Self-adaptive fault diagnosis method for open set field of water chilling unit

Abstract

The invention discloses an open set field adaptive fault diagnosis method for a water chilling unit, and belongs to the technical field of intelligent operation and maintenance and industrial equipment fault diagnosis. The invention provides an OSDA diagnosis framework combining a multi-level feature fusion lightweight network MFFLN and an extreme value theory EVT. Comprising the following steps: firstly, mapping one-dimensional sequence features of the water chilling unit into a grayscale image so as to enhance spatial structure expression; then, MFFLN is constructed to extract multi-scale features of an image, local details and global structure expressions are considered, and discriminant features between known and unknown faults are learned through end-to-end back propagation OSBP. In order to improve the detection capability of unknown categories, an EVT model is introduced to carry out modeling and identification on abnormal samples outside a discrimination boundary. Meanwhile, a pseudo-label weighted adversarial training strategy is adopted to optimize the feature alignment effect between the source domain and the target domain, so that the classification performance of known faults is improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance and industrial equipment fault diagnosis technology, and specifically designs an open set domain adaptive chiller unit fault diagnosis method. Background Technology

[0002] Heating, ventilation, and air-conditioning (HVAC) systems are a major source of building energy consumption, with chillers serving as core equipment. Their operational efficiency directly determines the overall building energy consumption level and system stability. However, due to complex operating environments and frequent changes in operating conditions, chillers are prone to various performance degradation and malfunctions during long-term operation, such as decreased heat exchange efficiency and refrigerant leakage. These problems not only lead to a significant decrease in unit energy efficiency but may also cause increased energy consumption, increased system vibration, and abnormal equipment shutdowns. Failure to identify and accurately locate faults in a timely manner will result in serious energy waste, equipment wear and tear, and increased maintenance costs. Therefore, developing a high-precision, highly generalizable chiller fault diagnosis technology suitable for edge deployment is of great significance for achieving efficient, safe, and low-carbon operation of equipment.

[0003] In recent years, intelligent fault diagnosis methods based on deep learning and transfer learning have made significant progress in the industrial field, especially demonstrating good adaptability in dealing with differences in data distribution between the source and target domains. However, most existing studies assume that the source and target domains share the same label space, i.e., the Closed-Set Domain Adaptation (CSDA) scenario. But in practical applications, due to factors such as equipment aging, external interference, and environmental changes, unknown fault types that were not seen during the training phase often occur during the operation of chiller units, thus violating the closed-set assumption. This situation is called the Open-Set Domain Adaptation (OSDA) problem. In OSDA, the diagnostic model must not only identify known faults but also have the ability to detect unknown faults; otherwise, a negative transfer effect will occur, leading to a significant decrease in diagnostic performance.

[0004] To address the OSDA problem, existing research has proposed various improved methods, such as adversarial domain adaptive models based on dual-auxiliary classifiers, unknown class detection algorithms incorporating extremum theory, and multi-adversarial training frameworks. These methods have achieved good results in the field of rotating machinery such as bearings and gears. However, compared with rotating machinery, chiller units are typical multi-component coupled systems, and their operating data are characterized by high dimensionality, nonlinearity, and strong correlation, with complex coupling relationships and time-varying characteristics between signals. These characteristics make it difficult for traditional OSDA models to fully capture the multi-level patterns and correlation structures of chiller unit fault data, thus limiting their performance in practical applications.

[0005] Furthermore, existing deep diagnostic models are generally complex in structure, have a large number of parameters, and are highly dependent on computing resources, making them unsuitable for deployment in edge computing or embedded environments. In industrial settings, chiller unit diagnostic systems often require real-time response and low-power operation, which traditional large-scale models struggle to meet. Although lightweight design methods such as knowledge distillation, network pruning, and the combination of lightweight convolutional networks and Transformers have emerged in recent years, most of these methods are still limited to rotating machinery scenarios. When facing complex, multivariable systems like chillers, their generalization ability and robustness remain insufficient.

