic model uses the Adam optimizer and employs batch normalization layers to accelerate network training. The selected activation function is the ReLU activation function.

[0029] S5, Determine the confidence threshold: After the diagnostic model is trained, determine the model's confidence index, and simultaneously determine the confidence threshold of the diagnostic model using a validation dataset. In this embodiment, the confidence index used is Entropy = - Where p represents the probability corresponding to a certain category. A higher entropy for a prediction indicates higher uncertainty and lower confidence in the prediction. The information entropy of the diagnostic model on the validation set is calculated, with a threshold set to 95% of the validation set data having an information entropy below this threshold. That is, in most cases, the information entropy of a prediction will not exceed this threshold. If the information entropy of online data exceeds this threshold, the prediction is considered inaccurate.

[0030] S6, construct as follows Figure 4 The autoencoder model shown is trained in parallel using the training dataset to learn the data distribution of the training data. The input of the autoencoder is the fault-sensitive features analyzed in S2, and the output is the fault-sensitive features reconstructed by the decoder.

[0031] The autoencoder model includes an encoder and a decoder, wherein the encoder reduces the original features to a lower-dimensional feature, and the decoder reconstructs the original features from the lower-dimensional feature; the autoencoder model is trained by minimizing the mean squared error loss, which can be viewed as the network learning the data distribution of the training data.

[0032] Both the encoder and decoder are 3-layer neural networks with 16 neurons per layer, using the ReLU activation function and the Adam optimizer. The reduced features are 4-dimensional.

[0033] S7, Determine the reconstruction error threshold: After the autoencoder is trained, the reconstruction error threshold is determined using a validation dataset. This reconstruction error is denoted as mse = ||xG d (G e (x))||2 2 , where: G e For encoder, G d For the decoder, the reconstruction error is calculated using the L2 norm between the original features and the reconstructed features. For the validation dataset, the reconstruction error is calculated for each data point, with a threshold set for the reconstruction error of 95% of the validation dataset whose reconstruction errors are below this value. The underlying principle is that when online data and training data belong to the same operating conditions, their reconstruction errors should also be very small; if the error exceeds this threshold, it is considered to belong to variable operating conditions and is rejected.

[0034] S8, Model Performance Validation: After determining the two models and their thresholds, the model performance is tested using a test dataset to verify whether it meets the requirements. In this embodiment, the performance metrics used to validate the model are accuracy, precision, and recall. The test dataset is divided into accept and reject portions. During performance validation, it is determined whether the accepted data portion meets the required metrics, such as an accuracy of over 90%. If the requirements are met, the process proceeds to the online fault prediction section; otherwise, the thresholds need to be readjusted, or the process returns to step S3 to re-divide the dataset and modify the determined hyperparameters. The hyperparameters and thresholds determined in this embodiment meet the accuracy requirements, achieving an accuracy of 97.4% on the test set.

[0035] S9, Online Fault Prediction: (e.g., ...) Figure 1 As shown, the real-time data from the chiller unit is first processed by an autoencoder model to determine if its reconstruction error exceeds a threshold. If it does, the data is placed in a rejection pool for further calibration. If the reconstruction error does not exceed the threshold, the data is then processed by a classifier for fault prediction. If the confidence level of the prediction result is less than the threshold, the data is again placed in the rejection pool for further calibration. If the confidence level is higher than the threshold, the diagnostic result is recorded, and the data is placed in the result pool. In this embodiment, data experiments verify that most of the data in the rejection pool consists of incorrect model predictions, while most of the data in the result pool consists of correctly predicted model predictions. This indicates that the rejection mechanism works effectively, filtering out incorrect predictions.

[0036] S10, Prediction Result Calibration: If the number of data entries in the recent rejection pool reaches 500 or the inconsistency of the fault prediction results in the recent result pool is high, then the following steps are performed: Figure 5 and Figure 6 The prediction calibration shown is as follows: By using a partial domain adaptive algorithm, the training set in the offline stage is used as the source domain, and the data in the rejection pool and the result pool are mixed as the target domain. A weighted domain adversarial neural network is used to narrow the distance between the data distribution of the source domain and the target domain, thereby providing more accurate diagnostic results.

