Water chilling unit small sample fault detection method and system based on transfer learning

Abstract

The invention discloses a water chilling unit small sample fault detection method and system based on transfer learning, and the method carries out the fault detection through an expansion causal convolution module, a dense neural network module and a classification layer which are connected in sequence. A sparse connection strategy is introduced into dense blocks of the fault detection model, and representative connection between far and near layers is only reserved in each dense block, so that the network complexity is reduced, the feature redundancy is reduced, and the feature multiplexing advantage is kept; meanwhile, the method utilizes a transfer learning strategy to transfer labeled data knowledge of a source domain to a target domain, and introduces a meta-learning thought in a fine tuning stage to carry out gradient updating. In addition, feature subsets with high discrimination ability are screened out from an original high-dimensional feature space in combination with importance scores and correlation analysis, so that the detection precision of model training is improved.

Description

Technical Field

[0001] This invention belongs to the field of chiller unit fault detection technology, specifically relating to a small-sample fault detection method and system for chiller units based on transfer learning. Background Technology

[0002] Chillers, as a key component of HVAC systems, are commonly used to control the temperature and humidity of building environments to meet indoor thermal comfort requirements. Statistics show that HVAC systems account for more than half of the total energy consumption in the building sector. Various chiller malfunctions can reduce the efficiency of HVAC systems by nearly 30%, resulting in significant energy waste. Therefore, timely handling of chiller malfunctions is crucial for reducing energy waste and maintaining indoor thermal comfort.

[0003] Deep learning-based fault detection methods have become a research hotspot due to their powerful automatic feature extraction capabilities. However, deep learning models typically require a large amount of labeled data for training. In practical applications, fault samples from chiller units are often scarce, and the distribution of fault features varies significantly under different operating conditions. This makes it difficult to directly apply a model trained for one operating condition to other conditions, thus limiting the model's versatility. Although generative modeling and metric learning methods alleviate the data shortage to some extent, they usually assume that the training and testing data distributions are similar. Therefore, when faced with distribution differences caused by different operating conditions, it is difficult to guarantee the model's accuracy and generalization ability. In contrast, transfer learning (TL) utilizes labeled data in the "source domain" to train a basic model with general fault feature learning capabilities. This model is then transferred to the "target domain" and fine-tuned using a small number of target domain samples. This effectively alleviates the model training difficulties caused by insufficient target domain samples, becoming an effective solution for the problem of small-sample fault detection.

[0004] Current transfer learning research often employs shallow network structures as the base model. However, in practical chiller unit fault detection tasks, due to the high diversity and complexity of fault types and data features, deep networks can significantly improve the model's feature representation capabilities. However, traditional neural networks are prone to problems such as vanishing gradients and training difficulties as the number of layers increases. Residual neural networks (ResNets) alleviate the vanishing gradient phenomenon by introducing residual connections, allowing information to be transferred across layers. However, their local inter-layer connections limit the efficiency of information flow and feature reuse. Furthermore, chiller unit operating data exhibits strong time-series characteristics, meaning the current operating state is influenced by previous states. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for detecting small-sample faults in chiller units based on transfer learning.

[0006] In a first aspect, the present invention provides a small-sample fault detection method for chiller units based on transfer learning, the method comprising:

[0007] The original features of the chiller unit under different state types under historical and current operating conditions are obtained, and high discriminative features are selected from the original features to construct a feature subset; the feature subset corresponding to the historical operating conditions is used as the source domain, and the feature subset corresponding to the current operating conditions is used as the target domain; the samples composed of high discriminative features in the source domain and the target domain are preprocessed respectively.

[0008] A fault detection model is constructed; the fault detection model processes and classifies the input samples sequentially through a dilated causal convolution module, a dense neural network module, and a classification layer to obtain the detection result of the state type;

[0009] The dense neural network module includes multiple dense blocks connected in sequence; a transition layer is connected in series between adjacent dense blocks; each dense block includes multiple bottleneck layers; in a dense block, except for the first bottleneck layer, the input of the remaining bottleneck layers is part or all of the output feature maps of the bottleneck layers preceding that bottleneck layer.