[0006] Therefore, the existing technologies have the following shortcomings: (1) the transfer learning method based on the closed set assumption cannot adapt to the open set scenario of unknown faults in chiller units; (2) the existing OSDA algorithm has insufficient feature representation ability for complex multi-source heterogeneous data and is difficult to accurately separate inter-class boundaries; (3) the traditional deep network model has a large computational load and is difficult to meet the real-time diagnostic needs of edge devices. Therefore, there is an urgent need for a lightweight open set domain adaptive diagnostic method for chiller units that can accurately identify known and unknown faults and can be efficiently deployed in resource-constrained environments to improve system energy efficiency and intelligent operation and maintenance. Summary of the Invention

[0007] To address the limitations of existing chiller unit fault diagnosis technologies in cross-domain generalization and unknown fault identification, this invention proposes an OSDA diagnostic framework that combines a Multi-level Feature Fusion Lightweight Network (MFFLN) with Extreme Value Theory (EVT). Specifically, the framework includes: first, mapping the one-dimensional sequence features of the chiller unit to grayscale images to enhance spatial structure representation; then, constructing an MFFLN to extract multi-scale features from the images, considering both local details and global structure representation, and learning discriminative features between known and unknown faults through end-to-end backpropagation (Open Set Domain Adaptation by Backpropagation, OSBP); and to improve the detection capability for unknown categories, introducing an EVT model to model and identify anomalous samples outside the discrimination boundary. Simultaneously, a pseudo-label weighted adversarial training strategy is employed to optimize feature alignment between the source and target domains, thereby improving the classification performance of known faults.

[0008] OSDA diagnostic methods based on MFFLN-EVT, such as Figure 1 As shown, the steps are as follows:

[0009] S1: Collect multi-source operating data of the chiller unit under different operating conditions. The data includes 64 feature variables, of which 48 are directly collected sensor data and 16 are derived variables obtained through simulation and calculation. These data mainly include temperature, pressure, flow rate, valve position, electrical power, and cooling capacity. Subsequently, the original one-dimensional data is normalized to construct source and target domain datasets.

[0010] S2: The autocorrelation matrix method is used to map one-dimensional data into a two-dimensional grayscale image to characterize the temporal correlation and spatial coupling between variables and enhance the structurality of feature expression;

[0011] S3: Construct the MFFLN model, extract multi-scale local and global features through depthwise separable convolution (DSC), channel attention and cross-layer feature fusion structure to achieve efficient feature representation and reduce computation.

[0012] S4: Input the grayscale image into MFFLN, perform open set adversarial training using OSBP, introduce the EVT model, perform tail distribution modeling for samples outside the classifier's discrimination boundary, establish an unknown fault detection threshold, and realize the identification of unknown class samples in the target domain.

[0013] S5: Generate pseudo-label samples for the target domain based on the output probability of the classifier, and use the intra-class adaptive confidence threshold to filter out low-confidence or potentially unknown samples to form a high-confidence pseudo-label sample set.

[0014] S6: Input the source domain dataset and the target domain high-confidence pseudo-label dataset into the model, optimize the loss function based on the pseudo-label weighting mechanism, realize the adaptive adjustment of the feature distribution of the source domain and the target domain, and improve the cross-domain diagnostic accuracy.

[0015] S7: After completing model training, the lightweight network is deployed on edge computing devices to achieve real-time monitoring of the chiller unit's operating status, fault identification, and alarm for unknown faults. Attached Figure Description

[0016] Figure 1 This is an overall flowchart of the MFFLN-EVT fault diagnosis method proposed in this invention.

[0017] Figure 2 This is the fault diagnosis framework for MFFLN-EVT.

[0018] Figure 3 The structure of the Squeeze-and-Excitation (SE) module and MFFB.

[0019] Figure 4 This is the network model structure proposed in this invention.

[0020] Figure 5 This is a schematic diagram of the experimental setup for the chiller unit in the ASHRAE1043-RP project.

[0021] Figure 6 This is an open-set diagnostic task on the ASHRAE1043RP dataset.

[0022] Figure 7 Autocorrelation grayscale images for different fault types.

[0023] Figure 8 This section compares the performance of model parameters using different methods.

[0024] Figure 9 The graph shows the diagnostic accuracy (%) of different methods on the ASHRAE dataset.

[0025] Figure 10 This is the confusion matrix for different methods in task A2 of the ASHRAE dataset.