[0037] The domain adversarial neural network includes a feature extractor, a classifier, and a discriminator. The feature extractor and classifier are taken from the deep learning diagnostic model trained in the offline stage, and the part before the last layer is regarded as the feature extractor, while the last layer is regarded as the classifier. The discriminator measures the difference in data distribution between online data and source domain data, and then the gradient inversion layer enables the feature extractor to minimize this difference in data distribution, so as to achieve the effect of domain adaptation.

[0038] The discriminator comprises a three-layer neural network, each with 128 neurons, and outputs a category corresponding to a domain label to determine whether the current data belongs to the source domain or the target domain.

[0039] The weights mentioned refer to: for example Figure 5 and Figure 6 The online data shown is first processed by a diagnostic network to determine which fault categories are most likely to exist in the current online data, i.e., a weight is assigned to each category. In the subsequent training process, for each training data, when calculating the domain adversarial loss and cross-entropy loss, the category weight γ is used to assign weights. Categories that are similar to the online data will be given larger weights, while those that are not will be given smaller weights.

[0040] The category weights Where γ is a multidimensional vector, representing the weight of each category, thus filtering out fault categories in the source domain data that are not similar to the current online data.

[0041] The loss function of the aforementioned domain adversarial neural network is: Where: Cross-entropy loss Domain confrontation loss Among them: G f G 嵠 G d These are feature extractor, classifier, and discriminator, respectively. d For domain adversarial losses, L 嵠 The cross-entropy loss is used, and the entire network updates its parameters through gradient descent to minimize the domain adversarial loss and cross-entropy loss.

[0042] Data experiments have verified that, under the above parameter settings, this method performs well on multiple chiller unit fault diagnosis datasets. After using the calibration mechanism, the average fault diagnosis accuracy improved by 18%; simultaneously, the false alarm rate of the model can be reduced to 0% through the rejection mechanism.

[0043] In summary, this invention trains an autoencoder model to learn the data distribution of the training data. If the data distribution of the online data is inconsistent with the training data, the predicted result is rejected, thereby reducing the false alarm rate of the diagnostic system. A confidence score is used to determine the diagnostic model's confidence in a prediction; when the prediction confidence is low, the prediction result is rejected, thus improving the reliability of the diagnostic system. Through a prediction calibration mechanism, the latest health status of the unit is obtained by maximizing the use of online data. Even if previous data is rejected, it can be fully utilized in subsequent calibration phases, improving the diagnostic efficiency of the diagnostic model.