[0010] The source domain is used to pre-train the fault detection model. After pre-training, some dense blocks in the fault detection model and the weights of the corresponding transition layers are frozen to obtain the base model. The target domain is divided into multiple meta-tasks according to the state type, and the base model is trained by using the divided meta-tasks in sequence.

[0011] The trained base model is used to detect faults in the chiller unit under test based on its high discriminative characteristics.

[0012] Preferably, the method for obtaining the input feature map of the bottleneck layer is as follows:

[0013] In a dense block, if the number of layers in the current bottleneck layer is less than or equal to the maximum number of connected layers N, then the current bottleneck layer uses the concatenation result of the output feature maps of all previous bottleneck layers as its input feature map; if the number of layers in the current bottleneck layer is greater than the maximum number of connected layers N, then the current bottleneck layer selects the concatenation result of the output feature maps of N bottleneck layers from all previous bottleneck layers as its input feature map.

[0014] As a preferred method, the selection of N bottleneck layers as input feature maps is as follows: if the maximum number of connected layers is even, then N / 2 bottleneck layers are selected sequentially from both the initial bottleneck layer and the current bottleneck layer; if the maximum number of connected layers is odd, then N / 2 bottleneck layers are selected sequentially from both the initial bottleneck layer and the current bottleneck layer. Bottleneck layers, select sequentially from one side of the current bottleneck layer. Bottleneck layer.

[0015] Preferably, the screening process for the highly discriminative features is as follows:

[0016] Importance scores of different original features are obtained, and multiple original features with high importance scores are selected as key features based on the importance scores. Feature subsets are constructed based on the key features. Correlation analysis is performed on the key features in the feature subsets to obtain the correlation between different key features in the feature subsets. Based on the correlation between different key features, the key features are further filtered to obtain highly discriminative features.

[0017] As a preferred option, the screening process for key features is as follows:

[0018] Obtain the correlation between different key features in the feature subset; if the absolute value of the correlation between a key feature and the other key features is greater than the correlation threshold, remove the key features with lower importance scores from the feature subset, and use the key features retained in the feature subset as high discriminative features.

[0019] Preferably, the importance score is obtained by processing the samples in the dataset using a fault classification model and obtaining the importance score by the difference between the classification results before and after adding perturbation.

[0020] Preferably, the bottleneck layer includes a first convolutional block and a second convolutional block connected in series; the first convolutional block and the second convolutional block have the same structure, both including a batch normalization layer, an activation function and a convolutional layer connected in sequence; the convolutional layers in the first convolutional block and the second convolutional block have different sizes.

[0021] Preferably, the transition layer includes a batch normalization layer, an activation function, a convolutional layer, and an average pooling layer connected in sequence.

[0022] Preferably, the classification layer includes a series of average pooling layers and linear layers; the output feature map of the dense neural network module is first passed to the average pooling layer, the feature map is reduced in dimensionality by the average pooling layer, and then the extracted features are integrated by the linear layer, and the probability distribution of each category is calculated, thereby completing the fault classification.

[0023] Secondly, the present invention provides a small-sample fault detection system for chiller units based on transfer learning, which is used to execute the aforementioned small-sample fault detection method for chiller units. The fault detection system includes a data acquisition module, a feature filtering module, a preprocessing module, and a fault detection module. The data acquisition module is used to acquire the original features of the chiller unit under different operating conditions. The feature filtering module is used to filter out high-discriminative features from the original features. The preprocessing module is used to perform data segmentation on the samples composed of high-discriminative features. The fault detection module is used to complete the fault detection of the chiller unit based on the high-discriminative features.

[0024] The beneficial effects of this invention are:

[0025] 1. This invention introduces a sparse connection strategy in the dense blocks of the fault detection model to adjust the connection relationship between different bottleneck layers. By retaining only representative inter-layer connections between near and far layers in each dense block, the network complexity is reduced and feature redundancy is decreased, while maintaining the advantage of feature reuse. At the same time, to improve the training stability of deep networks, the SELU activation function is introduced in the dense blocks and transition layers. Its self-normalization property is used to keep the output mean close to zero and stabilize the gradient, thereby enhancing the training efficiency of the model in deep structures.