[0026] Figure 11 Visualization of different methods for task A2 in the ASHRAE dataset. Detailed Implementation

[0027] This invention proposes a fault diagnosis method for MFFLN-EVT chiller units. The main process of the method is as follows: Figure 2 The following is a detailed description of specific embodiments of the present invention:

[0028] Step S1, Problem Definition: The source domain dataset is ,in This represents the i-th source domain sample. The samples are from a distribution with corresponding labels. , contains 100 labeled samples. The target domain dataset is 100 samples. ,in Represents the j-th target domain sample, from the distribution , contains Unlabeled samples, and The label spaces of the source and target domains are denoted as follows: and Under OSDA settings, the following conditions are met. The known fault categories are: Unknown fault category is If the source domain contains If there are 1 category, then the target domain contains 1 category. There are several categories. To measure the proportion of unknown fault categories, an openness index is defined. This indicator reflects the proportion of unknown categories in the target domain. The larger the value, the higher the difficulty of the task and the stronger the recognition challenge.

[0029] Step S2, Autocorrelation grayscale image construction: as follows Figure 2 As shown in step 1, this invention employs an autocorrelation matrix construction method to map the one-dimensional sequence features of the chiller unit into an image matrix, characterizing the correlation and dynamic features between system variables. Assuming the sequence feature values ​​of the chiller unit are... ,in This indicates that the data was collected by different sensors. Let be the sequence length. Based on this, construct the autocorrelation matrix. :

[0030] (1)

[0031] Each element in the matrix The calculation formula is:

[0032] (2)

[0033] in, It is the mean of the sequence feature values, calculated using the following formula: .

[0034] Step S3, MFFLN model construction: This invention designs a multi-level feature fusion module (Multi-level Feature Fusion Block, MFFB) (e.g. Figure 3 As shown in the diagram, this module integrates DSC, channel attention mechanisms, and skip connections, aiming to improve feature representation capabilities while meeting the requirements of lightweight deployment. MFFB consists of three different sizes of depthwise convolutional kernels, each used to extract feature information under different receptive fields. It has a large receptive field, which helps to extract global features and long-distance dependencies; It can capture feature information within a medium range, taking into account both local and global characteristics; This module is responsible for information exchange between channels, improving the expressive power of channel fusion and feature space correlation. To further enhance key information, a Squeeze-and-Excitation (SE) module is introduced to dynamically weight and model the channel dimensions, improving the discriminative power of the expression by constructing a channel attention distribution. The SE module has a lightweight structure and low computational cost, making it suitable for embedding in lightweight network structures. Simultaneously, skip connections are used to preserve low-level features, preventing feature information loss during deep convolution. The aforementioned multi-scale features are fused into a unified feature tensor through a concatenation operation. Let the input features be... Where H and W represent the spatial dimensions of the features, C1 is the number of input channels, and the fused intermediate features can be represented as:

[0035] (3)

[0036] in, This represents the feature output after multi-branch fusion. Finally, batch normalization (BN) and the ReLU activation function are used to normalize and non-linearly map the fused features, reducing distribution differences between features and improving model training stability. Then, pointwise convolution (PC) is used to reduce the feature dimension, reducing computational overhead and preserving key features. The final feature output is represented as follows:

[0037] (4)

[0038] in, This represents the final optimized feature output. The MFFLN network structure and its parameter configuration proposed in this invention are as follows: Figure 4 As shown, this includes the network architecture and parameter settings for the feature extractor G, the neighborhood classifier D, and the health status classifier C.

[0039] Step S4, adversarial training and unknown fault detection: such as Figure 2 As shown in step 2, this invention proposes an open-set fault diagnosis framework that integrates OSBP and EVT to improve the model's ability to identify unknown faults under complex working conditions. First, training set samples from the source and target domains are selected, and the feature extractor G and health status classifier C are trained using the OSBP method according to formulas (5) to (7).