[0044] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A robust fault prediction method for chiller units based on rejection and calibration mechanisms, characterized in that, In the offline phase, historical operating data of the chiller unit under typical fault categories is collected and a training set is built for training the deep learning diagnostic model and the autoencoder model. In the online phase, the trained autoencoder model is first used to determine whether the real-time operating data needs calibration, and then the deep learning diagnostic model is used for real-time fault prediction and prediction calibration. The fault categories mentioned include: refrigerant leakage, refrigerant overcharging, evaporator scaling, and reduced water flow. The deep learning diagnostic model includes an input layer, five hidden layers, and an output layer. The input layer extracts the current fault symptoms of the equipment based on the information from the temperature and pressure sensors of the chiller unit. Then, it performs feature transformation through the five hidden layers. Finally, the output layer outputs the probability that the current equipment belongs to each fault category. The autoencoder model includes an encoder and a decoder, wherein the encoder reduces the original features to a lower-dimensional feature, and the decoder reconstructs the original features from the lower-dimensional feature; The robust fault prediction for the chiller unit specifically includes: Step 1) Fault data collection and preprocessing: By conducting fault experiments, the operating data of the chiller unit under typical fault categories are collected, and the collected data is subjected to steady-state filtering and outlier removal to obtain a cleaned dataset; the dataset is then divided into training set, validation set and test set. Step 2) Construct a deep learning diagnostic model: Using the training dataset, construct a deep learning classifier; the output of the classifier is the probability of each fault category occurring under the current operating state of the unit, and at the same time, determine the confidence threshold; after the diagnostic model is trained, determine the confidence index of the model, and at the same time, determine the confidence threshold of the diagnostic model through the validation dataset. Step 3) Construct an autoencoder model: Using the training dataset, train an autoencoder model in parallel to learn the data distribution of the training data; the input of the autoencoder is the fault-sensitive features, and the output is the fault-sensitive features reconstructed by the decoder; at the same time, determine the reconstruction error threshold: after the autoencoder is trained, determine the threshold of the autoencoder reconstruction error by using the validation dataset; Step 4) Online Fault Prediction: The real-time data of the chiller unit will first be used by the autoencoder model to determine whether its reconstruction error exceeds the threshold. If it exceeds the threshold, the data will be entered into the rejection pool for subsequent calibration. If the reconstruction error does not exceed the threshold, it will be entered into the classifier for fault prediction. At this time, if the confidence of the prediction result is less than the threshold, the data will also be entered into the rejection pool for subsequent calibration. If the confidence is higher than the threshold, the diagnosis result of the data will be recorded, and the data will be entered into the result pool. When the amount of data in the recent rejection pool reaches 500, prediction calibration is performed: using a partial domain adaptive algorithm, the training set in the offline stage is used as the source domain, and the data in the rejection pool and the result pool are mixed as the target domain. By using a weighted domain adversarial neural network, the difference in data distribution between the source domain and the target domain is reduced, thereby providing more accurate diagnostic results. The domain adversarial neural network includes a feature extractor, a classifier, and a discriminator. The feature extractor and classifier are taken from the deep learning diagnostic model trained in the offline stage, and the part before the last layer is regarded as the feature extractor, while the last layer is regarded as the classifier. The discriminator measures the difference in data distribution between online data and source domain data, and then the gradient reversal layer enables the feature extractor to minimize this difference in data distribution, so as to obtain the domain adaptation effect. The weights refer to the following: For online data, the diagnostic network first calculates the most likely fault categories for the current online data, assigning a weight to each category; during subsequent training, for each training data point, the category weights are used when calculating the domain adversarial loss and cross-entropy loss. Weighting is applied to categories that are similar to online data; categories that are not similar to online data are assigned higher weights, while those that are not similar are assigned lower weights. The category weights ,in: It is a multi-dimensional vector, with weights for each category, thus filtering out fault categories in the source domain data that are not similar to the current online data.

2. The robust fault prediction method for chiller units based on rejection and calibration mechanism according to claim 1, characterized in that, The loss function of the aforementioned domain adversarial neural network is: Where: cross-entropy loss Domain confrontation loss ,in: These are feature extractor, classifier, and discriminator, respectively. For domain confrontation losses, The cross-entropy loss is used, and the entire network updates its parameters through gradient descent to minimize the domain adversarial loss and cross-entropy loss.