[0026] 2. This invention combines importance scoring and correlation analysis to screen original features. The importance scoring is used to initially screen original features and to assist in the removal of features in the correlation analysis. This allows for the selection of a subset of features with strong discriminative power from the original high-dimensional feature space, thus solving the interference caused by feature redundancy.

[0027] 3. This invention processes the input data through a dilated causal convolution module before the dense neural network module, overcoming the limitations of the dense neural network module in capturing temporal dependencies over long periods of time, expanding the receptive field of the convolution kernel, ensuring the causality of information transmission, and significantly enhancing the model's ability to process time series data.

[0028] 4. This invention employs a transfer learning strategy to train the fault detection model, transferring labeled data knowledge from the source domain to the target domain. This addresses the problem of scarce fault samples in chiller units and significant differences in fault characteristics under different operating conditions. Furthermore, this invention introduces meta-learning in the fine-tuning stage of transfer learning. By dividing a small number of samples in the target domain into multiple meta-tasks and updating gradients on each meta-task, the generalization performance and detection accuracy of the model are further improved. Attached Figure Description

[0029] Figure 1 This is the overall flowchart of the present invention.

[0030] Figure 2 This is a schematic diagram of time series sample partitioning in this invention.

[0031] Figure 3 This is a structural diagram of the fault detection model in this invention.

[0032] Figure 4 The graphs show a comparison of model loss curves under different activation functions; where (a) is a schematic diagram of the ReLU activation function and (b) is a schematic diagram of the SELU activation function.

[0033] Figure 5 This is a schematic diagram of the connection method between the dense block of the present invention and the bottleneck layer in the existing dense block; wherein, (a) is a schematic diagram of the connection method of the bottleneck layer in the existing dense block; and (a) is a schematic diagram of the connection method of the bottleneck layer in the dense block of the present invention.

[0034] Figure 6 This is a schematic diagram of the transfer learning strategy in this invention.

[0035] Figure 7 This is a schematic diagram of the meta-learning partitioning task in this invention.

[0036] Figure 8 This is a schematic diagram illustrating the importance scores of different original features in this invention.

[0037] Figure 9 This is a schematic diagram illustrating the correlation between different key features in this invention.

[0038] Figure 10 This is a comparison chart of the accuracy of each method under different conditions. Detailed Implementation

[0039] The present invention will be further described below with reference to the accompanying drawings.

[0040] A small-sample fault detection method for chiller units based on transfer learning is proposed. The fault detection system includes a data acquisition module, a feature selection module, a preprocessing module, and a fault detection module. The data acquisition module collects the raw features of the chiller unit under different operating conditions. The feature selection module filters out high-discriminative features from the raw features. The preprocessing module partitions the samples composed of high-discriminative features. The fault detection module performs fault detection on the chiller unit based on the high-discriminative features.

[0041] like Figure 1 As shown, the small-sample fault detection method for this chiller unit includes the following steps:

[0042] Step 1: Building the dataset

[0043] Operating data of the chiller unit under different operating conditions (normal and fault states) was collected using sensors. The data selected in this embodiment comes from the open-source fault experiment dataset in the ASHRAE 1043-RP project. This dataset uses a 316kW (90rt) centrifugal chiller unit for fault experiments, testing the chiller unit under 27 operating conditions in both normal and fault states (including 7 fault types), and studying the four fault severity levels (SL) for each fault test. The obtained operating data contains 64 raw features, of which 48 were directly measured by sensors, and the remaining 16 were calculated. Descriptions of different state types (including one normal state and 7 fault types) are shown in Table 1.

[0044] Table 1 Status Types

[0045]

[0046] The original dataset was constructed by selecting operating data corresponding to normal and fault states under the SL1 fault level of the chiller unit to verify the effectiveness of the proposed method in fault detection. Include Each timestamp and its corresponding running data is used as a raw sample, as shown in Table 2.