[0040] (5)

[0041] (6)

[0042] (7)

[0043] The source domain training set and the target domain test set are input into the trained G and C to obtain the features of the training set samples and the test set samples in the feature space, denoted as [features of G and C]. and The features of correctly classified samples for each class in the source domain training set are represented as follows: ,in Calculate the mean feature value of correctly classified samples in each class c, and denote it as the class center. :

[0044] (8)

[0045] Next, the features of correctly classified samples in each category c and their class centers M are calculated. c Euclidean distance:

[0046] (9)

[0047] Subsequently, the distance distribution for each category was fitted using a Weibull distribution. It returns the fitted Weibull model parameters:

[0048] (10)

[0049] in This refers to the tail size. , , These represent the position parameter, shape parameter, and scale parameter, respectively. During the testing phase, based on the predicted labels of the test samples, the features of each test sample and the class center of its corresponding category are calculated. The Euclidean distance between them is given by the following formula:

[0050] (11)

[0051] in Let be the feature of the j-th test sample in class c. Then, the probability that the sample belongs to the unknown fault category is calculated using the cumulative distribution function (CDF) of the Weibull distribution:

[0052] (12)

[0053] Finally, if If so, the sample is classified as an unknown fault sample.

[0054] Step S5, Pseudo-label weighted adversarial training: To improve the alignment effect of known fault category features, such as... Figure 2 As shown in step 3, based on the approach proposed in step S4, this invention further introduces a pseudo-label weighted adversarial training strategy. First, using the already trained feature extractor G and classifier C, pseudo-labels are applied to the target domain training set samples in each training cycle, generating the predicted label with the highest confidence as the pseudo-label.

[0055] (13)

[0056] in, Indicates a pseudo tag. Indicates its confidence level. This represents the index and maximum value of the Softmax output. To avoid introducing unknown faults and low-confidence pseudo-labels, only labels with a confidence level higher than the class adaptive threshold are selected. The samples constitute the pseudo-label sample set:

[0057] (14)

[0058] Among them, the pseudo-label threshold An intra-class adaptive strategy is used to determine the mean confidence score of all pseudo-labeled samples belonging to class A of the target domain in each mini-batch. This is used as the threshold for that class. A sample is only retained if its confidence level is higher than the within-class mean. The pseudo-labeled samples are merged with the source domain samples to form the augmented training set. During each training epoch, the pseudo-label set is dynamically updated based on the prediction results. Therefore, the classification loss function during training is:

[0059] (15)

[0060] in, yes The number of samples in the middle yes The j-th sample in the dataset.

[0061] Furthermore, to improve the effectiveness of adversarial training, weights need to be assigned to the target domain samples to balance the uncertainty differences between known and unknown class samples. (Information entropy) Information entropy is used as a weighting metric to measure the uncertainty of prediction. Known fault samples have a distribution close to the source domain, resulting in more accurate predictions and lower information entropy; unknown fault samples show greater variation, leading to higher prediction uncertainty and higher information entropy. The formula for calculating information entropy is as follows:

[0062] (16)

[0063] in, Let be the predicted probability for class 𝑘. The information entropy, after normalization, is... The loss function for domain classification is:

[0064] (17)

[0065] in, and These represent the domain labels. Finally, the model's total loss function is defined as:

[0066] (18)

[0067] in, To balance the hyperparameters, the network's parameter update rules are as follows:

[0068] (19)

[0069] Implementation Cases

[0070] This invention is experimentally validated using the ASHRAE1043-RP dataset published by the American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHEAE). The ASHRAE1043-RP dataset is widely used in chiller fault diagnosis research, often employed to verify the applicability and effectiveness of related technologies in performance evaluation and fault identification. The data originates from a simulation experiment of a 90-ton (316 kW) centrifugal chiller using R134a refrigerant, covering various operating conditions and states. The data for the centrifugal chiller was collected through simulation experiments, covering multiple operating conditions and typical running states. The operating status includes one normal state (NO) and seven typical fault types: Excess Oil (EO), Non-condensable Gas (NC), Refrigerant Leakage (RL), Refrigerant Overcharge (RO), Insufficient Evaporator-side Water Flow (FWE), Insufficient Condenser-side Water Flow (FWC), and Condenser Fouling (CF). Figure 5 As shown, the experiment simulated 27 different operating conditions by adjusting three main operating condition variables: evaporator water temperature, condenser water temperature, and cooling load, to comprehensively evaluate the performance of the chiller unit under different operating conditions. Data acquisition was performed at 10-second intervals, collecting system status data from 48 sensors and 16 real-time calculated parameters, including information on temperature, pressure, power, and flow rate. Detailed parameter descriptions can be found in the literature. Condenser water temperature, as an important variable with a high frequency of change during the chiller unit's operating life, was selected as one of the key indicators of operating conditions. To simulate complex differences in operating conditions, the experiment specifically set three different condensate water temperatures (40°C, 45°C, and 50°C) to represent different operating conditions, in order to test the robustness and adaptability of the chiller unit under different operating conditions.