3. The robust fault prediction method for chiller units based on rejection and calibration mechanisms according to claim 1 or 2, characterized in that, specifically... include: S1. For screw chiller units, fault simulation experiments were conducted to collect operating data of the chiller units under typical fault categories. Three faults were collected: refrigerant leakage, reduced evaporator water flow, and reduced condenser water flow. Refrigerant leakage was simulated by artificially reducing the system refrigerant charge, with three fault levels corresponding to 90% (Level 1), 80% (Level 2), and 70% (Level 3) of the nominal charge. The other two faults were simulated by reducing the evaporator and condenser water flow, each with two fault levels corresponding to 85% (Level 1) and 70% (Level 2) of the nominal water flow. Operating data of the unit under 27 operating conditions was collected by adjusting the chiller unit's cooling load, cooling water inlet temperature, and chilled water outlet temperature. Each operating condition was run for at least 30 minutes, with at least 10 minutes of steady-state operation. Steady-state discrimination and outlier removal were performed on the original dataset to remove dynamic and outlier data to avoid impacting model performance. The steady-state discrimination uses steady-state window recognition. By sliding the window, the standard deviation of a specified feature within each window is calculated. When the standard deviation is less than a threshold, the data in that window is considered to be in a steady state, and 10% of the data in that window is randomly selected and added to the training dataset. When the standard deviation is higher than the threshold, the process moves directly to the next window. The outlier removal method used is the quartile outlier removal method. The 25th percentile and 75th percentile of a specified feature are calculated. The upper and lower bounds are determined based on these two quantiles. When a data point exceeds this range for the feature, it is considered an outlier and the data point is removed from the dataset. The cleaned dataset is obtained through steady-state discrimination and outlier removal. S2. Using the cleaned dataset, perform fault-sensitive feature analysis. Specifically, first select data from the normal category in the dataset to construct a fault-free model. The inputs to the fault-free model are cooling load, chilled water outlet temperature, and cooling water inlet temperature. The outputs are fault symptom variables, including condensing pressure, compressor power, subcooling, evaporator temperature difference, and condenser temperature difference. For each fault category in the dataset, the actual value is compared with the baseline value to obtain the residual value. By analyzing the residual value, the fault category-sensitive features can be selected, including: subcooling, condensing pressure, expansion valve opening, evaporator temperature difference, condenser temperature difference, chilled water temperature difference, and cooling water temperature difference. S3, randomly divide the cleaned dataset into training dataset, validation dataset and test dataset in a ratio of 7:1:2; S4. A deep learning classifier is built using the training dataset. The output of the classifier is the probability of each fault category occurring under the current operating state of the unit. The input of the model is the fault-sensitive features obtained in S2. S5, determine the model's confidence metric, and simultaneously determine the threshold for the diagnostic model's confidence using the validation dataset. Specifically, this involves using information entropy. As a confidence index, where p is the probability corresponding to category i, the information entropy of the diagnostic model on the validation set is calculated. The threshold is set to 95% of the validation set data whose information entropy is lower than the threshold. That is, in most cases, the information entropy of a prediction result will not exceed the threshold. When the information entropy of online data exceeds the threshold, the prediction is considered to be inaccurate. S6 uses the training dataset to train an autoencoder model in parallel to learn the data distribution of the training data; the input of the autoencoder is the fault-sensitive features analyzed in S2, and the output is the fault-sensitive features reconstructed by the decoder. S7, determine the threshold of the autoencoder reconstruction error by validating the dataset; this reconstruction error is... ,in: For encoder, For the decoder, the reconstruction error is calculated as the L2 norm between the original features and the reconstructed features. For the validation dataset, the reconstruction error of each data point is calculated, and the threshold for the reconstruction error is set to 95% of the validation dataset. S8. After the two models and their thresholds are determined, the performance of the models is tested using a test dataset to see if they meet the requirements. The test dataset is divided into two parts: acceptance and rejection. When verifying the performance, it is determined whether the indicators of the acceptance data part meet the requirements. If they meet the requirements, the online fault prediction part is entered. Otherwise, the thresholds are readjusted or the process is returned to step S3 to re-divide the dataset and modify the determined hyperparameters. S9, the real-time data of the chiller unit will first be judged by the automatic encoder model to see if its reconstruction error exceeds the threshold. If it exceeds the threshold, the data will enter the rejection pool for subsequent calibration. If the reconstruction error does not exceed the threshold, it will enter the classifier for fault prediction. If the confidence of the prediction result is less than the threshold, the data will also enter the rejection pool for subsequent calibration. If the confidence is higher than the threshold, the diagnosis result of the data will be recorded and the data will enter the result pool. S10, Prediction Result Calibration: If the amount of data in the recent rejection pool reaches 500, prediction calibration is performed: Using a partial domain adaptive algorithm, the training set in the offline stage is used as the source domain, and the data in the rejection pool and the result pool are mixed as the target domain. A weighted domain adversarial neural network is used to narrow the distance between the data distribution of the source domain and the target domain, thereby providing more accurate diagnostic results.

Images