[0047] Table 2. Time series representation of the dataset

[0048]

[0049] The original samples in the original dataset are sorted according to time series, and the sorted original dataset D is represented as follows: ;in, Represents a time step sequence; For the i-th original sample, ; Let j be the j-th original feature in the i-th original sample; n is the original number of samples in the dataset; ; This represents the number of original features in the running data.

[0050] Step 2: Feature Filtering

[0051] Step 2-1: Construct a feature subset

[0052] A random forest model is used to process the samples in the dataset. The actual contribution of different original features is evaluated through a performance degradation metric, and the importance score of each original feature is obtained. Its expression is:

[0053]

[0054] in, The label is a status type The proportion of the sample in all samples; Indicates the random forest model in terms of state type The classification accuracy is as follows; Indicating in relation to the first After perturbing the original features, the random forest model changes the state type. The classification accuracy is as follows; ; This represents the number of fault types.

[0055] Sort the importance scores of all original features in the dataset in descending order, select the top d original features with the highest importance scores as key features, and construct a feature subset S. initial :

[0056]

[0057] In this embodiment, there are 30 key features.

[0058] Step 2-2: Filter the feature subset

[0059] Pearson correlation coefficient was used to perform correlation analysis on key features in the feature subset to obtain the correlation between different key features in the feature subset. Its expression is:

[0060]

[0061] in, and The key features of the i-th original sample; and The mean of the key features in the dataset; .

[0062] Set the relevance threshold to θ; if key features With other key features If the absolute value of the correlation between features exceeds a correlation threshold θ, then key features with lower importance scores are removed from the feature subset, and the key features retained in the feature subset are used as high-discriminative features. The original samples in the feature subset are composed of different high-discriminative features. The feature subset is divided into a source domain and a target domain according to different operating conditions. The source domain is historical operating condition data, which includes a large amount of labeled data, while the target domain is current operating condition data, which contains only a small amount of labeled data and is insufficient to support the complete training of the model.

[0063] In this embodiment, the correlation threshold θ = 0.9; the number of highly discriminative features is 21.

[0064] Step 3: Data Preprocessing

[0065] Step 3-1: Normalization

[0066] The L2 normalization method is used to normalize the original samples in the feature subset, so that the norm of the normalized vector is 1. The original samples after normalization. Represented as:

[0067]

[0068] Where x is the original sample before normalization.

[0069] Step 3-2, Data Segmentation

[0070] like Figure 2 As shown, a fixed-size time window (W in size and S in step size) is used to segment the original samples in the feature subset, resulting in M ​​input samples. ( ); each input sample Let be a W×J matrix, which includes elements from t i to t i+W-1 Data over a continuous time period is represented as:

[0071]

[0072] Where J represents the number of highly discriminative features.

[0073] Input Sample Total number for:

[0074]

[0075] in, This indicates rounding up to the nearest integer.

[0076] In this embodiment, the time window W is 10; the sliding step S is 5.

[0077] Step 3-3, Label Alignment

[0078] In fault detection tasks, labels typically represent the device status (normal or faulty) within an input sample; therefore, each input sample x m Each will correspond to a label y m ,Right now:

[0079]

[0080] in, Let be the label corresponding to the m-th input sample.

[0081] In this example, the category remains consistent within each sample, and each sample is labeled accordingly.

[0082] Step 4: Construct a fault detection model

[0083] like Figure 3 As shown, the fault detection model includes a dilated causal convolutional module (DCC), a dense neural network module (Dense Net), and a classification layer. The fault detection model processes the input samples sequentially through the dilated causal convolutional module, the dense neural network module, and the classification layer to detect the fault type.

[0084] The dilated causal convolution module is used for preliminary feature extraction from the multidimensional time-series data of the input samples. This module employs a causal convolution structure, ensuring that the output of the fault detection model at each time point depends only on current and past information, avoiding the leakage of future information. Simultaneously, a dilation rate is introduced into the dilated causal convolution module, causing the convolution kernel to sample at intervals along the time dimension, thus expanding the receptive field. Its convolution operation is shown below:

[0085]

[0086] in, Indicates time The output value; Indicates time The input value; The first convolution kernel Each weight; The size of the convolution kernel; The dilation rate determines the sampling interval of the convolution kernel on the input sequence. This refers to traditional convolution.