[0071] like Figure 6As shown, this dataset includes seven open-set diagnostic tasks, with 200 samples for each health state. All data from the source domain are used for model training, 90% of the target domain data is used for model training, and the remaining 10% is used for model testing. To reduce the impact of random factors and ensure the reliability of performance evaluation, each task is independently repeated ten times, and the final result is the average of the experimental data from each run as the evaluation metric. The input is a one-dimensional sequence feature length, and the constructed grayscale image is 64×64 as shown. Figure 7 As shown.

[0072] The present invention uses the following four evaluation indicators to comprehensively evaluate the model performance, namely: (1) OS* represents the classification accuracy of known fault categories, and its calculation formula is OS*=M k / N k M k N represents the number of correctly classified samples among the known fault categories. k (2) UNK represents the classification accuracy for unknown fault categories, calculated as UNK=M u / N u M u N represents the number of samples correctly classified within the unknown fault category. u (3) OS represents the total number of samples with unknown fault categories, calculated as (M k +M u ) / (N k +N u (4) H-score is the harmonic mean of the classification accuracy of known fault categories and unknown fault categories. Its calculation formula is H-score=(2×OS*×UNK) / (OS*+UNK).

[0073] To comprehensively evaluate the performance of the MFFLN-EVT method proposed in this paper, this invention selects several advanced intelligent fault diagnosis methods for comparative analysis.

[0074] (1)Liconvformer: A lightweight model that combines Transformer and convolutional neural network. It does not perform domain-adaptive training and is tested directly on the target domain after training in the source domain.

[0075] (2) DANN: A classic unsupervised domain adaptation method used for feature alignment. G uses a basic convolutional neural network architecture, while C and D both use... Figure 4 The structure shown.

[0076] (3) MK-MMD: An improved method based on maximum mean difference, used to measure the global difference between the feature distributions of the source and target domains. G uses a basic convolutional neural network architecture, and C also uses... Figure 4 The structure.

[0077] (4) Basic: Uses the basic model designed in this paper, where both G and C are used. Figure 4 The structure was not trained using DA; it was trained in the source domain and then tested directly in the target domain.

[0078] (5) OSBP: An adversarial training method specifically designed for OSDA problems. G uses a basic convolutional neural network architecture, and C uses... Figure 4 The structure.

[0079] (6) BWAN-EVT: A domain-adaptive model of BWAN combined with EVT for unknown fault category detection. It can effectively detect unknown fault categories and ensure the diagnostic accuracy of known fault categories.

[0080] (7)MALDA: A multi-adversarial learning model for cross-domain adaptive fault diagnosis that can simultaneously identify known fault categories and unknown fault categories.

[0081] This invention compares the proposed MFFLN model with five advanced intelligent diagnostic models, including ResNet, MobileNetV2, SqueezeNet, UL-GoogLeNet, and MPNet, with the latter four models employing lightweight designs. Task A1 was selected as the test object. In the experiments, all comparison models were used as feature extractors, and the models were comprehensively evaluated from five aspects, including: computational complexity (FLOPs), parameter size (Parameters), overall performance index (H-score), and time efficiency (training time and testing time). Figure 8As shown, MFFLN exhibits a significant lightweight advantage in terms of parameter size, with only 0.89MB of model parameters, a reduction of over 95% compared to ResNet (18.62MB). Compared to other models, MFFLN maintains excellent diagnostic performance while significantly reducing model complexity. In terms of diagnostic accuracy, MFFLN achieves the highest accuracy, fully demonstrating its strong ability to discriminate abnormal patterns in complex industrial scenarios, achieving an effective balance between lightweight structure and high diagnostic performance. Although the MFFB introduced in MFFLN increases computational complexity to some extent, the overall increase in computational cost is limited and remains within an acceptable range for engineering deployment. Regarding computational efficiency, MFFLN's training and testing times are comparable to other lightweight models, indicating its good time performance and deployability. In contrast, while UL-GoogLeNet and MPNet perform well on typical rolling bearing datasets, their diagnostic accuracy significantly decreases in industrial systems such as chiller units with strong variable coupling and complex operating conditions. These results further validate the robustness and domain adaptability of MFFLN under complex operating conditions, demonstrating its application potential in practical industrial fault diagnosis.