[0087] In this embodiment, the kernel size in the dilated causal convolution module is 3×1; the dilation rate is 2.

[0088] The dense neural network module consists of four sequentially connected dense blocks; transition layers are connected in series between adjacent dense blocks. Each dense block includes multiple bottleneck layers; each bottleneck layer consists of a first convolutional block and a second convolutional block connected in series. The first convolutional block consists of a batch normalization layer, a SELU activation function, and a 1×1 convolutional layer connected in series; the second convolutional block consists of a batch normalization layer, a SELU activation function, and a 3×1 convolutional layer connected in series. The transition layers consist of a batch normalization layer, a SELU activation function, a 1×1 convolutional layer, and a 2×1 average pooling layer (with a stride of 2) connected in series. In the transition layers, convolution and average pooling operations are used to reduce the number and size of feature maps, thereby processing and compressing the feature information from the previous dense block, making it more efficiently input into the next dense block.

[0089] In this embodiment, the number of bottleneck layers in the four dense blocks are 6, 12, 24, and 16, respectively.

[0090] By introducing the SELU activation function into the bottleneck and transition layers, and leveraging its self-normalization properties, the mean output of each layer is made close to zero while maintaining gradient stability, thereby enhancing the training efficiency of the model in deep structures. Figure 4 As shown. From Figure 4 As can be seen, the loss curves of this invention using the SELU activation function are smoother and converge faster than those using the ReLU activation function, indicating that the SELU activation function can effectively improve the training stability of deep models. The formula for the SELU activation function is as follows:

[0091]

[0092] Where x is the input to the SELU excitation function; λ is the scaling factor; and α is the negative half-axis exponent coefficient.

[0093] In this embodiment, the scaling factor λ is 1.0507; the negative half-axis exponent factor α is 1.6733.

[0094] like Figure 5 As shown, multiple bottleneck layers in the dense block introduce a sparse connection strategy on top of dense connections, retaining only some representative inter-layer connections to reduce parameter redundancy and improve feature transfer efficiency. The formula is expressed as:

[0095]

[0096]

[0097] in, Indicates the first The number of input feature maps for the bottleneck layer; The growth rate represents the number of feature maps added in each layer; N represents the preset maximum number of connection layers. Indicates the first Output feature map of the bottleneck layer; Indicates the first Bottleneck layer.

[0098] In this embodiment, the growth rate It is 32.

[0099] The classification layer consists of a series of average pooling layers and linear layers. The output feature map of the dense neural network module is first passed to the average pooling layer to perform dimensionality reduction. Then, the extracted features are integrated through the linear layer, and the probability distribution of each category is calculated to complete the fault classification.

[0100] Step 5: Train the model

[0101] like Figure 6 As shown, the source domain is used to pre-train the fault detection model. After pre-training, the weights of the first two dense blocks and the corresponding transition layers of the fault detection model are frozen, and only the subsequent dense blocks, transition layers and classification layers are retained to obtain the basic model.

[0102] like Figure 7 As shown, the target domain is partitioned using an N-way K-shot approach to construct multiple meta-tasks. Each meta-task contains N different categories, and 2K input samples are randomly selected from each category, divided into a support set and a query set. The support set contains K samples from each category (…). The query set contains an additional K samples for each category ( The support set is used for rapid model updates, and the query set is used to evaluate the performance of the updated model. The base model's parameters are updated using the support set S, and the updated parameters θ are... t Represented as:

[0103]

[0104] in, These are global model parameters; A learning rate specific to the task; This is the loss function based on the support set.

[0105] Using query sets Calculate the updated model parameters The loss is calculated, and the global parameters are optimized through backpropagation of the query set loss of all meta-tasks. :

[0106]

[0107]

[0108] Among them, Lmeta The total task loss for the query set; This represents the total number of meta-tasks; For query set-based task loss; It is the global learning rate of meta-learning.