[0082] To comprehensively evaluate the diagnostic performance of the proposed method, comparative experiments were conducted using the MFFLN-EVT method with other methods. The experimental results are as follows: Figure 9As shown in the results, the proposed method exhibits the best diagnostic performance, with a final average classification accuracy of 88.28%. From an engineering application perspective, the superior performance of MFFLN-EVT means that the model can more accurately identify the performance degradation and potential faults of chillers in the early stages, thereby effectively preventing them from evolving into serious energy efficiency losses or sudden shutdown events, and improving the operational reliability and energy utilization efficiency of chillers. As a control method without domain adaptive training, Basic improves the average classification accuracy of known fault categories by 6.79% compared to LicovFormer, demonstrating a significant performance advantage. Furthermore, compared with the traditional CSDA method, Basic performs essentially on par in terms of classification accuracy of known fault categories, which further verifies that the basic model designed in this paper has a strong feature extraction capability. Even without domain adaptive training, the model can still extract cross-domain shared features of known fault categories from the source domain quite well. The comparison results between Basic and MFFLN-EVT show that MFFLN-EVT improves the average classification accuracy of known fault categories by 13.72%, demonstrating a significant advantage. This result fully verifies the effectiveness of the pseudo-label weighted adversarial training method proposed in this invention for known fault identification. This method enhances the discriminative feature learning of known fault categories in the target domain by introducing high-confidence pseudo-label samples. Simultaneously, the information entropy-based weighting mechanism effectively suppresses interference from uncertain samples, improving the model's ability to distinguish fine-grained feature differences. Furthermore, assigning higher weights to known fault category samples allows the model to focus more on key features during training, thereby improving diagnostic accuracy under complex operating conditions of chiller units. Comparison results with three OSDA methods—OSBP, BWAN-EVT, and MALDA—show that MFFLN-EVT has significant performance advantages in OS*, UNK, and OS metrics, especially compared to OSBP, where its improvement is particularly pronounced. Although the difficulty of the diagnostic task increases with the number of unknown fault categories, MFFLN-EVT consistently demonstrates relatively stable diagnostic performance in experiments. Moreover, the latest methods such as LicovFormer, BWAN-EVT, and MALDA exhibit excellent performance in fault diagnosis of relatively simple equipment such as rolling bearings. However, because these methods use one-dimensional sequences as input, they struggle to fully extract the correlation information and spatial structural features among multiple variables. When faced with complex systems like chillers, which are highly coupled and exhibit strong dynamic nonlinearity, their feature extraction capabilities are limited, leading to a significant decrease in diagnostic performance. In contrast, MFFLN-EVT, by introducing grayscale encoding and a multi-level feature fusion mechanism, effectively enhances the model's ability to model the complex feature distribution of chillers, demonstrating stronger cross-domain adaptability and diagnostic accuracy.

[0083] To more intuitively evaluate the diagnostic performance of different models, task A2 was selected to plot the classification confusion matrix, and the t-SNE algorithm was used to visualize the features of the test samples, such as... Figure 10 , Figure 11 As shown in the diagram, for methods (a) to (d), a large number of UNK class samples are misclassified as known fault categories (mainly CF class), and there is significant overlap in feature distribution between categories (primarily between UNK and CF classes), leading to classification confusion and a lack of ability to identify unknown fault categories. For the three OSDA methods (e) to (g), they can identify unknown fault categories, but some UNK class samples are still misclassified as CF class, and feature overlap between UNK and CF classes still exists. However, compared to methods (a) to (d), the distribution of known fault category samples is clearer, the distance between classes is significantly increased, and the discrimination effect is significantly improved. The MFFLN-EVT method shows completely correct results in the classification of all samples. Samples of different classes are clearly separated in the feature space, with no overlap between classes. UNK class samples are distributed in independent regions and are completely separated from known fault categories.

[0084] In summary, experimental results on a cross-domain fault diagnosis dataset for chiller units demonstrate that the lightweight open-set domain adaptive method based on MFFLN-EVT proposed in this invention significantly improves diagnostic accuracy and stability even when the source and target domains are inconsistent and unknown fault categories exist. It outperforms existing comparative models across different chiller units and multiple operating conditions, achieving an average diagnostic accuracy of 88.28%, exhibiting excellent generalization performance and cross-domain adaptability. Furthermore, this method, while maintaining model accuracy, meets the real-time and low-power requirements of edge computing environments, possessing good engineering practicality and promotional value.

Claims

1. An open set domain adaptive fault diagnosis method for water chiller units, characterized in that, The implementation steps of the method include the following: S1: Collecting multiple-source operation data of the water chilling unit under different working conditions, specific data including 64 characteristic variables, 48 of which are directly collected sensor data, and the other 16 are derived variables obtained through simulation and calculation; including temperature, pressure, flow, valve position, electric power, refrigeration capacity; Then, the original one-dimensional data is normalized to construct the source domain and target domain data sets; S2: Using the autocorrelation matrix method to map one-dimensional data to two-dimensional gray images to depict the time correlation and spatial coupling relationship between variables and enhance the structural representation of features; S3: Constructing an MFFLN model to extract multi-scale local and global features through deep separable convolution DSC, channel attention, and cross-layer feature fusion structure, realizing efficient feature representation and reducing computational complexity; S4: Inputting the gray image into the MFFLN and using the OSBP method for open set adversarial training, introducing the EVT model to model the tail distribution of the classifier's discriminant boundary samples, establishing an unknown fault detection threshold, and realizing the identification of unknown class samples in the target domain; S5: Generating target domain pseudo-label samples according to the classifier output probability, using intra-class adaptive confidence threshold to filter out low-confidence or potentially unknown samples, and forming a high-confidence pseudo-label sample set; S6: Inputting the source domain data set and the target domain high-confidence pseudo-label data set into the model, optimizing the loss function based on the pseudo-label weighting mechanism, realizing the adaptive adjustment of the feature distribution between the source domain and the target domain, and improving the cross-domain diagnosis accuracy; S7: After completing the model training, deploying the lightweight network on the edge computing device to realize real-time monitoring, fault identification, and unknown fault alarm of the water chilling unit operation state.

2. The open set domain adaptive fault diagnosis method for water chillers according to claim 1, wherein, A source domain dataset is wherein represents an i-th source domain sample, is a corresponding label thereof, and the samples are from a distribution , and the source domain dataset contains a total of labeled samples; and a target domain dataset is wherein represents a j-th target domain sample, from a distribution , and the target domain dataset contains a total of unlabeled samples, and ; The label spaces of the source domain and the target domain are denoted as and respectively; under the OSDA setting, it satisfies where the known fault classes are and the unknown fault classes are ; if the source domain contains classes, the target domain contains classes; to measure the proportion of unknown fault classes, the openness index is defined as .

3. The open set domain adaptive fault diagnosis method for water chillers according to claim 2, wherein, Autocorrelation gray image construction: using the autocorrelation matrix construction method to map one-dimensional sequence features of the water chilling unit to image matrices to depict the correlation and dynamic characteristics between system variables; Assume the sequence eigenvalues of the water chiller are where denote the data collected by different sensors, is the sequence length; construct the autocorrelation matrix : (1) The calculation formula of each element in the matrix is: The calculation formula of each element in the matrix is: (2) wherein is the mean of the sequence characteristic values, calculated as .