[0109] By updating task-specific parameters and optimizing global parameters across different tasks, the global model parameters are ultimately optimized. convergence.

[0110] During training, the Adam optimizer is used for parameter optimization, and cross-entropy loss is used as the loss function. In pre-training, the number of iterations is set to 100, and the batch size is 16. In the meta-training phase, the number of iterations is set to 10, the number of tasks is 1000, and multiple meta-tasks are constructed using an 8-way, 1-shot approach. In some implementations, the task partitioning method can be adjusted according to specific circumstances, such as using different numbers of classes (ways) or different numbers of samples per class (shots), to adapt to different application scenarios and data characteristics.

[0111] Step Six: Model Evaluation

[0112] 6-1. To verify the advantages of the feature selection method (PCC-RF-MDA) in this invention, comparative experiments were conducted on CNN and LSTM models respectively. The final importance scores and correlation coefficients of the selected features are ranked as follows: Figure 8 and Figure 9 As shown in Table 3, the results of the comparative experiments are presented.

[0113] Table 3 Feature selection performance comparison

[0114]

[0115] As shown in Table 3, the feature selection method of this invention exhibits significant comprehensive advantages in practical applications. On 1DCNN and LSTM models, its accuracy is improved by 3.0% and 5.7% respectively, while training time is significantly reduced by nearly 50%. It maintains high detection performance while significantly improving computational efficiency, demonstrating superior overall performance compared to other methods.

[0116] 6-2. The present invention (DCC-Dense Net) is compared with existing fault detection methods (1DCNN, LSTM, DenseNet). The comparison results are as follows: Figure 10 As shown in Table 4.

[0117] Table 4 Comparison of detection results of different models

[0118]

[0119] As shown in Table 4, DenseNet generally exhibits high and stable accuracy across all fault state categories, achieving an overall accuracy of 97.8% and an F1 score of 97.7%. This demonstrates DenseNet's superior performance and stability in feature extraction. In contrast, 1D CNN performs well in certain specific categories such as FWC and FWE, achieving 100% detection accuracy. However, its performance fluctuates significantly in other categories, particularly for NOR, where the accuracy is only 91.0%, and its overall accuracy is lower than DenseNet. This reflects the insufficient adaptability of 1D CNN in handling certain states, resulting in poor overall stability. LSTM, on the other hand, performs significantly worse than DenseNet and is more unbalanced overall.

[0120] Although DenseNet outperforms other models overall, its kernel size limits its ability to capture long-term dependencies in time series data, thus limiting its performance in certain specific situations. This invention introduces a dilated causal convolution module to enhance DenseNet's ability to extract temporal features. To verify the effectiveness of this approach, the classification performance of this invention (DCC-Dense net) and common composite models were obtained, and the results are shown in Table 5.

[0121] Table 5. Comparison of overall performance and efficiency of different models

[0122]

[0123] As shown in Table 5, both common composite models outperform the basic Dense Net in classification performance. Dense Net-LSTM achieves the highest accuracy (99.6%) and the best performance across all evaluation metrics, but its training time is the longest (170.5s), reflecting its advantage in capturing temporal dependencies while also incurring higher computational costs. While DenseNet-GRU shortens its training time (130.2s), its accuracy drops significantly. In contrast, DCC-DenseNet maintains near-Dense Net-LSTM high accuracy (99.4%) while having a training time of only 112.8s, approximately 30% shorter than DenseNet-LSTM and superior to Dense Net-GRU, demonstrating its ability to effectively control computational costs while improving temporal feature extraction capabilities.

[0124] Figure 10The figure further illustrates the detection accuracy of the six models on the test set for eight states. The data in parentheses in the legend represent the model's average accuracy and training time. As shown in the figure, DCC-Dense Net maintains high training efficiency while exhibiting excellent overall performance. Although slightly inferior to Dense Net-LSTM, it significantly outperforms the other comparative models. The introduction of DCC improves the accuracy of Dense Net across multiple state categories without significantly increasing computational overhead. This is particularly evident for RO and RL categories, where accuracy increases by 3.9% and 3.6% respectively, resulting in an overall accuracy of 99.4%.