4. The open set domain adaptive fault diagnosis method for water chillers according to claim 3, wherein, MFFLN model construction: a multi-level feature fusion module MFFB is designed to fuse DSC, channel attention mechanism and skip connection; MFFB is composed of three different size deep convolution kernels, which are used to extract feature information under different receptive fields; SE module is introduced to dynamically weight the channel dimension, and the discriminative expression is improved by constructing channel attention distribution; the structure of SE module is light and the computational overhead is low, which is suitable for embedding in light network structure; multi-scale features are fused into a unified feature tensor through splicing operation; let the input feature be where H and W represent the spatial dimensions of the feature, C1 is the input channel number, and the fused intermediate feature can be represented as: (3) wherein, represents the multi-branch fused feature output; finally, batch normalization BN and ReLU activation function are used to normalize and nonlinearly map the fused features, and the feature dimension is reduced by point convolution PC to reduce the calculation overhead and retain key features; the final feature output is represented as: (4) wherein, represents the final optimized feature output; the MFFLN network structure and its parameter configuration include the network architecture and parameter settings of the feature extractor G, the domain classifier D, and the health state classifier C.

5. The open set domain adaptive fault diagnosis method for water chiller units according to claim 4, wherein, Adversarial training and unknown fault detection: using an open set fault diagnosis framework that combines OSBP and EVT to improve the model's ability to identify unknown faults under complex working conditions; first, select the training set samples of the source domain and the target domain, and use the OSBP method to train the feature extractor G and the health state classifier C according to formulas (5) to (7); (5) (6) (7) The source domain training set and the target domain test set are input into the trained G and C to obtain the features of the training set samples and the test set samples in the feature space, denoted as and respectively; the feature of each correctly classified sample in the source domain training set is denoted as , wherein ; the feature mean of the correctly classified samples in each category c is calculated, denoted as the category center : (8) Next, the Euclidean distance between the feature of a correctly classified sample in each class c and its class center M c is calculated: (9) Subsequently, the distance distribution for each class is fitted using a Weibull distribution and the fitted Weibull model parameters are returned: (10) wherein is a tail size, , , respectively denote a location parameter, a shape parameter and a scale parameter; During the testing phase, based on the predicted labels of the test samples, the features of each test sample and the class center of its corresponding category are calculated. The Euclidean distance between them is given by the following formula: (11) wherein is the feature of the jth test sample in class c; then, the probability that the sample belongs to an unknown failure class is calculated using the cumulative distribution function CDF of the Weibull distribution: (12) Finally, if then the sample is determined to be an unknown failure sample.

6. The open set domain adaptive fault diagnosis method for water chiller units according to claim 5, wherein, Pseudo-label weighted adversarial training: based on step S4, introduce the pseudo-label weighted adversarial training strategy; first, use the trained feature extractor G and classifier C to label the target domain training set samples with the highest confidence prediction label as pseudo-labels in each training cycle: (13) wherein, represents a pseudo label, represents a confidence thereof, represents an index of a maximum value of a Softmax output and the maximum value; to avoid introducing unknown faults and low-confidence pseudo labels, only samples whose confidence is higher than a class adaptive threshold constitute a pseudo label sample set: (14) wherein the pseudo-label threshold The in-class adaptive strategy is adopted to determine that, in each mini-batch, the mean of the confidence of all pseudo-label samples belonging to the c-th class of the target domain is used as the class threshold; when the sample confidence is higher than the in-class mean, the sample is retained; the pseudo-label samples and the source domain samples are merged to form an enhanced training set ; in each training period, the pseudo-label set is dynamically updated with the prediction result; during the training process, the classification loss function is: (15) wherein is the number of samples in is the jth sample in To improve the effect of adversarial training, weights are assigned to target domain samples to balance the uncertainty difference between known and unknown class samples; information entropy is taken as a weight indicator to measure the uncertainty of prediction; the distribution of known fault samples is close to the source domain, the prediction is more accurate, and the information entropy is lower; the difference of unknown fault samples is large, the prediction uncertainty is high, and the information entropy is higher; the calculation formula of information entropy is as follows: (16) wherein, is the prediction probability of the kth category; the information entropy is normalized as ; and the loss function of the domain classification is (17) where, and denote domain labels; finally, the total loss function of the model is defined as: (18) wherein, is the balance hyperparameter; the parameter update rule of the network is as follows: (19)。

Images