[0125] 6-4. To verify the effectiveness of the multi-task learning (MTL) method in this invention, the operating conditions of the chiller unit were controlled by changing key operating parameters, including three operating conditions with evaporator outlet water temperature (TWEO) set to 40F, 45F, and 50F, and five operating conditions with condenser inlet water temperature (TWCI) set to 62F, 65F, 70F, 75F, and 85F, totaling 27 different operating conditions. This study selected the operating conditions with TWEO of 40F, 45F, and 50F, and denoted them as A, B, and C, respectively. Six transfer tasks were constructed as shown in Table 6.

[0126] Table 6 Dataset partitioning under different migration tasks

[0127]

[0128] Based on the different transfer strategies in Table 6, the model accuracy and F1 score for different sample sizes were calculated, and the results are shown in Table 7.

[0129] Table 7 Comparison of fault detection performance under different migration tasks and sample sizes

[0130]

[0131] The baseline in Table 7 represents the model training of this invention using only the target domain training set. As shown in Table 7, reducing the number of samples significantly impacts the model's classification performance. When the number of training samples per class in the target domain decreases to 5, the model's accuracy and F1-score both decline significantly due to overfitting in small-sample scenarios and a lack of task adaptability; the accuracy reaches only 68.9%. However, after adopting the transfer strategy, the proposed method significantly outperforms the baseline under all sample conditions. Furthermore, transfer strategies with smaller differences in operating conditions (such as A→B and B→C) perform significantly better than transfer strategies with larger differences in operating conditions (such as A→C). In addition, multi-source transfer (A,B→C and A,C→B) achieves the best performance across all sample sizes, indicating that fusing information from multiple operating conditions helps improve the model's generalization ability and fault identification accuracy.

[0132] The fault detection methods of the present invention (MTL-DCC-Dense Net) were compared with other small-sample fault detection methods. To reduce the impact of randomness on the experimental results, each group of experiments was independently repeated 5 times. The results are shown in Table 8.

[0133] Table 8 Comparison of different small sample fault detection methods

[0134]

[0135] As can be seen from Table 8, the present invention (MTL-DCC-Dense Net) shows significant advantages under different small sample conditions (5 / 10 / 20 samples per class in the target domain), and its performance is better than other small sample fault detection methods.

[0136] To make the objectives, technical solutions, and advantages of this invention clearer, those skilled in the art should understand that the above embodiments are merely examples and do not imply that the scope of this disclosure (including the claims) is limited to these examples. Within the framework of this invention, the parameters (such as the number of network layers), hyperparameters (such as the number of tasks), and feature selection algorithms in the embodiments can be adjusted according to actual needs, and various aspects of this invention can have multiple variations and modifications, which are not all listed in the details.

[0137] This invention is intended to cover all alternatives, modifications and variations that fall within the broad scope of the appended claims; therefore, all such changes are within the protection scope of this invention.

Claims

1. A method for detecting small-sample faults in chiller units based on transfer learning, characterized in that: The method includes: The original features corresponding to different state types of the chiller unit under historical and current operating conditions are obtained, and high discriminative features are selected from the original features to construct a feature subset; the feature subsets corresponding to historical and current operating conditions are used as the source domain and target domain, respectively; the samples composed of high discriminative features in the source domain and target domain are preprocessed respectively. A fault detection model is constructed; the fault detection model processes and classifies the input samples sequentially through a dilated causal convolution module, a dense neural network module, and a classification layer to obtain the detection result of the state type; The dense neural network module includes multiple dense blocks connected in sequence; a transition layer is connected in series between adjacent dense blocks; each dense block includes multiple bottleneck layers; in a dense block, except for the first bottleneck layer, the input of the remaining bottleneck layers is part or all of the output feature maps of the bottleneck layers preceding that bottleneck layer. The source domain is used to pre-train the fault detection model. After pre-training, some dense blocks in the fault detection model and the weights of the corresponding transition layers are frozen to obtain the base model. The target domain is divided into multiple meta-tasks according to the state type, and the base model is trained by using the divided meta-tasks in sequence. The trained base model is used to detect faults in the chiller unit under test based on its high discriminative characteristics.

2. The method for small-sample fault detection of chiller units based on transfer learning according to claim 1, characterized in that: The method for obtaining the input feature map of the bottleneck layer is as follows: In a dense block, if the number of the current bottleneck layer is less than or equal to the maximum number of connected layers N, then the current bottleneck layer uses the concatenation result of the output feature maps of all previous bottleneck layers as the input feature map. If the number of layers in the current bottleneck layer is greater than the maximum number of connected layers N, then the current bottleneck layer selects the concatenation result of the output feature maps of N previous bottleneck layers as the input feature map.

3. The method for small-sample fault detection of chiller units based on transfer learning according to claim 2, characterized in that: The method for selecting N bottleneck layers as input feature maps is as follows: if the maximum number of connected layers is even, then select N / 2 bottleneck layers sequentially from both the initial bottleneck layer and the current bottleneck layer; if the maximum number of connected layers is odd, then select N / 2 bottleneck layers sequentially from both the initial bottleneck layer and the current bottleneck layer. Bottleneck layers, select sequentially from one side of the current bottleneck layer. Bottleneck layer.

4. The method for detecting small-sample faults in chiller units based on transfer learning according to claim 1, characterized in that: The screening process for the highly discriminative features is as follows: The importance scores of different original features are obtained respectively, and multiple original features with high importance scores are selected as key features based on the importance scores, and feature subsets are constructed based on the key features; Correlation analysis is performed on key features in the feature subset to obtain the correlation between different key features in the feature subset. Based on the correlation between different key features, the key features are further filtered to obtain highly discriminative features.

5. The method for small-sample fault detection of chiller units based on transfer learning according to claim 4, characterized in that: The process for selecting key features is as follows: Obtain the correlation between different key features in the feature subset; if the absolute value of the correlation between a key feature and the other key features is greater than the correlation threshold, remove the key features with lower importance scores from the feature subset, and use the key features retained in the feature subset as high discriminative features.

6. The method for detecting small-sample faults in chiller units based on transfer learning according to claim 4, characterized in that: The importance score is obtained by processing the samples in the dataset using a fault classification model and obtaining the importance score by the difference between the classification results before and after adding perturbation.

7. The method for small-sample fault detection of chiller units based on transfer learning according to claim 1, characterized in that: The bottleneck layer includes a first convolutional block and a second convolutional block connected in series; the first convolutional block and the second convolutional block have the same structure, both including a batch normalization layer, an activation function and a convolutional layer connected in sequence; the convolutional layers in the first convolutional block and the second convolutional block have different sizes.

8. The method for small-sample fault detection of chiller units based on transfer learning according to claim 1, characterized in that: The transition layer comprises a batch normalization layer, an activation function, a convolutional layer, and an average pooling layer connected in sequence.

9. The method for small-sample fault detection of chiller units based on transfer learning according to claim 1, characterized in that: The classification layer includes a series of average pooling layers and linear layers. The output feature map of the dense neural network module is first passed to the average pooling layer to perform dimensionality reduction. Then, the extracted features are integrated through the linear layer, and the probability distribution of each category is calculated to complete the fault classification.

10. A small-sample fault detection system for chiller units based on transfer learning, characterized in that: This system is used to execute the small-sample fault detection method for chiller units based on transfer learning as described in claim 1. The fault detection system includes a data acquisition module, a feature filtering module, a preprocessing module, and a fault detection module. The data acquisition module is used to acquire the original features of the chiller unit under different operating conditions. The feature filtering module is used to filter out highly discriminative features from the original features. The preprocessing module is used to segment the data of samples composed of high discriminative features; the fault detection module is used to complete the fault detection of the chiller unit based on the high discriminative features.

Images